refactor: move model-specific args into model parsers (#1757)

Co-authored-by: phil2sat <phil2sat@users.noreply.github.com>

.github/workflows/build.yml

@@ -15,7 +15,7 @@ on:
       [
         ".github/workflows/**",
         ".dockerignore",
-        "Dockerfile*",
+        "docker/Dockerfile*",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -32,7 +32,7 @@ on:
       [
         ".github/workflows/**",
         ".dockerignore",
-        "Dockerfile*",
+        "docker/Dockerfile*",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -83,7 +83,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON
+          cmake .. -DSD_BUILD_SHARED_LIBS=ON -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_INSTALL_RPATH='$ORIGIN'
           cmake --build . --config Release
 
       - name: Get commit hash
@@ -146,7 +146,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DSD_BUILD_SHARED_LIBS=ON -DSD_VULKAN=ON
+          cmake .. -DSD_BUILD_SHARED_LIBS=ON -DSD_VULKAN=ON -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_INSTALL_RPATH='$ORIGIN'
           cmake --build . --config Release
 
       - name: Get commit hash
@@ -207,6 +207,7 @@ jobs:
               UBUNTU_VERSION=24.04
               CUDA_ARCHITECTURES=121
               GGML_CUDA_FA_ALL_QUANTS=ON
+              GGML_CUDA_ENABLE_DYNAMIC_CPU_BACKENDS=OFF
 
     env:
       REGISTRY: ghcr.io
@@ -263,7 +264,7 @@ jobs:
           context: .
           platforms: ${{ matrix.platform }}
           push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-          file: Dockerfile.${{ matrix.variant }}
+          file: docker/Dockerfile.${{ matrix.variant }}
           tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}${{ matrix.tag-suffix }}
           labels: ${{ steps.meta.outputs.labels }}
           annotations: ${{ steps.meta.outputs.annotations }}
@@ -341,18 +342,12 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: "noavx"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "avx2"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "avx"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "avx512"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "cpu"
+            defines: "-DGGML_NATIVE=OFF -DSD_BUILD_SHARED_LIBS=ON -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON"
           - build: "cuda12"
-            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120' -DCMAKE_CUDA_FLAGS='-Xcudafe \"--diag_suppress=177\" -Xcudafe \"--diag_suppress=550\"'"
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120' -DCMAKE_CUDA_FLAGS='-Xcudafe \"--diag_suppress=177\" -Xcudafe \"--diag_suppress=550\"' -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON"
           - build: "vulkan"
-            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
+            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON"
     steps:
       - name: Clone
         id: checkout
@@ -400,19 +395,6 @@ jobs:
           cmake .. -DCMAKE_CXX_FLAGS='/bigobj' -G Ninja -DCMAKE_C_COMPILER=cl.exe -DCMAKE_CXX_COMPILER=cl.exe -DCMAKE_BUILD_TYPE=Release ${{ matrix.defines }}
           cmake --build .
 
-      - name: Check AVX512F support
-        id: check_avx512f
-        if: ${{ matrix.build == 'avx512' }}
-        continue-on-error: true
-        run: |
-          cd build
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
-          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
-          & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
-          .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
-
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -722,6 +704,25 @@ jobs:
         run: |
           sudo apt install -y build-essential cmake wget zip ninja-build
 
+      - name: Free disk space
+        run: |
+          df -h
+
+          # Remove preinstalled SDKs and caches not needed for this job before
+          # installing ROCm. The legacy ROCm apt packages are large enough to
+          # exhaust ubuntu-latest if cleanup runs after installation.
+          sudo rm -rf /usr/share/dotnet || true
+          sudo rm -rf /usr/local/lib/android || true
+          sudo rm -rf /opt/ghc || true
+          sudo rm -rf /usr/local/.ghcup || true
+          sudo rm -rf /opt/hostedtoolcache || true
+          sudo rm -rf /usr/share/swift || true
+          sudo rm -rf /usr/local/share/boost || true
+          docker system prune -af || true
+
+          sudo apt clean
+          df -h
+
       - name: Setup Legacy ROCm
         if: matrix.ROCM_VERSION == '7.2.1'
         id: legacy_env
@@ -743,19 +744,6 @@ jobs:
           sudo apt update
           sudo apt-get install -y libssl-dev rocm-hip-sdk
 
-      - name: Free disk space
-        run: |
-          # Remove preinstalled SDKs and caches not needed for this job
-          sudo rm -rf /usr/share/dotnet || true
-          sudo rm -rf /usr/local/lib/android || true
-          sudo rm -rf /opt/ghc || true
-          sudo rm -rf /usr/local/.ghcup || true
-          sudo rm -rf /opt/hostedtoolcache || true
-
-          # Remove old package lists and caches
-          sudo rm -rf /var/lib/apt/lists/* || true
-          sudo apt clean
-
       - name: Setup TheRock
         if: matrix.ROCM_VERSION != '7.2.1'
         id: therock_env
@@ -794,6 +782,11 @@ jobs:
             -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DGGML_NATIVE=OFF \
+            -DSD_BUILD_SHARED_GGML_LIB=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DSD_BUILD_SHARED_LIBS=ON
           cmake --build . --config Release
 

CMakeLists.txt

@@ -331,7 +331,8 @@ endif()
 
 add_subdirectory(thirdparty)
 
-target_link_libraries(${SD_LIB} PUBLIC ggml zip)
+target_sources(${SD_LIB} PRIVATE $<TARGET_OBJECTS:zip>)
+target_link_libraries(${SD_LIB} PUBLIC ggml)
 target_include_directories(${SD_LIB} PUBLIC . src include)
 target_include_directories(${SD_LIB} PRIVATE src/core)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
@@ -342,7 +343,58 @@ if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
-set(SD_PUBLIC_HEADERS include/stable-diffusion.h)
-set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
 
-install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)
+
+#
+# install
+#
+
+include(CMakePackageConfigHelpers)
+include(GNUInstallDirs)
+
+set(SD_INSTALL_VERSION "${SDCPP_BUILD_VERSION}")
+set(SD_INSTALL_COMMIT  "${SDCPP_BUILD_COMMIT}")
+set(SD_SHARED_LIB      ${SD_BUILD_SHARED_LIBS})
+
+set(SD_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(SD_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(SD_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
+
+set(SD_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/stable-diffusion.h)
+
+set_target_properties(${SD_LIB}
+    PROPERTIES
+        PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+
+install(TARGETS ${SD_LIB}
+        ARCHIVE
+        LIBRARY
+        RUNTIME
+        PUBLIC_HEADER)
+
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/stable-diffusion-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/stable-diffusion
+    PATH_VARS SD_INCLUDE_INSTALL_DIR
+              SD_LIB_INSTALL_DIR
+              SD_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+        ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-version.cmake
+    VERSION ${SD_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/stable-diffusion)
+
+configure_file(cmake/stable-diffusion.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion.pc"
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)

CONTRIBUTING.md

@@ -27,7 +27,7 @@ PRs should include:
 
 Format code according to the repository style before submitting changes.
 
-Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `format-code.sh` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
+Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `scripts/format-code.sh` or `scripts/format-code.ps1` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
 
 Naming conventions:
 

README.md

@@ -15,6 +15,7 @@ API and command-line option may change frequently.***
 
 ## 🔥Important News
 
+* **2026/06/25** 🚀 stable-diffusion.cpp now supports **Krea2**
 * **2026/06/04** 🚀 stable-diffusion.cpp now supports **Ideogram4**
 * **2026/05/31** 🚀 stable-diffusion.cpp now supports **PiD**
 * **2026/05/27** 🚀 stable-diffusion.cpp now supports **Lens**
@@ -34,8 +35,8 @@ API and command-line option may change frequently.***
 - Super lightweight and without external dependencies
 - Supported models
   - Image Models
-    - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
-    - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
+    - [SD1.x, SD2.x, SD-Turbo](./docs/sd.md)
+    - [SDXL, SDXL-Turbo](./docs/sd.md)
     - [Some SD1.x and SDXL distilled models](./docs/distilled_sd.md)
     - [SD3/SD3.5](./docs/sd3.md)
     - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
@@ -47,24 +48,29 @@ API and command-line option may change frequently.***
     - [PiD](./docs/pid.md)
     - [LongCat Image](./docs/longcat_image.md)
     - [Z-Image](./docs/z_image.md)
+    - [MiniT2I](./docs/minit2i.md)
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
+    - [Boogu Image](./docs/boogu_image.md)
+    - [Krea2](./docs/krea2.md)
+    - [SeFi-Image](./docs/sefi_image.md)
     - [HiDream-O1-Image](./docs/hidream_o1_image.md)
     - [Ideogram4](./docs/ideogram4.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
     - [LongCat Image Edit](./docs/longcat_image.md)
+    - [Boogu Image Edit](./docs/boogu_image.md)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
     - [LTX-2.3](./docs/ltx2.md)
-  - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
+  - [PhotoMaker](./docs/photo_maker.md) support.
   - Control Net support with SD 1.5
   - LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
   - Latent Consistency Models support (LCM/LCM-LoRA)
-  - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
-  - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
+  - Faster and memory efficient latent decoding with [TAESD](./docs/taesd.md)
+  - Upscale images generated with [ESRGAN](./docs/esrgan.md)
 - Supported backends
   - CPU (AVX, AVX2 and AVX512 support for x86 architectures)
   - CUDA
@@ -133,28 +139,9 @@ For runtime and parameter backend placement, see the [backend selection guide](.
 ## More Guides
 
 - [Backend selection](./docs/backend.md)
-- [SD1.x/SD2.x/SDXL](./docs/sd.md)
-- [SD3/SD3.5](./docs/sd3.md)
-- [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
-- [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
-- [FLUX.1-Kontext-dev](./docs/kontext.md)
-- [Chroma](./docs/chroma.md)
-- [🔥Qwen Image](./docs/qwen_image.md)
-- [🔥Qwen Image Edit series](./docs/qwen_image_edit.md)
-- [🔥Wan2.1/Wan2.2](./docs/wan.md)
-- [🔥LTX-2.3](./docs/ltx2.md)
-- [🔥Z-Image](./docs/z_image.md)
-- [Ovis-Image](./docs/ovis_image.md)
-- [Anima](./docs/anima.md)
-- [ERNIE-Image](./docs/ernie_image.md)
-- [HiDream-O1-Image](./docs/hidream_o1_image.md)
-- [Lens](./docs/lens.md)
-- [LongCat Image / LongCat Image Edit](./docs/longcat_image.md)
+- [RPC](./docs/rpc.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
-- [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
-- [Using ESRGAN to upscale results](./docs/esrgan.md)
-- [Using TAESD to faster decoding](./docs/taesd.md)
 - [Docker](./docs/docker.md)
 - [Quantization and GGUF](./docs/quantization_and_gguf.md)
 - [Inference acceleration via caching](./docs/caching.md)
@@ -193,19 +180,3 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=leejet/stable-diffusion.cpp&type=Date)](https://star-history.com/#leejet/stable-diffusion.cpp&Date)
-
-## References
-
-- [ggml](https://github.com/ggml-org/ggml)
-- [diffusers](https://github.com/huggingface/diffusers)
-- [stable-diffusion](https://github.com/CompVis/stable-diffusion)
-- [sd3-ref](https://github.com/Stability-AI/sd3-ref)
-- [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
-- [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui)
-- [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
-- [k-diffusion](https://github.com/crowsonkb/k-diffusion)
-- [latent-consistency-model](https://github.com/luosiallen/latent-consistency-model)
-- [generative-models](https://github.com/Stability-AI/generative-models/)
-- [PhotoMaker](https://github.com/TencentARC/PhotoMaker)
-- [Wan2.1](https://github.com/Wan-Video/Wan2.1)
-- [Wan2.2](https://github.com/Wan-Video/Wan2.2)

cmake/stable-diffusion-config.cmake.in

@@ -0,0 +1,37 @@
+set(SD_VERSION      "@SD_INSTALL_VERSION@")
+set(SD_BUILD_COMMIT "@SD_INSTALL_COMMIT@")
+set(SD_SHARED_LIB    @SD_SHARED_LIB@)
+
+@PACKAGE_INIT@
+
+set_and_check(SD_INCLUDE_DIR "@PACKAGE_SD_INCLUDE_INSTALL_DIR@")
+set_and_check(SD_LIB_DIR     "@PACKAGE_SD_LIB_INSTALL_DIR@")
+set(SD_BIN_DIR "@PACKAGE_SD_BIN_INSTALL_DIR@")
+
+include(CMakeFindDependencyMacro)
+find_dependency(ggml REQUIRED HINTS "${SD_LIB_DIR}/cmake")
+
+if(NOT TARGET stable-diffusion)
+    find_library(stable-diffusion_LIBRARY stable-diffusion
+        REQUIRED
+        HINTS "${SD_LIB_DIR}"
+        NO_CMAKE_FIND_ROOT_PATH
+    )
+
+    add_library(stable-diffusion UNKNOWN IMPORTED)
+    set_target_properties(stable-diffusion
+        PROPERTIES
+            INTERFACE_INCLUDE_DIRECTORIES "${SD_INCLUDE_DIR}"
+            INTERFACE_LINK_LIBRARIES "ggml::ggml"
+            IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+            IMPORTED_LOCATION "${stable-diffusion_LIBRARY}"
+            INTERFACE_COMPILE_FEATURES "c_std_11;cxx_std_17"
+            POSITION_INDEPENDENT_CODE ON)
+
+    if(SD_SHARED_LIB)
+        target_compile_definitions(stable-diffusion
+            INTERFACE SD_BUILD_SHARED_LIB)
+    endif()
+endif()
+
+check_required_components(stable-diffusion)

cmake/stable-diffusion.pc.in

@@ -0,0 +1,11 @@
+prefix=@CMAKE_INSTALL_PREFIX@
+exec_prefix=${prefix}
+libdir=@CMAKE_INSTALL_FULL_LIBDIR@
+includedir=@CMAKE_INSTALL_FULL_INCLUDEDIR@
+
+Name: stable-diffusion
+Description: Diffusion model(SD,Flux,Wan,Qwen Image,Z-Image,...) inference in pure C/C++
+Version: @SDCPP_BUILD_VERSION@
+Libs: -L${libdir} -lstable-diffusion
+Libs.private: -lggml -lggml-base
+Cflags: -I${includedir}

docker/Dockerfile

@@ -19,7 +19,14 @@ WORKDIR /sd.cpp
 
 COPY . .
 
-RUN cmake . -B ./build
+RUN cmake . -B ./build \
+    -DSD_BUILD_SHARED_LIBS=ON \
+    -DGGML_NATIVE=OFF \
+    -DSD_BUILD_SHARED_GGML_LIB=ON \
+    -DGGML_BACKEND_DL=ON \
+    -DGGML_CPU_ALL_VARIANTS=ON \
+    -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+    -DCMAKE_INSTALL_RPATH='$ORIGIN'
 RUN cmake --build ./build --config Release --parallel
 
 FROM ubuntu:$UBUNTU_VERSION AS runtime
@@ -28,7 +35,9 @@ RUN apt-get update && \
     apt-get install --yes --no-install-recommends libgomp1 && \
     apt-get clean
 
-COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
-COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
 
 ENTRYPOINT [ "/sd-cli" ]

docker/Dockerfile.cuda

@@ -23,11 +23,27 @@ COPY . .
 ARG CUDACXX=/usr/local/cuda/bin/nvcc
 ARG CUDA_ARCHITECTURES=""
 ARG GGML_CUDA_FA_ALL_QUANTS=""
+ARG GGML_CUDA_ENABLE_DYNAMIC_CPU_BACKENDS=ON
 
-RUN cmake . -B ./build \
-    -DSD_CUDA=ON \
-    ${CUDA_ARCHITECTURES:+-DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}"} \
-    ${GGML_CUDA_FA_ALL_QUANTS:+-DGGML_CUDA_FA_ALL_QUANTS=${GGML_CUDA_FA_ALL_QUANTS}}
+RUN set -- \
+        -DSD_CUDA=ON; \
+    if [ "${GGML_CUDA_ENABLE_DYNAMIC_CPU_BACKENDS}" = "ON" ]; then \
+        set -- "$@" \
+            -DSD_BUILD_SHARED_LIBS=ON \
+            -DGGML_NATIVE=OFF \
+            -DSD_BUILD_SHARED_GGML_LIB=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            "-DCMAKE_INSTALL_RPATH=\$ORIGIN"; \
+    fi; \
+    if [ -n "${CUDA_ARCHITECTURES}" ]; then \
+        set -- "$@" "-DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHITECTURES}"; \
+    fi; \
+    if [ -n "${GGML_CUDA_FA_ALL_QUANTS}" ]; then \
+        set -- "$@" "-DGGML_CUDA_FA_ALL_QUANTS=${GGML_CUDA_FA_ALL_QUANTS}"; \
+    fi; \
+    cmake . -B ./build "$@"
 RUN cmake --build ./build --config Release -j$(nproc)
 
 FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime
@@ -36,7 +52,9 @@ RUN apt-get update && \
     apt-get install --yes --no-install-recommends libgomp1 && \
     apt-get clean
 
-COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
-COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
 
 ENTRYPOINT [ "/sd-cli" ]

docker/Dockerfile.musa

@@ -24,12 +24,22 @@ RUN mkdir build && cd build && \
     cmake .. -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
         -DCMAKE_C_FLAGS="${CMAKE_C_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
         -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
-        -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release && \
+        -DSD_MUSA=ON \
+        -DSD_BUILD_SHARED_LIBS=ON \
+        -DGGML_NATIVE=OFF \
+        -DSD_BUILD_SHARED_GGML_LIB=ON \
+        -DGGML_BACKEND_DL=ON \
+        -DGGML_CPU_ALL_VARIANTS=ON \
+        -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+        -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+        -DCMAKE_BUILD_TYPE=Release && \
     cmake --build . --config Release
 
 FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runtime
 
-COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
-COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
 
 ENTRYPOINT [ "/sd-cli" ]

docker/Dockerfile.sycl

@@ -29,4 +29,4 @@ FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
 COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
 COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
 
-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd-cli" ]

docker/Dockerfile.vulkan

@@ -19,8 +19,16 @@ WORKDIR /sd.cpp
 
 COPY . .
 
-RUN cmake . -B ./build -DSD_VULKAN=ON
-RUN cmake --build ./build --config Release --parallel
+RUN cmake . -B ./build \
+    -DSD_VULKAN=ON \
+    -DSD_BUILD_SHARED_LIBS=ON \
+    -DGGML_NATIVE=OFF \
+    -DSD_BUILD_SHARED_GGML_LIB=ON \
+    -DGGML_BACKEND_DL=ON \
+    -DGGML_CPU_ALL_VARIANTS=ON \
+    -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+    -DCMAKE_INSTALL_RPATH='$ORIGIN'
+RUN cmake --build ./build --config Release -j$(nproc)
 
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
@@ -28,7 +36,9 @@ RUN apt-get update && \
     apt-get install --yes --no-install-recommends libgomp1 libvulkan1 mesa-vulkan-drivers && \
     apt-get clean
 
-COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
-COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
 
 ENTRYPOINT [ "/sd-cli" ]

docs/anima.md

@@ -15,7 +15,7 @@
 ## Examples
 
 ```sh
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\anima-preview.safetensors --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="anima image example" src="../assets/anima/example.png" />

docs/backend.md

@@ -35,6 +35,14 @@ sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend te=cpu,v
 sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend disk
 ```
 
+`--max-vram` can target resolved backend/device names:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend diffusion=cuda0,vae=vulkan0 --max-vram cuda0=6,vulkan0=2
+```
+
+The budget applies to every module running on that backend.
+
 Module names are case-insensitive. Hyphens and underscores in module names are ignored, so `clip_vision`, `clip-vision`, and `clipvision` are equivalent.
 
 `all=`, `default=`, and `*=` can be used to set the default backend inside a mixed assignment:
@@ -43,6 +51,97 @@ Module names are case-insensitive. Hyphens and underscores in module names are i
 sd-cli -m model.safetensors -p "a cat" --backend all=cuda0,te=cpu
 ```
 
+## Multiple devices per module (layer split)
+
+A `--backend` module assignment can list several devices separated by `&`:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1"
+```
+
+The module's transformer blocks are then distributed across the listed devices
+in contiguous ranges sized proportionally to each device's free memory (minus a
+compute-buffer headroom of about 2 GiB per device), and the
+module's graphs are executed with a `ggml_backend_sched` that runs each block
+on the device holding its weights, copying the residual stream at the range
+boundaries. The first device in the list is the module's main device: it also
+holds the non-block tensors (embeddings, final norms, small sub-runners such as
+CLIP models or projectors) and the graph inputs/outputs.
+
+Layer split is supported for the `diffusion` and `te` modules. For `te` it
+applies to the dominant text encoder (`t5xxl` or the LLM); other modules accept
+only a single device. If the module has no recognizable transformer blocks, the
+assignment falls back to the first listed device.
+
+`--params-backend` accepts no device lists. If the module has no explicit
+params assignment, each block range's parameters are loaded directly to (and,
+with `--params-backend diffusion=disk`, released directly from) its own device;
+an explicit assignment such as `te=cpu` keeps the parameters on that backend
+and stages each range to its device on demand.
+
+Layer split cannot be combined with `--max-vram` graph-cut segmentation or
+`--stream-layers` for the split module; those are single-device mechanisms and
+are disabled for it.
+
+Use `--list-devices` to see the device names available on the system.
+
+### Row split (`--split-mode row`)
+
+`--split-mode` selects how a multi-device module distributes its weights:
+`layer` (the default, described above) or `row`. It accepts a single mode or
+per-module assignments:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1" --split-mode row
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1,te=cuda0&cuda1" --split-mode diffusion=row,te=layer
+```
+
+In row mode the module keeps executing on its main (first listed) device, but
+its transformer-block matmul weights are allocated in the backend's row-split
+buffer type, which slices each weight's rows across the listed devices in
+proportion to free memory and runs those matmuls on all devices in parallel.
+Compared to a layer split this uses all GPUs within every layer (instead of
+sequentially device by device) at the cost of a cross-device reduction per
+matmul - usually the faster option when the devices have fast interconnect.
+
+Row split requires backend support for split buffers and is currently
+available on CUDA only; on other backends (or when the listed devices belong
+to different backend registries) the module falls back to a layer split.
+Embeddings, normalization weights, biases and other non-block tensors stay in
+regular buffers on the main device.
+
+Direct ("immediately") LoRA application cannot patch row-split tensors; with
+`--split-mode row` the automatic LoRA mode selects runtime application, and an
+explicit `--lora-apply-mode immediately` skips the split tensors with a
+warning.
+
+## Automatic placement (`--auto-fit`)
+
+`--auto-fit` derives the `diffusion` / `te` / `vae` placements from the model
+metadata and the per-device memory budgets, then feeds them into the same
+backend assignment mechanism described above (the chosen specs are printed).
+`--backend` and `--params-backend` are ignored while auto-fit is enabled.
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --auto-fit
+sd-cli -m model.safetensors -p "a cat" --auto-fit --max-vram cuda0=8,cuda1=14
+sd-cli -m model.safetensors -p "a cat" --auto-fit --split-mode row
+```
+
+Budgets reuse `--max-vram`: a positive per-device value caps what auto-fit
+plans with on that device, a negative value means "free memory minus that many
+GiB", and with no budget set each device's free memory minus a 512 MiB margin
+is used. (The same values still drive graph-cut segmented execution for
+modules that end up on a single device.)
+
+When everything fits resident, components are simply spread across the
+available GPUs. When it does not, auto-fit switches to time-share mode: the
+heavy components get `disk` params residency (loaded for their phase, freed
+after), and a component too large for any single device is split across all
+GPUs with the layer/row split mechanism (`--split-mode` selects which, layer
+by default). Components that fit nowhere fall back to the CPU. If a VAE decode
+still runs out of memory, tiling is enabled and the decode retried once.
+
 ## Modules
 
 | Module | Purpose | Accepted names |

docs/boogu_image.md

@@ -0,0 +1,31 @@
+# How to Use
+
+Boogu Image uses a Boogu diffusion transformer, the FLUX VAE, and Qwen3-VL as the LLM text and vision encoder.
+
+## Download weights
+
+- Download Boogu Image
+    - safetensors: https://huggingface.co/Comfy-Org/Boogu-Image/tree/main/diffusion_models
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download Qwen3-VL 8B
+    - gguf: https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/tree/main
+        - For image editing with GGUF text encoders, also download the matching mmproj file and pass it with `--llm_vision`.
+
+## Examples
+
+### Boogu Image Base
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\boogu_image_base_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\models\vae\ae.sft -p "a lovely cat" --diffusion-fa -v --offload-to-cpu
+```
+
+<img width="256" alt="Boogu Image Base example" src="../assets/boogu/example.png" />
+
+### Boogu Image Edit
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\boogu_image_edit_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\mmproj-Qwen3VL-8B-Instruct-F16.gguf --vae ..\models\vae\ae.sft --diffusion-fa -v --offload-to-cpu -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'boogu.cpp'"
+```
+
+<img width="256" alt="Boogu Image Edit example" src="../assets/boogu/edit_example.png" />

docs/chroma.md

@@ -15,7 +15,7 @@ You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4G
 You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF), this way you don't have to do the conversion yourself.
 
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
 ```
 
 ## Run
@@ -24,10 +24,9 @@ You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](h
 For example:
 
 ```
- .\bin\Release\sd-cli.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
+ .\bin\Release\sd-cli.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --model-args chroma_use_dit_mask=false --clip-on-cpu
 ```
 
 ![](../assets/flux/chroma_v40.png)
 
 
-

docs/chroma_radiance.md

@@ -12,7 +12,7 @@
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\..\ComfyUI\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
 ```
 
 <img alt="Chroma1-Radiance" src="../assets/flux/chroma1-radiance.png" />

docs/docker.md

@@ -19,7 +19,7 @@ docker run --rm --init -v /path/to/models:/models -v /path/to/output/:/output -p
 ## Building using Docker
 
 ```shell
-docker build -t sd .
+docker build -f docker/Dockerfile -t sd .
 ```
 
 ## Building variants using Docker
@@ -27,7 +27,7 @@ docker build -t sd .
 Vulkan:
 
 ```shell
-docker build -f Dockerfile.vulkan -t sd .
+docker build -f docker/Dockerfile.vulkan -t sd .
 ```
 
 ## Run locally built image's CLI

docs/ernie_image.md

@@ -21,7 +21,7 @@ You can run ERNIE-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — o
 ### ERNIE-Image-Turbo
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-turbo.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 8 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\ernie-image-turbo.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 8 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img width="256" alt="ERNIE-Image Turbo example" src="../assets/ernie_image/turbo_example.png" />
@@ -29,7 +29,7 @@ You can run ERNIE-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — o
 ### ERNIE-Image
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\ernie-image-UD-Q4_K_M.gguf --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img width="256" alt="ERNIE-Image example" src="../assets/ernie_image/example.png" />

docs/flux.md

@@ -17,7 +17,7 @@ You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://hu
 
 For example:
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
 ```
 
 ## Run

docs/flux2.md

@@ -16,7 +16,7 @@
 ### Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
 ```
 
 <img alt="flux2 example" src="../assets/flux2/example.png" />
@@ -42,19 +42,19 @@
 ### Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-4b" src="../assets/flux2/flux2-klein-4b.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
 ```
 
 <img alt="flux2-klein-4b-edit" src="../assets/flux2/flux2-klein-4b-edit.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-base-4b" src="../assets/flux2/flux2-klein-base-4b.png" />
@@ -78,19 +78,19 @@
 ### Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-9b" src="../assets/flux2/flux2-klein-9b.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
 ```
 
 <img alt="flux2-klein-9b-edit" src="../assets/flux2/flux2-klein-9b-edit.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-base-9b" src="../assets/flux2/flux2-klein-base-9b.png" />

docs/hidream_o1_image.md

@@ -12,7 +12,7 @@
 ### HiDream-O1-Image-Dev
 
 ```
-.\bin\Release\sd-cli.exe -m  ..\..\ComfyUI\models\diffusion_models\hidream_o1_image_dev_bf16.safetensors -p "a lovely cat holding a sign says 
+.\bin\Release\sd-cli.exe -m  ..\models\diffusion_models\hidream_o1_image_dev_bf16.safetensors -p "a lovely cat holding a sign says 
 'hidream o1 cpp'" --cfg-scale 1.0  -v -H 1024 -W 1024
 ```
 

docs/ideogram4.md

@@ -34,7 +34,7 @@ If you want lower VRAM usage, you can change the quantization from q8_0 to a low
 ## Examples
 
 ```sh
-.\bin\Release\sd-cli.exe --diffusion-model ideogram4-Q8_0.gguf --uncond-diffusion-model ideogram4_uncond-Q8_0.gguf --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors -p '{"high_level_description":"A square 1024 x 1024 luxury fashion magazine cover featuring exactly one short chubby fluffy cat as the main model. The cat sits on a soft ivory studio floor, facing the viewer with a stylish calm expression, wearing tiny black sunglasses, a red silk scarf, and a small gold collar charm. In front of the cat on the floor is a wide horizontal luxury nameplate that clearly reads ideogram4.cpp. The whole design feels premium, fashionable, clean, and editorial.","style_description":{"aesthetics":"luxury fashion magazine cover, high-end pet couture campaign, minimalist editorial design, elegant studio photography, soft paper texture, refined typography, fashionable and polished","lighting":"Soft diffused studio lighting, gentle spotlight on the cat, subtle floor shadow, warm ivory highlights, clean separation between subject and background","photo":"high-resolution fashion editorial photography look, front-facing cat portrait, crisp fur details, glossy sunglasses, clear readable nameplate text, shallow depth of field","medium":"mixed media fashion photography and premium editorial graphic design","color_palette":["#F4EFE7","#111111","#D8B56D","#B73A3A","#FFFFFF","#8A7A6A"]},"compositional_deconstruction":{"canvas":"Square 1024 x 1024 canvas with a normal upright orientation. Do not rotate the poster or any text. Use a clean fashion magazine cover layout.","background":"Warm ivory studio backdrop with subtle paper grain, a soft spotlight gradient, faint floor shadow, and a few minimal gold editorial lines. The background is spacious, premium, and uncluttered.","layout":"Top center has a small elegant headline. Center area features one cat as the main fashion model. Lower foreground has a wide horizontal luxury nameplate placed on the floor in front of the cat. Bottom center has a small footer. All text is horizontal, upright, and readable left to right.","elements":[{"type":"text","desc":"Top center headline reading LOOK WHAT I FOUND in a refined high-fashion serif font. The headline is horizontal, centered, elegant, and secondary to the nameplate text."},{"type":"obj","desc":"Exactly one short chubby fluffy cat sitting in the center like a luxury fashion model. The cat has a large round head, compact body, short legs, soft detailed fur, expressive eyes, and a calm confident pose. The cat is cute and rounded, not tall, not stretched, not duplicated."},{"type":"obj","desc":"Tiny glossy black sunglasses worn naturally by the cat, slightly oversized but still showing the cat face clearly. The sunglasses add a chic fashion-editorial attitude."},{"type":"obj","desc":"A red silk scarf tied neatly around the cat neck, with soft folds and a couture feeling. The scarf must not cover the cat face or the nameplate."},{"type":"obj","desc":"A small gold collar charm or fashion accessory under the scarf, subtle and premium, adding a luxury campaign detail."},{"type":"obj","desc":"In the lower foreground, place a wide horizontal luxury nameplate on the floor in front of the cat. The nameplate is low, flat, landscape-oriented, much wider than tall, like a fashion show seat card or premium display plaque. It is centered, front-facing, level, and fully visible. It must not become vertical, tall, standing, rotated, or side-facing."},{"type":"text","desc":"Print the exact text ideogram4.cpp only on the wide horizontal nameplate. Use clean bold black lettering, perfectly spelled, lowercase, with the number 4 and .cpp extension. The text must fit completely inside the nameplate, stay horizontal, and be readable from left to right."},{"type":"obj","desc":"Add sparse premium editorial accents around the edges: thin gold lines, small code brackets, tiny cursor marks, subtle dots, and minimal geometric details. No extra cats, no stickers, no animal faces, no busy decorations."},{"type":"text","desc":"Bottom center footer reading tiny paws, big compile energy in a small refined monospace or editorial font. The footer is horizontal, centered, understated, and much smaller than the nameplate text."}]}}'  --diffusion-fa -v --offload-to-cpu -H 1024 -W 1024
+.\bin\Release\sd-cli.exe --diffusion-model ideogram4-Q8_0.gguf --uncond-diffusion-model ideogram4_uncond-Q8_0.gguf --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\models\vae\flux2_ae.safetensors -p '{"high_level_description":"A square 1024 x 1024 luxury fashion magazine cover featuring exactly one short chubby fluffy cat as the main model. The cat sits on a soft ivory studio floor, facing the viewer with a stylish calm expression, wearing tiny black sunglasses, a red silk scarf, and a small gold collar charm. In front of the cat on the floor is a wide horizontal luxury nameplate that clearly reads ideogram4.cpp. The whole design feels premium, fashionable, clean, and editorial.","style_description":{"aesthetics":"luxury fashion magazine cover, high-end pet couture campaign, minimalist editorial design, elegant studio photography, soft paper texture, refined typography, fashionable and polished","lighting":"Soft diffused studio lighting, gentle spotlight on the cat, subtle floor shadow, warm ivory highlights, clean separation between subject and background","photo":"high-resolution fashion editorial photography look, front-facing cat portrait, crisp fur details, glossy sunglasses, clear readable nameplate text, shallow depth of field","medium":"mixed media fashion photography and premium editorial graphic design","color_palette":["#F4EFE7","#111111","#D8B56D","#B73A3A","#FFFFFF","#8A7A6A"]},"compositional_deconstruction":{"canvas":"Square 1024 x 1024 canvas with a normal upright orientation. Do not rotate the poster or any text. Use a clean fashion magazine cover layout.","background":"Warm ivory studio backdrop with subtle paper grain, a soft spotlight gradient, faint floor shadow, and a few minimal gold editorial lines. The background is spacious, premium, and uncluttered.","layout":"Top center has a small elegant headline. Center area features one cat as the main fashion model. Lower foreground has a wide horizontal luxury nameplate placed on the floor in front of the cat. Bottom center has a small footer. All text is horizontal, upright, and readable left to right.","elements":[{"type":"text","desc":"Top center headline reading LOOK WHAT I FOUND in a refined high-fashion serif font. The headline is horizontal, centered, elegant, and secondary to the nameplate text."},{"type":"obj","desc":"Exactly one short chubby fluffy cat sitting in the center like a luxury fashion model. The cat has a large round head, compact body, short legs, soft detailed fur, expressive eyes, and a calm confident pose. The cat is cute and rounded, not tall, not stretched, not duplicated."},{"type":"obj","desc":"Tiny glossy black sunglasses worn naturally by the cat, slightly oversized but still showing the cat face clearly. The sunglasses add a chic fashion-editorial attitude."},{"type":"obj","desc":"A red silk scarf tied neatly around the cat neck, with soft folds and a couture feeling. The scarf must not cover the cat face or the nameplate."},{"type":"obj","desc":"A small gold collar charm or fashion accessory under the scarf, subtle and premium, adding a luxury campaign detail."},{"type":"obj","desc":"In the lower foreground, place a wide horizontal luxury nameplate on the floor in front of the cat. The nameplate is low, flat, landscape-oriented, much wider than tall, like a fashion show seat card or premium display plaque. It is centered, front-facing, level, and fully visible. It must not become vertical, tall, standing, rotated, or side-facing."},{"type":"text","desc":"Print the exact text ideogram4.cpp only on the wide horizontal nameplate. Use clean bold black lettering, perfectly spelled, lowercase, with the number 4 and .cpp extension. The text must fit completely inside the nameplate, stay horizontal, and be readable from left to right."},{"type":"obj","desc":"Add sparse premium editorial accents around the edges: thin gold lines, small code brackets, tiny cursor marks, subtle dots, and minimal geometric details. No extra cats, no stickers, no animal faces, no busy decorations."},{"type":"text","desc":"Bottom center footer reading tiny paws, big compile energy in a small refined monospace or editorial font. The footer is horizontal, centered, understated, and much smaller than the nameplate text."}]}}'  --diffusion-fa -v --offload-to-cpu -H 1024 -W 1024
 ```
 
 <img alt="ideogram4 image example" src="../assets/ideogram4/example.png" />

docs/imatrix.md

@@ -0,0 +1,59 @@
+# Importance Matrix (imatrix) Quantization
+
+## What is an Importance Matrix?
+
+Quantization reduces the precision of a model's weights, decreasing its size and computational requirements. However, this can lead to a loss of quality. An importance matrix helps mitigate this by identifying which weights are *most* important for the model's performance. During quantization, these important weights are preserved with higher precision, while less important weights are quantized more aggressively.  This allows for better overall quality at a given quantization level.
+
+This originates from work done with language models in [llama.cpp](https://github.com/ggml-org/llama.cpp/blob/master/tools/imatrix/README.md).
+
+## Usage
+
+The imatrix feature involves two main steps: *training* the matrix and *using* it during quantization.
+
+### Training the Importance Matrix
+
+To generate an imatrix, run stable-diffusion.cpp with the `--imat-out` flag, specifying the output filename.  This process runs alongside normal image generation.
+
+```bash
+sd.exe [same exact parameters as normal generation] --imat-out imatrix.dat
+```
+
+*   **`[same exact parameters as normal generation]`**:  Use the same command-line arguments you would normally use for image generation (e.g., prompt, dimensions, sampling method, etc.).
+*   **`--imat-out imatrix.dat`**: Specifies the output file for the generated imatrix.
+
+You can generate multiple images at once using the `-b` flag to speed up the training process.
+
+### Continuing Training an Existing Matrix
+
+If you want to refine an existing imatrix, use the `--imat-in` flag *in addition* to `--imat-out`. This will load the existing matrix and continue training it.
+
+```bash
+sd.exe [same exact parameters as normal generation] --imat-out imatrix.dat --imat-in imatrix.dat
+```
+With that, you can train and refine the imatrix while generating images like you'd normally do. 
+
+### Using Multiple Matrices
+
+You can load and merge multiple imatrices together:
+
+```bash
+sd.exe [same exact parameters as normal generation] --imat-out imatrix.dat --imat-in imatrix.dat --imat-in imatrix2.dat
+```
+
+### Quantizing with an Importance Matrix
+
+To quantize a model using a trained imatrix, use the `-M convert` option (or equivalent quantization command) and the `--imat-in` flag, specifying the imatrix file.
+
+```bash
+sd.exe -M convert [same exact parameters as normal quantization] --imat-in imatrix.dat
+```
+
+*   **`[same exact parameters as normal quantization]`**: Use the same command-line arguments you would normally use for quantization (e.g., target quantization method, input/output filenames).
+*   **`--imat-in imatrix.dat`**: Specifies the imatrix file to use during quantization.  You can specify multiple `--imat-in` flags to combine multiple matrices.
+
+## Important Considerations
+
+*   The quality of the imatrix depends on the prompts and settings used during training. Use prompts and settings representative of the types of images you intend to generate for the best results.
+*   Experiment with different training parameters (e.g., number of images, prompt variations) to optimize the imatrix for your specific use case.
+*   The performance impact of training an imatrix during image generation or using an imatrix for quantization is negligible.
+*   Using already quantized models to train the imatrix seems to be working fine.

docs/kontext.md

@@ -16,7 +16,7 @@ You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4
 You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF), this way you don't have to do the conversion yourself.
 
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
 ```
 
 ## Run

docs/krea2.md

@@ -0,0 +1,27 @@
+# How to Use
+
+Krea2 uses a Krea2 diffusion transformer, the Wan2.1 VAE, and Qwen3-VL 4B as the LLM text encoder.
+
+## Download weights
+
+- Download Krea2 Raw
+    - safetensors: https://huggingface.co/krea/Krea-2-Raw/tree/main
+    - gguf: https://huggingface.co/realrebelai/KREA-2_GGUFs/tree/main/BASE
+- Download Krea2 Turbo
+    - safetensors: https://huggingface.co/krea/Krea-2-Turbo/tree/main
+    - gguf: https://huggingface.co/realrebelai/KREA-2_GGUFs/tree/main/TURBO
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
+- Download Qwen3-VL 4B
+    - safetensors: https://huggingface.co/Comfy-Org/Krea-2/tree/main/text_encoders
+    - gguf: https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/tree/main
+
+## Examples
+
+### Krea2
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\Krea-2-Raw-Q8_0.gguf --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf --vae ..\models\vae\wan_2.1_vae.safetensors -p "a lovely cat holding a sign says 'krea2.cpp'"  --diffusion-fa -v --offload-to-cpu
+```
+
+<img width="256" alt="Krea2 Raw example" src="../assets/krea2/example.png" />

docs/lens.md

@@ -18,7 +18,7 @@ Lens uses a Lens diffusion transformer, the FLUX.2 VAE, and GPT-OSS-20B as the L
 ### Lens
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 5.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\lens_bf16.safetensors --llm "..\models\text_encoders\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\models\vae\flux2_ae.safetensors --cfg-scale 5.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v
 ```
 
 <img width="256" alt="Lens example" src="../assets/lens/example.png" />
@@ -26,7 +26,7 @@ Lens uses a Lens diffusion transformer, the FLUX.2 VAE, and GPT-OSS-20B as the L
 ### Lens Turbo
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_turbo_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 1.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v --steps 4
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\lens_turbo_bf16.safetensors --llm "..\models\text_encoders\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\models\vae\flux2_ae.safetensors --cfg-scale 1.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v --steps 4
 ```
 
 <img width="256" alt="Lens Turbo example" src="../assets/lens/turbo_example.png" />

docs/longcat_image.md

@@ -24,7 +24,7 @@ LongCat uses quoted text for character-level text rendering. Put target text ins
 ### LongCat Image
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\LongCat-Image-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ae.sft --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf -p "a lovely cat holding a sign says 'longcat.cpp'" --cfg-scale 5.0 --sampling-method euler --flow-shift 3 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\LongCat-Image-Q4_K_M.gguf --vae ..\models\vae\ae.sft --llm ..\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf -p "a lovely cat holding a sign says 'longcat.cpp'" --cfg-scale 5.0 --sampling-method euler --flow-shift 3 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="longcat example" src="../assets/longcat/example.png" />

docs/ltx2.md

@@ -21,7 +21,7 @@
 ### LTX-2.3 dev T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "worst quality, low quality, blurry, distorted, artifacts" -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --fps 24 -o t2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "worst quality, low quality, blurry, distorted, artifacts" -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --fps 24 -o t2v.webm
 ```
 
 <video
@@ -33,7 +33,7 @@
 ### LTX-2.3 dev I2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o i2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o i2v.webm
 ```
 
 <video
@@ -45,7 +45,7 @@
 ### LTX-2.3 dev FLF2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png -o flf2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png -o flf2v.webm
 ```
 
 <video
@@ -61,7 +61,7 @@ LTX spatial latent upscale runs a model-backed x2 latent upsampler between the l
 Put `ltx-2.3-spatial-upscaler-x2-1.1.safetensors` under the directory passed to `--hires-upscalers-dir`, then use the model name without path or extension in `--hires-upscaler`.
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors --hires-upscalers-dir ..\..\ComfyUI\models\latent_upscale_models --hires-upscaler ltx-2.3-spatial-upscaler-x2-1.1 --hires --hires-steps 4 -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 640 -H 360 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o hires_i2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors --hires-upscalers-dir ..\models\latent_upscale_models --hires-upscaler ltx-2.3-spatial-upscaler-x2-1.1 --hires --hires-steps 4 -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 640 -H 360 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o hires_i2v.webm
 ```
 
 By default, the hires refine pass uses the main sampler and scheduler, then trims the second-pass sigma schedule by `--hires-denoising-strength` (`0.7` by default). To reproduce a ComfyUI-style explicit refine schedule, pass custom hires sigmas:

docs/minit2i.md

@@ -0,0 +1,48 @@
+# How to Use
+
+MiniT2I uses a MiniT2I diffusion transformer and `google/flan-t5-large` as the text encoder.
+
+## Download weights
+
+- Download MiniT2I diffusion model
+    - safetensors: https://huggingface.co/MiniT2I/MiniT2I/tree/main/minit2i-b-16/transformer (`diffusion_pytorch_model.safetensors`)
+- Download flan-t5-large text encoder
+    - safetensors: https://huggingface.co/google/flan-t5-large/tree/main (`model.safetensors`)
+
+## Examples
+
+### Mac Metal
+
+```
+./bin/sd-cli \
+  --backend metal \
+  --diffusion-model ../models/minit2i/diffusion_pytorch_model.safetensors \
+  --t5xxl ../models/flan-t5-large/model.safetensors \
+  --prompt "a cat" \
+  --steps 100 \
+  --cfg-scale 6 \
+  --width 512 \
+  --height 512 \
+  --seed 42 \
+  --sampling-method euler \
+  --rng cpu \
+  --output minit2i_metal.png \
+  --threads 8
+```
+
+### CUDA with diffusion flash attention
+
+```
+./bin/sd-cli \
+  --diffusion-model ../models/minit2i/diffusion_pytorch_model.safetensors \
+  --t5xxl ../models/flan-t5-large/model.safetensors \
+  --prompt "a cat" \
+  --steps 100 \
+  --cfg-scale 6 \
+  --width 512 \
+  --height 512 \
+  --seed 42 \
+  --sampling-method euler \
+  --diffusion-fa \
+  --output minit2i_cuda.png
+```

docs/ovis_image.md

@@ -13,7 +13,7 @@
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ovis_image-Q4_0.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\ovis_2.5.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ovis_image-Q4_0.gguf --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\ovis_2.5.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="ovis image example" src="../assets/ovis_image/example.png" />

docs/performance.md

@@ -53,6 +53,26 @@ Per-module assignments can target only the largest modules:
 
 See [backend selection](./backend.md) for full syntax.
 
+## Run models that don't fit in VRAM (CPU streaming).
+
+`--offload-to-cpu` alone keeps every parameter in system RAM and stages it to the runtime backend on first use, then leaves it resident there. If the diffusion model is larger than the runtime backend's free memory (e.g. Flux dev at bf16 on an 8 GiB GPU), that residency stops fitting during the sampling loop and generation fails. Two additional flags make it fit by trading a small amount of speed for room:
+
+- `--max-vram <GiB>` sets a VRAM budget the graph-cut segmenter respects. It cuts each forward pass into segments sized to fit the budget, running them in sequence and freeing intermediate activations between them. Negative values auto-detect free VRAM and spare the given amount (`--max-vram -1` uses most of the free VRAM and keeps ~1 GiB headroom), a positive value caps the budget, `0` disables segmentation.
+- `--stream-layers` streams the diffusion model's transformer blocks one at a time. Each block's parameters are copied from the CPU to the runtime backend just before it runs and evicted when the residency budget is reached. Prefetching hides most of the copy latency behind compute. This flag only takes effect when the diffusion params backend is CPU, so it must be combined with `--offload-to-cpu` (or an explicit `--params-backend diffusion=cpu`); a warning is logged and the flag is ignored otherwise.
+
+The three flags stack. The recommended shape for "biggest model my card can host":
+
+```shell
+sd-cli --diffusion-model flux1-dev.safetensors ... \
+       --offload-to-cpu --max-vram -1 --stream-layers
+```
+
+- `--offload-to-cpu`: params in RAM, staged as needed.
+- `--max-vram -1`: use most of the free VRAM as the compute budget, spare 1 GiB headroom, let the graph-cut segmenter split each forward pass to fit.
+- `--stream-layers`: on top of the segmenter, stream individual transformer blocks so their weights don't all need to be resident at once.
+
+Ordered from fastest to smallest-VRAM: no flags → `--offload-to-cpu` → `--offload-to-cpu --max-vram <N>` → `--offload-to-cpu --max-vram <N> --stream-layers`. Each step down costs a few percent of throughput to buy more room; combined they can run models roughly 3-4x larger than the raw VRAM would allow.
+
 ## Use quantization to reduce memory usage.
 
 [quantization](./quantization_and_gguf.md)

docs/pid.md

@@ -22,7 +22,7 @@ The official PiD model card should be checked before use. At the time of the ini
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\pid_flux1_512_to_2048_4step_bf16.safetensors --llm "..\..\ComfyUI\models\text_encoders\gemma_2_2b_it_elm_bf16.safetensors" --vae ..\..\ComfyUI\models\vae\ae.sft --vae-format flux --cfg-scale 1.0  -p "a lovely cat" -r ..\assets\ernie_image\turbo_example.png --diffusion-fa -v --steps 4 -H 2048 -W 2048 --rng cpu
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\pid_flux1_512_to_2048_4step_bf16.safetensors --llm "..\models\text_encoders\gemma_2_2b_it_elm_bf16.safetensors" --vae ..\models\vae\ae.sft --vae-format flux --cfg-scale 1.0  -p "a lovely cat" -r ..\assets\ernie_image\turbo_example.png --diffusion-fa -v --steps 4 -H 2048 -W 2048 --rng cpu
 ```
 
 Before:

docs/pulid.md

@@ -0,0 +1,196 @@
+# PuLID-Flux face-identity preservation
+
+stable-diffusion.cpp supports the [PuLID-Flux](https://github.com/ToTheBeginning/PuLID)
+identity-injection technique on top of Flux.1 (schnell or dev) models.
+Given a single source portrait, PuLID-Flux produces new generations that
+preserve the source person's face across arbitrary scenes, poses, and
+prompts.
+
+Unlike PhotoMaker (which extracts the identity inside the inference
+process from a directory of images), PuLID-Flux's identity extractor is
+a heavy stack (insightface ArcFace + EVA-CLIP-L + IDFormer encoder) that
+is impractical to port to C++/ggml. To keep this implementation small and
+cross-vendor, **stable-diffusion.cpp consumes a precomputed identity
+embedding** produced by an external Python tool that runs once per source
+portrait. Everything downstream of that one-shot extraction is C++ and
+runs on any backend (Vulkan, CUDA, Metal, ROCm, CPU).
+
+## Architecture summary
+
+The PuLID-Flux contribution to the Flux denoise loop is a stack of 20
+small cross-attention modules (`PerceiverAttentionCA`) inserted between
+the Flux transformer blocks:
+
+- After every 2nd of the 19 double-stream blocks (10 hook points)
+- After every 4th of the 38 single-stream blocks (10 hook points)
+
+Each cross-attention layer takes the current image tokens as query, the
+32-token / 2048-dim identity embedding as key+value, and adds its output
+(scaled by `id_weight`, typically 1.0) back to the image tokens.
+
+## Required weights
+
+Three files in addition to the standard Flux weight set:
+
+1. **Flux base** (transformer + VAE + clip_l + t5xxl) -- exactly as
+   [docs/flux.md](flux.md) describes.
+2. **PuLID weights** -- download from
+   [guozinan/PuLID](https://huggingface.co/guozinan/PuLID):
+   - `pulid_flux_v0.9.0.safetensors` or `pulid_flux_v0.9.1.safetensors`
+     (recommended; this implementation is verified against v0.9.1)
+   - **v1.1 (`pulid_v1.1.safetensors`) is NOT yet supported** -- it uses
+     renamed keys (`id_adapter_attn_layers.*` instead of `pulid_ca.*`)
+     and possibly different module structure. Future PR.
+3. **Identity embedding (.pulidembd)** -- produced by the precompute
+   tool below.
+
+## Precompute the identity embedding
+
+The precompute tool runs the PyTorch identity-extraction stack on a
+single portrait image and writes the resulting `(32, 2048)` embedding
+to a `.pulidembd` binary file (about 131 KB). Run it once per source
+person; the same file is reused for any number of generations.
+
+A reference Python script is provided alongside this docs file at
+[`scripts/pulid_extract_id.py`](../scripts/pulid_extract_id.py). It
+requires:
+- A working CUDA / CPU PyTorch stack
+- `insightface`, `facexlib`, `eva-clip`, `torchvision`, `opencv-python`,
+  `huggingface_hub`, `gguf`
+- The PuLID weights file (same one stable-diffusion.cpp will load below)
+- The ToTheBeginning/PuLID repo's `pulid/` package (including
+  `pulid/pipeline_flux.py`) and `eva_clip/` package on `PYTHONPATH`; `flux/`
+  is not needed for embedding extraction
+
+Run it as:
+
+```
+python pulid_extract_id.py \
+  --portrait /path/to/source-photo.jpg \
+  --pulid-weights /path/to/pulid_flux_v0.9.1.safetensors \
+  --out /path/to/source.pulidembd
+```
+
+## Format (gguf)
+
+The embedding is a standard **gguf** container holding a single tensor:
+
+```
+tensor name : "pulid_id"
+shape       : [token_dim, num_tokens]   (ggml order; typically [2048, 32])
+type        : F16 (also accepts F32 / BF16)
+metadata    : general.architecture = "pulid", pulid.version = 1
+```
+
+stable-diffusion.cpp loads it with the normal gguf reader
+(`gguf_init_from_file`) and converts to fp32 at load time -- no bespoke
+parser. Total file size for the typical (32, 2048, fp16) case is ~131 KB.
+
+## Command-line usage
+
+```
+.\bin\Release\sd-cli.exe \
+  --diffusion-model     models\flux1-schnell-Q4_K_S.gguf \
+  --vae                 models\ae.safetensors \
+  --clip_l              models\clip_l.safetensors \
+  --t5xxl               models\t5xxl_fp16.safetensors \
+  --pulid-weights       models\pulid_flux_v0.9.1.safetensors \
+  --pulid-id-embedding  source.pulidembd \
+  --pulid-id-weight     1.0 \
+  -p "candid photograph of a young woman on a beach at sunset" \
+  --cfg-scale 1.0 --sampling-method euler --steps 4 -W 512 -H 512 \
+  --seed 42 --clip-on-cpu \
+  -o out.png
+```
+
+For Flux Dev (instead of Schnell), add `--guidance 3.5` and `--steps 20`.
+
+## Flags
+
+| Flag                       | Purpose                                                           |
+|----------------------------|-------------------------------------------------------------------|
+| `--pulid-weights <path>`   | Path to `pulid_flux_v0.9.x.safetensors`. Loaded with the model.   |
+| `--pulid-id-embedding <p>` | Path to a `.pulidembd` binary produced by the precompute tool.    |
+| `--pulid-id-weight <f>`    | Identity-injection strength. Typical 0.7-1.2; default 1.0.        |
+
+All three flags must be set together to activate PuLID. Setting only
+`--pulid-weights` (no embedding) loads the weights but disables injection
+at runtime. Setting `--pulid-id-weight 0` zeros out the contribution
+(useful for falsification testing: outputs should be byte-identical to
+a no-PuLID run with the same seed).
+
+## Memory budget
+
+At 512x512, 4 steps (Schnell), the 20 cross-attention layers add roughly
+10% to denoise time and almost nothing to peak VRAM. Tested on a 12 GB
+consumer card alongside Flux Schnell Q4 GGUF + CPU-offloaded clip_l and
+t5xxl + GPU-resident VAE.
+
+At 1024x1024 with Flux Dev Q4 + 20 steps + PuLID, the VAE decode compute
+buffer doesn't fit on a 12 GB card even with `--vae-on-cpu`. Workaround:
+explicitly route VAE to the CPU backend instead of the offload flag:
+
+```
+--backend "diffusion=vulkan0,vae=cpu"
+```
+
+The `--vae-on-cpu` flag offloads VAE weights but leaves the compute graph
+on the default backend; this is existing stable-diffusion.cpp behavior,
+not a PuLID-specific issue. Documented here because anyone running PuLID
+at 1024 will hit it.
+
+## Backend selection
+
+The standard `--backend` flag works as documented. Common patterns:
+
+```
+# AMD Vulkan
+--backend "diffusion=vulkan0,vae=cpu"
+
+# NVIDIA Vulkan
+--backend "diffusion=vulkan1,vae=cpu"
+
+# CUDA
+--backend "diffusion=cuda0,vae=cpu"
+```
+
+The PuLID cross-attention layers run on the same backend as the main
+diffusion model. They have not yet been independently profiled on every
+backend; only Vulkan and CPU have been tested by the original contributor.
+
+## Verification
+
+A three-way SHA-256 check is the recommended sanity test when bringing up
+a new combination of model + backend + hardware:
+
+| Run                                          | Expected hash relation             |
+|----------------------------------------------|------------------------------------|
+| A: no `--pulid-*` flags                      | baseline                           |
+| B: PuLID flags, `--pulid-id-weight 0.0`      | **byte-identical to A**            |
+| C: PuLID flags, `--pulid-id-weight 1.0`      | **different from A,B**, preserves source identity |
+
+If A and C differ but A and B differ too, the injection is allocating
+or computing something even at zero weight -- likely a bug.
+
+## Limitations / not yet supported
+
+- **`--skip-layers` (skip-layer-guidance / SLG) combined with PuLID** is not
+  supported. The `pulid_ca` index advances per non-skipped block, so a
+  skipped block silently misaligns the cross-attention weight assignment
+  vs. the trained intervals. The reference PyTorch implementation does
+  not have SLG either, so there is no well-defined behavior to emulate.
+  Use either feature alone.
+- **PuLID v1.1 weights** (`pulid_v1.1.safetensors`, renamed key layout).
+- **Multiple ID images.** The reference PyTorch implementation can fuse
+  several portraits into one embedding for stronger identity. This
+  implementation accepts a single embedding produced from one or more
+  images by the external precompute tool.
+- **Negative-prompt branch of CFG.** PuLID only injects on the positive
+  conditioning path in the published reference, and the implementation
+  here follows that. Flux's distilled guidance doesn't run a separate
+  uncond branch in normal use, so this matters only for `--true-cfg`
+  workflows that aren't standard for Flux.
+- **Backends other than Vulkan and CPU** are untested by the original
+  contributor. The implementation is pure-ggml and should work on CUDA,
+  ROCm, and Metal, but verification by users on those backends is
+  welcomed.

docs/qwen_image.md

@@ -14,7 +14,7 @@
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
 ```
 
 <img alt="qwen example" src="../assets/qwen/example.png" />

docs/qwen_image_edit.md

@@ -23,7 +23,7 @@
 ### Qwen Image Edit
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
 ```
 
 <img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
@@ -32,17 +32,17 @@
 ### Qwen Image Edit 2509
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
 ```
 
 <img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />
 
 ### Qwen Image Edit 2511
 
-To use the new Qwen Image Edit 2511 mode, the  `--qwen-image-zero-cond-t` flag must be enabled; otherwise, image editing quality will degrade significantly.
+To use the new Qwen Image Edit 2511 mode, `--model-args qwen_image_zero_cond_t=true` must be enabled; otherwise, image editing quality will degrade significantly.
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-edit-2511-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'"  --qwen-image-zero-cond-t
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\qwen-image-edit-2511-Q4_K_M.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --model-args qwen_image_zero_cond_t=true
 ```
 
-<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2511.png" />
+<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2511.png" />

docs/sefi_image.md

@@ -0,0 +1,50 @@
+# How to Use
+
+SeFi-Image uses a Flux2-style dual-time transformer (semantic + texture streams), the standard Flux2 VAE, and Qwen3-VL as the LLM text encoder. Tech report: [arXiv:2606.22568](https://arxiv.org/abs/2606.22568).
+
+## Download weights
+
+The SeFi-Image family ships in three scales (1B / 2B / 5B) and three families (Base / RL / turbo), all gated on Hugging Face under https://huggingface.co/SeFi-Image.
+
+- 1B and 2B variants pair with Qwen3-VL-2B-Instruct.
+- 5B variants pair with Qwen3-VL-4B-Instruct.
+- All variants use the standard Flux2 VAE (`flux2_ae.safetensors` from https://huggingface.co/black-forest-labs/FLUX.2-dev).
+
+Convert the transformer and text encoder to sd.cpp safetensors:
+
+```bash
+python3 scripts/convert_sefi.py     <hf_repo_dir>                          <out_dir>/sefi_<scale>_<family>.safetensors
+python3 scripts/convert_qwen3_vl.py  <hf_repo_dir>/Qwen3-VL-XB-Instruct    <out_dir>/qwen3_vl_<X>b.safetensors
+```
+
+## Variant defaults
+
+| Family | timestep_shift_alpha | steps | cfg-scale |
+|---|---|---|---|
+| Base | 0.3 | 50 | 4.0 |
+| RL | 0.3 | 50 | 4.0 |
+| turbo | 1.0 | 4 | 1.0 |
+
+The dispatcher picks `alpha` from the filename (`turbo` substring => 1.0, otherwise 0.3). Override via `--extra-sample-args sefi_alpha=<value>` or `sefi_delta_t=<value>`.
+
+## Examples
+
+### 1B / 2B turbo
+
+```
+./build/bin/sd-cli --diffusion-model /path/to/sefi_1b_turbo.safetensors --vae /path/to/flux2_ae.safetensors --llm /path/to/qwen3_vl_2b.safetensors -p "a photograph of an orange tabby cat sitting on a couch" --cfg-scale 1.0 --steps 4 -W 1024 -H 1024 -s 42 --diffusion-fa --offload-to-cpu -o out.png
+```
+
+### 1B / 2B base
+
+```
+./build/bin/sd-cli --diffusion-model /path/to/sefi_1b_base.safetensors --vae /path/to/flux2_ae.safetensors --llm /path/to/qwen3_vl_2b.safetensors -p "a photograph of an orange tabby cat sitting on a couch" --cfg-scale 4.0 --steps 50 -W 1024 -H 1024 -s 42 --diffusion-fa --offload-to-cpu -o out.png
+```
+
+### 5B (needs streaming on 12 GiB VRAM)
+
+```
+./build/bin/sd-cli --diffusion-model /path/to/sefi_5b_turbo.safetensors --vae /path/to/flux2_ae.safetensors --llm /path/to/qwen3_vl_4b.safetensors -p "a photograph of an orange tabby cat sitting on a couch" --cfg-scale 1.0 --steps 4 -W 1024 -H 1024 -s 42 --diffusion-fa --max-vram 8 --stream-layers --offload-to-cpu -o out.png
+```
+
+<img alt="SeFi-Image 5B turbo example" src="../assets/sefi_image/example.png" />

docs/wan.md

@@ -55,7 +55,7 @@
 ### Wan2.1 T2V 1.3B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -63,7 +63,7 @@
 ### Wan2.1 T2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.1_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -73,7 +73,7 @@
 ### Wan2.1 I2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.1_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -81,7 +81,7 @@
 ### Wan2.2 T2V A14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -89,7 +89,7 @@
 ### Wan2.2 I2V A14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -97,7 +97,7 @@
 ### Wan2.2 T2V A14B T2I
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
 ```
 
 <img width="832" height="480" alt="Wan2 2_14B_t2i" src="../assets/wan/Wan2.2_14B_t2i.png" />
@@ -105,7 +105,7 @@
 ### Wan2.2 T2V 14B with Lora
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\models\loras --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_t2v_lora.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -117,7 +117,7 @@
 #### T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\models\vae\wan2.2_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_5B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -125,7 +125,7 @@
 #### I2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\models\vae\wan2.2_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_5B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -133,7 +133,7 @@
 ### Wan2.1 FLF2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```
 
 
@@ -142,7 +142,7 @@
 ### Wan2.2 FLF2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -152,7 +152,7 @@
 #### T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -161,7 +161,7 @@
 #### R2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -172,7 +172,7 @@
 ```
 mkdir post+depth
 ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -182,7 +182,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -191,7 +191,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### R2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -201,7 +201,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### V2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>

docs/z_image.md

@@ -21,7 +21,7 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 ### Z-Image-Turbo
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512 --steps 8
+.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512 --steps 8
 ```
 
 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
@@ -29,7 +29,7 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 ### Z-Image-Base
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\z_image_bf16.safetensors --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\z_image_bf16.safetensors --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
 ```
 
 <img width="256" alt="z-image example" src="../assets/z_image/base_bf16.png" />

examples/cli/main.cpp

@@ -1,6 +1,7 @@
 #include <stdio.h>
 #include <string.h>
 #include <time.h>
+#include <algorithm>
 #include <cctype>
 #include <filesystem>
 #include <functional>
@@ -53,6 +54,9 @@ struct SDCliParams {
     bool metadata_brief      = false;
     bool metadata_all        = false;
 
+    std::string imatrix_out;
+    std::vector<std::string> imatrix_in;
+
     bool normal_exit = false;
 
     ArgOptions get_options() {
@@ -62,19 +66,28 @@ struct SDCliParams {
             {"-o",
              "--output",
              "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs support .avi, .webm, and animated .webp",
+             0,
              &output_path},
             {"",
              "--image",
              "path to the image to inspect (for metadata mode)",
+             0,
              &image_path},
             {"",
              "--metadata-format",
              "metadata output format, one of [text, json] (default: text)",
+             0,
              &metadata_format},
             {"",
              "--preview-path",
              "path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp",
+             0,
              &preview_path},
+            {"",
+             "--imat-out",
+             "compute the imatrix for this run and save it to the provided path",
+             0,
+             &imatrix_out},
         };
 
         options.int_options = {
@@ -175,6 +188,14 @@ struct SDCliParams {
             return -1;
         };
 
+        auto on_imatrix_in_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
+            }
+            imatrix_in.push_back(argv[index]);
+            return 1;
+        };
+
         options.manual_options = {
             {"-M",
              "--mode",
@@ -188,6 +209,10 @@ struct SDCliParams {
              "--help",
              "show this help message and exit",
              on_help_arg},
+            {"",
+             "--imat-in",
+             "load an imatrix file for quantization or continued collection; can be specified multiple times",
+             on_imatrix_in_arg},
         };
 
         return options;
@@ -249,6 +274,7 @@ struct SDCliParams {
             << "  preview_fps: " << preview_fps << ",\n"
             << "  taesd_preview: " << (taesd_preview ? "true" : "false") << ",\n"
             << "  preview_noisy: " << (preview_noisy ? "true" : "false") << ",\n"
+            << "  imatrix_out: \"" << imatrix_out << "\",\n"
             << "  metadata_raw: " << (metadata_raw ? "true" : "false") << ",\n"
             << "  metadata_brief: " << (metadata_brief ? "true" : "false") << ",\n"
             << "  metadata_all: " << (metadata_all ? "true" : "false") << "\n"
@@ -455,7 +481,8 @@ bool save_results(const SDCliParams& cli_params,
         if (!img.data)
             return false;
 
-        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx;
+        int images_per_batch        = gen_params.batch_count > 0 ? std::max(1, num_results / gen_params.batch_count) : 1;
+        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx / images_per_batch;
         std::string params          = gen_params.embed_image_metadata
                                           ? get_image_params(ctx_params, gen_params, metadata_seed, cli_params.mode)
                                           : "";
@@ -601,13 +628,33 @@ int main(int argc, const char* argv[]) {
     LOG_DEBUG("%s", ctx_params.to_string().c_str());
     LOG_DEBUG("%s", gen_params.to_string().c_str());
 
+    if (!cli_params.imatrix_out.empty()) {
+        if (fs::exists(cli_params.imatrix_out) &&
+            std::find(cli_params.imatrix_in.begin(), cli_params.imatrix_in.end(), cli_params.imatrix_out) == cli_params.imatrix_in.end()) {
+            LOG_WARN("imatrix file '%s' already exists and will be overwritten", cli_params.imatrix_out.c_str());
+        }
+        enable_imatrix_collection();
+    }
+
+    for (const auto& in_file : cli_params.imatrix_in) {
+        LOG_INFO("loading imatrix from '%s'", in_file.c_str());
+        if (!load_imatrix(in_file.c_str())) {
+            LOG_WARN("failed to load imatrix from '%s'", in_file.c_str());
+        }
+    }
+
     if (cli_params.mode == CONVERT) {
-        bool success = convert(ctx_params.model_path.c_str(),
-                               ctx_params.vae_path.c_str(),
-                               cli_params.output_path.c_str(),
-                               ctx_params.wtype,
-                               ctx_params.tensor_type_rules.c_str(),
-                               cli_params.convert_name);
+        bool success = convert_with_components(ctx_params.model_path.c_str(),
+                                               ctx_params.clip_l_path.c_str(),
+                                               ctx_params.clip_g_path.c_str(),
+                                               ctx_params.t5xxl_path.c_str(),
+                                               ctx_params.diffusion_model_path.c_str(),
+                                               ctx_params.vae_path.c_str(),
+                                               cli_params.output_path.c_str(),
+                                               ctx_params.wtype,
+                                               ctx_params.tensor_type_rules.c_str(),
+                                               cli_params.convert_name,
+                                               ctx_params.n_threads);
         if (!success) {
             LOG_ERROR("convert '%s'/'%s' to '%s' failed",
                       ctx_params.model_path.c_str(),
@@ -762,8 +809,12 @@ int main(int argc, const char* argv[]) {
         if (cli_params.mode == IMG_GEN) {
             sd_img_gen_params_t img_gen_params = gen_params.to_sd_img_gen_params_t();
 
-            num_results = gen_params.batch_count;
-            results.adopt(generate_image(sd_ctx.get(), &img_gen_params), num_results);
+            sd_image_t* generated_images = nullptr;
+            if (!generate_image(sd_ctx.get(), &img_gen_params, &generated_images, &num_results)) {
+                generated_images = nullptr;
+                num_results      = 0;
+            }
+            results.adopt(generated_images, num_results);
         } else if (cli_params.mode == VID_GEN) {
             sd_vid_gen_params_t vid_gen_params = gen_params.to_sd_vid_gen_params_t();
             sd_image_t* generated_video        = nullptr;
@@ -798,12 +849,22 @@ int main(int argc, const char* argv[]) {
                 SDImageOwner current_image(results[i]);
                 results[i] = {0, 0, 0, nullptr};
                 for (int u = 0; u < gen_params.upscale_repeats; ++u) {
-                    SDImageOwner upscaled_image(upscale(upscaler_ctx.get(), current_image.get(), upscale_factor));
-                    if (upscaled_image.get().data == nullptr) {
+                    sd_image_t* upscaled_images = nullptr;
+                    int upscaled_count          = 0;
+                    bool upscale_ok             = upscale(upscaler_ctx.get(),
+                                                          current_image.get(),
+                                                          upscale_factor,
+                                                          &upscaled_images,
+                                                          &upscaled_count);
+                    if (!upscale_ok || upscaled_count <= 0 || upscaled_images[0].data == nullptr) {
+                        free_sd_images(upscaled_images, upscaled_count);
                         LOG_ERROR("upscale failed");
                         break;
                     }
-                    current_image = std::move(upscaled_image);
+                    sd_image_t upscaled_image = upscaled_images[0];
+                    upscaled_images[0]        = {0, 0, 0, nullptr};
+                    free_sd_images(upscaled_images, upscaled_count);
+                    current_image.reset(upscaled_image);
                 }
                 results[i] = current_image.release();  // Set the final upscaled image as the result
             }
@@ -815,6 +876,11 @@ int main(int argc, const char* argv[]) {
         return 1;
     }
 
+    if (!cli_params.imatrix_out.empty()) {
+        LOG_INFO("saving imatrix to '%s'", cli_params.imatrix_out.c_str());
+        save_imatrix(cli_params.imatrix_out.c_str());
+    }
+
     free_sd_audio(generated_audio);
 
     return 0;

examples/common/common.cpp

@@ -6,6 +6,7 @@
 #include <cstdlib>
 #include <ctime>
 #include <filesystem>
+#include <fstream>
 #include <iomanip>
 #include <iostream>
 #include <regex>
@@ -260,8 +261,15 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
                         invalid_arg = true;
                         return;
                     }
-                    *option.target = argv_to_utf8(i, argv);
-                    found_arg      = true;
+                    if (option.concat && !option.target->empty()) {
+                        if (option.concat > 0 && option.concat <= 0xff) {
+                            *option.target += static_cast<char>(option.concat);
+                        }
+                        *option.target += argv_to_utf8(i, argv);
+                    } else {
+                        *option.target = argv_to_utf8(i, argv);
+                    }
+                    found_arg = true;
                 }))
                 break;
 
@@ -324,113 +332,165 @@ ArgOptions SDContextParams::get_options() {
         {"-m",
          "--model",
          "path to full model",
+         0,
          &model_path},
         {"",
          "--clip_l",
-         "path to the clip-l text encoder", &clip_l_path},
+         "path to the clip-l text encoder",
+         0,
+         &clip_l_path},
         {"", "--clip_g",
          "path to the clip-g text encoder",
+         0,
          &clip_g_path},
         {"",
          "--clip_vision",
          "path to the clip-vision encoder",
+         0,
          &clip_vision_path},
         {"",
          "--t5xxl",
          "path to the t5xxl text encoder",
+         0,
          &t5xxl_path},
         {"",
          "--llm",
          "path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)",
+         0,
          &llm_path},
         {"",
          "--llm_vision",
          "path to the llm vit",
+         0,
          &llm_vision_path},
         {"",
          "--qwen2vl",
          "alias of --llm. Deprecated.",
+         0,
          &llm_path},
         {"",
          "--qwen2vl_vision",
          "alias of --llm_vision. Deprecated.",
+         0,
          &llm_vision_path},
         {"",
          "--diffusion-model",
          "path to the standalone diffusion model",
+         0,
          &diffusion_model_path},
         {"",
          "--high-noise-diffusion-model",
          "path to the standalone high noise diffusion model",
+         0,
          &high_noise_diffusion_model_path},
         {"",
          "--uncond-diffusion-model",
          "path to the standalone unconditional diffusion model, currently used by Ideogram4 CFG",
+         0,
          &uncond_diffusion_model_path},
         {"",
          "--embeddings-connectors",
          "path to LTXAV embeddings connectors",
+         0,
          &embeddings_connectors_path},
         {"",
          "--vae",
          "path to standalone vae model",
+         0,
          &vae_path},
         {"",
          "--vae-format",
          "VAE latent format override: auto, flux, sd3, or flux2 (default: auto)",
+         0,
          &vae_format},
         {"",
          "--audio-vae",
          "path to standalone LTX audio vae model",
+         0,
          &audio_vae_path},
         {"",
          "--taesd",
          "path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
+         0,
          &taesd_path},
         {"",
          "--tae",
          "alias of --taesd",
+         0,
          &taesd_path},
         {"",
          "--control-net",
          "path to control net model",
+         0,
          &control_net_path},
         {"",
          "--embd-dir",
          "embeddings directory",
+         0,
          &embedding_dir},
         {"",
          "--lora-model-dir",
          "lora model directory",
+         0,
          &lora_model_dir},
         {"",
          "--hires-upscalers-dir",
          "highres fix upscaler model directory",
+         0,
          &hires_upscalers_dir},
         {"",
          "--tensor-type-rules",
          "weight type per tensor pattern (example: \"^vae\\.=f16,model\\.=q8_0\")",
+         (int)',',
          &tensor_type_rules},
+        {"",
+         "--model-args",
+         "extra model args, key=value list. Supports chroma_use_dit_mask, chroma_use_t5_mask, "
+         "chroma_t5_mask_pad, qwen_image_zero_cond_t",
+         (int)',',
+         &model_args},
         {"",
          "--photo-maker",
          "path to PHOTOMAKER model",
+         0,
          &photo_maker_path},
+        {"",
+         "--pulid-weights",
+         "path to PuLID Flux weights",
+         0,
+         &pulid_weights_path},
         {"",
          "--upscale-model",
          "path to esrgan model.",
+         0,
          &esrgan_path},
         {"",
          "--backend",
          "runtime backend assignment, e.g. cpu or clip=cpu,vae=cuda0,diffusion=vulkan0",
+         (int)',',
          &backend},
         {"",
          "--params-backend",
          "parameter backend assignment, e.g. disk, cpu, or diffusion=disk,clip=cpu",
+         (int)',',
          &params_backend},
+        {"",
+         "--split-mode",
+         "weight distribution for modules assigned multiple devices (--backend \"diffusion=cuda0&cuda1\"): "
+         "layer (whole transformer blocks per device, default) or row (matmul rows split across devices, CUDA only). "
+         "Accepts a single mode or per-module assignments, e.g. row or diffusion=row,te=layer",
+         (int)',',
+         &split_mode},
         {"",
          "--rpc-servers",
          "comma-separated list of RPC servers to connect to for offloading, in the format host:port, e.g. localhost:50052,192.168.1.3:50052",
+         (int)',',
          &rpc_servers},
+        {"",
+         "--max-vram",
+         "maximum VRAM budget in GiB for graph-cut segmented execution. Accepts a single value or assignments by backend/device, e.g. 6 or cuda0=6,vulkan0=4. 0 disables graph splitting; a negative value auto-detects free VRAM, sparing the specified value",
+         0,
+         &max_vram},
     };
 
     options.int_options = {
@@ -439,17 +499,6 @@ ArgOptions SDContextParams::get_options() {
          "number of threads to use during computation (default: -1). "
          "If threads <= 0, then threads will be set to the number of CPU physical cores",
          &n_threads},
-        {"",
-         "--chroma-t5-mask-pad",
-         "t5 mask pad size of chroma",
-         &chroma_t5_mask_pad},
-    };
-
-    options.float_options = {
-        {"",
-         "--max-vram",
-         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; a negative value auto-detects free VRAM, sparing the specified value (e.g. -0.5 will keep at least 0.5 GiB free)",
-         &max_vram},
     };
 
     options.bool_options = {
@@ -457,6 +506,16 @@ ArgOptions SDContextParams::get_options() {
          "--stream-layers",
          "enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram; defaults to false)",
          true, &stream_layers},
+        {"",
+         "--eager-load",
+         "load all params into the params backend at model-load time instead of lazily on first use (defaults to false)",
+         true, &eager_load},
+        {"",
+         "--auto-fit",
+         "pick the diffusion/te/vae device placements automatically from the model size and the per-device "
+         "memory budgets (--max-vram; defaults to free memory minus a small margin). Overrides --backend and "
+         "--params-backend; may split modules across GPUs (--split-mode still selects layer or row)",
+         true, &auto_fit},
         {"",
          "--force-sdxl-vae-conv-scale",
          "force use of conv scale on sdxl vae",
@@ -497,30 +556,6 @@ ArgOptions SDContextParams::get_options() {
          "--vae-conv-direct",
          "use ggml_conv2d_direct in the vae model",
          true, &vae_conv_direct},
-        {"",
-         "--circular",
-         "enable circular padding for convolutions",
-         true, &circular},
-        {"",
-         "--circularx",
-         "enable circular RoPE wrapping on x-axis (width) only",
-         true, &circular_x},
-        {"",
-         "--circulary",
-         "enable circular RoPE wrapping on y-axis (height) only",
-         true, &circular_y},
-        {"",
-         "--chroma-disable-dit-mask",
-         "disable dit mask for chroma",
-         false, &chroma_use_dit_mask},
-        {"",
-         "--qwen-image-zero-cond-t",
-         "enable zero_cond_t for qwen image",
-         true, &qwen_image_zero_cond_t},
-        {"",
-         "--chroma-enable-t5-mask",
-         "enable t5 mask for chroma",
-         true, &chroma_use_t5_mask},
     };
 
     auto on_type_arg = [&](int argc, const char** argv, int index) {
@@ -609,7 +644,7 @@ ArgOptions SDContextParams::get_options() {
          on_sampler_rng_arg},
         {"",
          "--prediction",
-         "prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]",
+         "prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, sefi_flow]",
          on_prediction_arg},
         {"",
          "--lora-apply-mode",
@@ -619,6 +654,18 @@ ArgOptions SDContextParams::get_options() {
          "but it usually offers faster inference speed and, in some cases, lower memory usage. "
          "The at_runtime mode, on the other hand, is exactly the opposite.",
          on_lora_apply_mode_arg},
+        {"",
+         "--list-devices",
+         "list available ggml backend devices (one 'name<TAB>description' per line) and exit; "
+         "the names are the device names accepted by --backend and --params-backend",
+         [](int /*argc*/, const char** /*argv*/, int /*index*/) {
+             size_t device_list_size = sd_list_devices(nullptr, 0);
+             std::vector<char> devices(device_list_size + 1);
+             sd_list_devices(devices.data(), devices.size());
+             fputs(devices.data(), stdout);
+             std::exit(0);
+             return 0;
+         }},
     };
 
     return options;
@@ -666,7 +713,18 @@ bool SDContextParams::resolve(SDMode mode) {
 }
 
 bool SDContextParams::validate(SDMode mode) {
-    if (mode != UPSCALE && mode != METADATA && model_path.length() == 0 && diffusion_model_path.length() == 0) {
+    if (mode == CONVERT) {
+        const bool has_convert_input = model_path.length() != 0 ||
+                                       clip_l_path.length() != 0 ||
+                                       clip_g_path.length() != 0 ||
+                                       t5xxl_path.length() != 0 ||
+                                       diffusion_model_path.length() != 0 ||
+                                       vae_path.length() != 0;
+        if (!has_convert_input) {
+            LOG_ERROR("error: convert mode needs at least one model input path\n");
+            return false;
+        }
+    } else if (mode != UPSCALE && mode != METADATA && model_path.length() == 0 && diffusion_model_path.length() == 0) {
         LOG_ERROR("error: the following arguments are required: model_path/diffusion_model\n");
         return false;
     }
@@ -758,10 +816,14 @@ std::string SDContextParams::to_string() const {
         << "  rng_type: " << sd_rng_type_name(rng_type) << ",\n"
         << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
         << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
-        << "  max_vram: " << max_vram << ",\n"
+        << "  max_vram: \"" << max_vram << "\",\n"
         << "  stream_layers: " << (stream_layers ? "true" : "false") << ",\n"
+        << "  eager_load: " << (eager_load ? "true" : "false") << ",\n"
         << "  backend: \"" << backend << "\",\n"
         << "  params_backend: \"" << params_backend << "\",\n"
+        << "  split_mode: \"" << split_mode << "\",\n"
+        << "  model_args: \"" << model_args << "\",\n"
+        << "  auto_fit: " << (auto_fit ? "true" : "false") << ",\n"
         << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
         << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
         << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
@@ -770,13 +832,6 @@ std::string SDContextParams::to_string() const {
         << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
         << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
         << "  vae_conv_direct: " << (vae_conv_direct ? "true" : "false") << ",\n"
-        << "  circular: " << (circular ? "true" : "false") << ",\n"
-        << "  circular_x: " << (circular_x ? "true" : "false") << ",\n"
-        << "  circular_y: " << (circular_y ? "true" : "false") << ",\n"
-        << "  chroma_use_dit_mask: " << (chroma_use_dit_mask ? "true" : "false") << ",\n"
-        << "  qwen_image_zero_cond_t: " << (qwen_image_zero_cond_t ? "true" : "false") << ",\n"
-        << "  chroma_use_t5_mask: " << (chroma_use_t5_mask ? "true" : "false") << ",\n"
-        << "  chroma_t5_mask_pad: " << chroma_t5_mask_pad << ",\n"
         << "  prediction: " << sd_prediction_name(prediction) << ",\n"
         << "  lora_apply_mode: " << sd_lora_apply_mode_name(lora_apply_mode) << ",\n"
         << "  force_sdxl_vae_conv_scale: " << (force_sdxl_vae_conv_scale ? "true" : "false") << "\n"
@@ -815,6 +870,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.embeddings                      = embedding_vec.data();
     sd_ctx_params.embedding_count                 = static_cast<uint32_t>(embedding_vec.size());
     sd_ctx_params.photo_maker_path                = photo_maker_path.c_str();
+    sd_ctx_params.pulid_weights_path              = pulid_weights_path.c_str();
     sd_ctx_params.tensor_type_rules               = tensor_type_rules.c_str();
     sd_ctx_params.n_threads                       = n_threads;
     sd_ctx_params.wtype                           = wtype;
@@ -828,19 +884,17 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.tae_preview_only                = taesd_preview;
     sd_ctx_params.diffusion_conv_direct           = diffusion_conv_direct;
     sd_ctx_params.vae_conv_direct                 = vae_conv_direct;
-    sd_ctx_params.circular_x                      = circular || circular_x;
-    sd_ctx_params.circular_y                      = circular || circular_y;
     sd_ctx_params.force_sdxl_vae_conv_scale       = force_sdxl_vae_conv_scale;
-    sd_ctx_params.chroma_use_dit_mask             = chroma_use_dit_mask;
-    sd_ctx_params.chroma_use_t5_mask              = chroma_use_t5_mask;
-    sd_ctx_params.chroma_t5_mask_pad              = chroma_t5_mask_pad;
-    sd_ctx_params.qwen_image_zero_cond_t          = qwen_image_zero_cond_t;
     sd_ctx_params.vae_format                      = str_to_vae_format(vae_format);
-    sd_ctx_params.max_vram                        = max_vram;
+    sd_ctx_params.max_vram                        = max_vram.c_str();
     sd_ctx_params.stream_layers                   = stream_layers;
+    sd_ctx_params.eager_load                      = eager_load;
     sd_ctx_params.backend                         = effective_backend.c_str();
     sd_ctx_params.params_backend                  = effective_params_backend.c_str();
+    sd_ctx_params.split_mode                      = split_mode.c_str();
+    sd_ctx_params.auto_fit                        = auto_fit;
     sd_ctx_params.rpc_servers                     = rpc_servers.c_str();
+    sd_ctx_params.model_args                      = model_args.empty() ? nullptr : model_args.c_str();
     return sd_ctx_params;
 }
 
@@ -855,54 +909,71 @@ ArgOptions SDGenerationParams::get_options() {
         {"-p",
          "--prompt",
          "the prompt to render",
+         0,
          &prompt},
         {"-n",
          "--negative-prompt",
          "the negative prompt (default: \"\")",
+         0,
          &negative_prompt},
         {"-i",
          "--init-img",
          "path to the init image",
+         0,
          &init_image_path},
         {"",
          "--end-img",
          "path to the end image, required by flf2v",
+         0,
          &end_image_path},
         {"",
          "--mask",
          "path to the mask image",
+         0,
          &mask_image_path},
         {"",
          "--control-image",
          "path to control image, control net",
+         0,
          &control_image_path},
         {"",
          "--control-video",
          "path to control video frames, It must be a directory path. The video frames inside should be stored as images in "
          "lexicographical (character) order. For example, if the control video path is `frames`, the directory contain images "
          "such as 00.png, 01.png, ... etc.",
+         0,
          &control_video_path},
         {"",
          "--pm-id-images-dir",
          "path to PHOTOMAKER input id images dir",
+         0,
          &pm_id_images_dir},
         {"",
          "--pm-id-embed-path",
          "path to PHOTOMAKER v2 id embed",
+         0,
          &pm_id_embed_path},
+        {"",
+         "--pulid-id-embedding",
+         "path to PuLID id embedding",
+         0,
+         &pulid_id_embedding_path},
         {"",
          "--hires-upscaler",
          "highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent (nearest-exact), "
          "Latent (antialiased), Latent (bicubic), Latent (bicubic antialiased), or a model name "
          "under --hires-upscalers-dir (default: Latent)",
+         0,
          &hires_upscaler},
         {"",
          "--extra-sample-args",
-         "extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma",
+         "extra sampler/scheduler/guidance args, key=value list. CFG supports guidance_schedule; APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; flux supports base_shift, max_shift; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma;; logit_normal supports mu, std, logsnr_min, logsnr_max, resolution_aware",
+         (int)',',
          &extra_sample_args},
         {"",
          "--extra-tiling-args",
          "extra VAE tiling args, key=value list. LTX video VAE supports temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)",
+         (int)',',
          &extra_tiling_args},
     };
 
@@ -932,6 +1003,10 @@ ArgOptions SDGenerationParams::get_options() {
          "--batch-count",
          "batch count",
          &batch_count},
+        {"",
+         "--qwen-image-layers",
+         "number of Qwen Image Layered layers; latent/output count is layers + 1 (default: 3)",
+         &qwen_image_layers},
         {"",
          "--video-frames",
          "video frames (default: 1)",
@@ -998,7 +1073,7 @@ ArgOptions SDGenerationParams::get_options() {
          &sample_params.guidance.slg.layer_end},
         {"",
          "--eta",
-         "noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)",
+         "noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde, dpm++2s_a, dpm++2m_sde and dpm++2m_sde_bt)",
          &sample_params.eta},
         {"",
          "--flow-shift",
@@ -1030,7 +1105,7 @@ ArgOptions SDGenerationParams::get_options() {
          &high_noise_sample_params.guidance.slg.layer_end},
         {"",
          "--high-noise-eta",
-         "(high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)",
+         "(high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde, dpm++2s_a, dpm++2m_sde and dpm++2m_sde_bt)",
          &high_noise_sample_params.eta},
         {"",
          "--strength",
@@ -1040,6 +1115,10 @@ ArgOptions SDGenerationParams::get_options() {
          "--pm-style-strength",
          "",
          &pm_style_strength},
+        {"",
+         "--pulid-id-weight",
+         "strength of PuLID identity injection",
+         &pulid_id_weight},
         {"",
          "--control-strength",
          "strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image",
@@ -1077,6 +1156,18 @@ ArgOptions SDGenerationParams::get_options() {
          "disable auto resize of ref images",
          false,
          &auto_resize_ref_image},
+        {"",
+         "--circular",
+         "enable circular padding on both axes for tileable output",
+         true, &circular},
+        {"",
+         "--circularx",
+         "enable circular padding on x-axis (width) only",
+         true, &circular_x},
+        {"",
+         "--circulary",
+         "enable circular padding on y-axis (height) only",
+         true, &circular_y},
         {"",
          "--disable-image-metadata",
          "do not embed generation metadata on image files",
@@ -1354,6 +1445,42 @@ ArgOptions SDGenerationParams::get_options() {
         return 1;
     };
 
+    auto on_prompt_file_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        const char* arg = argv[index];
+        std::ifstream f(arg, std::ios::binary);
+        try {
+            prompt = std::string(std::istreambuf_iterator<char>{f}, {});
+        } catch (const std::ios_base::failure&) {
+            f.setstate(std::ios_base::failbit);
+        }
+        if (f.fail()) {
+            LOG_ERROR("error: failed to read prompt file '%s'\n", arg);
+            return -1;
+        }
+        return 1;
+    };
+
+    auto on_negative_prompt_file_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        const char* arg = argv[index];
+        std::ifstream f(arg, std::ios::binary);
+        try {
+            negative_prompt = std::string(std::istreambuf_iterator<char>{f}, {});
+        } catch (const std::ios_base::failure&) {
+            f.setstate(std::ios_base::failbit);
+        }
+        if (f.fail()) {
+            LOG_ERROR("error: failed to read negative prompt file '%s'\n", arg);
+            return -1;
+        }
+        return 1;
+    };
+
     options.manual_options = {
         {"-s",
          "--seed",
@@ -1361,17 +1488,17 @@ ArgOptions SDGenerationParams::get_options() {
          on_seed_arg},
         {"",
          "--sampling-method",
-         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
+         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, dpm++2m_sde, dpm++2m_sde_bt, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
          on_sample_method_arg},
         {"",
          "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
+         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, dpm++2m_sde, dpm++2m_sde_bt, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          " default: euler for Flux/SD3/Wan, euler_a otherwise",
          on_high_noise_sample_method_arg},
         {"",
          "--scheduler",
-         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default: model-specific",
+         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2, logit_normal, flux2, flux, beta], alias: normal=discrete, default: model-specific",
          on_scheduler_arg},
         {"",
          "--sigmas",
@@ -1417,6 +1544,14 @@ ArgOptions SDGenerationParams::get_options() {
          "--vae-relative-tile-size",
          "relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)",
          on_relative_tile_size_arg},
+        {"",
+         "--prompt-file",
+         "path to the file containing the prompt to render",
+         on_prompt_file_arg},
+        {"",
+         "--negative-prompt-file",
+         "path to the file containing the negative prompt",
+         on_negative_prompt_file_arg},
 
     };
 
@@ -1704,6 +1839,7 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("width", width);
     load_if_exists("height", height);
     load_if_exists("batch_count", batch_count);
+    load_if_exists("qwen_image_layers", qwen_image_layers);
     load_if_exists("video_frames", video_frames);
     load_if_exists("fps", fps);
     load_if_exists("upscale_repeats", upscale_repeats);
@@ -2128,6 +2264,11 @@ bool SDGenerationParams::validate(SDMode mode) {
         return false;
     }
 
+    if (qwen_image_layers < 0) {
+        LOG_ERROR("error: qwen_image_layers must be non-negative");
+        return false;
+    }
+
     if (sample_params.sample_steps <= 0) {
         LOG_ERROR("error: the sample_steps must be greater than 0\n");
         return false;
@@ -2272,6 +2413,11 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
         pm_style_strength,
     };
 
+    sd_pulid_params_t pulid_params = {
+        pulid_id_embedding_path.empty() ? nullptr : pulid_id_embedding_path.c_str(),
+        pulid_id_weight,
+    };
+
     params.loras                 = lora_vec.empty() ? nullptr : lora_vec.data();
     params.lora_count            = static_cast<uint32_t>(lora_vec.size());
     params.prompt                = prompt.c_str();
@@ -2289,9 +2435,11 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.strength              = strength;
     params.seed                  = seed;
     params.batch_count           = batch_count;
+    params.qwen_image_layers     = qwen_image_layers;
     params.control_image         = control_image.get();
     params.control_strength      = control_strength;
     params.pm_params             = pm_params;
+    params.pulid_params          = pulid_params;
     params.vae_tiling_params     = vae_tiling_params;
     params.cache                 = cache_params;
 
@@ -2306,6 +2454,8 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.hires.upscale_tile_size   = hires_upscale_tile_size;
     params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
     params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
+    params.circular_x                = circular || circular_x;
+    params.circular_y                = circular || circular_y;
     return params;
 }
 
@@ -2371,6 +2521,8 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     params.hires.upscale_tile_size   = hires_upscale_tile_size;
     params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
     params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
+    params.circular_x                = circular || circular_x;
+    params.circular_y                = circular || circular_y;
     return params;
 }
 
@@ -2413,6 +2565,7 @@ std::string SDGenerationParams::to_string() const {
         << "  width: " << width << ",\n"
         << "  height: " << height << ",\n"
         << "  batch_count: " << batch_count << ",\n"
+        << "  qwen_image_layers: " << qwen_image_layers << ",\n"
         << "  init_image_path: \"" << init_image_path << "\",\n"
         << "  end_image_path: \"" << end_image_path << "\",\n"
         << "  mask_image_path: \"" << mask_image_path << "\",\n"

examples/common/common.h

@@ -31,6 +31,7 @@ struct StringOption {
     std::string short_name;
     std::string long_name;
     std::string desc;
+    int concat;
     std::string* target;
 };
 
@@ -133,6 +134,7 @@ struct SDContextParams {
     std::string control_net_path;
     std::string embedding_dir;
     std::string photo_maker_path;
+    std::string pulid_weights_path;
     sd_type_t wtype = SD_TYPE_COUNT;
     std::string tensor_type_rules;
     std::string lora_model_dir = ".";
@@ -144,10 +146,14 @@ struct SDContextParams {
     rng_type_t rng_type         = CUDA_RNG;
     rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
     bool offload_params_to_cpu  = false;
-    float max_vram              = 0.f;
+    std::string max_vram        = "0";
     bool stream_layers          = false;
+    bool eager_load             = false;
     std::string backend;
     std::string params_backend;
+    std::string split_mode;
+    std::string model_args;
+    bool auto_fit = false;
     std::string rpc_servers;
     std::string effective_backend;
     std::string effective_params_backend;
@@ -160,16 +166,6 @@ struct SDContextParams {
     bool diffusion_conv_direct = false;
     bool vae_conv_direct       = false;
 
-    bool circular   = false;
-    bool circular_x = false;
-    bool circular_y = false;
-
-    bool chroma_use_dit_mask = true;
-    bool chroma_use_t5_mask  = false;
-    int chroma_t5_mask_pad   = 1;
-
-    bool qwen_image_zero_cond_t = false;
-
     prediction_t prediction           = PREDICTION_COUNT;
     lora_apply_mode_t lora_apply_mode = LORA_APPLY_AUTO;
 
@@ -194,6 +190,7 @@ struct SDGenerationParams {
     int width                  = -1;
     int height                 = -1;
     int batch_count            = 1;
+    int qwen_image_layers      = 3;
     int64_t seed               = 42;
     float strength             = 0.75f;
     float control_strength     = 0.9f;
@@ -234,9 +231,16 @@ struct SDGenerationParams {
     std::string pm_id_embed_path;
     float pm_style_strength = 20.f;
 
+    std::string pulid_id_embedding_path;
+    float pulid_id_weight = 1.0f;
+
     int upscale_repeats   = 1;
     int upscale_tile_size = 128;
 
+    bool circular   = false;
+    bool circular_x = false;
+    bool circular_y = false;
+
     bool hires_enabled         = false;
     std::string hires_upscaler = "Latent";
     std::string hires_upscaler_model_path;

examples/server/async_jobs.cpp

@@ -2,6 +2,7 @@
 
 #include "async_jobs.h"
 
+#include <algorithm>
 #include <iomanip>
 #include <sstream>
 
@@ -173,8 +174,13 @@ bool execute_img_gen_job(ServerRuntime& runtime,
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &params);
-        results.adopt(raw_results, params.batch_count);
+        sd_image_t* raw_results = nullptr;
+        int num_results         = 0;
+        if (!generate_image(runtime.sd_ctx, &params, &raw_results, &num_results)) {
+            raw_results = nullptr;
+            num_results = 0;
+        }
+        results.adopt(raw_results, num_results);
     }
 
     const int num_results = results.count();
@@ -190,6 +196,8 @@ bool execute_img_gen_job(ServerRuntime& runtime,
         encoded_format = EncodedImageFormat::WEBP;
     }
 
+    int batch_count      = job.img_gen.gen_params.batch_count;
+    int images_per_batch = batch_count > 0 ? std::max(1, num_results / batch_count) : 1;
     for (int i = 0; i < num_results; ++i) {
         if (results[i].data == nullptr) {
             continue;
@@ -198,7 +206,7 @@ bool execute_img_gen_job(ServerRuntime& runtime,
         const std::string metadata = job.img_gen.gen_params.embed_image_metadata
                                          ? get_image_params(*runtime.ctx_params,
                                                             job.img_gen.gen_params,
-                                                            job.img_gen.gen_params.seed + i)
+                                                            job.img_gen.gen_params.seed + i / images_per_batch)
                                          : "";
         auto image_bytes           = encode_image_to_vector(encoded_format,
                                                             results[i].data,

examples/server/routes_openai.cpp

@@ -229,8 +229,11 @@ static bool execute_sync_img_gen_request(ServerRuntime& runtime,
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &img_gen_params);
-        num_results             = request.gen_params.batch_count;
+        sd_image_t* raw_results = nullptr;
+        if (!generate_image(runtime.sd_ctx, &img_gen_params, &raw_results, &num_results)) {
+            raw_results = nullptr;
+            num_results = 0;
+        }
         results.adopt(raw_results, num_results);
     }
 
@@ -281,14 +284,16 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
             out["data"]          = json::array();
             out["output_format"] = request.output_format;
 
-            for (int i = 0; i < request.gen_params.batch_count; ++i) {
+            int result_count     = results.count();
+            int images_per_batch = request.gen_params.batch_count > 0 ? std::max(1, result_count / request.gen_params.batch_count) : 1;
+            for (int i = 0; i < result_count; ++i) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
                 std::string params = request.gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params,
                                                             request.gen_params,
-                                                            request.gen_params.seed + i)
+                                                            request.gen_params.seed + i / images_per_batch)
                                          : "";
                 auto image_bytes   = encode_image_to_vector(request.output_format == "jpeg"
                                                                 ? EncodedImageFormat::JPEG
@@ -353,14 +358,16 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
             out["data"]          = json::array();
             out["output_format"] = request.output_format;
 
-            for (int i = 0; i < request.gen_params.batch_count; ++i) {
+            int result_count     = results.count();
+            int images_per_batch = request.gen_params.batch_count > 0 ? std::max(1, result_count / request.gen_params.batch_count) : 1;
+            for (int i = 0; i < result_count; ++i) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
                 std::string params = request.gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params,
                                                             request.gen_params,
-                                                            request.gen_params.seed + i)
+                                                            request.gen_params.seed + i / images_per_batch)
                                          : "";
                 auto image_bytes   = encode_image_to_vector(request.output_format == "jpeg" ? EncodedImageFormat::JPEG : EncodedImageFormat::PNG,
                                                           results[i].data,

examples/server/routes_sdapi.cpp

@@ -63,6 +63,10 @@ static enum sample_method_t get_sdapi_sample_method(std::string name) {
         {"ddim", DDIM_TRAILING_SAMPLE_METHOD},
         {"dpm++ 2m", DPMPP2M_SAMPLE_METHOD},
         {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD},
+        {"dpm++ 2m sde", DPMPP2M_SDE_SAMPLE_METHOD},
+        {"k_dpmpp_2m_sde", DPMPP2M_SDE_SAMPLE_METHOD},
+        {"dpm++ 2m sde gpu", DPMPP2M_SDE_BT_SAMPLE_METHOD},
+        {"k_dpmpp_2m_sde_gpu", DPMPP2M_SDE_BT_SAMPLE_METHOD},
         {"res multistep", RES_MULTISTEP_SAMPLE_METHOD},
         {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
         {"res 2s", RES_2S_SAMPLE_METHOD},
@@ -259,6 +263,48 @@ static bool build_sdapi_img_gen_request(const json& j,
     return true;
 }
 
+static nlohmann::json prepare_info_field(const SDContextParams& ctx_params,
+                                         const SDGenerationParams& gen_params,
+                                         bool img2img) {
+    nlohmann::json jsoninfo = nlohmann::json::object();
+    jsoninfo["prompt"]      = gen_params.prompt;
+    if (!gen_params.negative_prompt.empty()) {
+        jsoninfo["negative_prompt"] = gen_params.negative_prompt;
+    }
+    jsoninfo["seed"]         = gen_params.seed;
+    jsoninfo["cfg_scale"]    = gen_params.sample_params.guidance.txt_cfg;
+    jsoninfo["width"]        = gen_params.get_resolved_width();
+    jsoninfo["height"]       = gen_params.get_resolved_height();
+    jsoninfo["steps"]        = gen_params.sample_params.sample_steps;
+    jsoninfo["sampler_name"] = sd_sample_method_name(gen_params.sample_params.sample_method);
+    if (gen_params.clip_skip != -1) {
+        jsoninfo["clip_skip"] = gen_params.clip_skip;
+    }
+    if (gen_params.sample_params.scheduler != scheduler_t::SCHEDULER_COUNT) {
+        jsoninfo["extra_generation_params"]                  = nlohmann::json::object();
+        jsoninfo["extra_generation_params"]["Schedule type"] = sd_scheduler_name(gen_params.sample_params.scheduler);
+    }
+    if (img2img) {
+        jsoninfo["denoising_strength"] = gen_params.strength;
+    }
+    // not clear what should happen if we have both model and diffusion_model
+    if (!ctx_params.diffusion_model_path.empty()) {
+        jsoninfo["sd_model_name"] = sd_basename(ctx_params.diffusion_model_path);
+    } else if (!ctx_params.model_path.empty()) {
+        jsoninfo["sd_model_name"] = sd_basename(ctx_params.model_path);
+    }
+    if (!ctx_params.vae_path.empty()) {
+        jsoninfo["sd_vae_name"] = sd_basename(ctx_params.vae_path);
+    }
+    jsoninfo["version"] = "stable-diffusion.cpp";
+
+    jsoninfo["infotexts"]            = nlohmann::json::array();
+    jsoninfo["all_prompts"]          = nlohmann::json::array();
+    jsoninfo["all_negative_prompts"] = nlohmann::json::array();
+    jsoninfo["all_seeds"]            = nlohmann::json::array();
+    return jsoninfo;
+}
+
 void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
     ServerRuntime* runtime = &rt;
 
@@ -292,8 +338,11 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             {
                 std::lock_guard<std::mutex> lock(*runtime->sd_ctx_mutex);
-                sd_image_t* raw_results = generate_image(runtime->sd_ctx, &img_gen_params);
-                num_results             = request.gen_params.batch_count;
+                sd_image_t* raw_results = nullptr;
+                if (!generate_image(runtime->sd_ctx, &img_gen_params, &raw_results, &num_results)) {
+                    raw_results = nullptr;
+                    num_results = 0;
+                }
                 results.adopt(raw_results, num_results);
             }
 
@@ -306,24 +355,26 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
             json out;
             out["images"]     = json::array();
             out["parameters"] = j;
-            out["info"]       = "";
+            json jsoninfo     = prepare_info_field(*runtime->ctx_params, request.gen_params, img2img);
 
+            int images_per_batch = request.gen_params.batch_count > 0 ? std::max(1, num_results / request.gen_params.batch_count) : 1;
             for (int i = 0; i < num_results; ++i) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
 
-                std::string params = request.gen_params.embed_image_metadata
-                                         ? get_image_params(*runtime->ctx_params,
-                                                            request.gen_params,
-                                                            request.gen_params.seed + i)
-                                         : "";
-                auto image_bytes   = encode_image_to_vector(EncodedImageFormat::PNG,
-                                                            results[i].data,
-                                                            results[i].width,
-                                                            results[i].height,
-                                                            results[i].channel,
-                                                            params);
+                bool embed_meta = request.gen_params.embed_image_metadata;
+
+                std::string params = get_image_params(*runtime->ctx_params,
+                                                      request.gen_params,
+                                                      request.gen_params.seed + i / images_per_batch);
+
+                auto image_bytes = encode_image_to_vector(EncodedImageFormat::PNG,
+                                                          results[i].data,
+                                                          results[i].width,
+                                                          results[i].height,
+                                                          results[i].channel,
+                                                          embed_meta ? params : "");
 
                 if (image_bytes.empty()) {
                     LOG_ERROR("write image to mem failed");
@@ -331,8 +382,16 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 }
 
                 out["images"].push_back(base64_encode(image_bytes));
+
+                jsoninfo["infotexts"][i]            = params;
+                jsoninfo["all_seeds"][i]            = request.gen_params.seed + i;
+                jsoninfo["all_prompts"][i]          = request.gen_params.prompt;
+                jsoninfo["all_negative_prompts"][i] = request.gen_params.negative_prompt;
             }
 
+            // not a mistake: it is supposed to be a string in json format
+            out["info"] = jsoninfo.dump();
+
             res.set_content(out.dump(), "application/json");
             res.status = 200;
 
@@ -438,6 +497,9 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         scheduler_names.push_back("default");
         for (int i = 0; i < SCHEDULER_COUNT; i++) {
             scheduler_names.push_back(sd_scheduler_name((scheduler_t)i));
+            if (i == DISCRETE_SCHEDULER) {
+                scheduler_names.push_back("normal");
+            }
         }
         json r = json::array();
         for (auto name : scheduler_names) {

examples/server/routes_sdcpp.cpp

@@ -126,6 +126,7 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
         {"strength", defaults.strength},
         {"seed", defaults.seed},
         {"batch_count", defaults.batch_count},
+        {"qwen_image_layers", defaults.qwen_image_layers},
         {"auto_resize_ref_image", defaults.auto_resize_ref_image},
         {"increase_ref_index", defaults.increase_ref_index},
         {"control_strength", defaults.control_strength},
@@ -219,6 +220,9 @@ static json make_capabilities_json(ServerRuntime& runtime) {
 
     for (int i = 0; i < SCHEDULER_COUNT; ++i) {
         schedulers.push_back(sd_scheduler_name((scheduler_t)i));
+        if (i == DISCRETE_SCHEDULER) {
+            schedulers.push_back("normal");
+        }
     }
 
     {

examples/server/runtime.cpp

@@ -190,8 +190,8 @@ ArgOptions SDSvrParams::get_options() {
     ArgOptions options;
 
     options.string_options = {
-        {"-l", "--listen-ip", "server listen ip (default: 127.0.0.1)", &listen_ip},
-        {"", "--serve-html-path", "path to HTML file to serve at root (optional)", &serve_html_path},
+        {"-l", "--listen-ip", "server listen ip (default: 127.0.0.1)", 0, &listen_ip},
+        {"", "--serve-html-path", "path to HTML file to serve at root (optional)", 0, &serve_html_path},
     };
 
     options.int_options = {

include/stable-diffusion.h

@@ -54,6 +54,8 @@ enum sample_method_t {
     EULER_CFG_PP_SAMPLE_METHOD,
     EULER_A_CFG_PP_SAMPLE_METHOD,
     EULER_GE_SAMPLE_METHOD,
+    DPMPP2M_SDE_SAMPLE_METHOD,
+    DPMPP2M_SDE_BT_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -70,6 +72,10 @@ enum scheduler_t {
     LCM_SCHEDULER,
     BONG_TANGENT_SCHEDULER,
     LTX2_SCHEDULER,
+    LOGIT_NORMAL_SCHEDULER,
+    FLUX2_SCHEDULER,
+    FLUX_SCHEDULER,
+    BETA_SCHEDULER,
     SCHEDULER_COUNT
 };
 
@@ -79,7 +85,8 @@ enum prediction_t {
     EDM_V_PRED,
     FLOW_PRED,
     FLUX_FLOW_PRED,
-    FLUX2_FLOW_PRED,
+    SEFI_FLOW_PRED,
+    MINIT2I_FLOW_PRED,
     PREDICTION_COUNT
 };
 
@@ -195,6 +202,7 @@ typedef struct {
     const sd_embedding_t* embeddings;
     uint32_t embedding_count;
     const char* photo_maker_path;
+    const char* pulid_weights_path;
     const char* tensor_type_rules;
     int n_threads;
     enum sd_type_t wtype;
@@ -208,19 +216,17 @@ typedef struct {
     bool tae_preview_only;
     bool diffusion_conv_direct;
     bool vae_conv_direct;
-    bool circular_x;
-    bool circular_y;
     bool force_sdxl_vae_conv_scale;
-    bool chroma_use_dit_mask;
-    bool chroma_use_t5_mask;
-    int chroma_t5_mask_pad;
-    bool qwen_image_zero_cond_t;
     enum sd_vae_format_t vae_format;
-    float max_vram;  // GiB budget for graph-cut segmented param offload (0 = disabled, -1 = auto free VRAM minus 1 GiB)
+    const char* max_vram;  // GiB budget or backend assignment spec for graph-cut segmented param offload (0 = disabled, -1 = auto)
     bool stream_layers;  // Enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram)
+    bool eager_load;  // Load all params into the params backend at model-load time instead of lazily on first use
     const char* backend;
     const char* params_backend;
+    const char* split_mode;  // weight distribution for multi-device modules: layer (default) or row, or per-module assignments e.g. "diffusion=row"
+    bool auto_fit;
     const char* rpc_servers;
+    const char* model_args;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -272,6 +278,11 @@ typedef struct {
     float style_strength;
 } sd_pm_params_t;  // photo maker
 
+typedef struct {
+    const char* id_embedding_path;
+    float id_weight;
+} sd_pulid_params_t;
+
 enum sd_cache_mode_t {
     SD_CACHE_DISABLED = 0,
     SD_CACHE_EASYCACHE,
@@ -364,9 +375,13 @@ typedef struct {
     sd_image_t control_image;
     float control_strength;
     sd_pm_params_t pm_params;
+    sd_pulid_params_t pulid_params;
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
     sd_hires_params_t hires;
+    int qwen_image_layers;
+    bool circular_x;
+    bool circular_y;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -392,17 +407,22 @@ typedef struct {
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
     sd_hires_params_t hires;
+    bool circular_x;
+    bool circular_y;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
+struct ggml_tensor;
 
 typedef void (*sd_log_cb_t)(enum sd_log_level_t level, const char* text, void* data);
 typedef void (*sd_progress_cb_t)(int step, int steps, float time, void* data);
 typedef void (*sd_preview_cb_t)(int step, int frame_count, sd_image_t* frames, bool is_noisy, void* data);
+typedef bool (*sd_graph_eval_callback_t)(struct ggml_tensor* t, bool ask, void* user_data);
 
 SD_API void sd_set_log_callback(sd_log_cb_t sd_log_cb, void* data);
 SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
 SD_API void sd_set_preview_callback(sd_preview_cb_t cb, enum preview_t mode, int interval, bool denoised, bool noisy, void* data);
+SD_API void sd_set_backend_eval_callback(sd_graph_eval_callback_t cb, void* data);
 SD_API int32_t sd_get_num_physical_cores();
 SD_API const char* sd_get_system_info();
 SD_API bool sd_ctx_supports_image_generation(const sd_ctx_t* sd_ctx);
@@ -443,7 +463,21 @@ SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx, enum sa
 
 SD_API void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params);
 SD_API char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params);
-SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params);
+SD_API bool generate_image(sd_ctx_t* sd_ctx,
+                           const sd_img_gen_params_t* sd_img_gen_params,
+                           sd_image_t** images_out,
+                           int* num_images_out);
+
+enum sd_cancel_mode_t {
+    // Stop the current generation as soon as possible.
+    SD_CANCEL_ALL,
+    // Finish the current image sample, then skip additional batch latents and return completed images.
+    SD_CANCEL_NEW_LATENTS,
+    // Clear a pending cancellation request.
+    SD_CANCEL_RESET
+};
+
+SD_API void sd_cancel_generation(sd_ctx_t* sd_ctx, enum sd_cancel_mode_t mode);
 
 SD_API void sd_vid_gen_params_init(sd_vid_gen_params_t* sd_vid_gen_params);
 SD_API bool generate_video(sd_ctx_t* sd_ctx,
@@ -462,9 +496,11 @@ SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
                                         const char* params_backend);
 SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
 
-SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,
-                          sd_image_t input_image,
-                          uint32_t upscale_factor);
+SD_API bool upscale(upscaler_ctx_t* upscaler_ctx,
+                    sd_image_t input_image,
+                    uint32_t upscale_factor,
+                    sd_image_t** images_out,
+                    int* num_images_out);
 
 SD_API int get_upscale_factor(upscaler_ctx_t* upscaler_ctx);
 
@@ -475,6 +511,18 @@ SD_API bool convert(const char* input_path,
                     const char* tensor_type_rules,
                     bool convert_name);
 
+SD_API bool convert_with_components(const char* model_path,
+                                    const char* clip_l_path,
+                                    const char* clip_g_path,
+                                    const char* t5xxl_path,
+                                    const char* diffusion_model_path,
+                                    const char* vae_path,
+                                    const char* output_path,
+                                    enum sd_type_t output_type,
+                                    const char* tensor_type_rules,
+                                    bool convert_name,
+                                    int n_threads);
+
 SD_API bool preprocess_canny(sd_image_t image,
                              float high_threshold,
                              float low_threshold,
@@ -482,9 +530,20 @@ SD_API bool preprocess_canny(sd_image_t image,
                              float strong,
                              bool inverse);
 
+SD_API bool load_imatrix(const char* imatrix_path);
+SD_API void save_imatrix(const char* imatrix_path);
+SD_API void enable_imatrix_collection(void);
+SD_API void disable_imatrix_collection(void);
+
 SD_API const char* sd_commit(void);
 SD_API const char* sd_version(void);
 
+// List available ggml backend devices, one `name<TAB>description` per line.
+// The names are the device names accepted by the --backend / --params-backend
+// assignment specs. Returns the number of bytes required, excluding the null
+// terminator. Passing nullptr or buffer_size 0 only queries the required size.
+SD_API size_t sd_list_devices(char* buffer, size_t buffer_size);
+
 // for C API, caller needs to call free_sd_images to free the memory after use
 // This helps avoid CRT problems on Windows when memory is allocated in the library but freed in the caller, which may use a different CRT.
 SD_API void free_sd_images(sd_image_t* result_images, int num_images);

scripts/convert_qwen3_vl.py

@@ -0,0 +1,112 @@
+#!/usr/bin/env python3
+"""Convert a Qwen3-VL HF safetensors checkpoint into a sd.cpp-loadable form.
+
+The HF dump prefixes text-tower keys with ``model.language_model.`` and
+vision-tower keys with ``model.visual.``. sd.cpp expects ``model.<rest>`` for
+the text side; the vision side is converted by sd.cpp's own
+``convert_qwen3_vl_vision_name`` and is left as-is here.
+
+Operates on raw safetensors bytes so any dtype (BF16/F16/F32) is preserved.
+
+Usage:
+    python3 scripts/convert_qwen3_vl.py <hf_qwen3_vl_dir_or_safetensors> <output.safetensors>
+"""
+
+import argparse
+import json
+import os
+import struct
+import sys
+
+
+def rewrite_key(key: str) -> str:
+    if key.startswith("model.language_model."):
+        return "model." + key[len("model.language_model."):]
+    return key
+
+
+def read_safetensors_header(path: str):
+    with open(path, "rb") as f:
+        hdr_len = struct.unpack("<Q", f.read(8))[0]
+        hdr_bytes = f.read(hdr_len)
+    return json.loads(hdr_bytes), 8 + hdr_len
+
+
+def collect_shard_paths(path: str):
+    if os.path.isdir(path):
+        index_path = os.path.join(path, "model.safetensors.index.json")
+        if os.path.isfile(index_path):
+            with open(index_path) as f:
+                idx = json.load(f)
+            return sorted({os.path.join(path, n) for n in idx["weight_map"].values()})
+        single = os.path.join(path, "model.safetensors")
+        if os.path.isfile(single):
+            return [single]
+        raise FileNotFoundError(f"No Qwen3-VL safetensors in {path}")
+    if os.path.isfile(path):
+        return [path]
+    raise FileNotFoundError(path)
+
+
+def stage_tensors(input_path: str):
+    entries = []
+    for shard_path in collect_shard_paths(input_path):
+        hdr, data_off = read_safetensors_header(shard_path)
+        for key, info in hdr.items():
+            if key == "__metadata__":
+                continue
+            entries.append((rewrite_key(key), shard_path, data_off, info))
+    return entries
+
+
+def write_consolidated(out_path: str, entries):
+    entries = sorted(entries, key=lambda e: e[0])
+
+    new_header = {}
+    cur_offset = 0
+    for new_key, shard_path, data_off, info in entries:
+        start, end = info["data_offsets"]
+        size = end - start
+        new_header[new_key] = {
+            "dtype": info["dtype"],
+            "shape": info["shape"],
+            "data_offsets": [cur_offset, cur_offset + size],
+        }
+        cur_offset += size
+
+    header_json = json.dumps(new_header, separators=(",", ":")).encode("utf-8")
+    pad = (-len(header_json)) % 8
+    header_json = header_json + (b" " * pad)
+
+    with open(out_path, "wb") as out:
+        out.write(struct.pack("<Q", len(header_json)))
+        out.write(header_json)
+        for new_key, shard_path, data_off, info in entries:
+            start, end = info["data_offsets"]
+            with open(shard_path, "rb") as src:
+                src.seek(data_off + start)
+                remaining = end - start
+                while remaining > 0:
+                    chunk = src.read(min(8 * 1024 * 1024, remaining))
+                    if not chunk:
+                        raise IOError(f"Truncated tensor in {shard_path}")
+                    out.write(chunk)
+                    remaining -= len(chunk)
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("input", help="HF Qwen3-VL directory or single safetensors file")
+    parser.add_argument("output", help="Output single safetensors path")
+    args = parser.parse_args()
+
+    entries = stage_tensors(args.input)
+    print(f"Tensors: {len(entries)}")
+    print(f"Writing -> {args.output}")
+    os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
+    write_consolidated(args.output, entries)
+    print(f"Done. Output size: {os.path.getsize(args.output) / 1e9:.2f} GB")
+
+
+if __name__ == "__main__":
+    main()

scripts/convert_sefi.py

@@ -0,0 +1,279 @@
+#!/usr/bin/env python3
+"""Convert a SeFi-Image diffusers checkpoint into a single sd.cpp-compatible safetensors.
+
+Operates on raw safetensors bytes so any dtype (BF16, F32, ...) is preserved exactly.
+No numpy or torch dependency required.
+
+Usage:
+    python3 scripts/convert_sefi.py <sefi_diffusers_dir> <output.safetensors>
+"""
+
+import argparse
+import json
+import os
+import re
+import struct
+import sys
+
+
+_LINEAR_TO_LIN = re.compile(r"\.linear\.")
+_SHARED_MOD_PREFIXES = (
+    "double_stream_modulation_img",
+    "double_stream_modulation_txt",
+    "single_stream_modulation",
+)
+
+
+def rewrite_transformer_key(key: str) -> str:
+    if key.startswith("backbone."):
+        key = key[len("backbone."):]
+    elif key.startswith("dual_time_embed."):
+        return key
+
+    if any(key.startswith(prefix + ".") for prefix in _SHARED_MOD_PREFIXES):
+        key = _LINEAR_TO_LIN.sub(".lin.", key, count=1)
+
+    if key == "context_embedder.weight":
+        return "txt_in.weight"
+    if key == "context_embedder.bias":
+        return "txt_in.bias"
+    if key == "x_embedder.weight":
+        return "img_in.weight"
+    if key == "x_embedder.bias":
+        return "img_in.bias"
+
+    if key == "proj_out.weight":
+        return "final_layer.linear.weight"
+    if key == "proj_out.bias":
+        return "final_layer.linear.bias"
+    if key == "norm_out.linear.weight":
+        return "final_layer.adaLN_modulation.1.weight"
+    if key == "norm_out.linear.bias":
+        return "final_layer.adaLN_modulation.1.bias"
+
+    m = re.match(r"transformer_blocks\.(\d+)\.(.*)$", key)
+    if m:
+        return _rewrite_double_stream(m.group(1), m.group(2))
+    m = re.match(r"single_transformer_blocks\.(\d+)\.(.*)$", key)
+    if m:
+        return _rewrite_single_stream(m.group(1), m.group(2))
+
+    return key
+
+
+def _rewrite_double_stream(idx: str, tail: str) -> str:
+    dst = f"double_blocks.{idx}."
+    mapping = {
+        "norm1.linear.weight":          "img_mod.lin.weight",
+        "norm1_context.linear.weight":  "txt_mod.lin.weight",
+        "attn.norm_q.weight":           "img_attn.norm.query_norm.scale",
+        "attn.norm_k.weight":           "img_attn.norm.key_norm.scale",
+        "attn.norm_added_q.weight":     "txt_attn.norm.query_norm.scale",
+        "attn.norm_added_k.weight":     "txt_attn.norm.key_norm.scale",
+        "attn.to_out.0.weight":         "img_attn.proj.weight",
+        "attn.to_add_out.weight":       "txt_attn.proj.weight",
+        "ff.net.0.proj.weight":         "img_mlp.0.weight",
+        "ff.net.2.weight":              "img_mlp.2.weight",
+        "ff_context.net.0.proj.weight": "txt_mlp.0.weight",
+        "ff_context.net.2.weight":      "txt_mlp.2.weight",
+        "ff.linear_in.weight":          "img_mlp.0.weight",
+        "ff.linear_out.weight":         "img_mlp.2.weight",
+        "ff_context.linear_in.weight":  "txt_mlp.0.weight",
+        "ff_context.linear_out.weight": "txt_mlp.2.weight",
+    }
+    return dst + mapping.get(tail, tail)
+
+
+# QKV triplets to fuse on output: source tails -> target fused tail.
+# Each tuple is (q_tail, k_tail, v_tail, fused_target_tail).
+QKV_DOUBLE_TRIPLETS = [
+    ("attn.to_q.weight",       "attn.to_k.weight",       "attn.to_v.weight",       "img_attn.qkv.weight"),
+    ("attn.add_q_proj.weight", "attn.add_k_proj.weight", "attn.add_v_proj.weight", "txt_attn.qkv.weight"),
+]
+
+
+def _rewrite_single_stream(idx: str, tail: str) -> str:
+    dst = f"single_blocks.{idx}."
+    mapping = {
+        "norm.linear.weight":          "modulation.lin.weight",
+        "attn.norm_q.weight":          "norm.query_norm.scale",
+        "attn.norm_k.weight":          "norm.key_norm.scale",
+        "attn.to_qkv_mlp_proj.weight": "linear1.weight",
+        "attn.to_out.weight":          "linear2.weight",
+    }
+    return dst + mapping.get(tail, tail)
+
+
+
+
+def read_safetensors_header(path: str):
+    """Return (header dict, data start byte offset)."""
+    with open(path, "rb") as f:
+        hdr_len = struct.unpack("<Q", f.read(8))[0]
+        hdr_bytes = f.read(hdr_len)
+    return json.loads(hdr_bytes), 8 + hdr_len
+
+
+def collect_shard_paths(directory: str, weight_pattern: str):
+    index_path = os.path.join(directory, f"{weight_pattern}.safetensors.index.json")
+    if os.path.isfile(index_path):
+        with open(index_path) as f:
+            idx = json.load(f)
+        return sorted({os.path.join(directory, n) for n in idx["weight_map"].values()})
+    single = os.path.join(directory, f"{weight_pattern}.safetensors")
+    if not os.path.isfile(single):
+        raise FileNotFoundError(f"No checkpoint at {directory}: missing {weight_pattern}")
+    return [single]
+
+
+def stage_tensors_for_section(section_dir: str, rewrite_fn):
+    """Return a list of (new_key, shard_path, data_start_offset, info_dict) entries.
+
+    A "qkv_fuse" pseudo-entry with three source descriptors is emitted when a
+    transformer_blocks.* split q/k/v triplet is found, so the writer can fuse
+    them into a single output tensor.
+    """
+    entries = []
+    # First, index all raw keys per shard so we can detect qkv triplets.
+    raw_by_block = {}  # block_idx -> {tail: (key, shard_path, data_off, info)}
+    raw_others = []
+    for shard_path in collect_shard_paths(section_dir, "diffusion_pytorch_model"):
+        hdr, data_off = read_safetensors_header(shard_path)
+        for key, info in hdr.items():
+            if key == "__metadata__":
+                continue
+            m = re.match(r"backbone\.transformer_blocks\.(\d+)\.(.*)$", key)
+            if m and any(m.group(2) in trip[:3] for trip in QKV_DOUBLE_TRIPLETS):
+                idx = m.group(1)
+                raw_by_block.setdefault(idx, {})[m.group(2)] = (key, shard_path, data_off, info)
+            else:
+                raw_others.append((key, shard_path, data_off, info))
+
+    for key, shard_path, data_off, info in raw_others:
+        new_key = rewrite_fn(key)
+        # Swap the (scale, shift) halves to (shift, scale) at conversion time so
+        # the on-disk weight matches BFL flux ordering and the runtime stays
+        # version-agnostic. norm_out.linear weight shape is [2*dim, dim] and bias
+        # is [2*dim]; both split along axis 0 (outermost == row-major outer).
+        if new_key in ("final_layer.adaLN_modulation.1.weight",
+                       "final_layer.adaLN_modulation.1.bias"):
+            info = dict(info)
+            info["_chunk_swap_halves"] = True
+        entries.append((new_key, shard_path, data_off, info))
+
+    for block_idx, tails in raw_by_block.items():
+        for q_tail, k_tail, v_tail, fused_tail in QKV_DOUBLE_TRIPLETS:
+            if q_tail in tails and k_tail in tails and v_tail in tails:
+                q = tails[q_tail]; k = tails[k_tail]; v = tails[v_tail]
+                # Validate shapes match.
+                q_shape = q[3]["shape"]; k_shape = k[3]["shape"]; v_shape = v[3]["shape"]
+                if q_shape != k_shape or q_shape != v_shape:
+                    raise ValueError(f"qkv shape mismatch at block {block_idx} {q_tail}: q={q_shape} k={k_shape} v={v_shape}")
+                fused_shape = [q_shape[0] * 3] + list(q_shape[1:])
+                fused_info = {
+                    "dtype": q[3]["dtype"],
+                    "shape": fused_shape,
+                    "_qkv_sources": [q, k, v],  # pseudo field consumed by writer
+                }
+                entries.append((f"double_blocks.{block_idx}.{fused_tail}",
+                                None, None, fused_info))
+                del tails[q_tail]; del tails[k_tail]; del tails[v_tail]
+        # Anything left in tails was an unmatched single - pass through.
+        for tail, payload in tails.items():
+            entries.append((rewrite_fn(payload[0]),) + payload[1:])
+    return entries
+
+
+_DTYPE_BYTES = {
+    "BF16": 2, "F16": 2, "F32": 4, "F64": 8,
+    "U8": 1, "I8": 1, "I16": 2, "I32": 4, "I64": 8,
+    "BOOL": 1,
+}
+
+
+def _total_bytes(info: dict) -> int:
+    if "_qkv_sources" in info:
+        elems = 1
+        for d in info["shape"]:
+            elems *= d
+        return elems * _DTYPE_BYTES[info["dtype"]]
+    start, end = info["data_offsets"]
+    return end - start
+
+
+def write_consolidated(out_path: str, entries):
+    """Write a single safetensors file by streaming raw bytes from each shard.
+
+    For qkv-fused entries, q/k/v are concatenated along axis 0 (row-major), so a
+    simple byte-level concatenation produces the correct fused layout for any
+    standard dtype.
+    """
+    entries = sorted(entries, key=lambda e: e[0])
+
+    new_header = {}
+    cur_offset = 0
+    for new_key, shard_path, data_off, info in entries:
+        size = _total_bytes(info)
+        new_header[new_key] = {
+            "dtype": info["dtype"],
+            "shape": info["shape"],
+            "data_offsets": [cur_offset, cur_offset + size],
+        }
+        cur_offset += size
+
+    header_json = json.dumps(new_header, separators=(",", ":")).encode("utf-8")
+    pad = (-len(header_json)) % 8
+    header_json = header_json + (b" " * pad)
+
+    def copy_range(src_path, src_data_off, src_info, out, byte_range=None):
+        start, end = src_info["data_offsets"]
+        if byte_range is not None:
+            sub_start, sub_end = byte_range
+            start, end = start + sub_start, start + sub_end
+        with open(src_path, "rb") as src:
+            src.seek(src_data_off + start)
+            remaining = end - start
+            while remaining > 0:
+                chunk = src.read(min(8 * 1024 * 1024, remaining))
+                if not chunk:
+                    raise IOError(f"Truncated tensor in {src_path}")
+                out.write(chunk)
+                remaining -= len(chunk)
+
+    with open(out_path, "wb") as out:
+        out.write(struct.pack("<Q", len(header_json)))
+        out.write(header_json)
+        for new_key, shard_path, data_off, info in entries:
+            if "_qkv_sources" in info:
+                for q_entry in info["_qkv_sources"]:
+                    _, src_path, src_data_off, src_info = q_entry
+                    copy_range(src_path, src_data_off, src_info, out)
+            elif info.get("_chunk_swap_halves"):
+                size = _total_bytes(info)
+                half = size // 2
+                if size != half * 2:
+                    raise ValueError(f"{new_key}: odd byte size {size} cannot be split into halves")
+                copy_range(shard_path, data_off, info, out, byte_range=(half, size))
+                copy_range(shard_path, data_off, info, out, byte_range=(0, half))
+            else:
+                copy_range(shard_path, data_off, info, out)
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("input_dir", help="SeFi diffusers checkpoint directory")
+    parser.add_argument("output", help="Output transformer safetensors path (load via --diffusion-model)")
+    args = parser.parse_args()
+
+    transformer_entries = stage_tensors_for_section(
+        os.path.join(args.input_dir, "transformer"), rewrite_transformer_key)
+
+    print(f"Transformer tensors: {len(transformer_entries)}")
+    print(f"Writing {len(transformer_entries)} tensors -> {args.output}")
+    os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
+    write_consolidated(args.output, transformer_entries)
+    print(f"Done. Output size: {os.path.getsize(args.output) / 1e9:.2f} GB")
+
+
+if __name__ == "__main__":
+    main()

scripts/format-code.ps1

@@ -31,24 +31,30 @@ $patterns = @(
     "examples/common/*.cpp"
 )
 
-$root = (Get-Location).Path
+Push-Location (Join-Path $PSScriptRoot "..")
 
-foreach ($pattern in $patterns) {
-    $files = Get-ChildItem -Path $pattern -File -ErrorAction SilentlyContinue | Sort-Object FullName
+try {
+    $root = (Get-Location).Path
 
-    foreach ($file in $files) {
-        $relativePath = $file.FullName.Substring($root.Length).TrimStart('\', '/') -replace '\\', '/'
+    foreach ($pattern in $patterns) {
+        $files = Get-ChildItem -Path $pattern -File -ErrorAction SilentlyContinue | Sort-Object FullName
 
-        if ($relativePath -like "vocab*") {
-            continue
+        foreach ($file in $files) {
+            $relativePath = $file.FullName.Substring($root.Length).TrimStart('\', '/') -replace '\\', '/'
+
+            if ($relativePath -like "vocab*") {
+                continue
+            }
+
+            Write-Host "formatting '$relativePath'"
+
+            # if ($relativePath -ne "stable-diffusion.h") {
+            #     clang-tidy -fix -p build_linux/ "$relativePath"
+            # }
+
+            & clang-format -style=file -i $relativePath
         }
-
-        Write-Host "formatting '$relativePath'"
-
-        # if ($relativePath -ne "stable-diffusion.h") {
-        #     clang-tidy -fix -p build_linux/ "$relativePath"
-        # }
-
-        & clang-format -style=file -i $relativePath
     }
+} finally {
+    Pop-Location
 }

scripts/format-code.sh

@@ -1,3 +1,10 @@
+#!/usr/bin/env bash
+
+set -e
+
+repo_root="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$repo_root"
+
 for f in src/*.cpp src/*.h src/*.hpp \
          src/conditioning/*.cpp src/conditioning/*.h src/conditioning/*.hpp \
          src/core/*.cpp src/core/*.h src/core/*.hpp \

scripts/pulid_extract_id.py

@@ -0,0 +1,134 @@
+"""
+Precompute a PuLID-Flux identity embedding from a single source portrait.
+
+Writes a gguf file (a single tensor `pulid_id`) that stable-diffusion.cpp's
+`--pulid-id-embedding` flag consumes.
+
+Dependencies (recommended: vendor rather than pip-install due to upstream
+packaging quirks):
+  - torch + safetensors
+  - The ToTheBeginning/PuLID repository's `pulid/` package and `eva_clip/`.
+    Put them on PYTHONPATH or sys.path before running this script.
+  - insightface, facexlib, torchvision, opencv-python, huggingface_hub, gguf
+  - numpy, Pillow
+
+Usage:
+  python scripts/pulid_extract_id.py \\
+    --portrait /path/to/source-photo.jpg \\
+    --pulid-weights /path/to/pulid_flux_v0.9.1.safetensors \\
+    --out /path/to/source.pulidembd
+
+The portrait must contain a clearly visible face. insightface's antelopev2
+detector will be auto-downloaded on first run.
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from types import SimpleNamespace
+
+
+def extract(portrait_path: str, pulid_weights: str) -> "torch.Tensor":
+    import numpy as np
+    import torch
+    from PIL import Image
+    from pulid.pipeline_flux import PuLIDPipeline
+
+    if torch.cuda.is_available():
+        device, onnx_provider = "cuda", "gpu"
+    else:
+        device, onnx_provider = "cpu", "cpu"
+
+    print(f"device={device}", flush=True)
+
+    # PuLIDPipeline only attaches pulid_ca attributes to `dit` during
+    # construction; get_id_embedding() never runs Flux, so a dummy object is
+    # enough and avoids importing/building a Flux skeleton.
+    print("instantiating PuLIDPipeline with a dummy Flux object", flush=True)
+    dit = SimpleNamespace()
+    pulid = PuLIDPipeline(dit=dit,
+                          device=device,
+                          weight_dtype=torch.bfloat16,
+                          onnx_provider=onnx_provider)
+
+    print(f"loading PuLID weights from {pulid_weights}", flush=True)
+    pulid.load_pretrain(pretrain_path=pulid_weights, version="v0.9.1")
+
+    print(f"extracting ID embedding from {portrait_path}", flush=True)
+    face_img = np.array(Image.open(portrait_path).convert("RGB"))
+    id_embedding, _ = pulid.get_id_embedding(face_img)
+    print(f"id embedding shape={tuple(id_embedding.shape)} dtype={id_embedding.dtype}",
+          flush=True)
+
+    if id_embedding.ndim == 3 and id_embedding.shape[0] == 1:
+        id_embedding = id_embedding[0]
+    return id_embedding
+
+
+def write_embd(tensor, out_path: str, dtype_choice: str) -> None:
+    import gguf
+    import torch
+
+    if tensor.ndim != 2:
+        raise ValueError(f"expected (num_tokens, token_dim); got {tuple(tensor.shape)}")
+    num_tokens, token_dim = tensor.shape
+
+    os.makedirs(os.path.dirname(out_path) or ".", exist_ok=True)
+
+    writer = gguf.GGUFWriter(out_path, arch="pulid")
+    writer.add_uint32("pulid.version", 1)
+
+    if dtype_choice == "fp16":
+        arr = tensor.to(torch.float16).contiguous().cpu().numpy()
+        writer.add_tensor("pulid_id", arr)
+    elif dtype_choice == "fp32":
+        arr = tensor.to(torch.float32).contiguous().cpu().numpy()
+        writer.add_tensor("pulid_id", arr)
+    elif dtype_choice == "bf16":
+        raw = tensor.to(torch.bfloat16).contiguous().view(torch.uint16).cpu().numpy()
+        writer.add_tensor("pulid_id", raw,
+                          raw_shape=(int(num_tokens), int(token_dim)),
+                          raw_dtype=gguf.GGMLQuantizationType.BF16)
+    else:
+        raise ValueError(f"unknown --dtype {dtype_choice}")
+
+    writer.write_header_to_file()
+    writer.write_kv_data_to_file()
+    writer.write_tensors_to_file()
+    writer.close()
+
+    print(f"wrote {out_path}: gguf, tensor pulid_id [{token_dim}, {num_tokens}] {dtype_choice}",
+          flush=True)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(
+        description=__doc__,
+        formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--portrait", required=True,
+                    help="Path to the source portrait image (JPG/PNG).")
+    ap.add_argument("--pulid-weights", required=True,
+                    help="Path to pulid_flux_v0.9.x.safetensors.")
+    ap.add_argument("--out", required=True,
+                    help="Output path for the .pulidembd binary.")
+    ap.add_argument("--dtype", default="fp16",
+                    choices=["fp16", "bf16", "fp32"],
+                    help="Storage dtype (default fp16; produces ~131 KB).")
+    args = ap.parse_args()
+
+    if not os.path.exists(args.portrait):
+        print(f"ERROR: portrait not found at {args.portrait}", file=sys.stderr)
+        return 2
+    if not os.path.exists(args.pulid_weights):
+        print(f"ERROR: PuLID weights not found at {args.pulid_weights}", file=sys.stderr)
+        return 3
+
+    embedding = extract(args.portrait, args.pulid_weights)
+    write_embd(embedding, args.out, args.dtype)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/remove_utf8_bom.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+"""Remove UTF-8 BOMs from files under a directory.
+
+By default this scans the current working directory recursively and skips
+repository areas that should not be touched by ordinary maintenance scripts.
+Only files whose first three bytes are the UTF-8 BOM are rewritten.
+"""
+
+import argparse
+import os
+import shutil
+import sys
+import tempfile
+from pathlib import Path
+
+
+UTF8_BOM = b"\xef\xbb\xbf"
+
+DEFAULT_EXCLUDED_DIR_NAMES = {
+    ".git",
+    ".hg",
+    ".svn",
+    ".mypy_cache",
+    ".pytest_cache",
+    "__pycache__",
+    "test",
+}
+
+DEFAULT_EXCLUDED_DIR_PREFIXES = {
+    "build",
+}
+
+DEFAULT_EXCLUDED_REL_DIRS = {
+    "examples/server/frontend",
+    "ggml",
+    "models",
+    "src/vocab",
+    "thirdparty",
+}
+
+
+def rel_posix(path: Path, root: Path) -> str:
+    try:
+        return path.relative_to(root).as_posix()
+    except ValueError:
+        return path.as_posix()
+
+
+def should_skip_dir(
+    path: Path,
+    root: Path,
+    excluded_rel_dirs: set[str],
+    excluded_names: set[str],
+    excluded_prefixes: set[str],
+) -> bool:
+    rel = rel_posix(path, root)
+    return (
+        path.name in excluded_names
+        or rel in excluded_rel_dirs
+        or any(path.name.startswith(prefix) for prefix in excluded_prefixes)
+    )
+
+
+def iter_files(
+    root: Path,
+    recursive: bool,
+    excluded_rel_dirs: set[str],
+    excluded_names: set[str],
+    excluded_prefixes: set[str],
+    follow_symlinks: bool,
+):
+    if recursive:
+        for dirpath, dirnames, filenames in os.walk(root, followlinks=follow_symlinks):
+            current_dir = Path(dirpath)
+            dirnames[:] = [
+                name
+                for name in dirnames
+                if not should_skip_dir(
+                    current_dir / name,
+                    root,
+                    excluded_rel_dirs,
+                    excluded_names,
+                    excluded_prefixes,
+                )
+            ]
+            for filename in filenames:
+                path = current_dir / filename
+                if path.is_symlink() and not follow_symlinks:
+                    continue
+                yield path
+    else:
+        for path in root.iterdir():
+            if path.is_file() and (follow_symlinks or not path.is_symlink()):
+                yield path
+
+
+def has_utf8_bom(path: Path) -> bool:
+    with path.open("rb") as f:
+        return f.read(len(UTF8_BOM)) == UTF8_BOM
+
+
+def strip_utf8_bom(path: Path) -> None:
+    tmp_path = None
+    try:
+        with path.open("rb") as src:
+            if src.read(len(UTF8_BOM)) != UTF8_BOM:
+                return
+
+            fd, tmp_name = tempfile.mkstemp(
+                prefix=f".{path.name}.",
+                suffix=".tmp",
+                dir=str(path.parent),
+            )
+            tmp_path = Path(tmp_name)
+            with os.fdopen(fd, "wb") as dst:
+                shutil.copyfileobj(src, dst, length=1024 * 1024)
+
+        shutil.copystat(path, tmp_path, follow_symlinks=False)
+        os.replace(tmp_path, path)
+        tmp_path = None
+    finally:
+        if tmp_path is not None:
+            try:
+                tmp_path.unlink()
+            except FileNotFoundError:
+                pass
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Scan files and convert UTF-8 BOM files to UTF-8 without BOM.",
+    )
+    parser.add_argument(
+        "root",
+        nargs="?",
+        default=".",
+        help="Directory to scan. Defaults to the current directory.",
+    )
+    parser.add_argument(
+        "-n",
+        "--dry-run",
+        action="store_true",
+        help="Only list files that would be converted.",
+    )
+    parser.add_argument(
+        "--no-recursive",
+        action="store_true",
+        help="Only scan files directly under root.",
+    )
+    parser.add_argument(
+        "--include-repo-excluded",
+        action="store_true",
+        help="Do not skip default repository excluded directories.",
+    )
+    parser.add_argument(
+        "--exclude-dir",
+        action="append",
+        default=[],
+        metavar="DIR",
+        help="Additional directory name or root-relative path to skip. Can be used multiple times.",
+    )
+    parser.add_argument(
+        "--follow-symlinks",
+        action="store_true",
+        help="Follow symlinked directories and files.",
+    )
+    parser.add_argument(
+        "-q",
+        "--quiet",
+        action="store_true",
+        help="Only print the final summary.",
+    )
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    root = Path(args.root).resolve()
+
+    if not root.is_dir():
+        print(f"error: not a directory: {root}", file=sys.stderr)
+        return 2
+
+    excluded_names = set()
+    excluded_rel_dirs = set()
+    excluded_prefixes = set()
+    if not args.include_repo_excluded:
+        excluded_names.update(DEFAULT_EXCLUDED_DIR_NAMES)
+        excluded_rel_dirs.update(DEFAULT_EXCLUDED_REL_DIRS)
+        excluded_prefixes.update(DEFAULT_EXCLUDED_DIR_PREFIXES)
+
+    for item in args.exclude_dir:
+        normalized = Path(item).as_posix().strip("/")
+        if "/" in normalized:
+            excluded_rel_dirs.add(normalized)
+        else:
+            excluded_names.add(normalized)
+
+    scanned = 0
+    converted = 0
+    errors = 0
+
+    for path in iter_files(
+        root=root,
+        recursive=not args.no_recursive,
+        excluded_rel_dirs=excluded_rel_dirs,
+        excluded_names=excluded_names,
+        excluded_prefixes=excluded_prefixes,
+        follow_symlinks=args.follow_symlinks,
+    ):
+        scanned += 1
+        try:
+            if not has_utf8_bom(path):
+                continue
+            converted += 1
+            rel = rel_posix(path, root)
+            if args.dry_run:
+                if not args.quiet:
+                    print(f"would convert: {rel}")
+            else:
+                strip_utf8_bom(path)
+                if not args.quiet:
+                    print(f"converted: {rel}")
+        except OSError as exc:
+            errors += 1
+            print(f"error: {rel_posix(path, root)}: {exc}", file=sys.stderr)
+
+    action = "would convert" if args.dry_run else "converted"
+    print(f"scanned {scanned} file(s), {action} {converted}, errors {errors}")
+    return 1 if errors else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/conditioning/conditioner.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_CONDITIONING_CONDITIONER_HPP__
+#ifndef __SD_CONDITIONING_CONDITIONER_HPP__
 #define __SD_CONDITIONING_CONDITIONER_HPP__
 
 #include <cmath>
@@ -6,6 +6,7 @@
 #include <optional>
 
 #include "core/tensor_ggml.hpp"
+#include "core/util.h"
 #include "model/te/clip.hpp"
 #include "model/te/llm.hpp"
 #include "model/te/t5.hpp"
@@ -116,6 +117,8 @@ public:
     virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)           = 0;
     virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
     virtual void set_stream_layers_enabled(bool enabled) {}
+    virtual void set_runtime_backends(const std::vector<ggml_backend_t>& backends) {}
+    virtual void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {}
     virtual void set_flash_attention_enabled(bool enabled) = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual void runner_done() {}
@@ -142,8 +145,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                       std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
         : version(version), tokenizer(sd_version_is_sd2(version) ? 0 : 49407) {
         for (const auto& kv : orig_embedding_map) {
-            std::string name = kv.first;
-            std::transform(name.begin(), name.end(), name.begin(), [](unsigned char c) { return std::tolower(c); });
+            std::string name    = normalize_embedding_name(kv.first);
             embedding_map[name] = kv.second;
             tokenizer.add_special_token(name);
         }
@@ -278,17 +280,23 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return true;
     }
 
+    static std::string normalize_embedding_name(std::string name) {
+        std::transform(name.begin(), name.end(), name.begin(), [](unsigned char c) { return std::tolower(c); });
+        return name;
+    }
+
+    bool append_embedding_tokens(std::string str, std::vector<int32_t>& bpe_tokens) {
+        std::string name = normalize_embedding_name(std::move(str));
+        auto iter        = embedding_map.find(name);
+        if (iter == embedding_map.end()) {
+            return false;
+        }
+        return load_embedding(name, iter->second, bpe_tokens);
+    }
+
     std::vector<int> convert_token_to_id(std::string text) {
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
-            auto iter = embedding_map.find(str);
-            if (iter == embedding_map.end()) {
-                return false;
-            }
-            std::string embedding_path = iter->second;
-            if (load_embedding(str, embedding_path, bpe_tokens)) {
-                return true;
-            }
-            return false;
+            return append_embedding_tokens(str, bpe_tokens);
         };
         std::vector<int> curr_tokens = tokenizer.encode(text, on_new_token_cb);
         return curr_tokens;
@@ -315,15 +323,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         }
 
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
-            auto iter = embedding_map.find(str);
-            if (iter == embedding_map.end()) {
-                return false;
-            }
-            std::string embedding_path = iter->second;
-            if (load_embedding(str, embedding_path, bpe_tokens)) {
-                return true;
-            }
-            return false;
+            return append_embedding_tokens(str, bpe_tokens);
         };
 
         std::vector<int> tokens;
@@ -638,6 +638,18 @@ struct SD3CLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -997,6 +1009,18 @@ struct FluxCLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1194,8 +1218,27 @@ struct T5CLIPEmbedder : public Conditioner {
                    bool use_mask                                       = false,
                    int mask_pad                                        = 0,
                    bool is_umt5                                        = false,
-                   std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+                   std::shared_ptr<RunnerWeightManager> weight_manager = nullptr,
+                   const char* model_args                              = nullptr)
         : use_mask(use_mask), mask_pad(mask_pad), t5_tokenizer(is_umt5) {
+        for (const auto& [key, value] : parse_key_value_args(model_args, "model arg")) {
+            if (key == "chroma_use_t5_mask") {
+                bool parsed = false;
+                if (parse_strict_bool(value, parsed)) {
+                    this->use_mask = parsed;
+                } else {
+                    LOG_WARN("ignoring invalid Chroma T5 model arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            } else if (key == "chroma_t5_mask_pad") {
+                int parsed = 0;
+                if (parse_strict_int(value, parsed)) {
+                    this->mask_pad = parsed;
+                } else {
+                    LOG_WARN("ignoring invalid Chroma T5 model arg '%s=%s'", key.c_str(), value.c_str());
+                }
+            }
+        }
+
         bool use_t5 = false;
         for (auto pair : tensor_storage_map) {
             if (pair.first.find("text_encoders.t5xxl") != std::string::npos) {
@@ -1229,6 +1272,18 @@ struct T5CLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (t5) {
             t5->set_flash_attention_enabled(enabled);
@@ -1381,6 +1436,113 @@ struct T5CLIPEmbedder : public Conditioner {
     }
 };
 
+struct MiniT2IConditioner : public Conditioner {
+    T5UniGramTokenizer tokenizer;
+    std::shared_ptr<T5Runner> t5;
+    size_t prompt_length = 256;
+
+    MiniT2IConditioner(ggml_backend_t backend,
+                       const String2TensorStorage& tensor_storage_map      = {},
+                       std::shared_ptr<RunnerWeightManager> weight_manager = nullptr) {
+        bool use_t5 = false;
+        for (const auto& pair : tensor_storage_map) {
+            if (pair.first.find("text_encoders.t5xxl") != std::string::npos) {
+                use_t5 = true;
+                break;
+            }
+        }
+        if (!use_t5) {
+            LOG_WARN("IMPORTANT NOTICE: No MiniT2I T5 text encoder provided, cannot process prompts!");
+            return;
+        }
+        t5 = std::make_shared<T5Runner>(backend, tensor_storage_map, "text_encoders.t5xxl.transformer", false, weight_manager);
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
+    void set_stream_layers_enabled(bool enabled) override {
+        if (t5) {
+            t5->set_stream_layers_enabled(enabled);
+        }
+    }
+
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
+    void set_flash_attention_enabled(bool enabled) override {
+        if (t5) {
+            t5->set_flash_attention_enabled(enabled);
+        }
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        if (t5) {
+            t5->set_weight_adapter(adapter);
+        }
+    }
+
+    void runner_done() override {
+        if (t5) {
+            t5->runner_done();
+        }
+    }
+
+    SDCondition get_learned_condition(int n_threads,
+                                      const ConditionerParams& conditioner_params) override {
+        SDCondition result;
+        if (!t5) {
+            result.c_crossattn = sd::Tensor<float>::zeros({1024, static_cast<int64_t>(prompt_length)});
+            result.c_vector    = sd::Tensor<float>::zeros({static_cast<int64_t>(prompt_length)});
+            return result;
+        }
+
+        std::vector<int> tokens = tokenizer.encode(conditioner_params.text);
+        if (tokens.size() > prompt_length) {
+            tokens.resize(prompt_length);
+        }
+        std::vector<float> mask(tokens.size(), 1.0f);
+        while (tokens.size() < prompt_length) {
+            tokens.push_back(tokenizer.PAD_TOKEN_ID);
+            mask.push_back(0.0f);
+        }
+
+        sd::Tensor<int32_t> input_ids({static_cast<int64_t>(tokens.size())}, tokens);
+        std::vector<float> t5_mask(mask.size(), 0.0f);
+        for (size_t i = 0; i < mask.size(); ++i) {
+            t5_mask[i] = mask[i] > 0.0f ? 0.0f : -HUGE_VALF;
+        }
+        sd::Tensor<float> hidden_states = t5->compute(n_threads,
+                                                      input_ids,
+                                                      sd::Tensor<float>::from_vector(t5_mask),
+                                                      false,
+                                                      true,
+                                                      true);
+        GGML_ASSERT(!hidden_states.empty());
+        result.c_crossattn = std::move(hidden_states);
+        result.c_vector    = sd::Tensor<float>::from_vector(mask);
+        return result;
+    }
+};
+
 struct AnimaConditioner : public Conditioner {
     std::shared_ptr<BPETokenizer> qwen_tokenizer;
     T5UniGramTokenizer t5_tokenizer;
@@ -1410,6 +1572,14 @@ struct AnimaConditioner : public Conditioner {
         llm->set_stream_layers_enabled(enabled);
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        llm->set_runtime_backends(backends);
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -1521,7 +1691,7 @@ struct LLMEmbedder : public Conditioner {
             arch = LLM::LLMArch::GPT_OSS_20B;
         } else if (sd_version_is_pid(version)) {
             arch = LLM::LLMArch::GEMMA2_2B;
-        } else if (sd_version_is_ideogram4(version)) {
+        } else if (sd_version_is_ideogram4(version) || sd_version_is_boogu_image(version) || sd_version_is_sefi_image(version) || sd_version_is_krea2(version)) {
             arch = LLM::LLMArch::QWEN3_VL;
         } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
             arch = LLM::LLMArch::QWEN3;
@@ -1555,6 +1725,14 @@ struct LLMEmbedder : public Conditioner {
         llm->set_stream_layers_enabled(enabled);
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        llm->set_runtime_backends(backends);
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -1781,6 +1959,76 @@ struct LLMEmbedder : public Conditioner {
 
                 prompt += "<|im_end|>\n<|im_start|>assistant\n";
             }
+        } else if (sd_version_is_boogu_image(version)) {
+            prompt_template_encode_start_idx = 0;
+
+            const std::string t2i_system_prompt =
+                "You are a helpful assistant that generates high-quality images based on user instructions. The instructions are as follows.";
+            const std::string edit_system_prompt =
+                "Describe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.";
+            const bool has_ref_images = llm->enable_vision && conditioner_params.ref_images != nullptr && !conditioner_params.ref_images->empty();
+            const bool text_empty     = conditioner_params.text.find_first_not_of(" \t\r\n") == std::string::npos;
+
+            if (has_ref_images) {
+                LOG_INFO("BooguImageEditPipeline");
+                const std::string prompt_prefix = "<|im_start|>system\n" + edit_system_prompt + "<|im_end|>\n<|im_start|>user\n";
+                std::string img_prompt;
+                const std::string placeholder = "<|image_pad|>";
+
+                for (int i = 0; i < conditioner_params.ref_images->size(); i++) {
+                    const auto& image = (*conditioner_params.ref_images)[i];
+                    double factor     = llm->config.vision.patch_size * llm->config.vision.spatial_merge_size;
+                    int height        = static_cast<int>(image.shape()[1]);
+                    int width         = static_cast<int>(image.shape()[0]);
+                    double beta       = std::sqrt((384.0 * 384.0) / (static_cast<double>(height) * static_cast<double>(width)));
+                    int h_bar         = std::max(static_cast<int>(factor),
+                                                 static_cast<int>(std::round(height * beta / factor)) * static_cast<int>(factor));
+                    int w_bar         = std::max(static_cast<int>(factor),
+                                                 static_cast<int>(std::round(width * beta / factor)) * static_cast<int>(factor));
+
+                    LOG_DEBUG("resize conditioner ref image %d from %dx%d to %dx%d", i, height, width, h_bar, w_bar);
+
+                    auto resized_image = clip_preprocess(image, w_bar, h_bar);
+                    auto image_embed   = llm->encode_image(n_threads, resized_image, false, true, true);
+                    GGML_ASSERT(!image_embed.empty());
+
+                    std::string image_prefix = prompt_prefix + img_prompt + "<|vision_start|>";
+                    int image_embed_idx      = static_cast<int>(tokenizer->encode(image_prefix, nullptr).size());
+                    image_embeds.emplace_back(image_embed_idx, image_embed);
+
+                    img_prompt += "<|vision_start|>";
+                    int64_t num_image_tokens = image_embed.shape()[1];
+                    img_prompt.reserve(img_prompt.size() + static_cast<size_t>(num_image_tokens) * placeholder.size() + 32);
+                    for (int j = 0; j < num_image_tokens; j++) {
+                        img_prompt += placeholder;
+                    }
+                    img_prompt += "<|vision_end|>";
+                }
+
+                prompt                  = prompt_prefix + img_prompt;
+                prompt_attn_range.first = static_cast<int>(prompt.size());
+                prompt += conditioner_params.text;
+                prompt_attn_range.second = static_cast<int>(prompt.size());
+                prompt += "<|im_end|>\n";
+            } else {
+                const std::string& system_prompt = text_empty ? edit_system_prompt : t2i_system_prompt;
+                prompt                           = "<|im_start|>system\n" + system_prompt + "<|im_end|>\n<|im_start|>user\n";
+                prompt_attn_range.first          = static_cast<int>(prompt.size());
+                prompt += conditioner_params.text;
+                prompt_attn_range.second = static_cast<int>(prompt.size());
+                prompt += "<|im_end|>\n";
+            }
+        } else if (sd_version_is_krea2(version)) {
+            prompt_template_encode_start_idx = 34;
+            out_layers                       = {2, 5, 8, 11, 14, 17, 20, 23, 26, 29, 32, 35};
+
+            prompt = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n";
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            prompt += "<|im_end|>\n<|im_start|>assistant\n";
         } else if (sd_version_is_longcat(version)) {
             spell_quotes = true;
 
@@ -1930,6 +2178,18 @@ struct LLMEmbedder : public Conditioner {
             prompt_attn_range.second = static_cast<int>(prompt.size());
 
             prompt += "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
+        } else if (sd_version_is_sefi_image(version)) {
+            prompt_template_encode_start_idx = 0;
+            min_length                       = 1024;
+            out_layers                       = {9, 18, 27};
+
+            prompt = "<|im_start|>user\n";
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            prompt += "<|im_end|>\n<|im_start|>assistant\n";
         } else if (version == VERSION_OVIS_IMAGE) {
             prompt_template_encode_start_idx = 28;
             min_length                       = prompt_template_encode_start_idx + 256;
@@ -2142,6 +2402,14 @@ struct LTXAVEmbedder : public Conditioner {
         projector->set_max_graph_vram_bytes(max_vram_bytes);
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        llm->set_runtime_backends(backends);
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         llm->set_weight_adapter(adapter);
         projector->set_weight_adapter(adapter);

src/convert.cpp

@@ -1,14 +1,33 @@
+#include <algorithm>
+#include <condition_variable>
+#include <cstdint>
 #include <cstring>
+#include <exception>
+#include <fstream>
+#include <memory>
 #include <mutex>
 #include <regex>
+#include <string>
+#include <thread>
 #include <vector>
 
+#include "core/util.h"
 #include "model_io/gguf_io.h"
 #include "model_io/safetensors_io.h"
+#include "model_io/streaming_writer.h"
 #include "model_loader.h"
-#include "util.h"
 
-#include "ggml_extend_backend.h"
+struct TensorExportInfo {
+    TensorStorage storage;
+    ggml_type type;
+};
+
+struct TensorExportJob {
+    TensorExportInfo info;
+    std::vector<uint8_t> data;
+    std::string error;
+    bool success = false;
+};
 
 static ggml_type get_export_tensor_type(ModelLoader& model_loader,
                                         const TensorStorage& tensor_storage,
@@ -33,106 +52,355 @@ static ggml_type get_export_tensor_type(ModelLoader& model_loader,
     return tensor_type;
 }
 
-static bool load_tensors_for_export(ModelLoader& model_loader,
-                                    ggml_context* ggml_ctx,
-                                    ggml_type type,
-                                    const TensorTypeRules& tensor_type_rules,
-                                    std::vector<TensorWriteInfo>& tensors) {
-    std::mutex tensor_mutex;
-    auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
-        const std::string& name = tensor_storage.name;
-        ggml_type tensor_type   = get_export_tensor_type(model_loader, tensor_storage, type, tensor_type_rules);
+static bool collect_tensors_for_export(ModelLoader& model_loader,
+                                       ggml_type type,
+                                       const TensorTypeRules& tensor_type_rules,
+                                       std::vector<TensorExportInfo>& tensors) {
+    tensors.clear();
+    tensors.reserve(model_loader.get_tensor_storage_map().size());
+    for (const auto& kv : model_loader.get_tensor_storage_map()) {
+        const TensorStorage& tensor_storage = kv.second;
+        TensorExportInfo info;
+        info.storage = tensor_storage;
+        info.type    = get_export_tensor_type(model_loader, tensor_storage, type, tensor_type_rules);
+        tensors.push_back(std::move(info));
+    }
+    LOG_INFO("collected %zu tensors for export", tensors.size());
+    return true;
+}
 
-        std::lock_guard<std::mutex> lock(tensor_mutex);
-        ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
-        if (tensor == nullptr) {
-            LOG_ERROR("ggml_new_tensor failed");
+static size_t export_tensor_nbytes(const TensorExportInfo& info) {
+    TensorStorage output_storage = info.storage;
+    output_storage.type          = info.type;
+    return static_cast<size_t>(output_storage.nbytes());
+}
+
+static TensorWritePlan tensor_write_plan_from_export_info(const TensorExportInfo& info) {
+    TensorWritePlan plan;
+    plan.name   = info.storage.name;
+    plan.type   = info.type;
+    plan.n_dims = info.storage.n_dims;
+    for (int i = 0; i < SD_MAX_DIMS; i++) {
+        plan.ne[i] = info.storage.ne[i];
+    }
+    return plan;
+}
+
+static std::vector<TensorWritePlan> tensor_write_plans_from_export_infos(const std::vector<TensorExportInfo>& tensors) {
+    std::vector<TensorWritePlan> plans;
+    plans.reserve(tensors.size());
+    for (const TensorExportInfo& info : tensors) {
+        plans.push_back(tensor_write_plan_from_export_info(info));
+    }
+    return plans;
+}
+
+static bool preallocate_output_file(const std::string& output_path, uint64_t file_size, std::string* error) {
+    if (file_size == 0) {
+        return true;
+    }
+
+    std::fstream file(output_path, std::ios::binary | std::ios::in | std::ios::out);
+    if (!file.is_open()) {
+        if (error != nullptr) {
+            *error = "failed to open output file '" + output_path + "' for preallocation";
+        }
+        return false;
+    }
+
+    // This portable fallback sets the final file size. A platform-specific
+    // posix_fallocate/ftruncate path can replace it later.
+    file.seekp(static_cast<std::streamoff>(file_size - 1), std::ios::beg);
+    file.put('\0');
+    file.flush();
+    if (!file) {
+        if (error != nullptr) {
+            *error = "failed to preallocate output file '" + output_path + "'";
+        }
+        return false;
+    }
+    return true;
+}
+
+static bool load_tensor_for_export(ModelLoader& model_loader, TensorExportJob& job) {
+    size_t mem_size = 1 * 1024 * 1024;
+    mem_size += ggml_tensor_overhead();
+    TensorStorage output_storage = job.info.storage;
+    output_storage.type          = job.info.type;
+    mem_size += static_cast<size_t>(output_storage.nbytes());
+
+    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
+    if (ggml_ctx == nullptr) {
+        job.error = "ggml_init failed for tensor '" + job.info.storage.name + "'";
+        return false;
+    }
+
+    ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, job.info.type, job.info.storage.n_dims, job.info.storage.ne);
+    if (tensor == nullptr) {
+        ggml_free(ggml_ctx);
+        job.error = "ggml_new_tensor failed for tensor '" + job.info.storage.name + "'";
+        return false;
+    }
+    ggml_set_name(tensor, job.info.storage.name.c_str());
+
+    const size_t tensor_nbytes = ggml_nbytes(tensor);
+    if (tensor_nbytes > 0 && !model_loader.load_tensor(job.info.storage, tensor)) {
+        ggml_free(ggml_ctx);
+        job.error = "failed to load tensor '" + job.info.storage.name + "'";
+        return false;
+    }
+
+    job.data.resize(tensor_nbytes);
+    if (tensor_nbytes > 0) {
+        memcpy(job.data.data(), tensor->data, tensor_nbytes);
+    }
+    ggml_free(ggml_ctx);
+    return true;
+}
+
+static bool stream_tensor_data(ModelLoader& model_loader,
+                               const std::string& output_path,
+                               const std::vector<TensorExportInfo>& tensors,
+                               const StreamingModelWriter& writer,
+                               int n_threads,
+                               std::string* error) {
+    n_threads = n_threads > 0 ? n_threads : sd_get_num_physical_cores();
+    n_threads = std::max(1, n_threads);
+    LOG_INFO("streaming convert with %d threads", n_threads);
+
+    int64_t start_time       = ggml_time_ms();
+    uint64_t bytes_written   = 0;
+    size_t tensors_written   = 0;
+    size_t next_tensor_index = 0;
+    bool failed              = false;
+    std::string failure;
+
+    const size_t memory_budget = 1024ull * 1024ull * 1024ull;
+    size_t reserved_bytes      = 0;
+
+    std::mutex work_mutex;
+    std::mutex progress_mutex;
+    std::condition_variable memory_cv;
+    std::vector<std::thread> workers;
+    workers.reserve(n_threads);
+
+    auto reserve_memory = [&](size_t bytes) -> bool {
+        std::unique_lock<std::mutex> lock(work_mutex);
+        memory_cv.wait(lock, [&]() {
+            return failed || reserved_bytes == 0 || reserved_bytes + bytes <= memory_budget;
+        });
+        if (failed) {
             return false;
         }
-        ggml_set_name(tensor, name.c_str());
-
-        if (!tensor->data) {
-            GGML_ASSERT(ggml_nelements(tensor) == 0);
-            // Avoid crashing writers by setting a dummy pointer for zero-sized tensors.
-            LOG_DEBUG("setting dummy pointer for zero-sized tensor %s", name.c_str());
-            tensor->data = ggml_get_mem_buffer(ggml_ctx);
-        }
-
-        TensorWriteInfo write_info;
-        write_info.tensor = tensor;
-        write_info.n_dims = tensor_storage.n_dims;
-        for (int i = 0; i < tensor_storage.n_dims; ++i) {
-            write_info.ne[i] = tensor_storage.ne[i];
-        }
-
-        *dst_tensor = tensor;
-        tensors.push_back(std::move(write_info));
-
+        reserved_bytes += bytes;
         return true;
     };
 
-    bool success = model_loader.load_tensors(on_new_tensor_cb);
-    LOG_INFO("load tensors done");
+    auto release_memory = [&](size_t bytes) {
+        {
+            std::lock_guard<std::mutex> lock(work_mutex);
+            reserved_bytes -= std::min(reserved_bytes, bytes);
+        }
+        memory_cv.notify_all();
+    };
+
+    auto fail = [&](const std::string& message) {
+        {
+            std::lock_guard<std::mutex> lock(work_mutex);
+            if (!failed) {
+                failed  = true;
+                failure = message;
+            }
+        }
+        memory_cv.notify_all();
+    };
+
+    for (int worker = 0; worker < n_threads; worker++) {
+        workers.emplace_back([&]() {
+            std::fstream output_file(output_path, std::ios::binary | std::ios::in | std::ios::out);
+            if (!output_file.is_open()) {
+                fail("failed to open output file '" + output_path + "' for tensor writing");
+                return;
+            }
+
+            while (true) {
+                size_t tensor_index = 0;
+                {
+                    std::lock_guard<std::mutex> lock(work_mutex);
+                    if (failed || next_tensor_index >= tensors.size()) {
+                        return;
+                    }
+                    tensor_index = next_tensor_index++;
+                }
+
+                const size_t tensor_bytes = export_tensor_nbytes(tensors[tensor_index]);
+                if (!reserve_memory(tensor_bytes)) {
+                    return;
+                }
+
+                TensorExportJob job;
+                job.info = tensors[tensor_index];
+                try {
+                    job.success = load_tensor_for_export(model_loader, job);
+                } catch (const std::exception& e) {
+                    job.error   = e.what();
+                    job.success = false;
+                }
+
+                if (!job.success) {
+                    release_memory(tensor_bytes);
+                    fail(job.error.empty() ? "streaming conversion failed" : job.error);
+                    return;
+                }
+
+                std::string write_error;
+                if (!writer.write_tensor(output_file,
+                                         tensor_index,
+                                         job.data.empty() ? nullptr : job.data.data(),
+                                         job.data.size(),
+                                         &write_error)) {
+                    release_memory(tensor_bytes);
+                    fail(write_error.empty() ? "streaming conversion write failed" : write_error);
+                    return;
+                }
+
+                {
+                    std::lock_guard<std::mutex> lock(progress_mutex);
+                    bytes_written += job.data.size();
+                    tensors_written++;
+                    float elapsed_seconds = (ggml_time_ms() - start_time) / 1000.0f;
+                    pretty_bytes_progress(static_cast<int>(tensors_written),
+                                          static_cast<int>(tensors.size()),
+                                          bytes_written,
+                                          elapsed_seconds);
+                }
+                release_memory(tensor_bytes);
+            }
+        });
+    }
+
+    for (auto& worker : workers) {
+        worker.join();
+    }
+    printf("\n");
+    if (failed) {
+        if (error != nullptr) {
+            *error = failure;
+        }
+        return false;
+    }
+    LOG_INFO("streaming conversion completed, taking %.2fs", (ggml_time_ms() - start_time) / 1000.f);
+    return true;
+}
+
+static bool write_model_file_streaming(ModelLoader& model_loader,
+                                       const std::string& output_path,
+                                       const std::vector<TensorExportInfo>& tensors,
+                                       StreamingModelWriter& writer,
+                                       int n_threads,
+                                       std::string* error) {
+    std::vector<TensorWritePlan> plans = tensor_write_plans_from_export_infos(tensors);
+    if (!writer.write_metadata(output_path, plans, error)) {
+        return false;
+    }
+    if (!preallocate_output_file(output_path, writer.file_size(), error)) {
+        return false;
+    }
+    model_loader.process_model_files(false, false);
+    return stream_tensor_data(model_loader, output_path, tensors, writer, n_threads, error);
+}
+
+static bool init_convert_path(ModelLoader& model_loader, const char* path, const char* prefix, bool& loaded_any) {
+    if (path == nullptr || strlen(path) == 0) {
+        return true;
+    }
+    if (!model_loader.init_from_file(path, prefix)) {
+        LOG_ERROR("init model loader from file failed: '%s'", path);
+        return false;
+    }
+    loaded_any = true;
+    return true;
+}
+
+static bool export_loaded_model(ModelLoader& model_loader,
+                                const char* output_path,
+                                sd_type_t output_type,
+                                const char* tensor_type_rules,
+                                int n_threads) {
+    ggml_type type             = sd_type_to_ggml_type(output_type);
+    bool output_is_safetensors = ends_with(output_path, ".safetensors");
+    TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);
+
+    std::vector<TensorExportInfo> tensors;
+    bool success = collect_tensors_for_export(model_loader, type, type_rules, tensors);
+    std::string error;
+    if (success) {
+        std::unique_ptr<StreamingModelWriter> writer;
+        if (output_is_safetensors) {
+            writer = std::make_unique<SafetensorsStreamingWriter>();
+        } else {
+            writer = std::make_unique<GGUFStreamingWriter>();
+        }
+        success = write_model_file_streaming(model_loader, output_path, tensors, *writer, n_threads, &error);
+    }
+
+    if (!success && !error.empty()) {
+        LOG_ERROR("%s", error.c_str());
+    }
+
     return success;
 }
 
+bool convert_with_components(const char* model_path,
+                             const char* clip_l_path,
+                             const char* clip_g_path,
+                             const char* t5xxl_path,
+                             const char* diffusion_model_path,
+                             const char* vae_path,
+                             const char* output_path,
+                             sd_type_t output_type,
+                             const char* tensor_type_rules,
+                             bool convert_name,
+                             int n_threads) {
+    ModelLoader model_loader;
+    bool loaded_any = false;
+
+    if (!init_convert_path(model_loader, model_path, "", loaded_any) ||
+        !init_convert_path(model_loader, clip_l_path, "text_encoders.clip_l.transformer.", loaded_any) ||
+        !init_convert_path(model_loader, clip_g_path, "text_encoders.clip_g.transformer.", loaded_any) ||
+        !init_convert_path(model_loader, t5xxl_path, "text_encoders.t5xxl.transformer.", loaded_any) ||
+        !init_convert_path(model_loader, diffusion_model_path, "model.diffusion_model.", loaded_any) ||
+        !init_convert_path(model_loader, vae_path, "vae.", loaded_any)) {
+        return false;
+    }
+
+    if (!loaded_any) {
+        LOG_ERROR("no input model path provided for convert");
+        return false;
+    }
+
+    if (convert_name) {
+        model_loader.convert_tensors_name();
+    }
+
+    return export_loaded_model(model_loader, output_path, output_type, tensor_type_rules, n_threads);
+}
+
 bool convert(const char* input_path,
              const char* vae_path,
              const char* output_path,
              sd_type_t output_type,
              const char* tensor_type_rules,
              bool convert_name) {
-    ModelLoader model_loader;
-
-    if (!model_loader.init_from_file(input_path)) {
-        LOG_ERROR("init model loader from file failed: '%s'", input_path);
-        return false;
-    }
-
-    if (vae_path != nullptr && strlen(vae_path) > 0) {
-        if (!model_loader.init_from_file(vae_path, "vae.")) {
-            LOG_ERROR("init model loader from file failed: '%s'", vae_path);
-            return false;
-        }
-    }
-    if (convert_name) {
-        model_loader.convert_tensors_name();
-    }
-
-    ggml_type type             = (ggml_type)output_type;
-    bool output_is_safetensors = ends_with(output_path, ".safetensors");
-    TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);
-
-    auto backend    = sd_backend_cpu_init();
-    size_t mem_size = 1 * 1024 * 1024;  // for padding
-    mem_size += model_loader.get_tensor_storage_map().size() * ggml_tensor_overhead();
-    mem_size += model_loader.get_params_mem_size(backend, type);
-    LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
-    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
-
-    if (ggml_ctx == nullptr) {
-        LOG_ERROR("ggml_init failed for converter");
-        ggml_backend_free(backend);
-        return false;
-    }
-
-    std::vector<TensorWriteInfo> tensors;
-    bool success = load_tensors_for_export(model_loader, ggml_ctx, type, type_rules, tensors);
-    ggml_backend_free(backend);
-
-    std::string error;
-    if (success) {
-        if (output_is_safetensors) {
-            success = write_safetensors_file(output_path, tensors, &error);
-        } else {
-            success = write_gguf_file(output_path, tensors, &error);
-        }
-    }
-
-    if (!success && !error.empty()) {
-        LOG_ERROR("%s", error.c_str());
-    }
-
-    ggml_free(ggml_ctx);
-    return success;
+    return convert_with_components(input_path,
+                                   nullptr,
+                                   nullptr,
+                                   nullptr,
+                                   nullptr,
+                                   vae_path,
+                                   output_path,
+                                   output_type,
+                                   tensor_type_rules,
+                                   convert_name,
+                                   0);
 }

src/core/backend_fit.cpp

@@ -0,0 +1,390 @@
+#include "backend_fit.h"
+
+#include <algorithm>
+#include <cctype>
+#include <cstdint>
+#include <utility>
+#include <vector>
+
+#include "core/ggml_extend_backend.h"
+#include "core/util.h"
+#include "ggml-backend.h"
+
+namespace sd::backend_fit {
+    namespace {
+
+        constexpr int64_t MiB = 1024ll * 1024;
+
+        enum class ComponentKind {
+            DIT         = 0,
+            VAE         = 1,
+            CONDITIONER = 2,
+        };
+
+        struct Component {
+            ComponentKind kind;
+            const char* name;
+            int64_t params_bytes  = 0;
+            int64_t reserve_bytes = 0;
+            bool splittable       = false;
+        };
+
+        struct Device {
+            ggml_backend_dev_t dev = nullptr;
+            std::string name;
+            std::string description;
+            int64_t free_bytes   = 0;
+            int64_t total_bytes  = 0;
+            int64_t budget_bytes = 0;
+        };
+
+        struct Decision {
+            ComponentKind kind;
+            bool on_cpu = false;
+            std::vector<size_t> device_idxs;
+        };
+
+        struct Plan {
+            bool valid      = false;
+            bool time_share = false;
+            std::vector<Decision> decisions;
+        };
+
+        bool classify_tensor(const std::string& name, ComponentKind& out) {
+            auto contains = [&](const char* s) { return name.find(s) != std::string::npos; };
+
+            if (contains("model.diffusion_model.") || contains("unet.")) {
+                out = ComponentKind::DIT;
+                return true;
+            }
+            if (contains("first_stage_model.") ||
+                name.rfind("vae.", 0) == 0 ||
+                name.rfind("tae.", 0) == 0) {
+                out = ComponentKind::VAE;
+                return true;
+            }
+            if (contains("text_encoders") ||
+                contains("cond_stage_model") ||
+                contains("te.text_model.") ||
+                contains("conditioner") ||
+                name.rfind("text_encoder.", 0) == 0 ||
+                name.rfind("text_embedding_projection.", 0) == 0 ||
+                contains(".aggregate_embed.")) {
+                out = ComponentKind::CONDITIONER;
+                return true;
+            }
+            return false;
+        }
+
+        std::vector<Component> estimate_components(ModelLoader& loader, ggml_type override_wtype) {
+            const auto& storage = loader.get_tensor_storage_map();
+
+            int64_t bytes[3] = {0, 0, 0};
+            for (const auto& [name, ts_const] : storage) {
+                TensorStorage ts = ts_const;
+                if (is_unused_tensor(ts.name)) {
+                    continue;
+                }
+                ComponentKind kind;
+                if (!classify_tensor(ts.name, kind)) {
+                    continue;
+                }
+                if (override_wtype != GGML_TYPE_COUNT &&
+                    loader.tensor_should_be_converted(ts, override_wtype)) {
+                    ts.type = override_wtype;
+                } else if (ts.expected_type != GGML_TYPE_COUNT && ts.expected_type != ts.type) {
+                    ts.type = ts.expected_type;
+                }
+                bytes[int(kind)] += (int64_t)ts.nbytes() + 64;
+            }
+
+            std::vector<Component> out;
+            out.push_back({ComponentKind::DIT, "DiT", bytes[int(ComponentKind::DIT)], 2048 * MiB, true});
+            out.push_back({ComponentKind::VAE, "VAE", bytes[int(ComponentKind::VAE)], 1024 * MiB, false});
+            out.push_back({ComponentKind::CONDITIONER, "Conditioner", bytes[int(ComponentKind::CONDITIONER)], 2048 * MiB, true});
+            return out;
+        }
+
+        std::vector<Device> enumerate_gpu_devices(const sd::ggml_graph_cut::MaxVramAssignment& budgets) {
+            std::vector<Device> out;
+            for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+                ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+                    continue;
+                }
+                Device d;
+                d.dev             = dev;
+                d.name            = ggml_backend_dev_name(dev);
+                d.description     = ggml_backend_dev_description(dev);
+                size_t free_bytes = 0, total_bytes = 0;
+                ggml_backend_dev_memory(dev, &free_bytes, &total_bytes);
+                d.free_bytes  = (int64_t)free_bytes;
+                d.total_bytes = (int64_t)total_bytes;
+
+                std::string budget_key = d.name;
+                std::transform(budget_key.begin(), budget_key.end(), budget_key.begin(),
+                               [](unsigned char c) { return (char)std::tolower(c); });
+                float gib = budgets.default_gib;
+                auto it   = budgets.backend_gib.find(budget_key);
+                if (it != budgets.backend_gib.end()) {
+                    gib = it->second;
+                }
+                if (gib > 0.f) {
+                    d.budget_bytes = std::min<int64_t>((int64_t)(gib * 1024.0 * 1024.0 * 1024.0), d.free_bytes);
+                } else if (gib < 0.f) {
+                    d.budget_bytes = d.free_bytes + (int64_t)(gib * 1024.0 * 1024.0 * 1024.0);
+                } else {
+                    d.budget_bytes = d.free_bytes - 512 * MiB;
+                }
+                d.budget_bytes = std::max<int64_t>(d.budget_bytes, 0);
+                out.push_back(d);
+            }
+            return out;
+        }
+
+        Plan compute_plan(const std::vector<Component>& components, const std::vector<Device>& devices) {
+            Plan plan;
+            if (devices.empty()) {
+                return plan;
+            }
+
+            std::vector<size_t> order(components.size());
+            for (size_t i = 0; i < order.size(); i++) {
+                order[i] = i;
+            }
+            std::sort(order.begin(), order.end(), [&](size_t a, size_t b) {
+                return components[a].params_bytes > components[b].params_bytes;
+            });
+
+            {
+                std::vector<int64_t> params_sum(devices.size(), 0);
+                std::vector<int64_t> max_reserve(devices.size(), 0);
+                std::vector<Decision> decisions(components.size());
+                bool ok = true;
+                for (size_t ci : order) {
+                    const Component& comp = components[ci];
+                    decisions[ci].kind    = comp.kind;
+                    if (comp.params_bytes == 0) {
+                        continue;
+                    }
+                    int best = -1;
+                    for (size_t di = 0; di < devices.size(); di++) {
+                        int64_t need = params_sum[di] + comp.params_bytes + std::max(max_reserve[di], comp.reserve_bytes);
+                        if (need <= devices[di].budget_bytes &&
+                            (best < 0 || devices[di].budget_bytes - params_sum[di] > devices[best].budget_bytes - params_sum[best])) {
+                            best = (int)di;
+                        }
+                    }
+                    if (best < 0) {
+                        ok = false;
+                        break;
+                    }
+                    params_sum[best] += comp.params_bytes;
+                    max_reserve[best] = std::max(max_reserve[best], comp.reserve_bytes);
+                    decisions[ci].device_idxs.push_back((size_t)best);
+                }
+                if (ok) {
+                    plan.valid      = true;
+                    plan.time_share = false;
+                    plan.decisions  = std::move(decisions);
+                    return plan;
+                }
+            }
+
+            plan.decisions.assign(components.size(), {});
+            for (size_t ci : order) {
+                const Component& comp = components[ci];
+                Decision& decision    = plan.decisions[ci];
+                decision.kind         = comp.kind;
+                if (comp.params_bytes == 0) {
+                    continue;
+                }
+                int best = -1;
+                for (size_t di = 0; di < devices.size(); di++) {
+                    if (comp.params_bytes + comp.reserve_bytes <= devices[di].budget_bytes &&
+                        (best < 0 || devices[di].budget_bytes > devices[best].budget_bytes)) {
+                        best = (int)di;
+                    }
+                }
+                if (best >= 0) {
+                    decision.device_idxs.push_back((size_t)best);
+                    continue;
+                }
+                if (comp.splittable && devices.size() > 1) {
+                    int64_t capacity = 0;
+                    for (const Device& d : devices) {
+                        capacity += std::max<int64_t>(d.budget_bytes - comp.reserve_bytes, 0);
+                    }
+                    if (comp.params_bytes <= capacity) {
+                        std::vector<size_t> idxs(devices.size());
+                        for (size_t i = 0; i < idxs.size(); i++) {
+                            idxs[i] = i;
+                        }
+                        std::sort(idxs.begin(), idxs.end(), [&](size_t a, size_t b) {
+                            return devices[a].budget_bytes > devices[b].budget_bytes;
+                        });
+                        decision.device_idxs = std::move(idxs);
+                        continue;
+                    }
+                }
+                decision.on_cpu = true;
+            }
+            plan.valid      = true;
+            plan.time_share = true;
+            return plan;
+        }
+
+        void print_plan(const Plan& plan,
+                        const std::vector<Component>& components,
+                        const std::vector<Device>& devices) {
+            LOG_INFO("auto-fit plan%s:", plan.time_share ? " (time-share: params load per phase and free after)" : "");
+            LOG_INFO("  devices:");
+            for (const Device& d : devices) {
+                LOG_INFO("    %-12s %-32s free %6lld MiB, budget %6lld MiB",
+                         d.name.c_str(), d.description.c_str(),
+                         (long long)(d.free_bytes / MiB), (long long)(d.budget_bytes / MiB));
+            }
+            LOG_INFO("  components:");
+            for (size_t ci = 0; ci < components.size(); ci++) {
+                const Component& comp    = components[ci];
+                const Decision& decision = plan.decisions[ci];
+                std::string target;
+                if (comp.params_bytes == 0) {
+                    target = "(not present)";
+                } else if (decision.on_cpu) {
+                    target = "CPU";
+                } else {
+                    for (size_t k = 0; k < decision.device_idxs.size(); k++) {
+                        if (k > 0) {
+                            target += " & ";
+                        }
+                        target += devices[decision.device_idxs[k]].name;
+                    }
+                    if (decision.device_idxs.size() > 1) {
+                        target += " (split)";
+                    }
+                }
+                LOG_INFO("    %-12s params %6lld MiB, compute reserve %5lld MiB -> %s",
+                         comp.name,
+                         (long long)(comp.params_bytes / MiB),
+                         (long long)(comp.reserve_bytes / MiB),
+                         target.c_str());
+            }
+        }
+
+        void append_assignment(std::string& spec, const char* key, const std::string& value) {
+            if (!spec.empty()) {
+                spec += ",";
+            }
+            spec += key;
+            spec += "=";
+            spec += value;
+        }
+
+        void append_component_decision(const std::vector<Component>& components,
+                                       const std::vector<Device>& devices,
+                                       const Plan& plan,
+                                       ComponentKind kind,
+                                       const char* module_key,
+                                       std::string& runtime_spec,
+                                       std::string& params_spec) {
+            for (size_t ci = 0; ci < components.size(); ci++) {
+                if (components[ci].kind != kind || components[ci].params_bytes == 0) {
+                    continue;
+                }
+                const Decision& decision = plan.decisions[ci];
+                if (decision.on_cpu) {
+                    append_assignment(runtime_spec, module_key, "cpu");
+                    return;
+                }
+                if (decision.device_idxs.empty()) {
+                    return;
+                }
+                std::string device_list;
+                for (size_t k = 0; k < decision.device_idxs.size(); k++) {
+                    if (k > 0) {
+                        device_list += "&";
+                    }
+                    device_list += devices[decision.device_idxs[k]].name;
+                }
+                append_assignment(runtime_spec, module_key, device_list);
+                if (plan.time_share) {
+                    append_assignment(params_spec, module_key, "disk");
+                }
+                return;
+            }
+        }
+
+    }  // namespace
+
+    bool derive_backend_specs(ModelLoader& loader,
+                              ggml_type override_wtype,
+                              sd::ggml_graph_cut::MaxVramAssignment& budgets,
+                              std::string& runtime_spec,
+                              std::string& params_spec) {
+        if (!runtime_spec.empty() || !params_spec.empty()) {
+            LOG_WARN("--auto-fit is enabled; ignoring --backend / --params-backend");
+        }
+
+        {
+            std::string error;
+            if (!budgets.canonicalize_backend_keys(&error)) {
+                LOG_ERROR("%s", error.c_str());
+                return false;
+            }
+        }
+
+        auto components = estimate_components(loader, override_wtype);
+        auto devices    = enumerate_gpu_devices(budgets);
+        auto plan       = compute_plan(components, devices);
+        if (!plan.valid) {
+            LOG_WARN("auto-fit: no usable GPU devices; using the default backend");
+            runtime_spec.clear();
+            params_spec.clear();
+            return true;
+        }
+
+        print_plan(plan, components, devices);
+
+        std::string derived_runtime_spec;
+        std::string derived_params_spec;
+        append_component_decision(components, devices, plan, ComponentKind::DIT, "diffusion", derived_runtime_spec, derived_params_spec);
+        append_component_decision(components, devices, plan, ComponentKind::CONDITIONER, "te", derived_runtime_spec, derived_params_spec);
+        append_component_decision(components, devices, plan, ComponentKind::VAE, "vae", derived_runtime_spec, derived_params_spec);
+
+        runtime_spec = std::move(derived_runtime_spec);
+        params_spec  = std::move(derived_params_spec);
+
+        LOG_INFO("auto-fit: --backend \"%s\"%s%s%s",
+                 runtime_spec.empty() ? "(default)" : runtime_spec.c_str(),
+                 params_spec.empty() ? "" : " --params-backend \"",
+                 params_spec.c_str(),
+                 params_spec.empty() ? "" : "\"");
+        return true;
+    }
+
+    bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params, bool prefer_temporal_tiling) {
+        if (prefer_temporal_tiling) {
+            if (tiling_params.temporal_tiling) {
+                return false;
+            }
+            tiling_params.temporal_tiling = true;
+        } else {
+            if (tiling_params.enabled) {
+                return false;
+            }
+            tiling_params.enabled = true;
+            if (tiling_params.tile_size_x <= 0) {
+                tiling_params.tile_size_x = 256;
+            }
+            if (tiling_params.tile_size_y <= 0) {
+                tiling_params.tile_size_y = 256;
+            }
+        }
+
+        LOG_WARN("auto-fit: VAE decode failed (likely out of memory); retrying with %s tiling",
+                 tiling_params.temporal_tiling ? "temporal" : "spatial");
+        return true;
+    }
+
+}  // namespace sd::backend_fit

src/core/backend_fit.h

@@ -0,0 +1,23 @@
+#ifndef __SD_BACKEND_FIT_H__
+#define __SD_BACKEND_FIT_H__
+
+#include <string>
+
+#include "core/ggml_graph_cut.h"
+#include "model_loader.h"
+#include "stable-diffusion.h"
+
+namespace sd::backend_fit {
+
+    bool derive_backend_specs(ModelLoader& loader,
+                              ggml_type override_wtype,
+                              sd::ggml_graph_cut::MaxVramAssignment& budgets,
+                              std::string& runtime_spec,
+                              std::string& params_spec);
+
+    bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params,
+                                         bool prefer_temporal_tiling);
+
+}  // namespace sd::backend_fit
+
+#endif  // __SD_BACKEND_FIT_H__

src/core/ggml_extend.hpp

@@ -391,7 +391,7 @@ __STATIC_INLINE__ uint8_t* ggml_tensor_to_sd_image(ggml_tensor* input, uint8_t*
     int64_t width    = input->ne[0];
     int64_t height   = input->ne[1];
     int64_t channels = input->ne[2];
-    GGML_ASSERT(channels == 3 && input->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32);
     if (image_data == nullptr) {
         image_data = (uint8_t*)malloc(width * height * channels);
     }
@@ -1038,6 +1038,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_linear(ggml_context* ctx,
 }
 
 __STATIC_INLINE__ ggml_tensor* ggml_ext_pad_ext(ggml_context* ctx,
+                                                ggml_backend_t backend,
                                                 ggml_tensor* x,
                                                 int lp0,
                                                 int rp0,
@@ -1063,7 +1064,17 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_pad_ext(ggml_context* ctx,
     }
 
     if (lp0 != 0 || rp0 != 0 || lp1 != 0 || rp1 != 0 || lp2 != 0 || rp2 != 0 || lp3 != 0 || rp3 != 0) {
-        x = ggml_pad_ext(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+        ggml_tensor* padded = ggml_pad_ext(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+        if (backend == nullptr || ggml_backend_supports_op(backend, padded)) {
+            x = padded;
+        } else {
+            // Some backends (e.g. Metal) only implement right-padding for
+            // GGML_OP_PAD (see #850): pad right by lp+rp instead, then roll
+            // the padding around to the left. shift < ne always holds because
+            // ne grew by lp+rp.
+            x = ggml_pad_ext(ctx, x, 0, lp0 + rp0, 0, lp1 + rp1, 0, lp2 + rp2, 0, lp3 + rp3);
+            x = ggml_roll(ctx, x, lp0, lp1, lp2, lp3);
+        }
     }
     return x;
 }
@@ -1076,7 +1087,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_pad(ggml_context* ctx,
                                             int p3          = 0,
                                             bool circular_x = false,
                                             bool circular_y = false) {
-    return ggml_ext_pad_ext(ctx, x, 0, p0, 0, p1, 0, p2, 0, p3, circular_x, circular_y);
+    return ggml_ext_pad_ext(ctx, nullptr, x, 0, p0, 0, p1, 0, p2, 0, p3, circular_x, circular_y);
 }
 
 // w: [OC，IC, KH, KW]
@@ -1105,7 +1116,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_2d(ggml_context* ctx,
     }
 
     if ((p0 != 0 || p1 != 0) && (circular_x || circular_y)) {
-        x  = ggml_ext_pad_ext(ctx, x, p0, p0, p1, p1, 0, 0, 0, 0, circular_x, circular_y);
+        x  = ggml_ext_pad_ext(ctx, nullptr, x, p0, p0, p1, p1, 0, 0, 0, 0, circular_x, circular_y);
         p0 = 0;
         p1 = 0;
     }
@@ -1130,6 +1141,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_2d(ggml_context* ctx,
 // b: [OC,]
 // result: [N*OC, OD, OH, OW]
 __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
+                                                ggml_backend_t backend,
                                                 ggml_tensor* x,
                                                 ggml_tensor* w,
                                                 ggml_tensor* b,
@@ -1159,7 +1171,21 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
         x          = ggml_cont(ctx, ggml_permute(ctx, x, 0, 1, 3, 2));
         x          = ggml_reshape_4d(ctx, x, im2col->ne[1], im2col->ne[2], OD, OC * N);
     } else {
-        x = ggml_conv_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2);
+        // ggml_conv_3d decomposes into GGML_OP_IM2COL_3D, which some backends
+        // (e.g. Metal, see #850) do not implement. Fall back to
+        // GGML_OP_CONV_3D on those backends.
+        bool im2col_3d_supported = true;
+        if (backend != nullptr) {
+            ggml_tensor* im2col = ggml_im2col_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, w->type);
+            im2col_3d_supported = ggml_backend_supports_op(backend, im2col);
+        }
+        if (im2col_3d_supported) {
+            x = ggml_conv_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2);
+        } else {
+            int64_t OC = w->ne[3] / IC;
+            int64_t N  = x->ne[3] / IC;
+            x          = ggml_conv_3d_direct(ctx, w, x, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
+        }
     }
 
     if (b != nullptr) {
@@ -1362,6 +1388,9 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         }
 
         auto out = ggml_flash_attn_ext(ctx, q_in, k_in, v_in, mask_in, scale / kv_scale, 0, 0);
+        if (!ggml_backend_supports_op(backend, out)) {
+            return nullptr;
+        }
         ggml_flash_attn_ext_set_prec(out, GGML_PREC_F32);
         if (kv_scale != 1.0f) {
             out = ggml_ext_scale(ctx, out, 1.0f / kv_scale);
@@ -1379,10 +1408,17 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
 
         if (can_use_flash_attn) {
             kqv = build_kqv(q, k, v, mask);
-            if (!ggml_backend_supports_op(backend, kqv)) {
-                kqv = nullptr;
-            } else {
-                kqv = ggml_view_3d(ctx, kqv, d_head, n_head, L_q, kqv->nb[1], kqv->nb[2], 0);
+            if (kqv != nullptr) {
+                kqv = ggml_view_4d(ctx,
+                                   kqv,
+                                   d_head,
+                                   n_head,
+                                   L_q,
+                                   N,
+                                   kqv->nb[1],
+                                   kqv->nb[2],
+                                   kqv->nb[1] * n_head,
+                                   0);
             }
         }
     }
@@ -1710,6 +1746,11 @@ protected:
     bool stream_layers_enabled            = false;
     size_t observed_max_effective_budget_ = 0;
 
+    std::vector<ggml_backend_t> extra_runtime_backends;  // borrowed (SDBackendManager-owned)
+    ggml_backend_sched_t sched             = nullptr;    // owned, multi-device only
+    ggml_backend_t cpu_fallback_backend    = nullptr;    // owned, sched requires a trailing CPU backend
+    bool multi_device_eval_callback_warned = false;
+
     std::shared_ptr<WeightAdapter> weight_adapter = nullptr;
     std::weak_ptr<RunnerWeightManager> weight_manager;
     std::unordered_set<const ggml_tensor*> kept_compute_param_tensor_set;
@@ -1977,7 +2018,121 @@ protected:
         return true;
     }
 
+    // Pass explicit buffer types: synthesized defaults can make CUDA devices
+    // report supporting each other's buffers and skip a required copy.
+    bool ensure_sched(ggml_cgraph* gf) {
+        if (sched != nullptr) {
+            return true;
+        }
+        std::vector<ggml_backend_t> backends;
+        backends.reserve(extra_runtime_backends.size() + 2);
+        backends.push_back(runtime_backend);
+        for (ggml_backend_t backend : extra_runtime_backends) {
+            backends.push_back(backend);
+        }
+        if (cpu_fallback_backend == nullptr && !sd_backend_is_cpu(runtime_backend)) {
+            cpu_fallback_backend = sd_backend_cpu_init();
+        }
+        if (cpu_fallback_backend != nullptr) {
+            backends.push_back(cpu_fallback_backend);
+        }
+
+        std::vector<ggml_backend_buffer_type_t> bufts;
+        bufts.reserve(backends.size());
+        ggml_backend_dev_t main_dev = ggml_backend_get_device(runtime_backend);
+        for (ggml_backend_t backend : backends) {
+            ggml_backend_buffer_type_t buft = nullptr;
+            if (backend == cpu_fallback_backend && main_dev != nullptr) {
+                buft = ggml_backend_dev_host_buffer_type(main_dev);
+            }
+            if (buft == nullptr) {
+                buft = ggml_backend_get_default_buffer_type(backend);
+            }
+            bufts.push_back(buft);
+        }
+
+        size_t graph_size = MAX_GRAPH_SIZE;
+        if (gf != nullptr) {
+            graph_size = std::max<size_t>(graph_size, (size_t)ggml_graph_n_nodes(gf));
+        }
+        sched = ggml_backend_sched_new(backends.data(),
+                                       bufts.data(),
+                                       (int)backends.size(),
+                                       graph_size,
+                                       /*parallel=*/false,
+                                       /*op_offload=*/false);
+        if (sched == nullptr) {
+            LOG_ERROR("%s: failed to create backend sched", get_desc().c_str());
+            return false;
+        }
+        return true;
+    }
+
+    ggml_backend_t backend_for_weight(const ggml_tensor* tensor) const {
+        if (tensor == nullptr || tensor->buffer == nullptr) {
+            return nullptr;
+        }
+        if (ggml_backend_buffer_get_usage(tensor->buffer) != GGML_BACKEND_BUFFER_USAGE_WEIGHTS ||
+            ggml_backend_buffer_is_host(tensor->buffer)) {
+            return nullptr;
+        }
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(ggml_backend_buffer_get_type(tensor->buffer));
+        if (dev == nullptr) {
+            return nullptr;
+        }
+        if (ggml_backend_get_device(runtime_backend) == dev) {
+            return runtime_backend;
+        }
+        for (ggml_backend_t backend : extra_runtime_backends) {
+            if (ggml_backend_get_device(backend) == dev) {
+                return backend;
+            }
+        }
+        return nullptr;
+    }
+
+    // Weightless ops have no scheduler anchor, so pin them to the most recent
+    // weight device. Views must stay unpinned or cross-device copies can be
+    // skipped for their consumers.
+    void pin_multi_device_nodes(ggml_cgraph* gf) {
+        if (sched == nullptr || gf == nullptr) {
+            return;
+        }
+        ggml_backend_t current = runtime_backend;
+        const int n_nodes      = ggml_graph_n_nodes(gf);
+        for (int i = 0; i < n_nodes; i++) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            for (int s = 0; s < GGML_MAX_SRC; s++) {
+                ggml_backend_t weight_backend = backend_for_weight(node->src[s]);
+                if (weight_backend != nullptr) {
+                    current = weight_backend;
+                }
+            }
+            if (node->op == GGML_OP_NONE || node->op == GGML_OP_VIEW || node->op == GGML_OP_RESHAPE ||
+                node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE) {
+                continue;
+            }
+            if (ggml_backend_supports_op(current, node)) {
+                ggml_backend_sched_set_tensor_backend(sched, node, current);
+            }
+        }
+    }
+
+    bool is_multi_device() const {
+        return !extra_runtime_backends.empty();
+    }
+
     bool alloc_compute_buffer(ggml_cgraph* gf) {
+        if (is_multi_device()) {
+            // The sched replaces the gallocr. Do NOT ggml_backend_sched_reserve
+            // the graph here: reserve runs split_graph, which rewires the
+            // graph's src pointers to sched-internal copy tensors, and the
+            // later ggml_backend_sched_alloc_graph would split the already
+            // rewired graph, silently corrupting every cross-backend input. A
+            // graph must be split at most once; the alloc in execute_graph
+            // performs the real allocation.
+            return ensure_sched(gf);
+        }
         if (compute_allocr != nullptr) {
             return true;
         }
@@ -2193,12 +2348,14 @@ protected:
                plan.valid &&
                max_graph_vram_bytes > 0 &&
                plan.segments.size() > 1 &&
-               !sd_backend_is_cpu(runtime_backend);
+               !sd_backend_is_cpu(runtime_backend) &&
+               !is_multi_device();
     }
 
     bool can_attempt_graph_cut_segmented_compute() const {
         return max_graph_vram_bytes > 0 &&
-               !sd_backend_is_cpu(runtime_backend);
+               !sd_backend_is_cpu(runtime_backend) &&
+               !is_multi_device();
     }
 
     bool resolve_graph_cut_plan(ggml_cgraph* gf,
@@ -2454,7 +2611,14 @@ protected:
         };
         ComputeBufferGuard compute_buffer_guard(this, free_compute_buffer);
 
-        if (!ggml_gallocr_alloc_graph(compute_allocr, gf)) {
+        if (is_multi_device()) {
+            ggml_backend_sched_reset(sched);
+            pin_multi_device_nodes(gf);  // reset clears the pins; re-apply before alloc
+            if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+                LOG_ERROR("%s sched alloc compute graph failed", get_desc().c_str());
+                return std::nullopt;
+            }
+        } else if (!ggml_gallocr_alloc_graph(compute_allocr, gf)) {
             LOG_ERROR("%s alloc compute graph failed", get_desc().c_str());
             return std::nullopt;
         }
@@ -2463,8 +2627,27 @@ protected:
         if (sd_backend_is_cpu(runtime_backend)) {
             sd_backend_cpu_set_n_threads(runtime_backend, n_threads);
         }
+        if (cpu_fallback_backend != nullptr) {
+            sd_backend_cpu_set_n_threads(cpu_fallback_backend, n_threads);
+        }
 
-        ggml_status status = ggml_backend_graph_compute(runtime_backend, gf);
+        ggml_status status;
+        if (is_multi_device()) {
+            if (sd_get_backend_eval_callback() != nullptr && !multi_device_eval_callback_warned) {
+                LOG_WARN("%s: eval callback is not supported with multiple runtime backends; ignoring",
+                         get_desc().c_str());
+                multi_device_eval_callback_warned = true;
+            }
+            status = ggml_backend_sched_graph_compute(sched, gf);
+            if (status == GGML_STATUS_SUCCESS) {
+                ggml_backend_sched_synchronize(sched);
+            }
+        } else {
+            status = sd_backend_graph_compute_with_eval_callback(runtime_backend,
+                                                                 gf,
+                                                                 sd_get_backend_eval_callback(),
+                                                                 sd_get_backend_eval_callback_data());
+        }
         if (status != GGML_STATUS_SUCCESS) {
             LOG_ERROR("%s compute failed: %s", get_desc().c_str(), ggml_status_to_string(status));
             return std::nullopt;
@@ -2641,6 +2824,10 @@ public:
         free_params_ctx();
         free_compute_ctx();
         free_cache_ctx_and_buffer();
+        if (cpu_fallback_backend != nullptr) {
+            ggml_backend_free(cpu_fallback_backend);
+            cpu_fallback_backend = nullptr;
+        }
     }
 
     virtual GGMLRunnerContext get_context() {
@@ -2681,10 +2868,20 @@ public:
             ggml_gallocr_free(compute_allocr);
             compute_allocr = nullptr;
         }
+        if (sched != nullptr) {
+            ggml_backend_sched_free(sched);
+            sched = nullptr;
+        }
     }
 
     // do copy after alloc graph
     void set_backend_tensor_data(ggml_tensor* tensor, const void* data) {
+        if (is_multi_device()) {
+            // The sched only assigns a backend (and thus a buffer) to tensors
+            // that participate in the graph; flag standalone data tensors as
+            // inputs so they get one.
+            ggml_set_input(tensor);
+        }
         backend_tensor_data_map[tensor] = data;
     }
 
@@ -2820,8 +3017,31 @@ public:
     }
 
     void set_stream_layers_enabled(bool enabled) {
+        if (enabled && is_multi_device()) {
+            LOG_WARN("%s: --stream-layers is not supported with multiple runtime backends; ignoring",
+                     get_desc().c_str());
+            return;
+        }
         stream_layers_enabled = enabled;
     }
+
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) {
+        extra_runtime_backends.clear();
+        for (ggml_backend_t backend : backends) {
+            if (backend == nullptr || backend == runtime_backend) {
+                continue;
+            }
+            if (std::find(extra_runtime_backends.begin(), extra_runtime_backends.end(), backend) ==
+                extra_runtime_backends.end()) {
+                extra_runtime_backends.push_back(backend);
+            }
+        }
+        if (is_multi_device() && stream_layers_enabled) {
+            LOG_WARN("%s: --stream-layers is not supported with multiple runtime backends; ignoring",
+                     get_desc().c_str());
+            stream_layers_enabled = false;
+        }
+    }
 };
 
 class GGMLBlock {
@@ -3356,7 +3576,7 @@ public:
                 b = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, b, prefix + "bias");
             }
         }
-        return ggml_ext_conv_3d(ctx->ggml_ctx, x, w, b, in_channels,
+        return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
                                 std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                 std::get<2>(padding), std::get<1>(padding), std::get<0>(padding),
                                 std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),

src/core/ggml_extend_backend.cpp

@@ -9,6 +9,7 @@
 #include <vector>
 
 #include "core/util.h"
+#include "ggml/src/ggml-impl.h"
 #include "stable-diffusion.h"
 
 static std::string trim_copy(const std::string& value) {
@@ -110,7 +111,67 @@ static std::string resolve_first_device_by_type(enum ggml_backend_dev_type type)
     if (dev == nullptr) {
         return "";
     }
-    return ggml_backend_dev_name(dev);
+    const char* dev_name = ggml_backend_dev_name(dev);
+    if (dev_name != nullptr && dev_name[0] != '\0') {
+        return dev_name;
+    }
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    const char* reg_name   = reg != nullptr ? ggml_backend_reg_name(reg) : nullptr;
+    return reg_name != nullptr ? reg_name : "";
+}
+
+static ggml_backend_dev_t resolve_first_device_by_registry_name(const std::string& name) {
+    std::string lower = lower_copy(trim_copy(name));
+    if (lower == "metal") {
+        lower = "mtl";
+    }
+    if (lower.empty()) {
+        return nullptr;
+    }
+
+    const size_t device_count = ggml_backend_dev_count();
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+        if (reg == nullptr) {
+            continue;
+        }
+        const char* reg_name = ggml_backend_reg_name(reg);
+        if (reg_name != nullptr && lower_copy(reg_name) == lower) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+static ggml_backend_dev_t resolve_device_by_name(const std::string& name) {
+    const std::string lower = lower_copy(trim_copy(name));
+    if (lower.empty()) {
+        return nullptr;
+    }
+
+    const size_t device_count = ggml_backend_dev_count();
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        const char* dev_name   = ggml_backend_dev_name(dev);
+        if (dev_name != nullptr && lower_copy(dev_name) == lower) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+static std::string backend_device_name(ggml_backend_dev_t dev) {
+    if (dev == nullptr) {
+        return "";
+    }
+    const char* name = ggml_backend_dev_name(dev);
+    if (name != nullptr && name[0] != '\0') {
+        return name;
+    }
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    const char* reg_name   = reg != nullptr ? ggml_backend_reg_name(reg) : nullptr;
+    return reg_name != nullptr ? reg_name : "";
 }
 
 static ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
@@ -280,7 +341,7 @@ static std::string get_default_backend_name() {
     return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
 }
 
-static std::string sd_resolve_backend_name(const std::string& name) {
+std::string sd_backend_resolve_name(const std::string& name) {
     ggml_backend_load_all_once();
     std::string requested = trim_copy(name);
     std::string lower     = lower_copy(requested);
@@ -296,6 +357,10 @@ static std::string sd_resolve_backend_name(const std::string& name) {
         return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
     }
 
+    if (ggml_backend_dev_t dev = resolve_first_device_by_registry_name(requested)) {
+        return backend_device_name(dev);
+    }
+
     const size_t device_count = ggml_backend_dev_count();
     for (size_t i = 0; i < device_count; ++i) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
@@ -318,7 +383,7 @@ static std::string sd_resolve_backend_name(const std::string& name) {
 }
 
 static bool backend_name_exists(const std::string& name) {
-    return !sd_resolve_backend_name(name).empty();
+    return !sd_backend_resolve_name(name).empty();
 }
 
 static ggml_backend_t init_named_backend(const std::string& name) {
@@ -328,7 +393,20 @@ static ggml_backend_t init_named_backend(const std::string& name) {
         return ggml_backend_init_best();
     }
 
-    std::string resolved = sd_resolve_backend_name(name);
+    if (ggml_backend_dev_t dev = resolve_device_by_name(name)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
+    if (ggml_backend_dev_t dev = resolve_first_device_by_registry_name(name)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
+
+    std::string resolved = sd_backend_resolve_name(name);
+    if (ggml_backend_dev_t dev = resolve_device_by_name(resolved)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
+    if (ggml_backend_dev_t dev = resolve_first_device_by_registry_name(resolved)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
     if (resolved.empty()) {
         return nullptr;
     }
@@ -364,6 +442,68 @@ bool sd_backend_cpu_set_n_threads(ggml_backend_t backend, int n_threads) {
     return false;
 }
 
+static ggml_cgraph sd_ggml_graph_view(ggml_cgraph* cgraph0, int i0, int i1) {
+    ggml_cgraph cgraph = {
+        /*.size             =*/0,
+        /*.n_nodes          =*/i1 - i0,
+        /*.n_leafs          =*/0,
+        /*.nodes            =*/cgraph0->nodes + i0,
+        /*.grads            =*/nullptr,
+        /*.grad_accs        =*/nullptr,
+        /*.leafs            =*/nullptr,
+        /*.use_counts       =*/cgraph0->use_counts,
+        /*.visited_hash_set =*/cgraph0->visited_hash_set,
+        /*.order            =*/cgraph0->order,
+        /*.uid              =*/0,
+    };
+    return cgraph;
+}
+
+ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,
+                                                        ggml_cgraph* gf,
+                                                        sd_graph_eval_callback_t callback_eval,
+                                                        void* callback_eval_user_data) {
+    if (callback_eval == nullptr) {
+        return ggml_backend_graph_compute(backend, gf);
+    }
+
+    ggml_status status = GGML_STATUS_SUCCESS;
+    const int n_nodes  = ggml_graph_n_nodes(gf);
+    bool stopped       = false;
+
+    for (int j0 = 0; j0 < n_nodes; ++j0) {
+        ggml_tensor* t = ggml_graph_node(gf, j0);
+        bool need      = callback_eval(t, true, callback_eval_user_data);
+        int j1         = j0;
+
+        while (!need && j1 < n_nodes - 1) {
+            t    = ggml_graph_node(gf, ++j1);
+            need = callback_eval(t, true, callback_eval_user_data);
+        }
+
+        ggml_cgraph gv = sd_ggml_graph_view(gf, j0, j1 + 1);
+        status         = ggml_backend_graph_compute_async(backend, &gv);
+        if (status != GGML_STATUS_SUCCESS) {
+            break;
+        }
+
+        ggml_backend_synchronize(backend);
+
+        if (need && !callback_eval(t, false, callback_eval_user_data)) {
+            stopped = true;
+            break;
+        }
+
+        j0 = j1;
+    }
+
+    ggml_backend_synchronize(backend);
+    if (stopped && status == GGML_STATUS_SUCCESS) {
+        status = GGML_STATUS_ABORTED;
+    }
+    return status;
+}
+
 const char* sd_get_system_info() {
     static std::string cache_info = []() -> std::string {
         ggml_backend_load_all_once();
@@ -525,12 +665,52 @@ SDBackendManager::~SDBackendManager() {
 
 void SDBackendManager::reset() {
     backends_.clear();
-    runtime_assignment_ = {};
-    params_assignment_  = {};
+    runtime_assignment_    = {};
+    params_assignment_     = {};
+    split_mode_assignment_ = {};
+}
+
+static std::vector<std::string> split_device_list(const std::string& value) {
+    std::vector<std::string> names;
+    for (const std::string& raw : split_copy(value, '&')) {
+        const std::string name = trim_copy(raw);
+        if (!name.empty()) {
+            names.push_back(name);
+        }
+    }
+    return names;
+}
+
+static std::string primary_device_name(const std::string& value) {
+    std::vector<std::string> names = split_device_list(value);
+    return names.empty() ? std::string() : names.front();
 }
 
 ggml_backend_t SDBackendManager::runtime_backend(SDBackendModule module) {
-    return init_cached_backend(runtime_assignment_.get(module));
+    return init_cached_backend(primary_device_name(runtime_assignment_.get(module)));
+}
+
+std::vector<ggml_backend_t> SDBackendManager::runtime_backends(SDBackendModule module) {
+    std::vector<ggml_backend_t> backends;
+    for (const std::string& name : split_device_list(runtime_assignment_.get(module))) {
+        ggml_backend_t backend = init_cached_backend(name);
+        if (backend == nullptr) {
+            LOG_ERROR("failed to initialize backend '%s' for module %s",
+                      name.c_str(),
+                      sd_backend_module_name(module));
+            continue;
+        }
+        if (std::find(backends.begin(), backends.end(), backend) == backends.end()) {
+            backends.push_back(backend);
+        }
+    }
+    if (backends.empty()) {
+        ggml_backend_t backend = runtime_backend(module);
+        if (backend != nullptr) {
+            backends.push_back(backend);
+        }
+    }
+    return backends;
 }
 
 ggml_backend_t SDBackendManager::params_backend(SDBackendModule module) {
@@ -556,6 +736,10 @@ bool SDBackendManager::params_backend_is_disk(SDBackendModule module) const {
     return is_disk_backend_token(params_assignment_.get(module));
 }
 
+bool SDBackendManager::params_backend_follows_runtime(SDBackendModule module) const {
+    return params_assignment_.get(module).empty();
+}
+
 bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule module) {
     ggml_backend_t backend = runtime_backend(module);
     if (backend == nullptr) {
@@ -575,6 +759,7 @@ bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule modu
 
 bool SDBackendManager::init(const char* backend_spec,
                             const char* params_backend_spec,
+                            const char* split_mode_spec,
                             std::string* error) {
     reset();
 
@@ -584,12 +769,53 @@ bool SDBackendManager::init(const char* backend_spec,
     if (!sd_parse_backend_assignment(SAFE_STR(params_backend_spec), &params_assignment_, error)) {
         return false;
     }
+    if (!sd_parse_backend_assignment(SAFE_STR(split_mode_spec), &split_mode_assignment_, error)) {
+        return false;
+    }
 
     return validate(error);
 }
 
+SDSplitMode SDBackendManager::split_mode(SDBackendModule module) const {
+    return lower_copy(trim_copy(split_mode_assignment_.get(module))) == "row" ? SDSplitMode::ROW
+                                                                              : SDSplitMode::LAYER;
+}
+
+ggml_backend_buffer_type_t SDBackendManager::split_buffer_type(ggml_backend_t backend,
+                                                               const std::vector<float>& tensor_split) {
+    if (backend == nullptr) {
+        return nullptr;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (dev == nullptr) {
+        return nullptr;
+    }
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    if (reg == nullptr) {
+        return nullptr;
+    }
+    auto fn = (ggml_backend_split_buffer_type_t)ggml_backend_reg_get_proc_address(reg, "ggml_backend_split_buffer_type");
+    if (fn == nullptr) {
+        return nullptr;
+    }
+    int main_device        = -1;
+    const size_t dev_count = ggml_backend_reg_dev_count(reg);
+    for (size_t i = 0; i < dev_count; ++i) {
+        if (ggml_backend_reg_dev_get(reg, i) == dev) {
+            main_device = (int)i;
+            break;
+        }
+    }
+    if (main_device < 0) {
+        return nullptr;
+    }
+    std::vector<float> padded_split(std::max<size_t>(tensor_split.size(), 64), 0.0f);
+    std::copy(tensor_split.begin(), tensor_split.end(), padded_split.begin());
+    return fn(main_device, padded_split.data());
+}
+
 bool SDBackendManager::validate(std::string* error) const {
-    auto validate_runtime_name = [&](const std::string& name) -> bool {
+    auto validate_single_runtime_name = [&](const std::string& name) -> bool {
         if (is_default_backend_token(name)) {
             return true;
         }
@@ -599,7 +825,7 @@ bool SDBackendManager::validate(std::string* error) const {
             }
             return false;
         }
-        if (!sd_resolve_backend_name(name).empty()) {
+        if (!sd_backend_resolve_name(name).empty() || resolve_first_device_by_registry_name(name) != nullptr) {
             return true;
         }
         if (error != nullptr) {
@@ -607,15 +833,56 @@ bool SDBackendManager::validate(std::string* error) const {
         }
         return false;
     };
+    auto validate_runtime_name = [&](const std::string& name) -> bool {
+        if (name.find('&') == std::string::npos) {
+            return validate_single_runtime_name(name);
+        }
+        std::vector<std::string> names = split_device_list(name);
+        if (names.empty()) {
+            if (error != nullptr) {
+                *error = "invalid backend device list '" + name + "'";
+            }
+            return false;
+        }
+        for (const std::string& entry : names) {
+            if (is_default_backend_token(entry)) {
+                if (error != nullptr) {
+                    *error = "default backend token is not allowed in a device list '" + name + "'";
+                }
+                return false;
+            }
+            if (!validate_single_runtime_name(entry)) {
+                return false;
+            }
+        }
+        return true;
+    };
     auto validate_params_name = [&](const std::string& name) -> bool {
         if (is_disk_backend_token(name)) {
             return true;
         }
-        return validate_runtime_name(name);
+        if (name.find('&') != std::string::npos) {
+            if (error != nullptr) {
+                *error = "params_backend does not accept device lists ('" + name + "')";
+            }
+            return false;
+        }
+        return validate_single_runtime_name(name);
+    };
+    auto validate_split_mode_name = [&](const std::string& name) -> bool {
+        const std::string lower = lower_copy(trim_copy(name));
+        if (lower.empty() || lower == "layer" || lower == "row") {
+            return true;
+        }
+        if (error != nullptr) {
+            *error = "invalid split mode '" + name + "' (expected layer or row)";
+        }
+        return false;
     };
 
     if (!validate_runtime_name(runtime_assignment_.default_name) ||
-        !validate_params_name(params_assignment_.default_name)) {
+        !validate_params_name(params_assignment_.default_name) ||
+        !validate_split_mode_name(split_mode_assignment_.default_name)) {
         return false;
     }
     for (const auto& kv : runtime_assignment_.module_names) {
@@ -628,11 +895,16 @@ bool SDBackendManager::validate(std::string* error) const {
             return false;
         }
     }
+    for (const auto& kv : split_mode_assignment_.module_names) {
+        if (!validate_split_mode_name(kv.second)) {
+            return false;
+        }
+    }
     return true;
 }
 
 ggml_backend_t SDBackendManager::init_cached_backend(const std::string& name) {
-    std::string resolved   = sd_resolve_backend_name(name);
+    std::string resolved   = sd_backend_resolve_name(name);
     std::string key        = lower_copy(resolved);
     ggml_backend_t backend = nullptr;
 

src/core/ggml_extend_backend.h

@@ -6,9 +6,11 @@
 #include <memory>
 #include <string>
 #include <unordered_map>
+#include <vector>
 
 #include "ggml-backend.h"
 #include "ggml.h"
+#include "stable-diffusion.h"
 
 enum class SDBackendModule {
     DIFFUSION,
@@ -36,10 +38,16 @@ struct SDBackendHandleDeleter {
 
 using SDBackendHandle = std::unique_ptr<struct ggml_backend, SDBackendHandleDeleter>;
 
+enum class SDSplitMode {
+    LAYER,
+    ROW,
+};
+
 class SDBackendManager {
 private:
     SDBackendAssignment runtime_assignment_;
     SDBackendAssignment params_assignment_;
+    SDBackendAssignment split_mode_assignment_;
     std::unordered_map<std::string, SDBackendHandle> backends_;
 
 public:
@@ -51,15 +59,23 @@ public:
 
     bool init(const char* backend_spec,
               const char* params_backend_spec,
+              const char* split_mode_spec,
               std::string* error);
     void reset();
 
     ggml_backend_t runtime_backend(SDBackendModule module);
     ggml_backend_t params_backend(SDBackendModule module);
 
+    std::vector<ggml_backend_t> runtime_backends(SDBackendModule module);
+
+    SDSplitMode split_mode(SDBackendModule module) const;
+    ggml_backend_buffer_type_t split_buffer_type(ggml_backend_t backend,
+                                                 const std::vector<float>& tensor_split);
+
     bool runtime_backend_is_cpu(SDBackendModule module);
     bool params_backend_is_cpu(SDBackendModule module);
     bool params_backend_is_disk(SDBackendModule module) const;
+    bool params_backend_follows_runtime(SDBackendModule module) const;
     bool runtime_backend_supports_host_buffer(SDBackendModule module);
 
 private:
@@ -71,6 +87,11 @@ bool sd_backend_is(ggml_backend_t backend, const std::string& name);
 bool sd_backend_is_cpu(ggml_backend_t backend);
 ggml_backend_t sd_backend_cpu_init();
 bool sd_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,
+                                                        ggml_cgraph* gf,
+                                                        sd_graph_eval_callback_t callback_eval,
+                                                        void* callback_eval_user_data);
+std::string sd_backend_resolve_name(const std::string& name);
 const char* sd_backend_module_name(SDBackendModule module);
 void ggml_ext_im_set_f32_1d(const struct ggml_tensor* tensor, int i, float value);
 bool add_rpc_devices(const std::string& servers);

src/core/ggml_graph_cut.cpp

@@ -1,6 +1,8 @@
 #include "core/ggml_graph_cut.h"
 
 #include <algorithm>
+#include <cctype>
+#include <cmath>
 #include <cstring>
 #include <map>
 #include <set>
@@ -8,6 +10,7 @@
 #include <stack>
 #include <unordered_map>
 
+#include "core/ggml_extend_backend.h"
 #include "core/util.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -83,6 +86,157 @@ namespace sd::ggml_graph_cut {
                segment.output_bytes;
     }
 
+    static std::string lower_ascii_copy(std::string value) {
+        std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+            return static_cast<char>(std::tolower(c));
+        });
+        return value;
+    }
+
+    static std::string normalize_backend_budget_key(const std::string& value) {
+        return lower_ascii_copy(trim(value));
+    }
+
+    static bool is_default_max_vram_key(const std::string& key) {
+        std::string normalized = normalize_backend_budget_key(key);
+        return normalized == "all" || normalized == "default" || normalized == "*";
+    }
+
+    static bool parse_max_vram_budget_value(const std::string& text, float* value, std::string* error) {
+        float parsed = 0.f;
+        if (!parse_strict_float(text, parsed) || !std::isfinite(parsed)) {
+            if (error != nullptr) {
+                *error = "invalid --max-vram value '" + text + "'";
+            }
+            return false;
+        }
+        *value = parsed;
+        return true;
+    }
+
+    static std::vector<std::string> backend_budget_keys(ggml_backend_t backend) {
+        std::vector<std::string> keys;
+        if (backend == nullptr) {
+            return keys;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        if (dev != nullptr) {
+            keys.push_back(normalize_backend_budget_key(ggml_backend_dev_name(dev)));
+        }
+        const char* backend_name = ggml_backend_name(backend);
+        if (backend_name != nullptr) {
+            keys.push_back(normalize_backend_budget_key(backend_name));
+        }
+        return keys;
+    }
+
+    void MaxVramAssignment::reset(float fallback_gib) {
+        default_gib = fallback_gib;
+        backend_gib.clear();
+        resolved_backend_bytes.clear();
+    }
+
+    bool MaxVramAssignment::parse(const std::string& raw_spec, std::string* error) {
+        const std::string in = trim(raw_spec);
+        if (in.empty()) {
+            return true;
+        }
+
+        for (const std::string& raw_part : split_string(in, ',')) {
+            const std::string part = trim(raw_part);
+            if (part.empty()) {
+                continue;
+            }
+
+            const size_t eq = part.find('=');
+            if (eq == std::string::npos) {
+                float value = 0.f;
+                if (!parse_max_vram_budget_value(part, &value, error)) {
+                    return false;
+                }
+                default_gib = value;
+                continue;
+            }
+
+            const std::string key        = trim(part.substr(0, eq));
+            const std::string value_text = trim(part.substr(eq + 1));
+            if (key.empty() || value_text.empty()) {
+                if (error != nullptr) {
+                    *error = "invalid --max-vram assignment '" + part + "'";
+                }
+                return false;
+            }
+
+            float value = 0.f;
+            if (!parse_max_vram_budget_value(value_text, &value, error)) {
+                return false;
+            }
+
+            if (is_default_max_vram_key(key)) {
+                default_gib = value;
+                continue;
+            }
+
+            const std::string backend_key = trim(key);
+            if (backend_key.empty()) {
+                if (error != nullptr) {
+                    *error = "invalid --max-vram backend key in '" + part + "'";
+                }
+                return false;
+            }
+            backend_gib[backend_key] = value;
+        }
+        resolved_backend_bytes.clear();
+        return true;
+    }
+
+    bool MaxVramAssignment::canonicalize_backend_keys(std::string* error) {
+        if (backend_gib.empty()) {
+            return true;
+        }
+
+        std::unordered_map<std::string, float> normalized;
+        for (const auto& kv : backend_gib) {
+            std::string resolved = sd_backend_resolve_name(kv.first);
+            if (resolved.empty()) {
+                if (error != nullptr) {
+                    *error = "unknown --max-vram backend '" + kv.first + "'";
+                }
+                return false;
+            }
+            normalized[normalize_backend_budget_key(resolved)] = kv.second;
+        }
+        backend_gib = std::move(normalized);
+        resolved_backend_bytes.clear();
+        return true;
+    }
+
+    size_t MaxVramAssignment::bytes_for_backend(ggml_backend_t backend) {
+        std::vector<std::string> keys = backend_budget_keys(backend);
+        const std::string cache_key   = keys.empty() ? std::string("<none>") : keys.front();
+        auto cached                   = resolved_backend_bytes.find(cache_key);
+        if (cached != resolved_backend_bytes.end()) {
+            return cached->second;
+        }
+
+        float budget_gib = default_gib;
+        if (!backend_gib.empty()) {
+            for (const std::string& key : keys) {
+                auto backend_it = backend_gib.find(key);
+                if (backend_it != backend_gib.end()) {
+                    budget_gib = backend_it->second;
+                    break;
+                }
+            }
+        }
+
+        const float resolved_gib          = resolve_max_vram_gib(budget_gib, backend);
+        const size_t bytes                = max_vram_gib_to_bytes(resolved_gib);
+        resolved_backend_bytes[cache_key] = bytes;
+        return bytes;
+    }
+
     size_t max_vram_gib_to_bytes(float max_vram) {
         if (max_vram <= 0.f) {
             return 0;

src/core/ggml_graph_cut.h

@@ -4,6 +4,7 @@
 #include <array>
 #include <cstdint>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -68,6 +69,17 @@ namespace sd::ggml_graph_cut {
 
     static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:";
 
+    struct MaxVramAssignment {
+        float default_gib = 0.f;
+        std::unordered_map<std::string, float> backend_gib;
+        std::unordered_map<std::string, size_t> resolved_backend_bytes;
+
+        void reset(float fallback_gib);
+        bool parse(const std::string& raw_spec, std::string* error);
+        bool canonicalize_backend_keys(std::string* error);
+        size_t bytes_for_backend(ggml_backend_t backend);
+    };
+
     bool is_graph_cut_tensor(const ggml_tensor* tensor);
     std::string make_graph_cut_name(const std::string& group, const std::string& output);
     void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output);

src/core/layer_split_partition.cpp

@@ -0,0 +1,221 @@
+#include "core/layer_split_partition.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+
+#include "core/util.h"
+
+namespace sd {
+
+    static bool layer_split_path_segment_starts_at(const std::string& name, size_t pos) {
+        return pos == 0 || name[pos - 1] == '.';
+    }
+
+    static bool layer_split_has_path_segment(const std::string& name, const char* segment) {
+        size_t pos = name.find(segment);
+        while (pos != std::string::npos) {
+            if (layer_split_path_segment_starts_at(name, pos)) {
+                return true;
+            }
+            pos = name.find(segment, pos + 1);
+        }
+        return false;
+    }
+
+    int layer_split_tensor_block_index(const std::string& name) {
+        static const char* unet_block_segments[] = {"input_blocks.", "output_blocks.", "middle_block.",
+                                                    "down_blocks.", "up_blocks.", "mid_block."};
+        for (const char* segment : unet_block_segments) {
+            if (layer_split_has_path_segment(name, segment)) {
+                return -1;
+            }
+        }
+
+        static const char* block_keywords[] = {"transformer_blocks.", "joint_blocks.", "double_blocks.",
+                                               "single_blocks.", "blocks.", "block.", "layers."};
+        for (const char* keyword : block_keywords) {
+            size_t pos = name.find(keyword);
+            while (pos != std::string::npos) {
+                if (!layer_split_path_segment_starts_at(name, pos)) {
+                    pos = name.find(keyword, pos + 1);
+                    continue;
+                }
+                pos += std::strlen(keyword);
+                size_t end = pos;
+                while (end < name.size() && name[end] >= '0' && name[end] <= '9') {
+                    end++;
+                }
+                if (end > pos && (end == name.size() || name[end] == '.')) {
+                    return std::atoi(name.substr(pos, end - pos).c_str());
+                }
+                break;
+            }
+        }
+        return -1;
+    }
+
+    std::string layer_split_backend_device_display_name(ggml_backend_t backend) {
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        const char* name       = dev != nullptr ? ggml_backend_dev_name(dev) : ggml_backend_name(backend);
+        return name != nullptr ? name : "unknown";
+    }
+
+    static bool layer_split_backend_supports_tensor(ggml_backend_t backend, const ggml_tensor* tensor) {
+        return backend != nullptr && tensor != nullptr && ggml_backend_supports_op(backend, tensor);
+    }
+
+    static size_t layer_split_supported_target(const std::string& desc,
+                                               const std::string& tensor_name,
+                                               const ggml_tensor* tensor,
+                                               const std::vector<ggml_backend_t>& backends,
+                                               size_t preferred) {
+        if (tensor == nullptr || backends.empty()) {
+            return preferred;
+        }
+        size_t preferred_safe = std::min(preferred, backends.size() - 1);
+        if (layer_split_backend_supports_tensor(backends[preferred_safe], tensor)) {
+            return preferred_safe;
+        }
+        for (size_t i = 0; i < backends.size(); i++) {
+            if (layer_split_backend_supports_tensor(backends[i], tensor)) {
+                LOG_WARN("%s layer split: moving tensor '%s' from %s to %s because the preferred backend cannot run op=%s type=%s nbytes=%.2f MB",
+                         desc.c_str(),
+                         tensor_name.c_str(),
+                         layer_split_backend_device_display_name(backends[preferred_safe]).c_str(),
+                         layer_split_backend_device_display_name(backends[i]).c_str(),
+                         ggml_op_name(tensor->op),
+                         ggml_type_name(tensor->type),
+                         ggml_nbytes(tensor) / (1024.0 * 1024.0));
+                return i;
+            }
+        }
+        LOG_WARN("%s layer split: tensor '%s' is not supported by any split backend: op=%s type=%s nbytes=%.2f MB",
+                 desc.c_str(),
+                 tensor_name.c_str(),
+                 ggml_op_name(tensor->op),
+                 ggml_type_name(tensor->type),
+                 ggml_nbytes(tensor) / (1024.0 * 1024.0));
+        return preferred_safe;
+    }
+
+    std::vector<std::map<std::string, ggml_tensor*>> partition_layer_split_tensors(
+        const std::string& desc,
+        const std::map<std::string, ggml_tensor*>& tensors,
+        const std::map<std::string, ggml_tensor*>& split_tensors,
+        const std::vector<ggml_backend_t>& backends) {
+        std::vector<std::map<std::string, ggml_tensor*>> partitions(backends.size());
+        if (backends.empty()) {
+            LOG_WARN("%s: no backend available for a layer split", desc.c_str());
+            return partitions;
+        }
+
+        std::map<int, int64_t> block_bytes;
+        std::map<std::string, size_t> non_block_targets;
+        std::vector<int64_t> other_bytes_by_backend(backends.size(), 0);
+        int64_t total_block_bytes = 0;
+        int64_t total_other_bytes = 0;
+        int n_blocks              = 0;
+        for (const auto& kv : tensors) {
+            int64_t bytes = (int64_t)ggml_nbytes(kv.second);
+            int idx       = split_tensors.count(kv.first) != 0 ? layer_split_tensor_block_index(kv.first) : -1;
+            if (idx >= 0) {
+                block_bytes[idx] += bytes;
+                total_block_bytes += bytes;
+                n_blocks = std::max(n_blocks, idx + 1);
+            } else {
+                size_t target               = layer_split_supported_target(desc, kv.first, kv.second, backends, 0);
+                non_block_targets[kv.first] = target;
+                other_bytes_by_backend[target] += bytes;
+                total_other_bytes += bytes;
+            }
+        }
+        if (n_blocks == 0) {
+            LOG_WARN("%s: no transformer blocks found for a layer split; keeping tensors on compatible backends starting from %s",
+                     desc.c_str(),
+                     layer_split_backend_device_display_name(backends[0]).c_str());
+            for (const auto& kv : tensors) {
+                size_t target  = 0;
+                auto target_it = non_block_targets.find(kv.first);
+                if (target_it != non_block_targets.end()) {
+                    target = target_it->second;
+                }
+                partitions[target][kv.first] = kv.second;
+            }
+            return partitions;
+        }
+
+        // Reserve compute headroom and subtract each device's actual non-block
+        // bytes from its block budget.
+        constexpr int64_t compute_headroom_bytes = 2ll * 1024 * 1024 * 1024;
+        std::vector<double> device_weights(backends.size(), 1.0);
+        double weight_sum = 0.0;
+        for (size_t i = 0; i < backends.size(); i++) {
+            ggml_backend_dev_t dev = ggml_backend_get_device(backends[i]);
+            size_t free_bytes = 0, total_bytes = 0;
+            if (dev != nullptr) {
+                ggml_backend_dev_memory(dev, &free_bytes, &total_bytes);
+            }
+            // Keep a small share even for tight devices instead of dropping them.
+            int64_t usable_bytes = std::max<int64_t>((int64_t)free_bytes - compute_headroom_bytes,
+                                                     (int64_t)free_bytes / 8);
+            device_weights[i]    = usable_bytes > 0 ? (double)usable_bytes : 1.0;
+            weight_sum += device_weights[i];
+        }
+
+        std::vector<int64_t> block_budgets(backends.size(), 0);
+        const int64_t total_bytes = total_block_bytes + total_other_bytes;
+        for (size_t i = 0; i < backends.size(); i++) {
+            int64_t budget   = (int64_t)((double)total_bytes * device_weights[i] / weight_sum);
+            budget           = std::max<int64_t>(budget - other_bytes_by_backend[i], 0);
+            block_budgets[i] = budget;
+        }
+
+        std::vector<int> boundaries(backends.size(), n_blocks);
+        size_t current = 0;
+        int64_t used   = 0;
+        for (int b = 0; b < n_blocks; b++) {
+            int64_t bytes = block_bytes.count(b) != 0 ? block_bytes[b] : 0;
+            if (current + 1 < backends.size() && used > 0 && used + bytes > block_budgets[current]) {
+                boundaries[current] = b;
+                current++;
+                used = 0;
+            }
+            used += bytes;
+        }
+
+        for (const auto& kv : tensors) {
+            size_t target = 0;
+            int idx       = split_tensors.count(kv.first) != 0 ? layer_split_tensor_block_index(kv.first) : -1;
+            if (idx >= 0) {
+                while (target < boundaries.size() && idx >= boundaries[target]) {
+                    target++;
+                }
+                target = std::min(target, backends.size() - 1);
+                target = layer_split_supported_target(desc, kv.first, kv.second, backends, target);
+            } else {
+                auto target_it = non_block_targets.find(kv.first);
+                if (target_it != non_block_targets.end()) {
+                    target = target_it->second;
+                }
+            }
+            partitions[target][kv.first] = kv.second;
+        }
+
+        int range_start = 0;
+        for (size_t i = 0; i < backends.size(); i++) {
+            int range_end                = boundaries[i];
+            const char* non_block_suffix = other_bytes_by_backend[i] > 0 ? " + non-block tensors" : "";
+            LOG_INFO("%s layer split: %s <- blocks [%d, %d)%s",
+                     desc.c_str(),
+                     layer_split_backend_device_display_name(backends[i]).c_str(),
+                     range_start,
+                     range_end,
+                     non_block_suffix);
+            range_start = range_end;
+        }
+        return partitions;
+    }
+
+}  // namespace sd

src/core/layer_split_partition.h

@@ -0,0 +1,24 @@
+#ifndef __SD_CORE_LAYER_SPLIT_PARTITION_H__
+#define __SD_CORE_LAYER_SPLIT_PARTITION_H__
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+namespace sd {
+
+    std::string layer_split_backend_device_display_name(ggml_backend_t backend);
+    int layer_split_tensor_block_index(const std::string& name);
+
+    std::vector<std::map<std::string, ggml_tensor*>> partition_layer_split_tensors(
+        const std::string& desc,
+        const std::map<std::string, ggml_tensor*>& tensors,
+        const std::map<std::string, ggml_tensor*>& split_tensors,
+        const std::vector<ggml_backend_t>& backends);
+
+}  // namespace sd
+
+#endif  // __SD_CORE_LAYER_SPLIT_PARTITION_H__

src/core/util.cpp

@@ -4,6 +4,8 @@
 #include <cmath>
 #include <codecvt>
 #include <cstdarg>
+#include <cstdlib>
+#include <cstring>
 #include <exception>
 #include <fstream>
 #include <locale>
@@ -25,6 +27,7 @@
 #include <unistd.h>
 #endif
 
+#include "ggml-backend.h"
 #include "ggml.h"
 #include "stable-diffusion.h"
 
@@ -346,6 +349,9 @@ int sd_preview_interval              = 1;
 bool sd_preview_denoised             = true;
 bool sd_preview_noisy                = false;
 
+static sd_graph_eval_callback_t sd_backend_eval_cb = nullptr;
+static void* sd_backend_eval_cb_data               = nullptr;
+
 std::u32string utf8_to_utf32(const std::string& utf8_str) {
     std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
     return converter.from_bytes(utf8_str);
@@ -406,6 +412,15 @@ std::vector<std::string> split_string(const std::string& str, char delimiter) {
     return result;
 }
 
+ggml_type sd_type_to_ggml_type(sd_type_t sdtype) {
+    const int type_value = static_cast<int>(sdtype);
+    if (type_value < std::min<int>(SD_TYPE_COUNT, GGML_TYPE_COUNT)) {
+        return static_cast<ggml_type>(type_value);
+    } else {
+        return GGML_TYPE_COUNT;
+    }
+}
+
 KeyValueArgs parse_key_value_args(const char* args, const char* context) {
     KeyValueArgs pairs;
 
@@ -620,6 +635,11 @@ void sd_set_preview_callback(sd_preview_cb_t cb, preview_t mode, int interval, b
     sd_preview_noisy    = noisy;
 }
 
+void sd_set_backend_eval_callback(sd_graph_eval_callback_t cb, void* data) {
+    sd_backend_eval_cb      = cb;
+    sd_backend_eval_cb_data = data;
+}
+
 sd_preview_cb_t sd_get_preview_callback() {
     return sd_preview_cb;
 }
@@ -640,6 +660,14 @@ bool sd_should_preview_noisy() {
     return sd_preview_noisy;
 }
 
+sd_graph_eval_callback_t sd_get_backend_eval_callback() {
+    return sd_backend_eval_cb;
+}
+
+void* sd_get_backend_eval_callback_data() {
+    return sd_backend_eval_cb_data;
+}
+
 sd_progress_cb_t sd_get_progress_callback() {
     return sd_progress_cb;
 }
@@ -972,3 +1000,26 @@ std::vector<std::pair<std::string, float>> split_quotation_attention(
     }
     return result;
 }
+
+size_t sd_list_devices(char* buffer, size_t buffer_size) {
+    if (ggml_backend_dev_count() == 0) {
+        // dynamic-backend builds discover their backend modules at runtime
+        ggml_backend_load_all();
+    }
+
+    std::ostringstream oss;
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        const char* name       = ggml_backend_dev_name(dev);
+        const char* desc       = ggml_backend_dev_description(dev);
+        oss << (name ? name : "") << '\t' << (desc ? desc : "") << '\n';
+    }
+
+    std::string devices = oss.str();
+    if (buffer != nullptr && buffer_size > 0) {
+        size_t copy_size = std::min(devices.size(), buffer_size - 1);
+        memcpy(buffer, devices.data(), copy_size);
+        buffer[copy_size] = '\0';
+    }
+    return devices.size();
+}

src/core/util.h

@@ -80,6 +80,8 @@ void pretty_bytes_progress(int step, int steps, uint64_t bytes_processed, float
 
 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);
 
+ggml_type sd_type_to_ggml_type(sd_type_t sdtype);
+
 std::string trim(const std::string& s);
 
 std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::string& text);
@@ -96,6 +98,9 @@ int sd_get_preview_interval();
 bool sd_should_preview_denoised();
 bool sd_should_preview_noisy();
 
+sd_graph_eval_callback_t sd_get_backend_eval_callback();
+void* sd_get_backend_eval_callback_data();
+
 // test if the backend is a specific one, e.g. "CUDA", "ROCm", "Vulkan" etc.
 bool sd_backend_is(ggml_backend_t backend, const std::string& name);
 

src/extensions/generation_extension.h

@@ -10,6 +10,7 @@
 
 #include "conditioning/conditioner.hpp"
 #include "core/ggml_extend_backend.h"
+#include "model/diffusion/model.hpp"
 #include "model_loader.h"
 #include "model_manager.h"
 #include "stable-diffusion.h"
@@ -30,6 +31,7 @@ struct GenerationExtensionConditionContext {
     Conditioner* conditioner;
     ConditionerParams& condition_params;
     const sd_pm_params_t& pm_params;
+    const sd_pulid_params_t& pulid_params;
     int n_threads;
     int total_steps;
 };
@@ -56,8 +58,20 @@ struct GenerationExtension {
                                                 const SDCondition& condition) const {
         return condition;
     }
+
+    // Called in the denoise loop for each enabled extension, after the per-step
+    // DiffusionParams (including its version-specific `extra`) has been built,
+    // but before diffusion_model->compute(). Lets an extension feed data into
+    // the diffusion forward that the conditioning-side hooks can't reach -- it
+    // can set/override fields on `params` (typically the architecture-specific
+    // `params.extra`, e.g. a guidance tensor, control payload, or an identity
+    // embedding for an adapter that injects inside the model's blocks). The
+    // extension targets whichever `extra` variant matches the active model.
+    // Mutates `params` only, never the extension. Default no-op.
+    virtual void before_diffusion(DiffusionParams& /*params*/, int /*step*/) const {}
 };
 
 std::shared_ptr<GenerationExtension> create_photomaker_extension();
+std::shared_ptr<GenerationExtension> create_pulid_extension();
 
 #endif

src/extensions/pulid_extension.cpp

@@ -0,0 +1,123 @@
+#include "extensions/generation_extension.h"
+
+#include <cstring>
+#include <variant>
+
+#include "core/tensor_ggml.hpp"
+#include "core/util.h"
+#include "gguf.h"
+
+static sd::Tensor<float> load_pulid_id_embedding(const char* path) {
+    sd::Tensor<float> empty;
+    if (path == nullptr || strlen(path) == 0) {
+        return empty;
+    }
+
+    struct ggml_context* ctx_data = nullptr;
+    struct gguf_init_params gp    = {/*.no_alloc =*/false, /*.ctx =*/&ctx_data};
+    struct gguf_context* gguf_ctx = gguf_init_from_file(path, gp);
+    if (gguf_ctx == nullptr || ctx_data == nullptr) {
+        LOG_WARN("PuLID id-embedding: cannot read gguf '%s'", path);
+        if (gguf_ctx != nullptr)
+            gguf_free(gguf_ctx);
+        if (ctx_data != nullptr)
+            ggml_free(ctx_data);
+        return empty;
+    }
+
+    struct ggml_tensor* t = ggml_get_tensor(ctx_data, "pulid_id");
+    if (t == nullptr) {
+        LOG_WARN("PuLID id-embedding: no 'pulid_id' tensor in '%s'", path);
+        gguf_free(gguf_ctx);
+        ggml_free(ctx_data);
+        return empty;
+    }
+
+    const int64_t token_dim  = t->ne[0];
+    const int64_t num_tokens = t->ne[1];
+    if (token_dim <= 0 || num_tokens <= 0 || token_dim > 65536 || num_tokens > 1024 ||
+        t->ne[2] != 1 || t->ne[3] != 1) {
+        LOG_WARN("PuLID id-embedding: implausible shape [%lld, %lld] in '%s'",
+                 (long long)token_dim, (long long)num_tokens, path);
+        gguf_free(gguf_ctx);
+        ggml_free(ctx_data);
+        return empty;
+    }
+
+    const size_t n_elem = (size_t)token_dim * (size_t)num_tokens;
+    sd::Tensor<float> out({token_dim, num_tokens, 1});
+    float* dst = out.data();
+    if (t->type == GGML_TYPE_F32) {
+        memcpy(dst, t->data, n_elem * sizeof(float));
+    } else if (t->type == GGML_TYPE_F16) {
+        const ggml_fp16_t* src = reinterpret_cast<const ggml_fp16_t*>(t->data);
+        for (size_t i = 0; i < n_elem; i++) {
+            dst[i] = ggml_fp16_to_fp32(src[i]);
+        }
+    } else if (t->type == GGML_TYPE_BF16) {
+        const ggml_bf16_t* src = reinterpret_cast<const ggml_bf16_t*>(t->data);
+        for (size_t i = 0; i < n_elem; i++) {
+            dst[i] = ggml_bf16_to_fp32(src[i]);
+        }
+    } else {
+        LOG_WARN("PuLID id-embedding: unsupported tensor type %s in '%s'",
+                 ggml_type_name(t->type), path);
+        gguf_free(gguf_ctx);
+        ggml_free(ctx_data);
+        return empty;
+    }
+
+    LOG_INFO("PuLID id-embedding: loaded [%lld, %lld] type=%s from '%s'",
+             (long long)token_dim, (long long)num_tokens, ggml_type_name(t->type), path);
+    gguf_free(gguf_ctx);
+    ggml_free(ctx_data);
+    return out;
+}
+
+struct PuLIDExtension : public GenerationExtension {
+    bool enabled = false;
+    sd::Tensor<float> id_embedding;
+    float id_weight = 1.0f;
+
+    const char* name() const override {
+        return "pulid";
+    }
+
+    bool is_enabled() const override {
+        return enabled;
+    }
+
+    bool init(const GenerationExtensionInitContext& ctx) override {
+        enabled = strlen(SAFE_STR(ctx.params->pulid_weights_path)) > 0;
+        return true;
+    }
+
+    void reset_runtime_condition() override {
+        id_embedding = {};
+        id_weight    = 1.0f;
+    }
+
+    bool prepare_condition(GenerationExtensionConditionContext& ctx) override {
+        reset_runtime_condition();
+        if (!enabled) {
+            return false;
+        }
+        id_embedding = load_pulid_id_embedding(ctx.pulid_params.id_embedding_path);
+        id_weight    = ctx.pulid_params.id_weight;
+        return false;  // PuLID does not modify the conditioning
+    }
+
+    void before_diffusion(DiffusionParams& params, int /*step*/) const override {
+        if (!enabled || id_embedding.empty()) {
+            return;
+        }
+        if (auto* flux_extra = std::get_if<FluxDiffusionExtra>(&params.extra)) {
+            flux_extra->pulid_id        = &id_embedding;
+            flux_extra->pulid_id_weight = id_weight;
+        }
+    }
+};
+
+std::shared_ptr<GenerationExtension> create_pulid_extension() {
+    return std::make_shared<PuLIDExtension>();
+}

src/model.h

@@ -36,18 +36,23 @@ enum SDVersion {
     VERSION_WAN2_2_I2V,
     VERSION_WAN2_2_TI2V,
     VERSION_QWEN_IMAGE,
+    VERSION_QWEN_IMAGE_LAYERED,
     VERSION_ANIMA,
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
     VERSION_LTXAV,
     VERSION_HIDREAM_O1,
     VERSION_Z_IMAGE,
+    VERSION_BOOGU_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
     VERSION_LENS,
+    VERSION_MINIT2I,
     VERSION_LONGCAT,
     VERSION_PID,
     VERSION_IDEOGRAM4,
+    VERSION_SEFI_IMAGE,
+    VERSION_KREA2,
     VERSION_ESRGAN,
     VERSION_COUNT,
 };
@@ -123,7 +128,7 @@ static inline bool sd_version_is_wan(SDVersion version) {
 }
 
 static inline bool sd_version_is_qwen_image(SDVersion version) {
-    if (version == VERSION_QWEN_IMAGE) {
+    if (version == VERSION_QWEN_IMAGE || version == VERSION_QWEN_IMAGE_LAYERED) {
         return true;
     }
     return false;
@@ -143,6 +148,13 @@ static inline bool sd_version_is_z_image(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_boogu_image(SDVersion version) {
+    if (version == VERSION_BOOGU_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_longcat(SDVersion version) {
     if (version == VERSION_LONGCAT) {
         return true;
@@ -164,6 +176,13 @@ static inline bool sd_version_is_lens(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_minit2i(SDVersion version) {
+    if (version == VERSION_MINIT2I) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_pid(SDVersion version) {
     if (version == VERSION_PID) {
         return true;
@@ -178,8 +197,36 @@ static inline bool sd_version_is_ideogram4(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_sefi_image(SDVersion version) {
+    if (version == VERSION_SEFI_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_krea2(SDVersion version) {
+    if (version == VERSION_KREA2) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_uses_flux_vae(SDVersion version) {
+    if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_boogu_image(version) || sd_version_is_longcat(version)) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_uses_flux2_vae(SDVersion version) {
-    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version)) {
+    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version) || sd_version_is_sefi_image(version)) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_uses_wan_vae(SDVersion version) {
+    if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_krea2(version) || sd_version_is_anima(version)) {
         return true;
     }
     return false;
@@ -206,11 +253,15 @@ static inline bool sd_version_is_dit(SDVersion version) {
         version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
+        sd_version_is_boogu_image(version) ||
         sd_version_is_ernie_image(version) ||
         sd_version_is_lens(version) ||
+        sd_version_is_minit2i(version) ||
         sd_version_is_longcat(version) ||
         sd_version_is_pid(version) ||
-        sd_version_is_ideogram4(version)) {
+        sd_version_is_ideogram4(version) ||
+        sd_version_is_sefi_image(version) ||
+        sd_version_is_krea2(version)) {
         return true;
     }
     return false;

src/model/adapter/pulid.hpp

@@ -0,0 +1,76 @@
+#ifndef __PULID_HPP__
+#define __PULID_HPP__
+
+#include "core/ggml_extend.hpp"
+#include "model/common/block.hpp"
+
+class PuLIDPerceiverAttentionCA : public GGMLBlock {
+public:
+    static constexpr int64_t DEFAULT_DIM      = 3072;  // Flux hidden size
+    static constexpr int64_t DEFAULT_DIM_HEAD = 128;
+    static constexpr int64_t DEFAULT_HEADS    = 16;
+    static constexpr int64_t DEFAULT_KV_DIM   = 2048;  // PuLID ID-embedding dim
+
+protected:
+    int64_t dim;
+    int64_t dim_head;
+    int64_t heads;
+    int64_t kv_dim;
+    int64_t inner_dim;
+
+public:
+    PuLIDPerceiverAttentionCA(int64_t dim      = DEFAULT_DIM,
+                              int64_t dim_head = DEFAULT_DIM_HEAD,
+                              int64_t heads    = DEFAULT_HEADS,
+                              int64_t kv_dim   = DEFAULT_KV_DIM)
+        : dim(dim),
+          dim_head(dim_head),
+          heads(heads),
+          kv_dim(kv_dim),
+          inner_dim(dim_head * heads) {
+        blocks["norm1"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(kv_dim));
+        blocks["norm2"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["to_q"]   = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim, /*bias=*/false));
+        blocks["to_kv"]  = std::shared_ptr<GGMLBlock>(new Linear(kv_dim, inner_dim * 2, /*bias=*/false));
+        blocks["to_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim, /*bias=*/false));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* id_embedding,
+                         ggml_tensor* image_tokens) {
+        auto norm1  = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto norm2  = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        auto to_q   = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+        auto to_kv  = std::dynamic_pointer_cast<Linear>(blocks["to_kv"]);
+        auto to_out = std::dynamic_pointer_cast<Linear>(blocks["to_out"]);
+
+        ggml_tensor* x_normed   = norm1->forward(ctx, id_embedding);
+        ggml_tensor* lat_normed = norm2->forward(ctx, image_tokens);
+
+        ggml_tensor* q  = to_q->forward(ctx, lat_normed);  // [N, T_img, 2048]
+        ggml_tensor* kv = to_kv->forward(ctx, x_normed);   // [N, T_img, 3072]
+
+        ggml_tensor* k = ggml_view_3d(ctx->ggml_ctx, kv,
+                                      inner_dim, kv->ne[1], kv->ne[2],
+                                      kv->nb[1], kv->nb[2],
+                                      /*offset=*/0);
+        ggml_tensor* v = ggml_view_3d(ctx->ggml_ctx, kv,
+                                      inner_dim, kv->ne[1], kv->ne[2],
+                                      kv->nb[1], kv->nb[2],
+                                      /*offset=*/inner_dim * ggml_element_size(kv));
+        k              = ggml_cont(ctx->ggml_ctx, k);
+        v              = ggml_cont(ctx->ggml_ctx, v);
+
+        ggml_tensor* attn_out = ggml_ext_attention_ext(
+            ctx->ggml_ctx, ctx->backend,
+            q, k, v,
+            heads,
+            /*mask=*/nullptr,
+            /*diag_mask_inf=*/false);
+
+        ggml_tensor* out = to_out->forward(ctx, attn_out);
+        return out;
+    }
+};
+
+#endif  // __PULID_HPP__

src/model/common/rope.hpp

@@ -12,6 +12,16 @@ namespace Rope {
         ErnieImage,
     };
 
+    enum class RefIndexMode {
+        FIXED,
+        INCREASE,
+        DECREASE,
+    };
+
+    __STATIC_INLINE__ RefIndexMode ref_index_mode_from_bool(bool increase_ref_index) {
+        return increase_ref_index ? RefIndexMode::INCREASE : RefIndexMode::FIXED;
+    }
+
     template <class T>
     __STATIC_INLINE__ std::vector<T> linspace(T start, T end, int num) {
         std::vector<T> result(num);
@@ -346,7 +356,7 @@ namespace Rope {
                                                                    int axes_dim_num,
                                                                    int start_index,
                                                                    const std::vector<ggml_tensor*>& ref_latents,
-                                                                   bool increase_ref_index,
+                                                                   RefIndexMode ref_index_mode,
                                                                    float ref_index_scale,
                                                                    bool scale_rope,
                                                                    int base_offset = 0) {
@@ -357,13 +367,15 @@ namespace Rope {
         for (ggml_tensor* ref : ref_latents) {
             int h_offset = 0;
             int w_offset = 0;
-            if (!increase_ref_index) {
+            if (ref_index_mode == RefIndexMode::FIXED) {
                 if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) {
                     w_offset = curr_w_offset;
                 } else {
                     h_offset = curr_h_offset;
                 }
                 scale_rope = false;
+            } else if (ref_index_mode == RefIndexMode::DECREASE) {
+                index--;
             }
 
             auto ref_ids = gen_flux_img_ids(static_cast<int>(ref->ne[1]),
@@ -377,7 +389,7 @@ namespace Rope {
                                             scale_rope);
             ids          = concat_ids(ids, ref_ids, bs);
 
-            if (increase_ref_index) {
+            if (ref_index_mode == RefIndexMode::INCREASE) {
                 index++;
             }
 
@@ -395,7 +407,7 @@ namespace Rope {
                                                                    int context_len,
                                                                    std::set<int> txt_arange_dims,
                                                                    const std::vector<ggml_tensor*>& ref_latents,
-                                                                   bool increase_ref_index,
+                                                                   RefIndexMode ref_index_mode,
                                                                    float ref_index_scale,
                                                                    bool is_longcat) {
         int x_index = is_longcat ? 1 : 0;
@@ -406,7 +418,7 @@ namespace Rope {
 
         auto ids = concat_ids(txt_ids, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, increase_ref_index, ref_index_scale, false, offset);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, ref_index_mode, ref_index_scale, false, offset);
             ids           = concat_ids(ids, refs_ids, bs);
         }
         return ids;
@@ -420,7 +432,7 @@ namespace Rope {
                                                      int context_len,
                                                      std::set<int> txt_arange_dims,
                                                      const std::vector<ggml_tensor*>& ref_latents,
-                                                     bool increase_ref_index,
+                                                     RefIndexMode ref_index_mode,
                                                      float ref_index_scale,
                                                      int theta,
                                                      bool circular_h,
@@ -435,7 +447,7 @@ namespace Rope {
                                                            context_len,
                                                            txt_arange_dims,
                                                            ref_latents,
-                                                           increase_ref_index,
+                                                           ref_index_mode,
                                                            ref_index_scale,
                                                            is_longcat);
         std::vector<std::vector<int>> wrap_dims;
@@ -481,17 +493,64 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
+                                                                  int h,
+                                                                  int w,
+                                                                  int pt,
+                                                                  int ph,
+                                                                  int pw,
+                                                                  int bs,
+                                                                  int t_offset    = 0,
+                                                                  int h_offset    = 0,
+                                                                  int w_offset    = 0,
+                                                                  bool scale_rope = false) {
+        int t_len = (t + (pt / 2)) / pt;
+        int h_len = (h + (ph / 2)) / ph;
+        int w_len = (w + (pw / 2)) / pw;
+
+        std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0));
+
+        if (scale_rope) {
+            h_offset -= h_len / 2;
+            w_offset -= w_len / 2;
+        }
+
+        std::vector<float> t_ids = linspace<float>(1.f * t_offset, 1.f * t_len - 1 + t_offset, t_len);
+        std::vector<float> h_ids = linspace<float>(1.f * h_offset, 1.f * h_len - 1 + h_offset, h_len);
+        std::vector<float> w_ids = linspace<float>(1.f * w_offset, 1.f * w_len - 1 + w_offset, w_len);
+
+        for (int i = 0; i < t_len; ++i) {
+            for (int j = 0; j < h_len; ++j) {
+                for (int k = 0; k < w_len; ++k) {
+                    int idx         = i * h_len * w_len + j * w_len + k;
+                    vid_ids[idx][0] = t_ids[i];
+                    vid_ids[idx][1] = h_ids[j];
+                    vid_ids[idx][2] = w_ids[k];
+                }
+            }
+        }
+
+        std::vector<std::vector<float>> vid_ids_repeated(bs * vid_ids.size(), std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < vid_ids.size(); ++j) {
+                vid_ids_repeated[i * vid_ids.size() + j] = vid_ids[j];
+            }
+        }
+        return vid_ids_repeated;
+    }
+
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int t,
+                                                                         int h,
                                                                          int w,
                                                                          int patch_size,
                                                                          int bs,
                                                                          int context_len,
                                                                          const std::vector<ggml_tensor*>& ref_latents,
-                                                                         bool increase_ref_index) {
+                                                                         RefIndexMode ref_index_mode) {
         int h_len        = (h + (patch_size / 2)) / patch_size;
         int w_len        = (w + (patch_size / 2)) / patch_size;
-        int txt_id_start = std::max(h_len, w_len);
-        auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * context_len + txt_id_start, context_len);
+        int txt_id_start = std::max(h_len, w_len) / 2;
+        auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * txt_id_start + context_len - 1, context_len);
         std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3));
         for (int i = 0; i < bs; ++i) {
             for (int j = 0; j < txt_ids.size(); ++j) {
@@ -499,28 +558,30 @@ namespace Rope {
             }
         }
         int axes_dim_num = 3;
-        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, 0, 0, 0, true);
+        auto img_ids     = gen_vid_ids(t, h, w, 1, patch_size, patch_size, bs, 0, 0, 0, true);
         auto ids         = concat_ids(txt_ids_repeated, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, 1, ref_latents, increase_ref_index, 1.f, true);
-            ids           = concat_ids(ids, refs_ids, bs);
+            int ref_start_index = ref_index_mode == RefIndexMode::DECREASE ? 0 : 1;
+            auto refs_ids       = gen_refs_ids(patch_size, bs, axes_dim_num, ref_start_index, ref_latents, ref_index_mode, 1.f, true);
+            ids                 = concat_ids(ids, refs_ids, bs);
         }
         return ids;
     }
 
     // Generate qwen_image positional embeddings
-    __STATIC_INLINE__ std::vector<float> gen_qwen_image_pe(int h,
+    __STATIC_INLINE__ std::vector<float> gen_qwen_image_pe(int t,
+                                                           int h,
                                                            int w,
                                                            int patch_size,
                                                            int bs,
                                                            int context_len,
                                                            const std::vector<ggml_tensor*>& ref_latents,
-                                                           bool increase_ref_index,
+                                                           RefIndexMode ref_index_mode,
                                                            int theta,
                                                            bool circular_h,
                                                            bool circular_w,
                                                            const std::vector<int>& axes_dim) {
-        std::vector<std::vector<float>> ids = gen_qwen_image_ids(h, w, patch_size, bs, context_len, ref_latents, increase_ref_index);
+        std::vector<std::vector<float>> ids = gen_qwen_image_ids(t, h, w, patch_size, bs, context_len, ref_latents, ref_index_mode);
         std::vector<std::vector<int>> wrap_dims;
         // This logic simply stores the (pad and patch_adjusted) sizes of images so we can make sure rope correctly tiles
         if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
@@ -533,7 +594,7 @@ namespace Rope {
                 // Track per-token wrap lengths for the row/column axes so only spatial tokens become periodic.
                 wrap_dims.assign(axes_dim.size(), std::vector<int>(total_tokens / bs, 0));
                 size_t cursor           = context_len;  // ignore text tokens
-                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                const size_t img_tokens = static_cast<size_t>(t) * static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
                 for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
                     if (circular_h) {
                         wrap_dims[1][cursor + token_i] = h_len;
@@ -684,46 +745,6 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims, EmbedNDLayout::ErnieImage);
     }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
-                                                                  int h,
-                                                                  int w,
-                                                                  int pt,
-                                                                  int ph,
-                                                                  int pw,
-                                                                  int bs,
-                                                                  int t_offset = 0,
-                                                                  int h_offset = 0,
-                                                                  int w_offset = 0) {
-        int t_len = (t + (pt / 2)) / pt;
-        int h_len = (h + (ph / 2)) / ph;
-        int w_len = (w + (pw / 2)) / pw;
-
-        std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0));
-
-        std::vector<float> t_ids = linspace<float>(1.f * t_offset, 1.f * t_len - 1 + t_offset, t_len);
-        std::vector<float> h_ids = linspace<float>(1.f * h_offset, 1.f * h_len - 1 + h_offset, h_len);
-        std::vector<float> w_ids = linspace<float>(1.f * w_offset, 1.f * w_len - 1 + w_offset, w_len);
-
-        for (int i = 0; i < t_len; ++i) {
-            for (int j = 0; j < h_len; ++j) {
-                for (int k = 0; k < w_len; ++k) {
-                    int idx         = i * h_len * w_len + j * w_len + k;
-                    vid_ids[idx][0] = t_ids[i];
-                    vid_ids[idx][1] = h_ids[j];
-                    vid_ids[idx][2] = w_ids[k];
-                }
-            }
-        }
-
-        std::vector<std::vector<float>> vid_ids_repeated(bs * vid_ids.size(), std::vector<float>(3));
-        for (int i = 0; i < bs; ++i) {
-            for (int j = 0; j < vid_ids.size(); ++j) {
-                vid_ids_repeated[i * vid_ids.size() + j] = vid_ids[j];
-            }
-        }
-        return vid_ids_repeated;
-    }
-
     // Generate wan positional embeddings
     __STATIC_INLINE__ std::vector<float> gen_wan_pe(int t,
                                                     int h,
@@ -785,7 +806,8 @@ namespace Rope {
                                                                       int context_len,
                                                                       int seq_multi_of,
                                                                       const std::vector<ggml_tensor*>& ref_latents,
-                                                                      bool increase_ref_index) {
+                                                                      RefIndexMode ref_index_mode) {
+        SD_UNUSED(ref_index_mode);
         int padded_context_len = context_len + bound_mod(context_len, seq_multi_of);
         auto txt_ids           = std::vector<std::vector<float>>(bs * padded_context_len, std::vector<float>(3, 0.0f));
         for (int i = 0; i < bs * padded_context_len; i++) {
@@ -816,12 +838,12 @@ namespace Rope {
                                                         int context_len,
                                                         int seq_multi_of,
                                                         const std::vector<ggml_tensor*>& ref_latents,
-                                                        bool increase_ref_index,
+                                                        RefIndexMode ref_index_mode,
                                                         int theta,
                                                         bool circular_h,
                                                         bool circular_w,
                                                         const std::vector<int>& axes_dim) {
-        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, increase_ref_index);
+        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, ref_index_mode);
         std::vector<std::vector<int>> wrap_dims;
         if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
             int pad_h = (patch_size - (h % patch_size)) % patch_size;
@@ -899,10 +921,12 @@ namespace Rope {
         // q,k,v: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2]
         // return: [N, L, n_head*d_head]
+        int64_t n_head = q->ne[1];
+
         q = apply_rope(ctx->ggml_ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
         k = apply_rope(ctx->ggml_ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
 
-        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, v->ne[1], mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
+        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
         return x;
     }
 };  // namespace Rope

src/model/diffusion/anima.hpp

@@ -612,7 +612,7 @@ namespace Anima {
                                           0,
                                           {},
                                           empty_ref_latents,
-                                          false,
+                                          Rope::RefIndexMode::FIXED,
                                           1.0f,
                                           false);
 

src/model/diffusion/boogu.hpp

@@ -0,0 +1,835 @@
+#ifndef __SD_MODEL_DIFFUSION_BOOGU_HPP__
+#define __SD_MODEL_DIFFUSION_BOOGU_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <tuple>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model/diffusion/qwen_image.hpp"
+#include "model_loader.h"
+
+namespace Boogu {
+    constexpr int BOOGU_GRAPH_SIZE = 65536;
+
+    struct BooguConfig {
+        int patch_size                   = 2;
+        int64_t in_channels              = 16;
+        int64_t out_channels             = 16;
+        int64_t hidden_size              = 3360;
+        int64_t num_layers               = 32;
+        int64_t num_double_stream_layers = 8;
+        int64_t num_refiner_layers       = 2;
+        int64_t num_attention_heads      = 28;
+        int64_t num_kv_heads             = 7;
+        int64_t head_dim                 = 120;
+        int64_t multiple_of              = 256;
+        int64_t instruction_feat_dim     = 4096;
+        int64_t timestep_embed_dim       = 1024;
+        int theta                        = 10000;
+        float timestep_scale             = 1000.0f;
+        float norm_eps                   = 1e-5f;
+        std::vector<int> axes_dim        = {40, 40, 40};
+        int64_t axes_dim_sum             = 120;
+
+        static int64_t count_blocks(const String2TensorStorage& tensor_storage_map,
+                                    const std::string& prefix,
+                                    const std::string& block_prefix) {
+            int64_t count = 0;
+            for (const auto& [name, _] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                size_t pos = name.find(block_prefix);
+                if (pos == std::string::npos) {
+                    continue;
+                }
+                auto items = split_string(name.substr(pos), '.');
+                if (items.size() > 1) {
+                    count = std::max<int64_t>(count, atoi(items[1].c_str()) + 1);
+                }
+            }
+            return count;
+        }
+
+        static BooguConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            BooguConfig config;
+            int64_t detected_head_dim = 0;
+            int64_t detected_kv_dim   = 0;
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "x_embedder.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area = config.patch_size * config.patch_size;
+                    config.in_channels = tensor_storage.ne[0] / patch_area;
+                    config.hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "time_caption_embed.caption_embedder.1.weight") && tensor_storage.n_dims == 2) {
+                    config.instruction_feat_dim = tensor_storage.ne[0];
+                    config.hidden_size          = tensor_storage.ne[1];
+                } else if (ends_with(name, "single_stream_layers.0.attn.norm_q.weight") && tensor_storage.n_dims == 1) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "double_stream_layers.0.img_self_attn.norm_q.weight") && tensor_storage.n_dims == 1) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "single_stream_layers.0.attn.to_k.weight") && tensor_storage.n_dims == 2) {
+                    detected_kv_dim = tensor_storage.ne[1];
+                } else if (ends_with(name, "double_stream_layers.0.img_instruct_attn.processor.img_to_k.weight") && tensor_storage.n_dims == 2) {
+                    detected_kv_dim = tensor_storage.ne[1];
+                } else if (ends_with(name, "norm_out.linear_2.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area  = config.patch_size * config.patch_size;
+                    config.out_channels = tensor_storage.ne[1] / patch_area;
+                }
+            }
+
+            config.num_layers               = std::max<int64_t>(1, count_blocks(tensor_storage_map, prefix, "single_stream_layers."));
+            config.num_double_stream_layers = std::max<int64_t>(0, count_blocks(tensor_storage_map, prefix, "double_stream_layers."));
+            int64_t noise_refiner_layers    = count_blocks(tensor_storage_map, prefix, "noise_refiner.");
+            int64_t ref_refiner_layers      = count_blocks(tensor_storage_map, prefix, "ref_image_refiner.");
+            int64_t context_refiner_layers  = count_blocks(tensor_storage_map, prefix, "context_refiner.");
+            config.num_refiner_layers       = std::max<int64_t>(1, std::max(noise_refiner_layers, std::max(ref_refiner_layers, context_refiner_layers)));
+
+            if (detected_head_dim > 0) {
+                config.head_dim            = detected_head_dim;
+                config.num_attention_heads = config.hidden_size / config.head_dim;
+                config.axes_dim_sum        = config.head_dim;
+                if (detected_kv_dim > 0) {
+                    config.num_kv_heads = detected_kv_dim / config.head_dim;
+                }
+                if (config.axes_dim_sum == 120) {
+                    config.axes_dim = {40, 40, 40};
+                } else if (config.axes_dim_sum % 3 == 0) {
+                    int axis        = static_cast<int>(config.axes_dim_sum / 3);
+                    config.axes_dim = {axis, axis, axis};
+                }
+            }
+            config.timestep_embed_dim = std::min<int64_t>(config.hidden_size, 1024);
+
+            LOG_DEBUG("boogu_image: layers=%" PRId64 ", double_stream_layers=%" PRId64 ", refiner_layers=%" PRId64 ", hidden=%" PRId64 ", heads=%" PRId64 ", kv_heads=%" PRId64 ", head_dim=%" PRId64 ", in_channels=%" PRId64 ", out_channels=%" PRId64,
+                      config.num_layers,
+                      config.num_double_stream_layers,
+                      config.num_refiner_layers,
+                      config.hidden_size,
+                      config.num_attention_heads,
+                      config.num_kv_heads,
+                      config.head_dim,
+                      config.in_channels,
+                      config.out_channels);
+            return config;
+        }
+    };
+
+    __STATIC_INLINE__ ggml_tensor* scale_modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* scale) {
+        scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);
+        return ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+    }
+
+    __STATIC_INLINE__ ggml_tensor* gate_residual(ggml_context* ctx, ggml_tensor* residual, ggml_tensor* x, ggml_tensor* gate) {
+        gate = ggml_tanh(ctx, gate);
+        gate = ggml_reshape_3d(ctx, gate, gate->ne[0], 1, gate->ne[1]);
+        x    = ggml_mul(ctx, x, gate);
+        return ggml_add(ctx, residual, x);
+    }
+
+    struct LuminaCombinedTimestepCaptionEmbedding : public GGMLBlock {
+        int64_t frequency_embedding_size;
+        float timestep_scale;
+
+        LuminaCombinedTimestepCaptionEmbedding(int64_t hidden_size,
+                                               int64_t instruction_feat_dim,
+                                               int64_t frequency_embedding_size,
+                                               float norm_eps,
+                                               float timestep_scale)
+            : frequency_embedding_size(frequency_embedding_size),
+              timestep_scale(timestep_scale) {
+            blocks["timestep_embedder"]  = std::make_shared<Qwen::TimestepEmbedding>(frequency_embedding_size, std::min<int64_t>(hidden_size, 1024));
+            blocks["caption_embedder.0"] = std::make_shared<RMSNorm>(instruction_feat_dim, norm_eps);
+            blocks["caption_embedder.1"] = std::make_shared<Linear>(instruction_feat_dim, hidden_size, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* timestep, ggml_tensor* text_hidden_states) {
+            auto timestep_embedder  = std::dynamic_pointer_cast<Qwen::TimestepEmbedding>(blocks["timestep_embedder"]);
+            auto caption_embedder_0 = std::dynamic_pointer_cast<RMSNorm>(blocks["caption_embedder.0"]);
+            auto caption_embedder_1 = std::dynamic_pointer_cast<Linear>(blocks["caption_embedder.1"]);
+
+            auto timestep_proj = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(frequency_embedding_size), 10000, timestep_scale);
+            auto time_embed    = timestep_embedder->forward(ctx, timestep_proj);
+            auto caption_embed = caption_embedder_1->forward(ctx, caption_embedder_0->forward(ctx, text_hidden_states));
+            return {time_embed, caption_embed};
+        }
+    };
+
+    struct LuminaRMSNormZero : public GGMLBlock {
+        LuminaRMSNormZero(int64_t embedding_dim, int64_t conditioning_embedding_dim, float norm_eps) {
+            blocks["linear"] = std::make_shared<Linear>(conditioning_embedding_dim, 4 * embedding_dim, true);
+            blocks["norm"]   = std::make_shared<RMSNorm>(embedding_dim, norm_eps);
+        }
+
+        std::tuple<ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* emb) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+
+            emb       = linear->forward(ctx, ggml_silu(ctx->ggml_ctx, emb));
+            auto mods = ggml_ext_chunk(ctx->ggml_ctx, emb, 4, 0);
+
+            auto scale_msa = mods[0];
+            auto gate_msa  = mods[1];
+            auto scale_mlp = mods[2];
+            auto gate_mlp  = mods[3];
+
+            x = scale_modulate(ctx->ggml_ctx, norm->forward(ctx, x), scale_msa);
+            return {x, gate_msa, scale_mlp, gate_mlp};
+        }
+    };
+
+    struct LuminaFeedForward : public GGMLBlock {
+        LuminaFeedForward(int64_t dim, int64_t inner_dim, int64_t multiple_of) {
+            inner_dim          = multiple_of * ((inner_dim + multiple_of - 1) / multiple_of);
+            blocks["linear_1"] = std::make_shared<Linear>(dim, inner_dim, false);
+            blocks["linear_2"] = std::make_shared<Linear>(inner_dim, dim, false);
+            blocks["linear_3"] = std::make_shared<Linear>(dim, inner_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+            auto linear_3 = std::dynamic_pointer_cast<Linear>(blocks["linear_3"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                linear_2->set_force_prec_f32(true);
+            }
+
+            auto h1 = linear_1->forward(ctx, x);
+            auto h2 = linear_3->forward(ctx, x);
+            x       = ggml_swiglu_split(ctx->ggml_ctx, h1, h2);
+            x       = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct LuminaLayerNormContinuous : public GGMLBlock {
+        LuminaLayerNormContinuous(int64_t embedding_dim,
+                                  int64_t conditioning_embedding_dim,
+                                  int64_t out_dim) {
+            blocks["linear_1"] = std::make_shared<Linear>(conditioning_embedding_dim, embedding_dim, true);
+            blocks["norm"]     = std::make_shared<LayerNorm>(embedding_dim, 1e-6f, false);
+            blocks["linear_2"] = std::make_shared<Linear>(embedding_dim, out_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* conditioning_embedding) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+
+            auto emb = linear_1->forward(ctx, ggml_silu(ctx->ggml_ctx, conditioning_embedding));
+            x        = scale_modulate(ctx->ggml_ctx, norm->forward(ctx, x), emb);
+            x        = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct Attention : public GGMLBlock {
+        int64_t dim_head;
+        int64_t heads;
+        int64_t kv_heads;
+
+        Attention(int64_t query_dim, int64_t dim_head, int64_t heads, int64_t kv_heads, float eps = 1e-5f)
+            : dim_head(dim_head), heads(heads), kv_heads(kv_heads) {
+            blocks["to_q"]     = std::make_shared<Linear>(query_dim, heads * dim_head, false);
+            blocks["to_k"]     = std::make_shared<Linear>(query_dim, kv_heads * dim_head, false);
+            blocks["to_v"]     = std::make_shared<Linear>(query_dim, kv_heads * dim_head, false);
+            blocks["norm_q"]   = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_k"]   = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["to_out.0"] = std::make_shared<Linear>(heads * dim_head, query_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* encoder_hidden_states,
+                             ggml_tensor* rotary_emb,
+                             ggml_tensor* attention_mask = nullptr) {
+            auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+            auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+            auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+            auto norm_q   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
+            int64_t N  = hidden_states->ne[2];
+            int64_t Lq = hidden_states->ne[1];
+            int64_t Lk = encoder_hidden_states->ne[1];
+
+            auto q = to_q->forward(ctx, hidden_states);
+            q      = ggml_reshape_4d(ctx->ggml_ctx, q, dim_head, heads, Lq, N);
+            auto k = to_k->forward(ctx, encoder_hidden_states);
+            k      = ggml_reshape_4d(ctx->ggml_ctx, k, dim_head, kv_heads, Lk, N);
+            auto v = to_v->forward(ctx, encoder_hidden_states);
+            v      = ggml_reshape_4d(ctx->ggml_ctx, v, dim_head, kv_heads, Lk, N);
+
+            q = norm_q->forward(ctx, q);
+            k = norm_k->forward(ctx, k);
+
+            auto out = Rope::attention(ctx, q, k, v, rotary_emb, attention_mask);
+            out      = to_out_0->forward(ctx, out);
+            return out;
+        }
+    };
+
+    struct BooguImageTransformerBlock : public GGMLBlock {
+        bool modulation;
+
+        BooguImageTransformerBlock(int64_t dim,
+                                   int64_t num_attention_heads,
+                                   int64_t num_kv_heads,
+                                   int64_t multiple_of,
+                                   float norm_eps,
+                                   bool modulation)
+            : modulation(modulation) {
+            int64_t head_dim       = dim / num_attention_heads;
+            blocks["attn"]         = std::make_shared<Attention>(dim, head_dim, num_attention_heads, num_kv_heads, 1e-5f);
+            blocks["feed_forward"] = std::make_shared<LuminaFeedForward>(dim, 4 * dim, multiple_of);
+            if (modulation) {
+                blocks["norm1"] = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            } else {
+                blocks["norm1"] = std::make_shared<RMSNorm>(dim, norm_eps);
+            }
+            blocks["ffn_norm1"] = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["norm2"]     = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["ffn_norm2"] = std::make_shared<RMSNorm>(dim, norm_eps);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* rotary_emb,
+                             ggml_tensor* temb           = nullptr,
+                             ggml_tensor* attention_mask = nullptr) {
+            auto attn         = std::dynamic_pointer_cast<Attention>(blocks["attn"]);
+            auto feed_forward = std::dynamic_pointer_cast<LuminaFeedForward>(blocks["feed_forward"]);
+            auto ffn_norm1    = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm1"]);
+            auto norm2        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto ffn_norm2    = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm2"]);
+
+            if (modulation) {
+                auto norm1 = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["norm1"]);
+                auto mods  = norm1->forward(ctx, hidden_states, temb);
+
+                auto norm_hidden_states = std::get<0>(mods);
+                auto gate_msa           = std::get<1>(mods);
+                auto scale_mlp          = std::get<2>(mods);
+                auto gate_mlp           = std::get<3>(mods);
+
+                auto attn_output = attn->forward(ctx, norm_hidden_states, norm_hidden_states, rotary_emb, attention_mask);
+                hidden_states    = gate_residual(ctx->ggml_ctx, hidden_states, norm2->forward(ctx, attn_output), gate_msa);
+
+                auto mlp_input  = scale_modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, hidden_states), scale_mlp);
+                auto mlp_output = feed_forward->forward(ctx, mlp_input);
+                hidden_states   = gate_residual(ctx->ggml_ctx, hidden_states, ffn_norm2->forward(ctx, mlp_output), gate_mlp);
+            } else {
+                auto norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+
+                auto norm_hidden_states = norm1->forward(ctx, hidden_states);
+                auto attn_output        = attn->forward(ctx, norm_hidden_states, norm_hidden_states, rotary_emb, attention_mask);
+                hidden_states           = ggml_add(ctx->ggml_ctx, hidden_states, norm2->forward(ctx, attn_output));
+
+                auto mlp_output = feed_forward->forward(ctx, ffn_norm1->forward(ctx, hidden_states));
+                hidden_states   = ggml_add(ctx->ggml_ctx, hidden_states, ffn_norm2->forward(ctx, mlp_output));
+            }
+            return hidden_states;
+        }
+    };
+
+    struct BooguImageJointAttention : public GGMLBlock {
+        int64_t dim_head;
+        int64_t heads;
+        int64_t kv_heads;
+
+        BooguImageJointAttention(int64_t dim, int64_t dim_head, int64_t heads, int64_t kv_heads)
+            : dim_head(dim_head), heads(heads), kv_heads(kv_heads) {
+            blocks["norm_q"]                  = std::make_shared<RMSNorm>(dim_head, 1e-5f);
+            blocks["norm_k"]                  = std::make_shared<RMSNorm>(dim_head, 1e-5f);
+            blocks["to_out.0"]                = std::make_shared<Linear>(heads * dim_head, dim, false);
+            blocks["processor.img_to_q"]      = std::make_shared<Linear>(dim, heads * dim_head, false);
+            blocks["processor.img_to_k"]      = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.img_to_v"]      = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.instruct_to_q"] = std::make_shared<Linear>(dim, heads * dim_head, false);
+            blocks["processor.instruct_to_k"] = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.instruct_to_v"] = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.instruct_out"]  = std::make_shared<Linear>(heads * dim_head, dim, false);
+            blocks["processor.img_out"]       = std::make_shared<Linear>(heads * dim_head, dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* img_hidden_states,
+                             ggml_tensor* instruct_hidden_states,
+                             ggml_tensor* rotary_emb,
+                             ggml_tensor* attention_mask = nullptr) {
+            auto norm_q        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto to_out_0      = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+            auto img_to_q      = std::dynamic_pointer_cast<Linear>(blocks["processor.img_to_q"]);
+            auto img_to_k      = std::dynamic_pointer_cast<Linear>(blocks["processor.img_to_k"]);
+            auto img_to_v      = std::dynamic_pointer_cast<Linear>(blocks["processor.img_to_v"]);
+            auto instruct_to_q = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_to_q"]);
+            auto instruct_to_k = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_to_k"]);
+            auto instruct_to_v = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_to_v"]);
+            auto instruct_out  = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_out"]);
+            auto img_out       = std::dynamic_pointer_cast<Linear>(blocks["processor.img_out"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
+            int64_t N          = img_hidden_states->ne[2];
+            int64_t L_img      = img_hidden_states->ne[1];
+            int64_t L_instruct = instruct_hidden_states->ne[1];
+
+            auto img_q = img_to_q->forward(ctx, img_hidden_states);
+            img_q      = ggml_reshape_4d(ctx->ggml_ctx, img_q, dim_head, heads, L_img, N);
+            auto img_k = img_to_k->forward(ctx, img_hidden_states);
+            img_k      = ggml_reshape_4d(ctx->ggml_ctx, img_k, dim_head, kv_heads, L_img, N);
+            auto img_v = img_to_v->forward(ctx, img_hidden_states);
+            img_v      = ggml_reshape_4d(ctx->ggml_ctx, img_v, dim_head, kv_heads, L_img, N);
+
+            auto instruct_q = instruct_to_q->forward(ctx, instruct_hidden_states);
+            instruct_q      = ggml_reshape_4d(ctx->ggml_ctx, instruct_q, dim_head, heads, L_instruct, N);
+            auto instruct_k = instruct_to_k->forward(ctx, instruct_hidden_states);
+            instruct_k      = ggml_reshape_4d(ctx->ggml_ctx, instruct_k, dim_head, kv_heads, L_instruct, N);
+            auto instruct_v = instruct_to_v->forward(ctx, instruct_hidden_states);
+            instruct_v      = ggml_reshape_4d(ctx->ggml_ctx, instruct_v, dim_head, kv_heads, L_instruct, N);
+
+            auto q = ggml_concat(ctx->ggml_ctx, instruct_q, img_q, 2);
+            auto k = ggml_concat(ctx->ggml_ctx, instruct_k, img_k, 2);
+            auto v = ggml_concat(ctx->ggml_ctx, instruct_v, img_v, 2);
+            q      = norm_q->forward(ctx, q);
+            k      = norm_k->forward(ctx, k);
+
+            auto hidden_states = Rope::attention(ctx, q, k, v, rotary_emb, attention_mask);
+            auto instruct_attn = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, 0, L_instruct);
+            auto img_attn      = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, L_instruct, L_instruct + L_img);
+
+            instruct_attn = instruct_out->forward(ctx, instruct_attn);
+            img_attn      = img_out->forward(ctx, img_attn);
+            hidden_states = ggml_concat(ctx->ggml_ctx, instruct_attn, img_attn, 1);
+            hidden_states = to_out_0->forward(ctx, hidden_states);
+            return hidden_states;
+        }
+    };
+
+    struct BooguImageDoubleStreamBlock : public GGMLBlock {
+        BooguImageDoubleStreamBlock(int64_t dim,
+                                    int64_t num_attention_heads,
+                                    int64_t num_kv_heads,
+                                    int64_t multiple_of,
+                                    float norm_eps) {
+            int64_t head_dim                = dim / num_attention_heads;
+            blocks["img_instruct_attn"]     = std::make_shared<BooguImageJointAttention>(dim, head_dim, num_attention_heads, num_kv_heads);
+            blocks["img_self_attn"]         = std::make_shared<Attention>(dim, head_dim, num_attention_heads, num_kv_heads, 1e-5f);
+            blocks["img_feed_forward"]      = std::make_shared<LuminaFeedForward>(dim, 4 * dim, multiple_of);
+            blocks["instruct_feed_forward"] = std::make_shared<LuminaFeedForward>(dim, 4 * dim, multiple_of);
+            blocks["img_norm1"]             = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["img_norm2"]             = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["img_norm3"]             = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["instruct_norm1"]        = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["instruct_norm2"]        = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["img_attn_norm"]         = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["img_self_attn_norm"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["img_ffn_norm1"]         = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["img_ffn_norm2"]         = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["instruct_attn_norm"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["instruct_ffn_norm1"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["instruct_ffn_norm2"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img_hidden_states,
+                                                      ggml_tensor* instruct_hidden_states,
+                                                      ggml_tensor* joint_rotary_emb,
+                                                      ggml_tensor* img_rotary_emb,
+                                                      ggml_tensor* temb) {
+            auto img_instruct_attn     = std::dynamic_pointer_cast<BooguImageJointAttention>(blocks["img_instruct_attn"]);
+            auto img_self_attn         = std::dynamic_pointer_cast<Attention>(blocks["img_self_attn"]);
+            auto img_feed_forward      = std::dynamic_pointer_cast<LuminaFeedForward>(blocks["img_feed_forward"]);
+            auto instruct_feed_forward = std::dynamic_pointer_cast<LuminaFeedForward>(blocks["instruct_feed_forward"]);
+            auto img_norm1             = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["img_norm1"]);
+            auto img_norm2             = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["img_norm2"]);
+            auto img_norm3             = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["img_norm3"]);
+            auto instruct_norm1        = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["instruct_norm1"]);
+            auto instruct_norm2        = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["instruct_norm2"]);
+            auto img_attn_norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["img_attn_norm"]);
+            auto img_self_attn_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["img_self_attn_norm"]);
+            auto img_ffn_norm1         = std::dynamic_pointer_cast<RMSNorm>(blocks["img_ffn_norm1"]);
+            auto img_ffn_norm2         = std::dynamic_pointer_cast<RMSNorm>(blocks["img_ffn_norm2"]);
+            auto instruct_attn_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["instruct_attn_norm"]);
+            auto instruct_ffn_norm1    = std::dynamic_pointer_cast<RMSNorm>(blocks["instruct_ffn_norm1"]);
+            auto instruct_ffn_norm2    = std::dynamic_pointer_cast<RMSNorm>(blocks["instruct_ffn_norm2"]);
+
+            int64_t L_instruct = instruct_hidden_states->ne[1];
+
+            auto img_norm1_out_vec      = img_norm1->forward(ctx, img_hidden_states, temb);
+            auto img_norm2_out_vec      = img_norm2->forward(ctx, img_hidden_states, temb);
+            auto img_norm3_out_vec      = img_norm3->forward(ctx, img_hidden_states, temb);
+            auto instruct_norm1_out_vec = instruct_norm1->forward(ctx, instruct_hidden_states, temb);
+            auto instruct_norm2_out_vec = instruct_norm2->forward(ctx, instruct_hidden_states, temb);
+
+            auto img_norm1_out = std::get<0>(img_norm1_out_vec);
+            auto img_gate_msa  = std::get<1>(img_norm1_out_vec);
+            auto img_scale_mlp = std::get<2>(img_norm1_out_vec);
+            auto img_gate_mlp  = std::get<3>(img_norm1_out_vec);
+
+            auto img_norm2_out = std::get<0>(img_norm2_out_vec);
+            auto img_shift_mlp = std::get<1>(img_norm2_out_vec);
+
+            auto img_norm3_out = std::get<0>(img_norm3_out_vec);
+            auto img_gate_self = std::get<1>(img_norm3_out_vec);
+
+            auto instruct_norm1_out = std::get<0>(instruct_norm1_out_vec);
+            auto instruct_gate_msa  = std::get<1>(instruct_norm1_out_vec);
+            auto instruct_scale_mlp = std::get<2>(instruct_norm1_out_vec);
+            auto instruct_gate_mlp  = std::get<3>(instruct_norm1_out_vec);
+
+            auto instruct_norm2_out = std::get<0>(instruct_norm2_out_vec);
+            auto instruct_shift_mlp = std::get<1>(instruct_norm2_out_vec);
+
+            auto joint_attn_out    = img_instruct_attn->forward(ctx, img_norm1_out, instruct_norm1_out, joint_rotary_emb);
+            auto instruct_attn_out = ggml_ext_slice(ctx->ggml_ctx, joint_attn_out, 1, 0, L_instruct);
+            auto img_attn_out      = ggml_ext_slice(ctx->ggml_ctx, joint_attn_out, 1, L_instruct, joint_attn_out->ne[1]);
+
+            auto img_self_attn_out = img_self_attn->forward(ctx, img_norm3_out, img_norm3_out, img_rotary_emb);
+
+            img_hidden_states = gate_residual(ctx->ggml_ctx, img_hidden_states, img_attn_norm->forward(ctx, img_attn_out), img_gate_msa);
+            img_hidden_states = gate_residual(ctx->ggml_ctx, img_hidden_states, img_self_attn_norm->forward(ctx, img_self_attn_out), img_gate_self);
+
+            auto img_mlp_input = scale_modulate(ctx->ggml_ctx, img_norm2_out, img_scale_mlp);
+            img_shift_mlp      = ggml_reshape_3d(ctx->ggml_ctx, img_shift_mlp, img_shift_mlp->ne[0], 1, img_shift_mlp->ne[1]);
+            img_mlp_input      = ggml_add(ctx->ggml_ctx, img_mlp_input, img_shift_mlp);
+            auto img_mlp_out   = img_feed_forward->forward(ctx, img_ffn_norm1->forward(ctx, img_mlp_input));
+            img_hidden_states  = gate_residual(ctx->ggml_ctx, img_hidden_states, img_ffn_norm2->forward(ctx, img_mlp_out), img_gate_mlp);
+
+            instruct_hidden_states  = gate_residual(ctx->ggml_ctx, instruct_hidden_states, instruct_attn_norm->forward(ctx, instruct_attn_out), instruct_gate_msa);
+            auto instruct_mlp_input = scale_modulate(ctx->ggml_ctx, instruct_norm2_out, instruct_scale_mlp);
+            instruct_shift_mlp      = ggml_reshape_3d(ctx->ggml_ctx, instruct_shift_mlp, instruct_shift_mlp->ne[0], 1, instruct_shift_mlp->ne[1]);
+            instruct_mlp_input      = ggml_add(ctx->ggml_ctx, instruct_mlp_input, instruct_shift_mlp);
+            auto instruct_mlp_out   = instruct_feed_forward->forward(ctx, instruct_ffn_norm1->forward(ctx, instruct_mlp_input));
+            instruct_hidden_states  = gate_residual(ctx->ggml_ctx, instruct_hidden_states, instruct_ffn_norm2->forward(ctx, instruct_mlp_out), instruct_gate_mlp);
+
+            return {img_hidden_states, instruct_hidden_states};
+        }
+    };
+
+    struct BooguImageModel : public GGMLBlock {
+        BooguConfig config;
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            GGML_UNUSED(prefix);
+            params["image_index_embedding"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.hidden_size, 5);
+        }
+
+        BooguImageModel() = default;
+        BooguImageModel(BooguConfig config)
+            : config(std::move(config)) {
+            blocks["x_embedder"]               = std::make_shared<Linear>(this->config.patch_size * this->config.patch_size * this->config.in_channels, this->config.hidden_size, true);
+            blocks["ref_image_patch_embedder"] = std::make_shared<Linear>(this->config.patch_size * this->config.patch_size * this->config.in_channels, this->config.hidden_size, true);
+            blocks["time_caption_embed"]       = std::make_shared<LuminaCombinedTimestepCaptionEmbedding>(this->config.hidden_size,
+                                                                                                    this->config.instruction_feat_dim,
+                                                                                                    256,
+                                                                                                    this->config.norm_eps,
+                                                                                                    this->config.timestep_scale);
+
+            for (int i = 0; i < this->config.num_refiner_layers; i++) {
+                blocks["noise_refiner." + std::to_string(i)]     = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                            this->config.num_attention_heads,
+                                                                                                            this->config.num_kv_heads,
+                                                                                                            this->config.multiple_of,
+                                                                                                            this->config.norm_eps,
+                                                                                                            true);
+                blocks["ref_image_refiner." + std::to_string(i)] = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                                this->config.num_attention_heads,
+                                                                                                                this->config.num_kv_heads,
+                                                                                                                this->config.multiple_of,
+                                                                                                                this->config.norm_eps,
+                                                                                                                true);
+                blocks["context_refiner." + std::to_string(i)]   = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                              this->config.num_attention_heads,
+                                                                                                              this->config.num_kv_heads,
+                                                                                                              this->config.multiple_of,
+                                                                                                              this->config.norm_eps,
+                                                                                                              false);
+            }
+
+            for (int i = 0; i < this->config.num_double_stream_layers; i++) {
+                blocks["double_stream_layers." + std::to_string(i)] = std::make_shared<BooguImageDoubleStreamBlock>(this->config.hidden_size,
+                                                                                                                    this->config.num_attention_heads,
+                                                                                                                    this->config.num_kv_heads,
+                                                                                                                    this->config.multiple_of,
+                                                                                                                    this->config.norm_eps);
+            }
+
+            for (int i = 0; i < this->config.num_layers; i++) {
+                blocks["single_stream_layers." + std::to_string(i)] = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                                   this->config.num_attention_heads,
+                                                                                                                   this->config.num_kv_heads,
+                                                                                                                   this->config.multiple_of,
+                                                                                                                   this->config.norm_eps,
+                                                                                                                   true);
+            }
+
+            blocks["norm_out"] = std::make_shared<LuminaLayerNormContinuous>(this->config.hidden_size,
+                                                                             this->config.timestep_embed_dim,
+                                                                             this->config.patch_size * this->config.patch_size * this->config.out_channels);
+        }
+
+        ggml_tensor* image_index_embedding(GGMLRunnerContext* ctx, int index) {
+            GGML_ASSERT(index >= 0 && index < 5);
+            auto embedding = params["image_index_embedding"];
+            auto out       = ggml_view_1d(ctx->ggml_ctx,
+                                          embedding,
+                                          config.hidden_size,
+                                          index * config.hidden_size * ggml_element_size(embedding));
+            out            = ggml_reshape_3d(ctx->ggml_ctx, out, config.hidden_size, 1, 1);
+            return out;
+        }
+
+        ggml_tensor* embed_refs(GGMLRunnerContext* ctx, const std::vector<ggml_tensor*>& ref_latents) {
+            if (ref_latents.empty()) {
+                return nullptr;
+            }
+            auto ref_image_patch_embedder = std::dynamic_pointer_cast<Linear>(blocks["ref_image_patch_embedder"]);
+
+            ggml_tensor* ref_img = nullptr;
+            for (int i = 0; i < static_cast<int>(ref_latents.size()); i++) {
+                auto ref = DiT::pad_and_patchify(ctx, ref_latents[i], config.patch_size, config.patch_size, false);
+                ref      = ref_image_patch_embedder->forward(ctx, ref);
+                ref      = ggml_add(ctx->ggml_ctx, ref, image_index_embedding(ctx, std::min(i, 4)));
+                ref_img  = ref_img == nullptr ? ref : ggml_concat(ctx->ggml_ctx, ref_img, ref, 1);
+            }
+            return ref_img;
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timesteps,
+                             ggml_tensor* context,
+                             ggml_tensor* pe,
+                             std::vector<ggml_tensor*> ref_latents = {}) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t N = x->ne[3];
+            GGML_ASSERT(N == 1);
+
+            auto x_embedder         = std::dynamic_pointer_cast<Linear>(blocks["x_embedder"]);
+            auto time_caption_embed = std::dynamic_pointer_cast<LuminaCombinedTimestepCaptionEmbedding>(blocks["time_caption_embed"]);
+            auto norm_out           = std::dynamic_pointer_cast<LuminaLayerNormContinuous>(blocks["norm_out"]);
+
+            auto timestep = ggml_sub(ctx->ggml_ctx, ggml_ext_ones_like(ctx->ggml_ctx, timesteps), timesteps);
+            auto embeds   = time_caption_embed->forward(ctx, timestep, context);
+            auto temb     = embeds.first;
+            auto txt      = embeds.second;
+
+            auto img        = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size, false);
+            int64_t img_len = img->ne[1];
+            img             = x_embedder->forward(ctx, img);
+            auto ref_img    = embed_refs(ctx, ref_latents);
+            int64_t ref_len = ref_img != nullptr ? ref_img->ne[1] : 0;
+            int64_t txt_len = txt->ne[1];
+
+            GGML_ASSERT(pe->ne[3] == txt_len + ref_len + img_len);
+            auto txt_pe   = ggml_ext_slice(ctx->ggml_ctx, pe, 3, 0, txt_len);
+            auto noise_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt_len + ref_len, txt_len + ref_len + img_len);
+
+            for (int i = 0; i < config.num_refiner_layers; i++) {
+                auto block = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
+                txt        = block->forward(ctx, txt, txt_pe);
+                sd::ggml_graph_cut::mark_graph_cut(txt, "boogu.context_refiner." + std::to_string(i), "txt");
+            }
+
+            for (int i = 0; i < config.num_refiner_layers; i++) {
+                auto block = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
+                img        = block->forward(ctx, img, noise_pe, temb);
+                sd::ggml_graph_cut::mark_graph_cut(img, "boogu.noise_refiner." + std::to_string(i), "img");
+            }
+
+            ggml_tensor* combined_img = img;
+            if (ref_img != nullptr) {
+                auto ref_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt_len, txt_len + ref_len);
+                for (int i = 0; i < config.num_refiner_layers; i++) {
+                    auto block = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["ref_image_refiner." + std::to_string(i)]);
+                    ref_img    = block->forward(ctx, ref_img, ref_pe, temb);
+                    sd::ggml_graph_cut::mark_graph_cut(ref_img, "boogu.ref_image_refiner." + std::to_string(i), "ref_img");
+                }
+                combined_img = ggml_concat(ctx->ggml_ctx, ref_img, img, 1);
+            }
+
+            auto img_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt_len, txt_len + combined_img->ne[1]);
+            for (int i = 0; i < config.num_double_stream_layers; i++) {
+                auto block   = std::dynamic_pointer_cast<BooguImageDoubleStreamBlock>(blocks["double_stream_layers." + std::to_string(i)]);
+                auto result  = block->forward(ctx, combined_img, txt, pe, img_pe, temb);
+                combined_img = result.first;
+                txt          = result.second;
+                sd::ggml_graph_cut::mark_graph_cut(combined_img, "boogu.double_stream_layers." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "boogu.double_stream_layers." + std::to_string(i), "txt");
+            }
+
+            auto hidden_states = ggml_concat(ctx->ggml_ctx, txt, combined_img, 1);
+            for (int i = 0; i < config.num_layers; i++) {
+                auto block    = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["single_stream_layers." + std::to_string(i)]);
+                hidden_states = block->forward(ctx, hidden_states, pe, temb);
+                sd::ggml_graph_cut::mark_graph_cut(hidden_states, "boogu.single_stream_layers." + std::to_string(i), "hidden_states");
+            }
+
+            hidden_states = norm_out->forward(ctx, hidden_states, temb);
+            hidden_states = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, hidden_states->ne[1] - img_len, hidden_states->ne[1]);
+            hidden_states = DiT::unpatchify_and_crop(ctx->ggml_ctx, hidden_states, H, W, config.patch_size, config.patch_size, false);
+            hidden_states = ggml_ext_scale(ctx->ggml_ctx, hidden_states, -1.f);
+            return hidden_states;
+        }
+    };
+
+    __STATIC_INLINE__ int patched_token_count(int64_t size, int patch_size) {
+        int pad = (patch_size - (static_cast<int>(size) % patch_size)) % patch_size;
+        return (static_cast<int>(size) + pad) / patch_size;
+    }
+
+    __STATIC_INLINE__ void append_spatial_ids(std::vector<std::vector<float>>& ids,
+                                              int bs,
+                                              int pe_shift,
+                                              int h_tokens,
+                                              int w_tokens) {
+        std::vector<std::vector<float>> image_ids(h_tokens * w_tokens, std::vector<float>(3, 0.0f));
+        for (int h = 0; h < h_tokens; h++) {
+            for (int w = 0; w < w_tokens; w++) {
+                image_ids[h * w_tokens + w][0] = static_cast<float>(pe_shift);
+                image_ids[h * w_tokens + w][1] = static_cast<float>(h);
+                image_ids[h * w_tokens + w][2] = static_cast<float>(w);
+            }
+        }
+        for (int b = 0; b < bs; b++) {
+            ids.insert(ids.end(), image_ids.begin(), image_ids.end());
+        }
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_boogu_pe(int h,
+                                                      int w,
+                                                      int patch_size,
+                                                      int bs,
+                                                      int context_len,
+                                                      const std::vector<ggml_tensor*>& ref_latents,
+                                                      int theta,
+                                                      const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids;
+        ids.reserve(static_cast<size_t>(bs) * context_len);
+        for (int b = 0; b < bs; b++) {
+            for (int i = 0; i < context_len; i++) {
+                float pos = static_cast<float>(i);
+                ids.push_back({pos, pos, pos});
+            }
+        }
+
+        int pe_shift = context_len;
+        for (ggml_tensor* ref : ref_latents) {
+            int ref_h_tokens = patched_token_count(ref->ne[1], patch_size);
+            int ref_w_tokens = patched_token_count(ref->ne[0], patch_size);
+            append_spatial_ids(ids, bs, pe_shift, ref_h_tokens, ref_w_tokens);
+            pe_shift += std::max(ref_h_tokens, ref_w_tokens);
+        }
+
+        int h_tokens = patched_token_count(h, patch_size);
+        int w_tokens = patched_token_count(w, patch_size);
+        append_spatial_ids(ids, bs, pe_shift, h_tokens, w_tokens);
+
+        return Rope::embed_nd(ids, bs, static_cast<float>(theta), axes_dim);
+    }
+
+    struct BooguImageRunner : public DiffusionModelRunner {
+        BooguConfig config;
+        BooguImageModel boogu;
+        std::vector<float> pe_vec;
+
+        BooguImageRunner(ggml_backend_t backend,
+                         const String2TensorStorage& tensor_storage_map      = {},
+                         const std::string prefix                            = "",
+                         SDVersion version                                   = VERSION_BOOGU_IMAGE,
+                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(BooguConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            boogu = BooguImageModel(config);
+            boogu.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "boogu_image";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            boogu.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const std::vector<sd::Tensor<float>>& ref_latents_tensor = {}) {
+            ggml_cgraph* gf        = new_graph_custom(BOOGU_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            std::vector<ggml_tensor*> ref_latents;
+            ref_latents.reserve(ref_latents_tensor.size());
+            for (const auto& ref_latent_tensor : ref_latents_tensor) {
+                ref_latents.push_back(make_input(ref_latent_tensor));
+            }
+
+            pe_vec      = gen_boogu_pe(static_cast<int>(x->ne[1]),
+                                       static_cast<int>(x->ne[0]),
+                                       config.patch_size,
+                                       static_cast<int>(x->ne[3]),
+                                       static_cast<int>(context->ne[1]),
+                                       ref_latents,
+                                       config.theta,
+                                       config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = boogu.forward(&runner_ctx, x, timesteps, context, pe, ref_latents);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const std::vector<sd::Tensor<float>>& ref_latents = {}) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, ref_latents);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents);
+        }
+    };
+}  // namespace Boogu
+
+#endif  // __SD_MODEL_DIFFUSION_BOOGU_HPP__

src/model/diffusion/control.hpp

@@ -1,11 +1,12 @@
-#ifndef __SD_MODEL_DIFFUSION_CONTROL_HPP__
+#ifndef __SD_MODEL_DIFFUSION_CONTROL_HPP__
 #define __SD_MODEL_DIFFUSION_CONTROL_HPP__
 
 #include "model/common/block.hpp"
 #include "model_loader.h"
 #include "model_manager.h"
 
-#define CONTROL_NET_GRAPH_SIZE 1536
+// Match main UNet's MAX_GRAPH_SIZE so SDXL ControlNet (transformer_depth={1,2,10}) fits.
+#define CONTROL_NET_GRAPH_SIZE MAX_GRAPH_SIZE
 
 /*
     =================================== ControlNet ===================================

src/model/diffusion/dit.hpp

@@ -135,23 +135,23 @@ namespace DiT {
         return x;
     }
 
-    inline ggml_tensor* unpatchify(ggml_context* ctx,
-                                   ggml_tensor* x,
-                                   int64_t t_len,
-                                   int64_t h_len,
-                                   int64_t w_len,
-                                   int pt,
-                                   int ph,
-                                   int pw) {
-        // x: [N, t_len*h_len*w_len, pt*ph*pw*C]
+    inline ggml_tensor* unpatchify_3d(ggml_context* ctx,
+                                      ggml_tensor* x,
+                                      int64_t t_len,
+                                      int64_t h_len,
+                                      int64_t w_len,
+                                      int pt,
+                                      int ph,
+                                      int pw) {
+        // x: [N, t_len*h_len*w_len, C*pt*ph*pw]
         // return: [N*C, t_len*pt, h_len*ph, w_len*pw]
-        int64_t N = x->ne[3];
+        int64_t N = x->ne[2];
         int64_t C = x->ne[0] / pt / ph / pw;
 
         GGML_ASSERT(C * pt * ph * pw == x->ne[0]);
 
-        x = ggml_reshape_4d(ctx, x, C, pw * ph * pt, w_len * h_len * t_len, N);  // [N, t_len*h_len*w_len, pt*ph*pw, C]
-        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3));      // [N, C, t_len*h_len*w_len, pt*ph*pw]
+        x = ggml_reshape_4d(ctx, x, pw * ph * pt, C, w_len * h_len * t_len, N);  // [N, t_len*h_len*w_len, C, pt*ph*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N, C, t_len*h_len*w_len, pt*ph*pw]
         x = ggml_reshape_4d(ctx, x, pw, ph * pt, w_len, h_len * t_len * C * N);  // [N*C*t_len*h_len, w_len, pt*ph, pw]
         x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N*C*t_len*h_len, pt*ph, w_len, pw]
         x = ggml_reshape_4d(ctx, x, pw * w_len, ph, pt, h_len * t_len * C * N);  // [N*C*t_len*h_len, pt, ph, w_len*pw]

src/model/diffusion/flux.hpp

@@ -4,9 +4,12 @@
 #include <memory>
 #include <vector>
 
+#include "core/util.h"
+#include "model/adapter/pulid.hpp"
 #include "model/common/rope.hpp"
 #include "model/diffusion/dit.hpp"
 #include "model/diffusion/model.hpp"
+#include "model/diffusion/sefi_image.hpp"
 #include "model_loader.h"
 
 #define FLUX_GRAPH_SIZE 10240
@@ -25,6 +28,9 @@ namespace Flux {
     struct FluxConfig {
         SDVersion version         = VERSION_FLUX;
         bool is_chroma            = false;
+        bool is_sefi              = false;
+        int64_t semantic_channels = 0;
+        float sefi_delta_t        = 0.1f;
         int patch_size            = 2;
         int64_t in_channels       = 64;
         int64_t out_channels      = 64;
@@ -49,6 +55,10 @@ namespace Flux {
         float ref_index_scale     = 1.f;
         ChromaRadianceConfig chroma_radiance_params;
 
+        bool pulid_enabled        = false;
+        int pulid_double_interval = 2;
+        int pulid_single_interval = 4;
+
         static FluxConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
                                               const std::string& prefix,
                                               SDVersion version = VERSION_FLUX) {
@@ -83,6 +93,21 @@ namespace Flux {
                 config.share_modulation = true;
                 config.ref_index_scale  = 10.f;
                 config.use_mlp_silu_act = true;
+            } else if (sd_version_is_sefi_image(version)) {
+                config.is_sefi           = true;
+                config.semantic_channels = 16;
+                config.in_channels       = 128 + config.semantic_channels;
+                config.patch_size        = 1;
+                config.out_channels      = 128 + config.semantic_channels;
+                config.mlp_ratio         = 3.f;
+                config.theta             = 2000;
+                config.axes_dim          = {32, 32, 32, 32};
+                config.vec_in_dim        = 0;
+                config.qkv_bias          = false;
+                config.disable_bias      = true;
+                config.share_modulation  = true;
+                config.ref_index_scale   = 10.f;
+                config.use_mlp_silu_act  = true;
             } else if (sd_version_is_longcat(version)) {
                 config.context_in_dim = 3584;
                 config.vec_in_dim     = 0;
@@ -138,6 +163,9 @@ namespace Flux {
                 if (ends_with(name, "double_blocks.0.txt_attn.norm.key_norm.scale")) {
                     head_dim = tensor_storage.ne[0];
                 }
+                if (name.find("pulid_ca.") != std::string::npos) {
+                    config.pulid_enabled = true;
+                }
             }
             if (actual_radiance_patch_size > 0 && actual_radiance_patch_size != config.patch_size) {
                 GGML_ASSERT(config.patch_size == 2 * actual_radiance_patch_size);
@@ -715,8 +743,8 @@ namespace Flux {
 
                 auto m     = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, 2 * hidden_size]
                 auto m_vec = ggml_ext_chunk(ctx->ggml_ctx, m, 2, 0);
-                shift      = m_vec[0];  // [N, hidden_size]
-                scale      = m_vec[1];  // [N, hidden_size]
+                shift      = m_vec[0];
+                scale      = m_vec[1];
             }
 
             x = Flux::modulate(ctx->ggml_ctx, norm_final->forward(ctx, x), shift, scale);
@@ -894,6 +922,8 @@ namespace Flux {
             }
             if (config.is_chroma) {
                 blocks["distilled_guidance_layer"] = std::make_shared<ChromaApproximator>(config.in_dim, config.hidden_size);
+            } else if (config.is_sefi) {
+                blocks["dual_time_embed"] = std::make_shared<SefiImage::SefiDualTimestepEmbeddings>(256, config.hidden_size);
             } else {
                 blocks["time_in"] = std::make_shared<MLPEmbedder>(256, config.hidden_size, !config.disable_bias);
                 if (config.vec_in_dim > 0) {
@@ -957,6 +987,20 @@ namespace Flux {
                 blocks["double_stream_modulation_txt"] = std::make_shared<Modulation>(config.hidden_size, true, !config.disable_bias);
                 blocks["single_stream_modulation"]     = std::make_shared<Modulation>(config.hidden_size, false, !config.disable_bias);
             }
+
+            if (config.pulid_enabled) {
+                int num_double_ca = (config.depth + config.pulid_double_interval - 1) / config.pulid_double_interval;
+                int num_single_ca = (config.depth_single_blocks + config.pulid_single_interval - 1) / config.pulid_single_interval;
+                int num_ca        = num_double_ca + num_single_ca;
+                for (int i = 0; i < num_ca; i++) {
+                    blocks["pulid_ca." + std::to_string(i)] =
+                        std::shared_ptr<GGMLBlock>(new PuLIDPerceiverAttentionCA(
+                            /*dim=*/config.hidden_size,
+                            /*dim_head=*/PuLIDPerceiverAttentionCA::DEFAULT_DIM_HEAD,
+                            /*heads=*/PuLIDPerceiverAttentionCA::DEFAULT_HEADS,
+                            /*kv_dim=*/PuLIDPerceiverAttentionCA::DEFAULT_KV_DIM));
+                }
+            }
         }
 
         ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
@@ -967,7 +1011,9 @@ namespace Flux {
                                   ggml_tensor* guidance,
                                   ggml_tensor* pe,
                                   ggml_tensor* mod_index_arange = nullptr,
-                                  std::vector<int> skip_layers  = {}) {
+                                  std::vector<int> skip_layers  = {},
+                                  ggml_tensor* pulid_id         = nullptr,
+                                  float pulid_id_weight         = 1.0f) {
             auto img_in      = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
             auto txt_in      = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
             auto final_layer = std::dynamic_pointer_cast<LastLayer>(blocks["final_layer"]);
@@ -1003,6 +1049,11 @@ namespace Flux {
                 if (y != nullptr) {
                     txt_img_mask = ggml_pad(ctx->ggml_ctx, y, static_cast<int>(img->ne[1]), 0, 0, 0);
                 }
+            } else if (config.is_sefi) {
+                auto dual_time_embed = std::dynamic_pointer_cast<SefiImage::SefiDualTimestepEmbeddings>(blocks["dual_time_embed"]);
+                auto timestep_sem    = ggml_view_1d(ctx->ggml_ctx, timesteps, 1, 0);
+                auto timestep_tex    = ggml_view_1d(ctx->ggml_ctx, timesteps, 1, ggml_element_size(timesteps));
+                vec                  = dual_time_embed->forward(ctx, timestep_sem, timestep_tex);
             } else {
                 auto time_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);
                 vec          = time_in->forward(ctx, ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 256, 10000, 1000.f));
@@ -1044,6 +1095,13 @@ namespace Flux {
             sd::ggml_graph_cut::mark_graph_cut(txt, "flux.prelude", "txt");
             sd::ggml_graph_cut::mark_graph_cut(vec, "flux.prelude", "vec");
 
+            const bool pulid_active = config.pulid_enabled && pulid_id != nullptr;
+            if (pulid_active && !skip_layers.empty()) {
+                LOG_WARN("PuLID + skip_layers is not supported; disabling PuLID for this generation.");
+            }
+            const bool pulid_run = pulid_active && skip_layers.empty();
+            int ca_idx           = 0;
+
             for (int i = 0; i < config.depth; i++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
                     continue;
@@ -1056,9 +1114,19 @@ namespace Flux {
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
                 sd::ggml_graph_cut::mark_graph_cut(img, "flux.double_blocks." + std::to_string(i), "img");
                 sd::ggml_graph_cut::mark_graph_cut(txt, "flux.double_blocks." + std::to_string(i), "txt");
+
+                if (pulid_run && (i % config.pulid_double_interval == 0)) {
+                    auto pulid_ca = std::dynamic_pointer_cast<PuLIDPerceiverAttentionCA>(
+                        blocks["pulid_ca." + std::to_string(ca_idx)]);
+                    ggml_tensor* ca_out = pulid_ca->forward(ctx, pulid_id, img);  // [N, n_img_token, hidden_size]
+                    img                 = ggml_add(ctx->ggml_ctx, img, ggml_scale(ctx->ggml_ctx, ca_out, pulid_id_weight));
+                    sd::ggml_graph_cut::mark_graph_cut(img, "flux.pulid_ca." + std::to_string(ca_idx), "img");
+                    ca_idx++;
+                }
             }
 
-            auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
+            auto txt_img            = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
+            const int64_t n_txt_tok = txt->ne[1];
             for (int i = 0; i < config.depth_single_blocks; i++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i + config.depth) != skip_layers.end()) {
                     continue;
@@ -1067,6 +1135,29 @@ namespace Flux {
 
                 txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask, ss_mods);
                 sd::ggml_graph_cut::mark_graph_cut(txt_img, "flux.single_blocks." + std::to_string(i), "txt_img");
+
+                if (pulid_run && (i % config.pulid_single_interval == 0)) {
+                    auto pulid_ca = std::dynamic_pointer_cast<PuLIDPerceiverAttentionCA>(
+                        blocks["pulid_ca." + std::to_string(ca_idx)]);
+                    ggml_tensor* txt_part = ggml_view_3d(ctx->ggml_ctx, txt_img,
+                                                         txt_img->ne[0], n_txt_tok, txt_img->ne[2],
+                                                         txt_img->nb[1], txt_img->nb[2],
+                                                         0);
+                    ggml_tensor* img_part = ggml_view_3d(ctx->ggml_ctx, txt_img,
+                                                         txt_img->ne[0],
+                                                         txt_img->ne[1] - n_txt_tok,
+                                                         txt_img->ne[2],
+                                                         txt_img->nb[1],
+                                                         txt_img->nb[2],
+                                                         n_txt_tok * txt_img->nb[1]);
+                    txt_part              = ggml_cont(ctx->ggml_ctx, txt_part);
+                    img_part              = ggml_cont(ctx->ggml_ctx, img_part);
+                    ggml_tensor* ca_out   = pulid_ca->forward(ctx, pulid_id, img_part);
+                    img_part              = ggml_add(ctx->ggml_ctx, img_part, ggml_scale(ctx->ggml_ctx, ca_out, pulid_id_weight));
+                    txt_img               = ggml_concat(ctx->ggml_ctx, txt_part, img_part, 1);
+                    sd::ggml_graph_cut::mark_graph_cut(txt_img, "flux.pulid_ca." + std::to_string(ca_idx), "txt_img");
+                    ca_idx++;
+                }
             }
 
             img = ggml_view_3d(ctx->ggml_ctx,
@@ -1105,7 +1196,9 @@ namespace Flux {
                                              ggml_tensor* mod_index_arange         = nullptr,
                                              ggml_tensor* dct                      = nullptr,
                                              std::vector<ggml_tensor*> ref_latents = {},
-                                             std::vector<int> skip_layers          = {}) {
+                                             std::vector<int> skip_layers          = {},
+                                             ggml_tensor* pulid_id                 = nullptr,
+                                             float pulid_id_weight                 = 1.0f) {
             GGML_ASSERT(x->ne[3] == 1);
 
             int64_t W      = x->ne[0];
@@ -1131,7 +1224,8 @@ namespace Flux {
             img = ggml_reshape_3d(ctx->ggml_ctx, img, img->ne[0] * img->ne[1], img->ne[2], img->ne[3]);  // [N, hidden_size, H/patch_size*W/patch_size]
             img = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img, 1, 0, 2, 3));      // [N, H/patch_size*W/patch_size, hidden_size]
 
-            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers);  // [N, n_img_token, hidden_size]
+            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers,
+                                    pulid_id, pulid_id_weight);  // [N, n_img_token, hidden_size]
 
             // nerf decode
             auto nerf_image_embedder   = std::dynamic_pointer_cast<NerfEmbedder>(blocks["nerf_image_embedder"]);
@@ -1179,7 +1273,9 @@ namespace Flux {
                                          ggml_tensor* mod_index_arange         = nullptr,
                                          ggml_tensor* dct                      = nullptr,
                                          std::vector<ggml_tensor*> ref_latents = {},
-                                         std::vector<int> skip_layers          = {}) {
+                                         std::vector<int> skip_layers          = {},
+                                         ggml_tensor* pulid_id                 = nullptr,
+                                         float pulid_id_weight                 = 1.0f) {
             GGML_ASSERT(x->ne[3] == 1);
 
             int64_t W      = x->ne[0];
@@ -1226,7 +1322,8 @@ namespace Flux {
                 }
             }
 
-            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers);  // [N, num_tokens, C * patch_size * patch_size]
+            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers,
+                                    pulid_id, pulid_id_weight);  // [N, num_tokens, C * patch_size * patch_size]
 
             if (out->ne[1] > img_tokens) {
                 out = ggml_view_3d(ctx->ggml_ctx, out, out->ne[0], img_tokens, out->ne[2], out->nb[1], out->nb[2], 0);
@@ -1248,7 +1345,9 @@ namespace Flux {
                              ggml_tensor* mod_index_arange         = nullptr,
                              ggml_tensor* dct                      = nullptr,
                              std::vector<ggml_tensor*> ref_latents = {},
-                             std::vector<int> skip_layers          = {}) {
+                             std::vector<int> skip_layers          = {},
+                             ggml_tensor* pulid_id                 = nullptr,
+                             float pulid_id_weight                 = 1.0f) {
             // Forward pass of DiT.
             // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
             // timestep: (N,) tensor of diffusion timesteps
@@ -1271,7 +1370,9 @@ namespace Flux {
                                                mod_index_arange,
                                                dct,
                                                ref_latents,
-                                               skip_layers);
+                                               skip_layers,
+                                               pulid_id,
+                                               pulid_id_weight);
             } else {
                 return forward_flux_chroma(ctx,
                                            x,
@@ -1284,7 +1385,9 @@ namespace Flux {
                                            mod_index_arange,
                                            dct,
                                            ref_latents,
-                                           skip_layers);
+                                           skip_layers,
+                                           pulid_id,
+                                           pulid_id_weight);
             }
         }
     };
@@ -1298,18 +1401,28 @@ namespace Flux {
         std::vector<float> dct_vec;
         sd::Tensor<float> guidance_tensor;
         SDVersion version;
-        bool use_mask = false;
+        bool use_mask = true;
 
         FluxRunner(ggml_backend_t backend,
                    const String2TensorStorage& tensor_storage_map      = {},
                    const std::string prefix                            = "",
                    SDVersion version                                   = VERSION_FLUX,
-                   bool use_mask                                       = false,
-                   std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+                   std::shared_ptr<RunnerWeightManager> weight_manager = nullptr,
+                   const char* model_args                              = nullptr)
             : DiffusionModelRunner(backend, prefix, weight_manager),
               config(FluxConfig::detect_from_weights(tensor_storage_map, prefix, version)),
-              version(version),
-              use_mask(use_mask) {
+              version(version) {
+            for (const auto& [key, value] : parse_key_value_args(model_args, "model arg")) {
+                if (key == "chroma_use_dit_mask") {
+                    bool parsed = true;
+                    if (parse_strict_bool(value, parsed)) {
+                        use_mask = parsed;
+                    } else {
+                        LOG_WARN("ignoring invalid Chroma DiT model arg '%s=%s'", key.c_str(), value.c_str());
+                    }
+                }
+            }
+
             if (config.is_chroma) {
                 LOG_INFO("Using pruned modulation (Chroma)");
             }
@@ -1383,8 +1496,10 @@ namespace Flux {
                                  const sd::Tensor<float>& y_tensor                        = {},
                                  const sd::Tensor<float>& guidance_tensor                 = {},
                                  const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
-                                 bool increase_ref_index                                  = false,
-                                 std::vector<int> skip_layers                             = {}) {
+                                 Rope::RefIndexMode ref_index_mode                        = Rope::RefIndexMode::FIXED,
+                                 std::vector<int> skip_layers                             = {},
+                                 const sd::Tensor<float>& pulid_id_tensor                 = {},
+                                 float pulid_id_weight                                    = 1.0f) {
             ggml_tensor* x         = make_input(x_tensor);
             ggml_tensor* timesteps = make_input(timesteps_tensor);
             ggml_tensor* context   = make_optional_input(context_tensor);
@@ -1422,9 +1537,9 @@ namespace Flux {
                 set_backend_tensor_data(mod_index_arange, mod_index_arange_vec.data());
             }
             std::set<int> txt_arange_dims;
-            if (sd_version_is_flux2(version)) {
-                txt_arange_dims    = {3};
-                increase_ref_index = true;
+            if (sd_version_is_flux2(version) || sd_version_is_sefi_image(version)) {
+                txt_arange_dims = {3};
+                ref_index_mode  = Rope::RefIndexMode::INCREASE;
             } else if (version == VERSION_OVIS_IMAGE) {
                 txt_arange_dims = {1, 2};
             }
@@ -1435,7 +1550,7 @@ namespace Flux {
                                             static_cast<int>(context->ne[1]),
                                             txt_arange_dims,
                                             ref_latents,
-                                            increase_ref_index,
+                                            ref_index_mode,
                                             config.ref_index_scale,
                                             config.theta,
                                             circular_y_enabled,
@@ -1461,6 +1576,10 @@ namespace Flux {
                 set_backend_tensor_data(dct, dct_vec.data());
             }
 
+            ggml_tensor* pulid_id = pulid_id_tensor.empty()
+                                        ? nullptr
+                                        : make_input(pulid_id_tensor);
+
             auto runner_ctx = get_context();
 
             ggml_tensor* out = flux.forward(&runner_ctx,
@@ -1474,7 +1593,9 @@ namespace Flux {
                                             mod_index_arange,
                                             dct,
                                             ref_latents,
-                                            skip_layers);
+                                            skip_layers,
+                                            pulid_id,
+                                            pulid_id_weight);
 
             ggml_build_forward_expand(gf, out);
 
@@ -1489,15 +1610,18 @@ namespace Flux {
                                   const sd::Tensor<float>& y                        = {},
                                   const sd::Tensor<float>& guidance                 = {},
                                   const std::vector<sd::Tensor<float>>& ref_latents = {},
-                                  bool increase_ref_index                           = false,
-                                  std::vector<int> skip_layers                      = std::vector<int>()) {
+                                  Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::FIXED,
+                                  std::vector<int> skip_layers                      = std::vector<int>(),
+                                  const sd::Tensor<float>& pulid_id                 = {},
+                                  float pulid_id_weight                             = 1.0f) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             // y: [N, adm_in_channels] or [1, adm_in_channels]
             // guidance: [N, ]
+            // pulid_id: empty (no injection) or [N, num_id_tokens=32, kv_dim=2048]
             auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, increase_ref_index, skip_layers);
+                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, ref_index_mode, skip_layers, pulid_id, pulid_id_weight);
             };
 
             auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
@@ -1519,8 +1643,10 @@ namespace Flux {
                            tensor_or_empty(diffusion_params.y),
                            tensor_or_empty(extra->guidance),
                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                           diffusion_params.increase_ref_index,
-                           extra->skip_layers ? *extra->skip_layers : empty_skip_layers);
+                           diffusion_params.ref_index_mode,
+                           extra->skip_layers ? *extra->skip_layers : empty_skip_layers,
+                           tensor_or_empty(extra->pulid_id),
+                           extra->pulid_id_weight);
         }
 
         void test() {
@@ -1568,7 +1694,7 @@ namespace Flux {
                                        {},
                                        guidance,
                                        {},
-                                       false);
+                                       Rope::RefIndexMode::FIXED);
                 int64_t t1   = ggml_time_ms();
 
                 GGML_ASSERT(!out_opt.empty());
@@ -1603,7 +1729,6 @@ namespace Flux {
                                                                             tensor_storage_map,
                                                                             "model.diffusion_model",
                                                                             VERSION_FLUX2,
-                                                                            false,
                                                                             model_manager);
 
             if (!model_manager->register_runner_params("Flux test",

src/model/diffusion/hidream_o1.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
+#ifndef __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
 #define __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
 
 #include <algorithm>

src/model/diffusion/krea2.hpp

@@ -0,0 +1,683 @@
+#ifndef __SD_MODEL_DIFFUSION_KREA2_HPP__
+#define __SD_MODEL_DIFFUSION_KREA2_HPP__
+
+#include <inttypes.h>
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <memory>
+#include <set>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "core/ggml_graph_cut.h"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/flux.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
+
+namespace Krea2 {
+    constexpr int KREA2_GRAPH_SIZE = 65536;
+
+    struct Krea2Config {
+        int patch_size            = 2;
+        int64_t in_channels       = 16;
+        int64_t out_channels      = 16;
+        int64_t features          = 6144;
+        int64_t timestep_dim      = 256;
+        int64_t text_dim          = 2560;
+        int64_t text_layers       = 12;
+        int64_t layers            = 28;
+        int64_t heads             = 48;
+        int64_t kv_heads          = 12;
+        int64_t text_heads        = 20;
+        int64_t text_kv_heads     = 20;
+        int64_t mlp_multiplier    = 4;
+        float theta               = 1000.f;
+        float norm_eps            = 1e-5f;
+        std::vector<int> axes_dim = {32, 48, 48};
+        int axes_dim_sum          = 128;
+
+        int64_t head_dim() const {
+            return features / heads;
+        }
+
+        static int64_t count_blocks(const String2TensorStorage& tensor_storage_map,
+                                    const std::string& prefix,
+                                    const std::string& block_prefix) {
+            int64_t count           = 0;
+            std::string full_prefix = prefix.empty() ? block_prefix : prefix + "." + block_prefix;
+            for (const auto& [name, _] : tensor_storage_map) {
+                if (!starts_with(name, full_prefix)) {
+                    continue;
+                }
+                std::string tail = name.substr(full_prefix.size());
+                size_t dot       = tail.find('.');
+                if (dot == std::string::npos) {
+                    continue;
+                }
+                int block_index = std::atoi(tail.substr(0, dot).c_str());
+                count           = std::max<int64_t>(count, block_index + 1);
+            }
+            return count;
+        }
+
+        void update_axes_dim() {
+            int64_t dim_head = head_dim();
+            int64_t unit     = dim_head / 16;
+            axes_dim         = {
+                        static_cast<int>(dim_head - 12 * unit),
+                        static_cast<int>(6 * unit),
+                        static_cast<int>(6 * unit),
+            };
+            axes_dim_sum = axes_dim[0] + axes_dim[1] + axes_dim[2];
+        }
+
+        static Krea2Config detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                               const std::string& prefix) {
+            Krea2Config config;
+            int64_t detected_head_dim      = 0;
+            int64_t detected_text_head_dim = 0;
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "first.weight") && tensor_storage.n_dims == 2) {
+                    config.in_channels  = tensor_storage.ne[0] / (config.patch_size * config.patch_size);
+                    config.out_channels = config.in_channels;
+                    config.features     = tensor_storage.ne[1];
+                } else if (ends_with(name, "blocks.0.attn.qknorm.qnorm.scale") && tensor_storage.n_dims == 1) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "blocks.0.attn.wq.weight") && tensor_storage.n_dims == 2) {
+                    if (detected_head_dim > 0) {
+                        config.heads = tensor_storage.ne[1] / detected_head_dim;
+                    }
+                } else if (ends_with(name, "blocks.0.attn.wk.weight") && tensor_storage.n_dims == 2) {
+                    if (detected_head_dim > 0) {
+                        config.kv_heads = tensor_storage.ne[1] / detected_head_dim;
+                    }
+                } else if (ends_with(name, "txtfusion.projector.weight") && tensor_storage.n_dims == 2) {
+                    config.text_layers = tensor_storage.ne[0];
+                } else if (ends_with(name, "txtfusion.layerwise_blocks.0.prenorm.scale") && tensor_storage.n_dims == 1) {
+                    config.text_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "txtfusion.layerwise_blocks.0.attn.qknorm.qnorm.scale") && tensor_storage.n_dims == 1) {
+                    detected_text_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "txtfusion.layerwise_blocks.0.attn.wq.weight") && tensor_storage.n_dims == 2) {
+                    if (detected_text_head_dim > 0) {
+                        config.text_heads = tensor_storage.ne[1] / detected_text_head_dim;
+                    }
+                } else if (ends_with(name, "txtfusion.layerwise_blocks.0.attn.wk.weight") && tensor_storage.n_dims == 2) {
+                    if (detected_text_head_dim > 0) {
+                        config.text_kv_heads = tensor_storage.ne[1] / detected_text_head_dim;
+                    }
+                } else if (ends_with(name, "last.linear.weight") && tensor_storage.n_dims == 2) {
+                    config.out_channels = tensor_storage.ne[1] / (config.patch_size * config.patch_size);
+                }
+            }
+
+            config.layers = std::max<int64_t>(1, count_blocks(tensor_storage_map, prefix, "blocks."));
+            if (detected_head_dim > 0 && config.features > 0) {
+                config.heads = config.features / detected_head_dim;
+            }
+            if (detected_head_dim > 0) {
+                std::string wk_name = prefix.empty() ? "blocks.0.attn.wk.weight" : prefix + ".blocks.0.attn.wk.weight";
+                auto it             = tensor_storage_map.find(wk_name);
+                if (it != tensor_storage_map.end() && it->second.n_dims == 2) {
+                    config.kv_heads = it->second.ne[1] / detected_head_dim;
+                }
+            }
+            if (detected_text_head_dim > 0 && config.text_dim > 0) {
+                config.text_heads = config.text_dim / detected_text_head_dim;
+            }
+            if (detected_text_head_dim > 0) {
+                std::string wk_name = prefix.empty() ? "txtfusion.layerwise_blocks.0.attn.wk.weight" : prefix + ".txtfusion.layerwise_blocks.0.attn.wk.weight";
+                auto it             = tensor_storage_map.find(wk_name);
+                if (it != tensor_storage_map.end() && it->second.n_dims == 2) {
+                    config.text_kv_heads = it->second.ne[1] / detected_text_head_dim;
+                }
+            }
+            config.update_axes_dim();
+
+            LOG_DEBUG("krea2: layers=%" PRId64 ", features=%" PRId64 ", heads=%" PRId64 ", kv_heads=%" PRId64 ", text_dim=%" PRId64 ", text_layers=%" PRId64 ", text_heads=%" PRId64 ", text_kv_heads=%" PRId64 ", channels=%" PRId64,
+                      config.layers,
+                      config.features,
+                      config.heads,
+                      config.kv_heads,
+                      config.text_dim,
+                      config.text_layers,
+                      config.text_heads,
+                      config.text_kv_heads,
+                      config.in_channels);
+            return config;
+        }
+    };
+
+    __STATIC_INLINE__ int64_t ceil_to_multiple(int64_t value, int64_t multiple) {
+        return ((value + multiple - 1) / multiple) * multiple;
+    }
+
+    class KreaRMSNorm : public UnaryBlock {
+    protected:
+        int64_t hidden_size;
+        float eps;
+        std::string prefix;
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            this->prefix    = prefix;
+            params["scale"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
+        }
+
+    public:
+        KreaRMSNorm(int64_t hidden_size, float eps = 1e-5f)
+            : hidden_size(hidden_size),
+              eps(eps) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            ggml_tensor* scale = params["scale"];
+            scale              = ggml_add(ctx->ggml_ctx, scale, ggml_ext_ones(ctx->ggml_ctx, scale->ne[0], 1, 1, 1));
+            x                  = ggml_rms_norm(ctx->ggml_ctx, x, eps);
+            x                  = ggml_mul_inplace(ctx->ggml_ctx, x, scale);
+            return x;
+        }
+    };
+
+    class KreaSwiGLU : public UnaryBlock {
+    public:
+        KreaSwiGLU(int64_t features, int64_t multiplier) {
+            int64_t mlp_dim = ceil_to_multiple(((2 * features) / 3) * multiplier, 128);
+            blocks["gate"]  = std::make_shared<Linear>(features, mlp_dim, false);
+            blocks["up"]    = std::make_shared<Linear>(features, mlp_dim, false);
+            blocks["down"]  = std::make_shared<Linear>(mlp_dim, features, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            auto gate = std::dynamic_pointer_cast<Linear>(blocks["gate"]);
+            auto up   = std::dynamic_pointer_cast<Linear>(blocks["up"]);
+            auto down = std::dynamic_pointer_cast<Linear>(blocks["down"]);
+
+            auto gated = ggml_silu(ctx->ggml_ctx, gate->forward(ctx, x));
+            auto up_x  = up->forward(ctx, x);
+            x          = ggml_mul(ctx->ggml_ctx, gated, up_x);
+            return down->forward(ctx, x);
+        }
+    };
+
+    class KreaAttention : public GGMLBlock {
+    protected:
+        int64_t features;
+        int64_t heads;
+        int64_t kv_heads;
+        int64_t head_dim_;
+
+        ggml_tensor* attention_no_rope(GGMLRunnerContext* ctx,
+                                       ggml_tensor* q,
+                                       ggml_tensor* k,
+                                       ggml_tensor* v,
+                                       ggml_tensor* mask) {
+            int64_t Lq = q->ne[2];
+            int64_t Lk = k->ne[2];
+            int64_t N  = q->ne[3];
+            q          = ggml_reshape_3d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, q), head_dim_ * heads, Lq, N);
+            k          = ggml_reshape_3d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, k), head_dim_ * kv_heads, Lk, N);
+            v          = ggml_reshape_3d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, v), head_dim_ * kv_heads, Lk, N);
+            return ggml_ext_attention_ext(ctx->ggml_ctx,
+                                          ctx->backend,
+                                          q,
+                                          k,
+                                          v,
+                                          heads,
+                                          mask,
+                                          false,
+                                          ctx->flash_attn_enabled);
+        }
+
+    public:
+        KreaAttention(int64_t features,
+                      int64_t heads,
+                      int64_t kv_heads,
+                      float eps = 1e-5f)
+            : features(features),
+              heads(heads),
+              kv_heads(kv_heads),
+              head_dim_(features / heads) {
+            blocks["wq"]           = std::make_shared<Linear>(features, heads * head_dim_, false);
+            blocks["wk"]           = std::make_shared<Linear>(features, kv_heads * head_dim_, false);
+            blocks["wv"]           = std::make_shared<Linear>(features, kv_heads * head_dim_, false);
+            blocks["gate"]         = std::make_shared<Linear>(features, features, false);
+            blocks["qknorm.qnorm"] = std::make_shared<KreaRMSNorm>(head_dim_, eps);
+            blocks["qknorm.knorm"] = std::make_shared<KreaRMSNorm>(head_dim_, eps);
+            blocks["wo"]           = std::make_shared<Linear>(features, features, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe   = nullptr,
+                             ggml_tensor* mask = nullptr) {
+            auto wq    = std::dynamic_pointer_cast<Linear>(blocks["wq"]);
+            auto wk    = std::dynamic_pointer_cast<Linear>(blocks["wk"]);
+            auto wv    = std::dynamic_pointer_cast<Linear>(blocks["wv"]);
+            auto gate  = std::dynamic_pointer_cast<Linear>(blocks["gate"]);
+            auto qnorm = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["qknorm.qnorm"]);
+            auto knorm = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["qknorm.knorm"]);
+            auto wo    = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                wo->set_force_prec_f32(true);
+            }
+
+            int64_t L = x->ne[1];
+            int64_t N = x->ne[2];
+
+            auto q = wq->forward(ctx, x);
+            q      = ggml_reshape_4d(ctx->ggml_ctx, q, head_dim_, heads, L, N);
+            auto k = wk->forward(ctx, x);
+            k      = ggml_reshape_4d(ctx->ggml_ctx, k, head_dim_, kv_heads, L, N);
+            auto v = wv->forward(ctx, x);
+            v      = ggml_reshape_4d(ctx->ggml_ctx, v, head_dim_, kv_heads, L, N);
+
+            q = qnorm->forward(ctx, q);
+            k = knorm->forward(ctx, k);
+
+            auto out = pe != nullptr ? Rope::attention(ctx, q, k, v, pe, mask)
+                                     : attention_no_rope(ctx, q, k, v, mask);
+            out      = ggml_mul(ctx->ggml_ctx, out, ggml_sigmoid(ctx->ggml_ctx, gate->forward(ctx, x)));
+            out      = wo->forward(ctx, out);
+            return out;
+        }
+    };
+
+    class KreaDoubleSharedModulation : public GGMLBlock {
+    protected:
+        int64_t dim;
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            GGML_UNUSED(prefix);
+            params["lin"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim * 6);
+        }
+
+    public:
+        KreaDoubleSharedModulation(int64_t dim)
+            : dim(dim) {}
+
+        std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* vec) {
+            auto lin = ggml_repeat(ctx->ggml_ctx, params["lin"], vec);
+            auto out = ggml_add(ctx->ggml_ctx, vec, lin);
+            return ggml_ext_chunk(ctx->ggml_ctx, out, 6, 0);
+        }
+    };
+
+    class KreaFinalModulation : public GGMLBlock {
+    protected:
+        int64_t dim;
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            GGML_UNUSED(prefix);
+            params["lin"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, dim, 2);
+        }
+
+    public:
+        KreaFinalModulation(int64_t dim)
+            : dim(dim) {}
+
+        std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* vec) {
+            auto out = ggml_add(ctx->ggml_ctx, params["lin"], vec);
+            return ggml_ext_chunk(ctx->ggml_ctx, out, 2, 1);
+        }
+    };
+
+    class KreaTextFusionBlock : public UnaryBlock {
+    public:
+        KreaTextFusionBlock(int64_t dim,
+                            int64_t heads,
+                            int64_t kv_heads,
+                            int64_t multiplier,
+                            float eps) {
+            blocks["prenorm"]  = std::make_shared<KreaRMSNorm>(dim, eps);
+            blocks["postnorm"] = std::make_shared<KreaRMSNorm>(dim, eps);
+            blocks["attn"]     = std::make_shared<KreaAttention>(dim, heads, kv_heads, eps);
+            blocks["mlp"]      = std::make_shared<KreaSwiGLU>(dim, multiplier);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            auto prenorm  = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["prenorm"]);
+            auto postnorm = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["postnorm"]);
+            auto attn     = std::dynamic_pointer_cast<KreaAttention>(blocks["attn"]);
+            auto mlp      = std::dynamic_pointer_cast<KreaSwiGLU>(blocks["mlp"]);
+
+            x = ggml_add(ctx->ggml_ctx, x, attn->forward(ctx, prenorm->forward(ctx, x)));
+            x = ggml_add(ctx->ggml_ctx, x, mlp->forward(ctx, postnorm->forward(ctx, x)));
+            return x;
+        }
+    };
+
+    class KreaTextFusionTransformer : public UnaryBlock {
+    protected:
+        Krea2Config config;
+
+    public:
+        explicit KreaTextFusionTransformer(Krea2Config config)
+            : config(std::move(config)) {
+            for (int i = 0; i < 2; ++i) {
+                blocks["layerwise_blocks." + std::to_string(i)] = std::make_shared<KreaTextFusionBlock>(this->config.text_dim,
+                                                                                                        this->config.text_heads,
+                                                                                                        this->config.text_kv_heads,
+                                                                                                        this->config.mlp_multiplier,
+                                                                                                        this->config.norm_eps);
+                blocks["refiner_blocks." + std::to_string(i)]   = std::make_shared<KreaTextFusionBlock>(this->config.text_dim,
+                                                                                                      this->config.text_heads,
+                                                                                                      this->config.text_kv_heads,
+                                                                                                      this->config.mlp_multiplier,
+                                                                                                      this->config.norm_eps);
+            }
+            blocks["projector"] = std::make_shared<Linear>(this->config.text_layers, 1, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* context) override {
+            int64_t text_tokens = context->ne[1];
+            int64_t batch       = context->ne[2];
+
+            context = ggml_reshape_3d(ctx->ggml_ctx,
+                                      context,
+                                      config.text_dim,
+                                      config.text_layers,
+                                      text_tokens * batch);
+
+            for (int i = 0; i < 2; ++i) {
+                auto block = std::dynamic_pointer_cast<KreaTextFusionBlock>(blocks["layerwise_blocks." + std::to_string(i)]);
+                context    = block->forward(ctx, context);
+            }
+
+            context        = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, context, 1, 0, 2, 3));
+            auto projector = std::dynamic_pointer_cast<Linear>(blocks["projector"]);
+            context        = projector->forward(ctx, context);
+            context        = ggml_reshape_3d(ctx->ggml_ctx, context, config.text_dim, text_tokens, batch);
+
+            for (int i = 0; i < 2; ++i) {
+                auto block = std::dynamic_pointer_cast<KreaTextFusionBlock>(blocks["refiner_blocks." + std::to_string(i)]);
+                context    = block->forward(ctx, context);
+            }
+            return context;
+        }
+    };
+
+    class KreaSingleStreamBlock : public UnaryBlock {
+    public:
+        explicit KreaSingleStreamBlock(Krea2Config config) {
+            blocks["mod"]      = std::make_shared<KreaDoubleSharedModulation>(config.features);
+            blocks["prenorm"]  = std::make_shared<KreaRMSNorm>(config.features, config.norm_eps);
+            blocks["postnorm"] = std::make_shared<KreaRMSNorm>(config.features, config.norm_eps);
+            blocks["attn"]     = std::make_shared<KreaAttention>(config.features, config.heads, config.kv_heads, config.norm_eps);
+            blocks["mlp"]      = std::make_shared<KreaSwiGLU>(config.features, config.mlp_multiplier);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* vec,
+                             ggml_tensor* pe) {
+            auto mod      = std::dynamic_pointer_cast<KreaDoubleSharedModulation>(blocks["mod"]);
+            auto prenorm  = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["prenorm"]);
+            auto postnorm = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["postnorm"]);
+            auto attn     = std::dynamic_pointer_cast<KreaAttention>(blocks["attn"]);
+            auto mlp      = std::dynamic_pointer_cast<KreaSwiGLU>(blocks["mlp"]);
+
+            auto mods       = mod->forward(ctx, vec);
+            auto attn_input = Flux::modulate(ctx->ggml_ctx,
+                                             prenorm->forward(ctx, x),
+                                             mods[1],
+                                             mods[0],
+                                             true);
+            auto attn_out   = attn->forward(ctx, attn_input, pe);
+            x               = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, mods[2]));
+
+            auto mlp_input = Flux::modulate(ctx->ggml_ctx,
+                                            postnorm->forward(ctx, x),
+                                            mods[4],
+                                            mods[3],
+                                            true);
+            auto mlp_out   = mlp->forward(ctx, mlp_input);
+            x              = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, mlp_out, mods[5]));
+            return x;
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            GGML_UNUSED(ctx);
+            GGML_UNUSED(x);
+            GGML_ABORT("KreaSingleStreamBlock requires conditioning");
+            return nullptr;
+        }
+    };
+
+    class KreaTimeMLP : public UnaryBlock {
+    public:
+        explicit KreaTimeMLP(Krea2Config config) {
+            blocks["0"] = std::make_shared<Linear>(config.timestep_dim, config.features, true);
+            blocks["2"] = std::make_shared<Linear>(config.features, config.features, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            auto linear_0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+            x             = linear_0->forward(ctx, x);
+            x             = ggml_ext_gelu(ctx->ggml_ctx, x, false);
+            x             = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class KreaTProj : public UnaryBlock {
+    public:
+        explicit KreaTProj(Krea2Config config) {
+            blocks["1"] = std::make_shared<Linear>(config.features, config.features * 6, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
+            x             = ggml_ext_gelu(ctx->ggml_ctx, x, false);
+            x             = linear_1->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class KreaTextMLP : public UnaryBlock {
+    public:
+        explicit KreaTextMLP(Krea2Config config) {
+            blocks["0"] = std::make_shared<KreaRMSNorm>(config.text_dim, config.norm_eps);
+            blocks["1"] = std::make_shared<Linear>(config.text_dim, config.features, true);
+            blocks["3"] = std::make_shared<Linear>(config.features, config.features, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            auto norm     = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["0"]);
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
+            auto linear_3 = std::dynamic_pointer_cast<Linear>(blocks["3"]);
+            x             = norm->forward(ctx, x);
+            x             = linear_1->forward(ctx, x);
+            x             = ggml_ext_gelu(ctx->ggml_ctx, x, true);
+            x             = linear_3->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class KreaLastLayer : public GGMLBlock {
+    public:
+        explicit KreaLastLayer(Krea2Config config) {
+            blocks["norm"]       = std::make_shared<KreaRMSNorm>(config.features, config.norm_eps);
+            blocks["linear"]     = std::make_shared<Linear>(config.features, config.patch_size * config.patch_size * config.out_channels, true);
+            blocks["modulation"] = std::make_shared<KreaFinalModulation>(config.features);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* vec) {
+            auto norm       = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["norm"]);
+            auto linear     = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto modulation = std::dynamic_pointer_cast<KreaFinalModulation>(blocks["modulation"]);
+
+            auto mods = modulation->forward(ctx, vec);
+            x         = Flux::modulate(ctx->ggml_ctx,
+                                       norm->forward(ctx, x),
+                                       mods[1],
+                                       mods[0],
+                                       true);
+            x         = linear->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Krea2Model : public GGMLBlock {
+    protected:
+        Krea2Config config;
+
+    public:
+        Krea2Model() = default;
+        explicit Krea2Model(Krea2Config config)
+            : config(std::move(config)) {
+            blocks["first"]     = std::make_shared<Linear>(this->config.patch_size * this->config.patch_size * this->config.in_channels,
+                                                       this->config.features,
+                                                       true);
+            blocks["tmlp"]      = std::make_shared<KreaTimeMLP>(this->config);
+            blocks["txtfusion"] = std::make_shared<KreaTextFusionTransformer>(this->config);
+            blocks["txtmlp"]    = std::make_shared<KreaTextMLP>(this->config);
+            blocks["tproj"]     = std::make_shared<KreaTProj>(this->config);
+            for (int i = 0; i < this->config.layers; ++i) {
+                blocks["blocks." + std::to_string(i)] = std::make_shared<KreaSingleStreamBlock>(this->config);
+            }
+            blocks["last"] = std::make_shared<KreaLastLayer>(this->config);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t N = x->ne[3];
+            GGML_ASSERT(N == 1);
+
+            auto first     = std::dynamic_pointer_cast<Linear>(blocks["first"]);
+            auto tmlp      = std::dynamic_pointer_cast<KreaTimeMLP>(blocks["tmlp"]);
+            auto txtfusion = std::dynamic_pointer_cast<KreaTextFusionTransformer>(blocks["txtfusion"]);
+            auto txtmlp    = std::dynamic_pointer_cast<KreaTextMLP>(blocks["txtmlp"]);
+            auto tproj     = std::dynamic_pointer_cast<KreaTProj>(blocks["tproj"]);
+            auto last      = std::dynamic_pointer_cast<KreaLastLayer>(blocks["last"]);
+
+            auto img        = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size, true);
+            int64_t img_len = img->ne[1];
+            img             = first->forward(ctx, img);
+
+            auto t    = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(config.timestep_dim), 10000, 1000.f);
+            t         = tmlp->forward(ctx, t);
+            t         = ggml_reshape_3d(ctx->ggml_ctx, t, t->ne[0], 1, t->ne[1]);
+            auto tvec = tproj->forward(ctx, t);
+
+            auto txt        = txtfusion->forward(ctx, context);
+            txt             = txtmlp->forward(ctx, txt);
+            int64_t txt_len = txt->ne[1];
+
+            auto hidden_states = ggml_concat(ctx->ggml_ctx, txt, img, 1);
+            for (int i = 0; i < config.layers; ++i) {
+                auto block    = std::dynamic_pointer_cast<KreaSingleStreamBlock>(blocks["blocks." + std::to_string(i)]);
+                hidden_states = block->forward(ctx, hidden_states, tvec, pe);
+                sd::ggml_graph_cut::mark_graph_cut(hidden_states, "krea2.blocks." + std::to_string(i), "hidden_states");
+            }
+
+            hidden_states = last->forward(ctx, hidden_states, t);
+            hidden_states = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, txt_len, txt_len + img_len);
+            hidden_states = DiT::unpatchify_and_crop(ctx->ggml_ctx, hidden_states, H, W, config.patch_size, config.patch_size, true);
+            return hidden_states;
+        }
+    };
+
+    __STATIC_INLINE__ std::vector<float> gen_krea2_pe(int h,
+                                                      int w,
+                                                      int patch_size,
+                                                      int bs,
+                                                      int context_len,
+                                                      float theta,
+                                                      const std::vector<int>& axes_dim) {
+        auto txt_ids = Rope::gen_flux_txt_ids(bs, context_len, 3, {});
+        auto img_ids = Rope::gen_flux_img_ids(h, w, patch_size, bs, 3, 0, 0, 0, false);
+        auto ids     = Rope::concat_ids(txt_ids, img_ids, bs);
+        return Rope::embed_nd(ids, bs, theta, axes_dim);
+    }
+
+    struct Krea2Runner : public DiffusionModelRunner {
+        Krea2Config config;
+        Krea2Model model;
+        std::vector<float> pe_vec;
+
+        Krea2Runner(ggml_backend_t backend,
+                    const String2TensorStorage& tensor_storage_map      = {},
+                    const std::string prefix                            = "",
+                    std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(Krea2Config::detect_from_weights(tensor_storage_map, prefix)) {
+            model = Krea2Model(config);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "krea2";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor) {
+            ggml_cgraph* gf        = new_graph_custom(KREA2_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            pe_vec      = gen_krea2_pe(static_cast<int>(x->ne[1]),
+                                       static_cast<int>(x->ne[0]),
+                                       config.patch_size,
+                                       static_cast<int>(x->ne[3]),
+                                       static_cast<int>(context->ne[1]),
+                                       config.theta,
+                                       config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = model.forward(&runner_ctx, x, timesteps, context, pe);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context));
+        }
+    };
+}  // namespace Krea2
+
+#endif  // __SD_MODEL_DIFFUSION_KREA2_HPP__

src/model/diffusion/minit2i.hpp

@@ -0,0 +1,611 @@
+#ifndef __SD_MODEL_DIFFUSION_MINIT2I_HPP__
+#define __SD_MODEL_DIFFUSION_MINIT2I_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
+
+namespace MiniT2I {
+    constexpr int MINIT2I_GRAPH_SIZE = 196608;
+
+    struct MiniT2IConfig {
+        int64_t image_size         = 512;
+        int64_t patch_size         = 16;
+        int64_t in_channels        = 3;
+        int64_t txt_input_size     = 1024;
+        int64_t hidden_size        = 768;
+        int64_t txt_hidden_size    = 768;
+        int64_t cond_vec_size      = 768;
+        int64_t depth_double       = 17;
+        int64_t txt_preamble_depth = 2;
+        int64_t num_heads          = 12;
+        int64_t head_dim           = 64;
+        float mlp_ratio            = 2.6667f;
+        int64_t pca_channels       = 128;
+        int64_t prompt_length      = 256;
+        int64_t n_T                = 100;
+        float cfg_interval_start   = 0.0f;
+        float cfg_interval_end     = 1.0f;
+
+        static MiniT2IConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            MiniT2IConfig config;
+            config.depth_double       = 0;
+            config.txt_preamble_depth = 0;
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "img_embedder.proj1.weight") && tensor_storage.n_dims == 4) {
+                    config.patch_size   = tensor_storage.ne[0];
+                    config.in_channels  = tensor_storage.ne[2];
+                    config.pca_channels = tensor_storage.ne[3];
+                } else if (ends_with(name, "img_embedder.proj2.weight") && tensor_storage.n_dims == 4) {
+                    config.pca_channels = tensor_storage.ne[2];
+                    config.hidden_size  = tensor_storage.ne[3];
+                } else if (ends_with(name, "txt_embedder.weight") && tensor_storage.n_dims == 2) {
+                    config.txt_input_size  = tensor_storage.ne[0];
+                    config.txt_hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "pooled_embedder.weight") && tensor_storage.n_dims == 2) {
+                    config.cond_vec_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "double_blocks.0.img_qkv.weight") && tensor_storage.n_dims == 2) {
+                    int64_t inner3     = tensor_storage.ne[1];
+                    int64_t inner      = inner3 / 3;
+                    config.hidden_size = tensor_storage.ne[0];
+                    if (config.hidden_size == 768) {
+                        config.num_heads = 12;
+                        config.head_dim  = 64;
+                    } else if (config.hidden_size == 1248) {
+                        config.num_heads = 24;
+                        config.head_dim  = 52;
+                    } else if (inner > 0) {
+                        config.head_dim  = 64;
+                        config.num_heads = std::max<int64_t>(1, inner / config.head_dim);
+                    }
+                } else if (ends_with(name, "final_layer.linear.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area = config.patch_size * config.patch_size;
+                    config.hidden_size = tensor_storage.ne[0];
+                    config.in_channels = patch_area > 0 ? tensor_storage.ne[1] / patch_area : config.in_channels;
+                } else if (ends_with(name, "mask_token") && tensor_storage.n_dims >= 2) {
+                    config.prompt_length = tensor_storage.ne[1];
+                }
+
+                size_t pos = name.find("double_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int64_t idx         = atoi(items[1].c_str());
+                        config.depth_double = std::max<int64_t>(config.depth_double, idx + 1);
+                    }
+                }
+                pos = name.find("txt_preamble_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int64_t idx               = atoi(items[1].c_str());
+                        config.txt_preamble_depth = std::max<int64_t>(config.txt_preamble_depth, idx + 1);
+                    }
+                }
+            }
+
+            if (config.depth_double <= 0) {
+                config.depth_double = config.hidden_size == 1248 ? 23 : 17;
+            }
+            if (config.txt_preamble_depth <= 0) {
+                config.txt_preamble_depth = 2;
+            }
+            if (config.head_dim <= 0 || config.num_heads <= 0) {
+                config.head_dim  = config.hidden_size == 1248 ? 52 : 64;
+                config.num_heads = config.hidden_size / config.head_dim;
+            }
+            LOG_DEBUG("minit2i: hidden_size=%" PRId64 ", txt_hidden_size=%" PRId64 ", heads=%" PRId64 ", head_dim=%" PRId64 ", double_blocks=%" PRId64 ", txt_blocks=%" PRId64 ", patch=%" PRId64 ", in_channels=%" PRId64,
+                      config.hidden_size,
+                      config.txt_hidden_size,
+                      config.num_heads,
+                      config.head_dim,
+                      config.depth_double,
+                      config.txt_preamble_depth,
+                      config.patch_size,
+                      config.in_channels);
+            return config;
+        }
+    };
+
+    inline std::vector<float> make_2d_sincos_pos_embed(int grid_size, int dim) {
+        GGML_ASSERT(dim % 4 == 0);
+        int half_dim = dim / 2;
+        int quarter  = half_dim / 2;
+        std::vector<float> out(static_cast<size_t>(grid_size) * grid_size * dim);
+        std::vector<float> omega(quarter);
+        for (int i = 0; i < quarter; ++i) {
+            omega[i] = 1.0f / std::pow(10000.0f, static_cast<float>(i) / static_cast<float>(quarter));
+        }
+        for (int y = 0; y < grid_size; ++y) {
+            for (int x = 0; x < grid_size; ++x) {
+                size_t base = static_cast<size_t>(y * grid_size + x) * dim;
+                for (int i = 0; i < quarter; ++i) {
+                    float ay                           = y * omega[i];
+                    float ax                           = x * omega[i];
+                    out[base + i]                      = std::sin(ax);
+                    out[base + quarter + i]            = std::cos(ax);
+                    out[base + half_dim + i]           = std::sin(ay);
+                    out[base + half_dim + quarter + i] = std::cos(ay);
+                }
+            }
+        }
+        return out;
+    }
+
+    inline std::vector<float> make_text_rope(int length, int head_dim) {
+        return Rope::flatten(Rope::rope(Rope::linspace(0.f, static_cast<float>(length - 1), length), head_dim, 10000.f));
+    }
+
+    inline std::vector<float> make_vision_rope(int side, int head_dim) {
+        GGML_ASSERT(head_dim % 4 == 0);
+        int dim     = head_dim / 2;
+        int quarter = dim / 2;
+        int length  = side * side;
+        std::vector<float> out(static_cast<size_t>(length) * (head_dim / 2) * 4);
+        std::vector<float> freqs(quarter);
+        for (int i = 0; i < quarter; ++i) {
+            freqs[i] = 1.0f / std::pow(10000.0f, static_cast<float>(2 * i) / static_cast<float>(dim));
+        }
+        for (int y = 0; y < side; ++y) {
+            for (int x = 0; x < side; ++x) {
+                int pos     = y * side + x;
+                size_t base = static_cast<size_t>(pos) * (head_dim / 2) * 4;
+                for (int i = 0; i < quarter; ++i) {
+                    float ay        = y * freqs[i];
+                    float ax        = x * freqs[i];
+                    float angles[2] = {ay, ax};
+                    for (int axis = 0; axis < 2; ++axis) {
+                        int j                 = axis * quarter + i;
+                        out[base + 4 * j]     = std::cos(angles[axis]);
+                        out[base + 4 * j + 1] = -std::sin(angles[axis]);
+                        out[base + 4 * j + 2] = std::sin(angles[axis]);
+                        out[base + 4 * j + 3] = std::cos(angles[axis]);
+                    }
+                }
+            }
+        }
+        return out;
+    }
+
+    struct SwiGLUMlp : public GGMLBlock {
+        SwiGLUMlp(int64_t in_features, int64_t hidden_features) {
+            int64_t hidden_dim = ((hidden_features + 7) / 8) * 8;
+            blocks["w1"]       = std::make_shared<Linear>(in_features, hidden_dim, false);
+            blocks["w3"]       = std::make_shared<Linear>(in_features, hidden_dim, false);
+            blocks["w2"]       = std::make_shared<Linear>(hidden_dim, in_features, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1   = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w3   = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+            auto w2   = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto gate = ggml_silu(ctx->ggml_ctx, w1->forward(ctx, x));
+            auto up   = w3->forward(ctx, x);
+            return w2->forward(ctx, ggml_mul(ctx->ggml_ctx, gate, up));
+        }
+    };
+
+    struct BottleneckPatchEmbed : public GGMLBlock {
+        int64_t patch_size;
+
+        BottleneckPatchEmbed(int64_t patch_size, int64_t in_channels, int64_t pca_channels, int64_t hidden_size)
+            : patch_size(patch_size) {
+            blocks["proj1"] = std::make_shared<Conv2d>(in_channels,
+                                                       pca_channels,
+                                                       std::pair<int, int>{static_cast<int>(patch_size), static_cast<int>(patch_size)},
+                                                       std::pair<int, int>{static_cast<int>(patch_size), static_cast<int>(patch_size)},
+                                                       std::pair<int, int>{0, 0},
+                                                       std::pair<int, int>{1, 1},
+                                                       false);
+            blocks["proj2"] = std::make_shared<Conv2d>(pca_channels,
+                                                       hidden_size,
+                                                       std::pair<int, int>{1, 1},
+                                                       std::pair<int, int>{1, 1},
+                                                       std::pair<int, int>{0, 0},
+                                                       std::pair<int, int>{1, 1},
+                                                       true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj1 = std::dynamic_pointer_cast<Conv2d>(blocks["proj1"]);
+            auto proj2 = std::dynamic_pointer_cast<Conv2d>(blocks["proj2"]);
+            x          = proj1->forward(ctx, x);
+            x          = proj2->forward(ctx, x);
+            x          = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0] * x->ne[1], x->ne[2], x->ne[3]);
+            x          = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
+            return x;
+        }
+    };
+
+    struct TimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size;
+
+        TimestepEmbedder(int64_t hidden_size, int frequency_embedding_size = 256)
+            : frequency_embedding_size(frequency_embedding_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
+            auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10000, 1.0f);
+            t_emb      = mlp_0->forward(ctx, t_emb);
+            t_emb      = ggml_silu_inplace(ctx->ggml_ctx, t_emb);
+            return mlp_2->forward(ctx, t_emb);
+        }
+    };
+
+    inline std::vector<ggml_tensor*> split_qkv(ggml_context* ctx, ggml_tensor* qkv, int64_t num_heads, int64_t head_dim) {
+        int64_t N = qkv->ne[2];
+        int64_t L = qkv->ne[1];
+        auto q    = ggml_view_4d(ctx, qkv, head_dim, num_heads, L, N,
+                                 qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], 0);
+        auto k    = ggml_view_4d(ctx, qkv, head_dim, num_heads, L, N,
+                                 qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], qkv->nb[0] * head_dim * num_heads);
+        auto v    = ggml_view_4d(ctx, qkv, head_dim, num_heads, L, N,
+                                 qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], qkv->nb[0] * head_dim * num_heads * 2);
+        return {q, k, v};
+    }
+
+    struct PlainTextTransformerBlock : public GGMLBlock {
+        int64_t num_heads;
+        int64_t head_dim;
+
+        PlainTextTransformerBlock(int64_t hidden_size, int64_t num_heads, int64_t head_dim, float mlp_ratio)
+            : num_heads(num_heads), head_dim(head_dim) {
+            int64_t inner_dim   = num_heads * head_dim;
+            blocks["norm1"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["norm2"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["qkv"]       = std::make_shared<Linear>(hidden_size, inner_dim * 3, true);
+            blocks["attn_proj"] = std::make_shared<Linear>(inner_dim, hidden_size, true);
+            blocks["mlp"]       = std::make_shared<SwiGLUMlp>(hidden_size, static_cast<int64_t>(hidden_size * mlp_ratio));
+            blocks["q_norm"]    = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"]    = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* txt, ggml_tensor* pe) {
+            auto norm1     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+            auto norm2     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto qkv_proj  = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto attn_proj = std::dynamic_pointer_cast<Linear>(blocks["attn_proj"]);
+            auto mlp       = std::dynamic_pointer_cast<SwiGLUMlp>(blocks["mlp"]);
+            auto q_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+
+            auto qkv = split_qkv(ctx->ggml_ctx, qkv_proj->forward(ctx, norm1->forward(ctx, txt)), num_heads, head_dim);
+            auto q   = q_norm->forward(ctx, qkv[0]);
+            auto k   = k_norm->forward(ctx, qkv[1]);
+            auto v   = qkv[2];
+            auto out = Rope::attention(ctx, q, k, v, pe, nullptr, 1.0f, false);
+            txt      = ggml_add(ctx->ggml_ctx, txt, attn_proj->forward(ctx, out));
+            txt      = ggml_add(ctx->ggml_ctx, txt, mlp->forward(ctx, norm2->forward(ctx, txt)));
+            return txt;
+        }
+    };
+
+    struct DoubleStreamDiTBlock : public GGMLBlock {
+        int64_t num_heads;
+        int64_t head_dim;
+
+        DoubleStreamDiTBlock(int64_t hidden_size, int64_t txt_hidden_size, int64_t num_heads, int64_t head_dim, float mlp_ratio)
+            : num_heads(num_heads), head_dim(head_dim) {
+            int64_t inner_dim       = num_heads * head_dim;
+            blocks["img_norm1"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["img_norm2"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["txt_norm1"]     = std::make_shared<RMSNorm>(txt_hidden_size, 1e-6f);
+            blocks["txt_norm2"]     = std::make_shared<RMSNorm>(txt_hidden_size, 1e-6f);
+            blocks["img_qkv"]       = std::make_shared<Linear>(hidden_size, inner_dim * 3, true);
+            blocks["txt_qkv"]       = std::make_shared<Linear>(txt_hidden_size, inner_dim * 3, true);
+            blocks["q_norm"]        = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"]        = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["img_attn_proj"] = std::make_shared<Linear>(inner_dim, hidden_size, true);
+            blocks["txt_attn_proj"] = std::make_shared<Linear>(inner_dim, txt_hidden_size, true);
+            blocks["img_mlp"]       = std::make_shared<SwiGLUMlp>(hidden_size, static_cast<int64_t>(hidden_size * mlp_ratio));
+            blocks["txt_mlp"]       = std::make_shared<SwiGLUMlp>(txt_hidden_size, static_cast<int64_t>(txt_hidden_size * mlp_ratio));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* pe) {
+            auto img_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm2"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm2"]);
+            auto img_qkv_p = std::dynamic_pointer_cast<Linear>(blocks["img_qkv"]);
+            auto txt_qkv_p = std::dynamic_pointer_cast<Linear>(blocks["txt_qkv"]);
+            auto q_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto img_proj  = std::dynamic_pointer_cast<Linear>(blocks["img_attn_proj"]);
+            auto txt_proj  = std::dynamic_pointer_cast<Linear>(blocks["txt_attn_proj"]);
+            auto img_mlp   = std::dynamic_pointer_cast<SwiGLUMlp>(blocks["img_mlp"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<SwiGLUMlp>(blocks["txt_mlp"]);
+
+            int64_t li = img->ne[1];
+            int64_t lt = txt->ne[1];
+
+            auto img_qkv = split_qkv(ctx->ggml_ctx, img_qkv_p->forward(ctx, img_norm1->forward(ctx, img)), num_heads, head_dim);
+            auto txt_qkv = split_qkv(ctx->ggml_ctx, txt_qkv_p->forward(ctx, txt_norm1->forward(ctx, txt)), num_heads, head_dim);
+
+            auto q = ggml_concat(ctx->ggml_ctx, q_norm->forward(ctx, txt_qkv[0]), q_norm->forward(ctx, img_qkv[0]), 2);
+            auto k = ggml_concat(ctx->ggml_ctx, k_norm->forward(ctx, txt_qkv[1]), k_norm->forward(ctx, img_qkv[1]), 2);
+            auto v = ggml_concat(ctx->ggml_ctx, txt_qkv[2], img_qkv[2], 2);
+
+            auto out     = Rope::attention(ctx, q, k, v, pe, nullptr, 1.0f, false);
+            auto out_txt = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, lt);
+            auto out_img = ggml_ext_slice(ctx->ggml_ctx, out, 1, lt, lt + li);
+
+            img = ggml_add(ctx->ggml_ctx, img, img_proj->forward(ctx, out_img));
+            txt = ggml_add(ctx->ggml_ctx, txt, txt_proj->forward(ctx, out_txt));
+            img = ggml_add(ctx->ggml_ctx, img, img_mlp->forward(ctx, img_norm2->forward(ctx, img)));
+            txt = ggml_add(ctx->ggml_ctx, txt, txt_mlp->forward(ctx, txt_norm2->forward(ctx, txt)));
+            return {img, txt};
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t patch_size, int64_t out_channels) {
+            blocks["norm_final"] = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["linear"]     = std::make_shared<Linear>(hidden_size, patch_size * patch_size * out_channels, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto norm_final = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_final"]);
+            auto linear     = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, norm_final->forward(ctx, x));
+        }
+    };
+
+    struct MMJiT : public GGMLBlock {
+        MiniT2IConfig config;
+
+        MMJiT(const MiniT2IConfig& config)
+            : config(config) {
+            blocks["img_embedder"]    = std::make_shared<BottleneckPatchEmbed>(config.patch_size, config.in_channels, config.pca_channels, config.hidden_size);
+            blocks["txt_embedder"]    = std::make_shared<Linear>(config.txt_input_size, config.txt_hidden_size, false);
+            blocks["t_embedder"]      = std::make_shared<TimestepEmbedder>(config.cond_vec_size);
+            blocks["pooled_embedder"] = std::make_shared<Linear>(config.txt_input_size, config.cond_vec_size, false);
+            for (int64_t i = 0; i < config.txt_preamble_depth; ++i) {
+                blocks["txt_preamble_blocks." + std::to_string(i)] = std::make_shared<PlainTextTransformerBlock>(config.txt_hidden_size, config.num_heads, config.head_dim, config.mlp_ratio);
+            }
+            for (int64_t i = 0; i < config.depth_double; ++i) {
+                blocks["double_blocks." + std::to_string(i)] = std::make_shared<DoubleStreamDiTBlock>(config.hidden_size, config.txt_hidden_size, config.num_heads, config.head_dim, config.mlp_ratio);
+            }
+            blocks["final_layer"] = std::make_shared<FinalLayer>(config.hidden_size, config.patch_size, config.in_channels);
+        }
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGMLBlock::init_params(ctx, tensor_storage_map, prefix);
+            enum ggml_type wtype = get_type(prefix + "mask_token", tensor_storage_map, GGML_TYPE_F32);
+            params["mask_token"] = ggml_new_tensor_3d(ctx, wtype, config.txt_input_size, 1, 1);
+        }
+
+        ggml_tensor* apply_text_mask(GGMLRunnerContext* ctx, ggml_tensor* context, ggml_tensor* mask) {
+            if (mask == nullptr) {
+                return context;
+            }
+            mask            = ggml_reshape_3d(ctx->ggml_ctx, mask, 1, mask->ne[0], mask->ne[1]);
+            mask            = ggml_repeat(ctx->ggml_ctx, mask, context);
+            auto keep       = ggml_mul(ctx->ggml_ctx, context, mask);
+            auto inv        = ggml_sub(ctx->ggml_ctx, ggml_ext_ones_like(ctx->ggml_ctx, mask), mask);
+            auto mask_token = ggml_repeat(ctx->ggml_ctx, params["mask_token"], context);
+            return ggml_add(ctx->ggml_ctx, keep, ggml_mul(ctx->ggml_ctx, mask_token, inv));
+        }
+
+        ggml_tensor* pool_context(GGMLRunnerContext* ctx, ggml_tensor* context) {
+            int64_t dim = context->ne[0];
+            int64_t len = context->ne[1];
+            int64_t N   = context->ne[2];
+            auto x      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, context, 1, 0, 2, 3));
+            x           = ggml_reshape_3d(ctx->ggml_ctx, x, len, dim, N);
+            x           = ggml_mean(ctx->ggml_ctx, x);
+            x           = ggml_reshape_2d(ctx->ggml_ctx, x, dim, N);
+            return x;
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* img,
+                             ggml_tensor* context,
+                             ggml_tensor* mask,
+                             ggml_tensor* pos_embed,
+                             ggml_tensor* txt_pe,
+                             ggml_tensor* joint_pe) {
+            auto img_embedder = std::dynamic_pointer_cast<BottleneckPatchEmbed>(blocks["img_embedder"]);
+            auto txt_embedder = std::dynamic_pointer_cast<Linear>(blocks["txt_embedder"]);
+            auto final_layer  = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
+
+            int64_t W  = img->ne[0];
+            int64_t H  = img->ne[1];
+            int64_t hp = H / config.patch_size;
+            int64_t wp = W / config.patch_size;
+
+            context = apply_text_mask(ctx, context, mask);
+            auto x  = img_embedder->forward(ctx, img);
+            x       = ggml_add(ctx->ggml_ctx, x, pos_embed);
+
+            auto txt = txt_embedder->forward(ctx, context);
+            for (int64_t i = 0; i < config.txt_preamble_depth; ++i) {
+                auto block = std::dynamic_pointer_cast<PlainTextTransformerBlock>(blocks["txt_preamble_blocks." + std::to_string(i)]);
+                txt        = block->forward(ctx, txt, txt_pe);
+                sd::ggml_graph_cut::mark_graph_cut(txt, "minit2i.txt_preamble_blocks." + std::to_string(i), "txt");
+            }
+            for (int64_t i = 0; i < config.depth_double; ++i) {
+                auto block = std::dynamic_pointer_cast<DoubleStreamDiTBlock>(blocks["double_blocks." + std::to_string(i)]);
+                auto out   = block->forward(ctx, x, txt, joint_pe);
+                x          = out.first;
+                txt        = out.second;
+                sd::ggml_graph_cut::mark_graph_cut(x, "minit2i.double_blocks." + std::to_string(i), "x");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "minit2i.double_blocks." + std::to_string(i), "txt");
+            }
+            auto combined = ggml_concat(ctx->ggml_ctx, txt, x, 1);
+            auto out      = final_layer->forward(ctx, combined);
+            auto img_out  = ggml_ext_slice(ctx->ggml_ctx, out, 1, txt->ne[1], txt->ne[1] + x->ne[1]);
+            return DiT::unpatchify(ctx->ggml_ctx, img_out, hp, wp, static_cast<int>(config.patch_size), static_cast<int>(config.patch_size), false);
+        }
+    };
+
+    struct MiniT2IRunner : public DiffusionModelRunner {
+        MiniT2IConfig config;
+        MMJiT model;
+        ggml_context* position_cache_ctx            = nullptr;
+        ggml_backend_buffer_t position_cache_buffer = nullptr;
+        ggml_tensor* cached_pos_embed               = nullptr;
+        ggml_tensor* cached_txt_pe                  = nullptr;
+        ggml_tensor* cached_joint_pe                = nullptr;
+        int64_t cached_img_side                     = -1;
+        int64_t cached_txt_len                      = -1;
+        int64_t cached_hidden_size                  = -1;
+        int64_t cached_head_dim                     = -1;
+
+        MiniT2IRunner(ggml_backend_t backend,
+                      const String2TensorStorage& tensor_storage_map      = {},
+                      const std::string prefix                            = "",
+                      std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(MiniT2IConfig::detect_from_weights(tensor_storage_map, this->prefix)),
+              model(config) {
+            model.init(params_ctx, tensor_storage_map, this->prefix);
+        }
+
+        ~MiniT2IRunner() override {
+            free_position_cache();
+        }
+
+        std::string get_desc() override {
+            return "MiniT2I";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        void free_position_cache() {
+            if (position_cache_buffer != nullptr) {
+                ggml_backend_buffer_free(position_cache_buffer);
+                position_cache_buffer = nullptr;
+            }
+            if (position_cache_ctx != nullptr) {
+                ggml_free(position_cache_ctx);
+                position_cache_ctx = nullptr;
+            }
+            cached_pos_embed   = nullptr;
+            cached_txt_pe      = nullptr;
+            cached_joint_pe    = nullptr;
+            cached_img_side    = -1;
+            cached_txt_len     = -1;
+            cached_hidden_size = -1;
+            cached_head_dim    = -1;
+        }
+
+        void ensure_position_cache(int64_t img_side, int64_t txt_len) {
+            if (cached_img_side == img_side &&
+                cached_txt_len == txt_len &&
+                cached_hidden_size == config.hidden_size &&
+                cached_head_dim == config.head_dim &&
+                cached_pos_embed != nullptr &&
+                cached_txt_pe != nullptr &&
+                cached_joint_pe != nullptr) {
+                return;
+            }
+
+            free_position_cache();
+
+            auto pos_embed_vec = make_2d_sincos_pos_embed(static_cast<int>(img_side), static_cast<int>(config.hidden_size));
+            auto txt_pe_vec    = make_text_rope(static_cast<int>(txt_len), static_cast<int>(config.head_dim));
+            auto img_pe_vec    = make_vision_rope(static_cast<int>(img_side), static_cast<int>(config.head_dim));
+            auto joint_pe_vec  = txt_pe_vec;
+            joint_pe_vec.insert(joint_pe_vec.end(), img_pe_vec.begin(), img_pe_vec.end());
+
+            ggml_init_params params;
+            params.mem_size    = static_cast<size_t>(3 * ggml_tensor_overhead());
+            params.mem_buffer  = nullptr;
+            params.no_alloc    = true;
+            position_cache_ctx = ggml_init(params);
+            GGML_ASSERT(position_cache_ctx != nullptr);
+
+            cached_pos_embed = ggml_new_tensor_3d(position_cache_ctx, GGML_TYPE_F32, config.hidden_size, img_side * img_side, 1);
+            ggml_set_name(cached_pos_embed, "minit2i.pos_embed");
+            cached_txt_pe = ggml_new_tensor_4d(position_cache_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, txt_len);
+            ggml_set_name(cached_txt_pe, "minit2i.txt_pe");
+            cached_joint_pe = ggml_new_tensor_4d(position_cache_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, txt_len + img_side * img_side);
+            ggml_set_name(cached_joint_pe, "minit2i.joint_pe");
+
+            position_cache_buffer = ggml_backend_alloc_ctx_tensors(position_cache_ctx, runtime_backend);
+            GGML_ASSERT(position_cache_buffer != nullptr);
+            ggml_backend_buffer_set_usage(position_cache_buffer, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+            ggml_backend_tensor_set(cached_pos_embed, pos_embed_vec.data(), 0, ggml_nbytes(cached_pos_embed));
+            ggml_backend_tensor_set(cached_txt_pe, txt_pe_vec.data(), 0, ggml_nbytes(cached_txt_pe));
+            ggml_backend_tensor_set(cached_joint_pe, joint_pe_vec.data(), 0, ggml_nbytes(cached_joint_pe));
+            ggml_backend_synchronize(runtime_backend);
+
+            cached_img_side    = img_side;
+            cached_txt_len     = txt_len;
+            cached_hidden_size = config.hidden_size;
+            cached_head_dim    = config.head_dim;
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const sd::Tensor<float>& mask_tensor) {
+            ggml_cgraph* gf      = new_graph_custom(MINIT2I_GRAPH_SIZE);
+            ggml_tensor* x       = make_input(x_tensor);
+            ggml_tensor* context = make_input(context_tensor);
+            ggml_tensor* mask    = make_input(mask_tensor);
+            SD_UNUSED(timesteps_tensor);
+
+            int64_t W        = x->ne[0];
+            int64_t H        = x->ne[1];
+            int64_t img_side = H / config.patch_size;
+            int64_t txt_len  = context->ne[1];
+            ensure_position_cache(img_side, txt_len);
+
+            auto runner_ctx = get_context();
+            auto out        = model.forward(&runner_ctx, x, context, mask, cached_pos_embed, cached_txt_pe, cached_joint_pe);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const sd::Tensor<float>& mask) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, mask);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            GGML_ASSERT(diffusion_params.context != nullptr);
+            const auto* extra = diffusion_extra_as<MiniT2IDiffusionExtra>(diffusion_params);
+            GGML_ASSERT(extra->mask != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *diffusion_params.context,
+                           *extra->mask);
+        }
+    };
+}  // namespace MiniT2I
+
+#endif  // __SD_MODEL_DIFFUSION_MINIT2I_HPP__

src/model/diffusion/model.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_DIFFUSION_MODEL_HPP__
+#ifndef __SD_MODEL_DIFFUSION_MODEL_HPP__
 #define __SD_MODEL_DIFFUSION_MODEL_HPP__
 
 #include <string>
@@ -7,6 +7,7 @@
 
 #include "core/ggml_extend.hpp"
 #include "core/tensor_ggml.hpp"
+#include "model/common/rope.hpp"
 #include "model_manager.h"
 
 struct UNetDiffusionExtra {
@@ -22,6 +23,8 @@ struct SkipLayerDiffusionExtra {
 struct FluxDiffusionExtra {
     const sd::Tensor<float>* guidance   = nullptr;
     const std::vector<int>* skip_layers = nullptr;
+    const sd::Tensor<float>* pulid_id   = nullptr;
+    float pulid_id_weight               = 1.0f;
 };
 
 struct AnimaDiffusionExtra {
@@ -50,6 +53,10 @@ struct LTXAVDiffusionExtra {
     const sd::Tensor<float>* video_positions = nullptr;
 };
 
+struct MiniT2IDiffusionExtra {
+    const sd::Tensor<float>* mask = nullptr;
+};
+
 using DiffusionExtraParams = std::variant<std::monostate,
                                           UNetDiffusionExtra,
                                           SkipLayerDiffusionExtra,
@@ -57,7 +64,8 @@ using DiffusionExtraParams = std::variant<std::monostate,
                                           AnimaDiffusionExtra,
                                           WanDiffusionExtra,
                                           HiDreamO1DiffusionExtra,
-                                          LTXAVDiffusionExtra>;
+                                          LTXAVDiffusionExtra,
+                                          MiniT2IDiffusionExtra>;
 
 struct DiffusionParams {
     const sd::Tensor<float>* x                        = nullptr;
@@ -66,7 +74,7 @@ struct DiffusionParams {
     const sd::Tensor<float>* c_concat                 = nullptr;
     const sd::Tensor<float>* y                        = nullptr;
     const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
-    bool increase_ref_index                           = false;
+    Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::FIXED;
     DiffusionExtraParams extra                        = std::monostate{};
 };
 

src/model/diffusion/qwen_image.hpp

@@ -3,7 +3,9 @@
 
 #include <memory>
 
+#include "core/util.h"
 #include "model/common/block.hpp"
+#include "model/diffusion/dit.hpp"
 #include "model/diffusion/flux.hpp"
 #include "model/diffusion/model.hpp"
 #include "model_loader.h"
@@ -23,6 +25,7 @@ namespace Qwen {
         std::vector<int> axes_dim   = {16, 56, 56};
         int axes_dim_sum            = 128;
         bool zero_cond_t            = false;
+        bool use_additional_t_cond  = false;
 
         static QwenImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
             QwenImageConfig config;
@@ -88,19 +91,33 @@ namespace Qwen {
     };
 
     struct QwenTimestepProjEmbeddings : public GGMLBlock {
+    protected:
+        bool use_additional_t_cond = false;
+
     public:
-        QwenTimestepProjEmbeddings(int64_t embedding_dim) {
+        QwenTimestepProjEmbeddings(int64_t embedding_dim, bool use_additional_t_cond = false)
+            : use_additional_t_cond(use_additional_t_cond) {
             blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
+            if (use_additional_t_cond) {
+                blocks["addition_t_embedding"] = std::shared_ptr<GGMLBlock>(new Embedding(2, embedding_dim));
+            }
         }
 
         ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* timesteps) {
+                             ggml_tensor* timesteps,
+                             ggml_tensor* addition_t_cond = nullptr) {
             // timesteps: [N,]
             // return: [N, embedding_dim]
             auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
 
             auto timesteps_proj = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 256, 10000, 1.f);
             auto timesteps_emb  = timestep_embedder->forward(ctx, timesteps_proj);
+            if (use_additional_t_cond) {
+                GGML_ASSERT(addition_t_cond != nullptr);
+                auto addition_t_embedding = std::dynamic_pointer_cast<Embedding>(blocks["addition_t_embedding"]);
+                auto addition_t_emb       = addition_t_embedding->forward(ctx, addition_t_cond);
+                timesteps_emb             = ggml_add(ctx->ggml_ctx, timesteps_emb, addition_t_emb);
+            }
             return timesteps_emb;
         }
     };
@@ -402,7 +419,7 @@ namespace Qwen {
         QwenImageModel(QwenImageConfig config)
             : config(config) {
             int64_t inner_dim         = config.num_attention_heads * config.attention_head_dim;
-            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim));
+            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim, config.use_additional_t_cond));
             blocks["txt_norm"]        = std::shared_ptr<GGMLBlock>(new RMSNorm(config.joint_attention_dim, 1e-6f));
             blocks["img_in"]          = std::shared_ptr<GGMLBlock>(new Linear(config.in_channels, inner_dim));
             blocks["txt_in"]          = std::shared_ptr<GGMLBlock>(new Linear(config.joint_attention_dim, inner_dim));
@@ -424,6 +441,7 @@ namespace Qwen {
         ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
                                   ggml_tensor* x,
                                   ggml_tensor* timestep,
+                                  ggml_tensor* addition_t_cond,
                                   ggml_tensor* context,
                                   ggml_tensor* pe,
                                   ggml_tensor* modulate_index = nullptr) {
@@ -434,9 +452,9 @@ namespace Qwen {
             auto norm_out        = std::dynamic_pointer_cast<AdaLayerNormContinuous>(blocks["norm_out"]);
             auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
 
-            auto t_emb = time_text_embed->forward(ctx, timestep);
+            auto t_emb = time_text_embed->forward(ctx, timestep, addition_t_cond);
             if (config.zero_cond_t) {
-                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep), addition_t_cond);
                 t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
             }
             auto img = img_in->forward(ctx, x);
@@ -469,33 +487,50 @@ namespace Qwen {
         ggml_tensor* forward(GGMLRunnerContext* ctx,
                              ggml_tensor* x,
                              ggml_tensor* timestep,
+                             ggml_tensor* addition_t_cond,
                              ggml_tensor* context,
                              ggml_tensor* pe,
                              std::vector<ggml_tensor*> ref_latents = {},
                              ggml_tensor* modulate_index           = nullptr) {
             // Forward pass of DiT.
-            // x: [N, C, H, W]
+            // x: [N, C, H, W] or [N*C, T, H, W]
             // timestep: [N,]
             // context: [N, L, D]
             // pe: [L, d_head/2, 2, 2]
-            // return: [N, C, H, W]
+            // return: [N, C, H, W] or [N*C, T, H, W]
 
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-            int64_t C = x->ne[2];
-            int64_t N = x->ne[3];
+            int64_t W          = x->ne[0];
+            int64_t H          = x->ne[1];
+            int64_t T          = 1;
+            int64_t N          = addition_t_cond != nullptr ? addition_t_cond->ne[0] : x->ne[3];
+            bool has_time_axis = false;
+            if (x->ne[3] != 1) {
+                T             = x->ne[2];
+                has_time_axis = true;
+            }
 
-            auto img           = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size);
+            auto patchify_input = [&](ggml_tensor* input) -> ggml_tensor* {
+                input = DiT::pad_to_patch_size(ctx, input, config.patch_size, config.patch_size);
+                if (!has_time_axis) {
+                    return DiT::patchify(ctx->ggml_ctx, input, config.patch_size, config.patch_size);
+                }
+                if (input->ne[3] == 1) {
+                    input = ggml_reshape_4d(ctx->ggml_ctx, input, input->ne[0], input->ne[1], 1, input->ne[2]);
+                }
+                return DiT::patchify(ctx->ggml_ctx, input, 1, config.patch_size, config.patch_size, N);
+            };
+
+            auto img           = patchify_input(x);
             int64_t img_tokens = img->ne[1];
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = DiT::pad_and_patchify(ctx, ref, config.patch_size, config.patch_size);
+                    ref = patchify_input(ref);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
 
-            auto out = forward_orig(ctx, img, timestep, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]
+            auto out = forward_orig(ctx, img, timestep, addition_t_cond, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]
 
             if (out->ne[1] > img_tokens) {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [num_tokens, N, C * patch_size * patch_size]
@@ -503,7 +538,17 @@ namespace Qwen {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
             }
 
-            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, config.patch_size, config.patch_size);  // [N, C, H, W]
+            if (has_time_axis) {
+                int pad_h = (config.patch_size - H % config.patch_size) % config.patch_size;
+                int pad_w = (config.patch_size - W % config.patch_size) % config.patch_size;
+                int h_len = static_cast<int>((H + pad_h) / config.patch_size);
+                int w_len = static_cast<int>((W + pad_w) / config.patch_size);
+                out       = DiT::unpatchify_3d(ctx->ggml_ctx, out, T, h_len, w_len, 1, config.patch_size, config.patch_size);
+                out       = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H);  // [N*C, T, H, W + pad_w]
+                out       = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W);  // [N*C, T, H, W]
+            } else {
+                out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, config.patch_size, config.patch_size);  // [N, C, H, W]
+            }
 
             return out;
         }
@@ -515,18 +560,32 @@ namespace Qwen {
         QwenImageModel qwen_image;
         std::vector<float> pe_vec;
         std::vector<float> modulate_index_vec;
+        std::vector<int32_t> additional_t_cond_vec;
         SDVersion version;
 
         QwenImageRunner(ggml_backend_t backend,
                         const String2TensorStorage& tensor_storage_map      = {},
                         const std::string prefix                            = "",
                         SDVersion version                                   = VERSION_QWEN_IMAGE,
-                        bool zero_cond_t                                    = false,
-                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr,
+                        const char* model_args                              = nullptr)
             : DiffusionModelRunner(backend, prefix, weight_manager),
-              config(QwenImageConfig::detect_from_weights(tensor_storage_map, prefix)) {
-            config.zero_cond_t = config.zero_cond_t || zero_cond_t;
-            qwen_image         = QwenImageModel(config);
+              config(QwenImageConfig::detect_from_weights(tensor_storage_map, prefix)),
+              version(version) {
+            for (const auto& [key, value] : parse_key_value_args(model_args, "model arg")) {
+                if (key == "qwen_image_zero_cond_t") {
+                    bool parsed = false;
+                    if (parse_strict_bool(value, parsed)) {
+                        config.zero_cond_t = config.zero_cond_t || parsed;
+                    } else {
+                        LOG_WARN("ignoring invalid Qwen Image model arg '%s=%s'", key.c_str(), value.c_str());
+                    }
+                }
+            }
+            if (version == VERSION_QWEN_IMAGE_LAYERED) {
+                config.use_additional_t_cond = true;
+            }
+            qwen_image = QwenImageModel(config);
             qwen_image.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -542,11 +601,11 @@ namespace Qwen {
                                  const sd::Tensor<float>& timesteps_tensor,
                                  const sd::Tensor<float>& context_tensor,
                                  const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
-                                 bool increase_ref_index                                  = false) {
+                                 Rope::RefIndexMode ref_index_mode                        = Rope::RefIndexMode::INCREASE) {
             ggml_cgraph* gf        = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE);
             ggml_tensor* x         = make_input(x_tensor);
             ggml_tensor* timesteps = make_input(timesteps_tensor);
-            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(x->ne[3] == 1 || x_tensor.dim() == 5);
             GGML_ASSERT(!context_tensor.empty());
             ggml_tensor* context = make_input(context_tensor);
             std::vector<ggml_tensor*> ref_latents;
@@ -555,13 +614,29 @@ namespace Qwen {
                 ref_latents.push_back(make_input(ref_latent_tensor));
             }
 
-            pe_vec      = Rope::gen_qwen_image_pe(static_cast<int>(x->ne[1]),
+            int batch_size = static_cast<int>(x->ne[3]);
+            int time_len   = 1;
+            if (x_tensor.dim() == 5) {
+                time_len   = static_cast<int>(x_tensor.shape()[2]);
+                batch_size = static_cast<int>(x_tensor.shape()[4]);
+            }
+
+            ggml_tensor* addition_t_cond = nullptr;
+            if (version == VERSION_QWEN_IMAGE_LAYERED) {
+                additional_t_cond_vec.assign(static_cast<size_t>(batch_size), 0);
+                addition_t_cond = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, batch_size);
+                set_backend_tensor_data(addition_t_cond, additional_t_cond_vec.data());
+                ref_index_mode = Rope::RefIndexMode::DECREASE;
+            }
+
+            pe_vec      = Rope::gen_qwen_image_pe(time_len,
+                                                  static_cast<int>(x->ne[1]),
                                                   static_cast<int>(x->ne[0]),
                                                   config.patch_size,
-                                                  static_cast<int>(x->ne[3]),
+                                                  batch_size,
                                                   static_cast<int>(context->ne[1]),
                                                   ref_latents,
-                                                  increase_ref_index,
+                                                  ref_index_mode,
                                                   config.theta,
                                                   circular_y_enabled,
                                                   circular_x_enabled,
@@ -604,6 +679,7 @@ namespace Qwen {
             ggml_tensor* out = qwen_image.forward(&runner_ctx,
                                                   x,
                                                   timesteps,
+                                                  addition_t_cond,
                                                   context,
                                                   pe,
                                                   ref_latents,
@@ -619,12 +695,12 @@ namespace Qwen {
                                   const sd::Tensor<float>& timesteps,
                                   const sd::Tensor<float>& context,
                                   const std::vector<sd::Tensor<float>>& ref_latents = {},
-                                  bool increase_ref_index                           = false) {
-            // x: [N, in_channels, h, w]
+                                  Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::INCREASE) {
+            // x: [N, C, H, W] or [N*C, T, H, W]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+                return build_graph(x, timesteps, context, ref_latents, ref_index_mode);
             };
 
             return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
@@ -640,7 +716,7 @@ namespace Qwen {
                            *diffusion_params.timesteps,
                            tensor_or_empty(diffusion_params.context),
                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                           diffusion_params.increase_ref_index);
+                           diffusion_params.ref_index_mode);
         }
 
         void test() {
@@ -674,7 +750,7 @@ namespace Qwen {
                                        timesteps,
                                        context,
                                        {},
-                                       false);
+                                       Rope::RefIndexMode::FIXED);
                 int64_t t1   = ggml_time_ms();
 
                 GGML_ASSERT(!out_opt.empty());
@@ -709,7 +785,6 @@ namespace Qwen {
                                                                                             tensor_storage_map,
                                                                                             "model.diffusion_model",
                                                                                             VERSION_QWEN_IMAGE,
-                                                                                            false,
                                                                                             model_manager);
 
             if (!model_manager->register_runner_params("Qwen image test",

src/model/diffusion/sefi_image.hpp

@@ -0,0 +1,91 @@
+#ifndef __SD_MODEL_DIFFUSION_SEFI_IMAGE_HPP__
+#define __SD_MODEL_DIFFUSION_SEFI_IMAGE_HPP__
+
+#include <memory>
+
+#include "model/common/block.hpp"
+
+namespace SefiImage {
+    struct SefiImageConfig {
+        int64_t semantic_channels        = 16;
+        int64_t texture_latent_channels  = 32;
+        int64_t timestep_guidance_in_dim = 256;
+        int64_t hidden_size              = 3072;
+        float timestep_shift_alpha       = 0.3f;
+        float delta_t                    = 0.1f;
+
+        int64_t packed_texture_channels(int patch_size) const {
+            return texture_latent_channels * patch_size * patch_size;
+        }
+
+        int64_t packed_input_channels(int patch_size) const {
+            return semantic_channels + packed_texture_channels(patch_size);
+        }
+
+        static SefiImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                                   const std::string& prefix) {
+            SefiImageConfig config;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "dual_time_embed.semantic_embedder.linear_1.weight") && tensor_storage.n_dims == 2) {
+                    config.timestep_guidance_in_dim = tensor_storage.ne[0];
+                    config.hidden_size              = tensor_storage.ne[1] * 2;
+                }
+            }
+            LOG_DEBUG("sefi_image: semantic_channels = %" PRId64 ", texture_latent_channels = %" PRId64 ", hidden_size = %" PRId64,
+                      config.semantic_channels,
+                      config.texture_latent_channels,
+                      config.hidden_size);
+            return config;
+        }
+    };
+
+    struct SefiTimestepEmbedding : public GGMLBlock {
+    public:
+        SefiTimestepEmbedding(int64_t in_channels, int64_t time_embed_dim) {
+            blocks["linear_1"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, time_embed_dim, false));
+            blocks["linear_2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim, false));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* sample) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+
+            sample = linear_1->forward(ctx, sample);
+            sample = ggml_silu_inplace(ctx->ggml_ctx, sample);
+            sample = linear_2->forward(ctx, sample);
+            return sample;
+        }
+    };
+
+    struct SefiDualTimestepEmbeddings : public GGMLBlock {
+    public:
+        SefiDualTimestepEmbeddings(int64_t in_channels, int64_t embedding_dim) {
+            GGML_ASSERT(embedding_dim % 2 == 0);
+            int64_t half_dim            = embedding_dim / 2;
+            blocks["semantic_embedder"] = std::make_shared<SefiTimestepEmbedding>(in_channels, half_dim);
+            blocks["texture_embedder"]  = std::make_shared<SefiTimestepEmbedding>(in_channels, half_dim);
+            timestep_guidance_in_dim    = in_channels;
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* timestep_sem,
+                             ggml_tensor* timestep_tex) {
+            auto semantic_embedder = std::dynamic_pointer_cast<SefiTimestepEmbedding>(blocks["semantic_embedder"]);
+            auto texture_embedder  = std::dynamic_pointer_cast<SefiTimestepEmbedding>(blocks["texture_embedder"]);
+
+            auto sem_proj = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep_sem, (int)timestep_guidance_in_dim, 10000, 1.f);
+            auto tex_proj = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep_tex, (int)timestep_guidance_in_dim, 10000, 1.f);
+            auto sem_emb  = semantic_embedder->forward(ctx, sem_proj);
+            auto tex_emb  = texture_embedder->forward(ctx, tex_proj);
+            return ggml_concat(ctx->ggml_ctx, sem_emb, tex_emb, 0);
+        }
+
+    private:
+        int64_t timestep_guidance_in_dim = 256;
+    };
+}  // namespace SefiImage
+
+#endif  // __SD_MODEL_DIFFUSION_SEFI_IMAGE_HPP__

src/model/diffusion/z_image.hpp

@@ -575,7 +575,7 @@ namespace ZImage {
                                  const sd::Tensor<float>& timesteps_tensor,
                                  const sd::Tensor<float>& context_tensor,
                                  const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
-                                 bool increase_ref_index                                  = false) {
+                                 Rope::RefIndexMode ref_index_mode                        = Rope::RefIndexMode::FIXED) {
             ggml_cgraph* gf        = new_graph_custom(Z_IMAGE_GRAPH_SIZE);
             ggml_tensor* x         = make_input(x_tensor);
             ggml_tensor* timesteps = make_input(timesteps_tensor);
@@ -595,7 +595,7 @@ namespace ZImage {
                                                static_cast<int>(context->ne[1]),
                                                SEQ_MULTI_OF,
                                                ref_latents,
-                                               increase_ref_index,
+                                               ref_index_mode,
                                                config.theta,
                                                circular_y_enabled,
                                                circular_x_enabled,
@@ -626,12 +626,12 @@ namespace ZImage {
                                   const sd::Tensor<float>& timesteps,
                                   const sd::Tensor<float>& context,
                                   const std::vector<sd::Tensor<float>>& ref_latents = {},
-                                  bool increase_ref_index                           = false) {
+                                  Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::FIXED) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+                return build_graph(x, timesteps, context, ref_latents, ref_index_mode);
             };
 
             return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
@@ -647,7 +647,7 @@ namespace ZImage {
                            *diffusion_params.timesteps,
                            tensor_or_empty(diffusion_params.context),
                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                           diffusion_params.increase_ref_index);
+                           diffusion_params.ref_index_mode);
         }
 
         void test() {
@@ -681,7 +681,7 @@ namespace ZImage {
                                        timesteps,
                                        context,
                                        {},
-                                       false);
+                                       Rope::RefIndexMode::FIXED);
                 int64_t t1   = ggml_time_ms();
 
                 GGML_ASSERT(!out_opt.empty());

src/model/te/clip.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_TE_CLIP_HPP__
+#ifndef __SD_MODEL_TE_CLIP_HPP__
 #define __SD_MODEL_TE_CLIP_HPP__
 
 #include "core/ggml_extend.hpp"

src/model/te/llm.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_TE_LLM_HPP__
+#ifndef __SD_MODEL_TE_LLM_HPP__
 #define __SD_MODEL_TE_LLM_HPP__
 
 #include <algorithm>
@@ -79,6 +79,7 @@ namespace LLM {
         int window_size                     = 112;
         int num_position_embeddings         = 0;
         std::set<int> fullatt_block_indexes = {7, 15, 23, 31};
+        bool split_patch_embed              = false;
     };
 
     struct LLMConfig {
@@ -179,7 +180,8 @@ namespace LLM {
                 config.num_experts_per_tok     = 4;
             }
 
-            config.num_layers = 0;
+            config.num_layers          = 0;
+            int detected_vision_layers = 0;
             for (const auto& [name, tensor_storage] : tensor_storage_map) {
                 if (!starts_with(name, prefix)) {
                     continue;
@@ -190,6 +192,38 @@ namespace LLM {
                     if (contains(name, "attn.q_proj")) {
                         config.llama_cpp_style = true;
                     }
+                    if (contains(name, "visual.patch_embed.proj.1.weight")) {
+                        config.vision.split_patch_embed = true;
+                    }
+                    if (contains(name, "visual.patch_embed.proj.0.weight")) {
+                        config.vision.patch_size  = static_cast<int>(tensor_storage.ne[0]);
+                        config.vision.in_channels = tensor_storage.ne[2];
+                        config.vision.hidden_size = tensor_storage.ne[3];
+                    }
+                    if (contains(name, "visual.patch_embed.bias")) {
+                        config.vision.hidden_size = tensor_storage.ne[0];
+                    }
+                    if (contains(name, "visual.pos_embed.weight")) {
+                        config.vision.hidden_size             = tensor_storage.ne[0];
+                        config.vision.num_position_embeddings = static_cast<int>(tensor_storage.ne[1]);
+                    }
+                    if (contains(name, "visual.blocks.")) {
+                        auto items = split_string(name.substr(pos), '.');
+                        if (items.size() > 2) {
+                            int block_index = atoi(items[2].c_str());
+                            if (block_index + 1 > detected_vision_layers) {
+                                detected_vision_layers = block_index + 1;
+                            }
+                        }
+                    }
+                    if (contains(name, "visual.blocks.0.mlp.linear_fc1.weight") ||
+                        contains(name, "visual.blocks.0.mlp.gate_proj.weight")) {
+                        config.vision.intermediate_size = tensor_storage.ne[1];
+                    }
+                    if (contains(name, "visual.merger.linear_fc2.weight") ||
+                        contains(name, "visual.merger.mlp.2.weight")) {
+                        config.vision.out_hidden_size = tensor_storage.ne[1];
+                    }
                     continue;
                 }
                 pos = name.find("layers.");
@@ -216,9 +250,12 @@ namespace LLM {
                     config.intermediate_size = tensor_storage.ne[1];
                 }
             }
-            if (arch == LLMArch::QWEN3 && config.num_layers == 28) {
+            if ((arch == LLMArch::QWEN3 || arch == LLMArch::QWEN3_VL) && config.num_layers == 28) {
                 config.num_heads = 16;
             }
+            if (detected_vision_layers > 0) {
+                config.vision.num_layers = detected_vision_layers;
+            }
             LOG_DEBUG("llm: num_layers = %" PRId64 ", vocab_size = %" PRId64 ", hidden_size = %" PRId64 ", intermediate_size = %" PRId64,
                       config.num_layers,
                       config.vocab_size,
@@ -539,40 +576,51 @@ namespace LLM {
 
     struct VisionPatchEmbed : public GGMLBlock {
     protected:
-        bool llama_cpp_style;
+        bool split_patch_embed;
+        bool bias;
         int patch_size;
         int temporal_patch_size;
         int64_t in_channels;
         int64_t embed_dim;
 
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            GGML_UNUSED(prefix);
+            if (split_patch_embed && bias) {
+                params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, embed_dim);
+            }
+        }
+
     public:
-        VisionPatchEmbed(bool llama_cpp_style,
+        VisionPatchEmbed(bool split_patch_embed,
                          LLMVisionArch arch,
                          int patch_size          = 14,
                          int temporal_patch_size = 2,
                          int64_t in_channels     = 3,
                          int64_t embed_dim       = 1152)
-            : llama_cpp_style(llama_cpp_style),
+            : split_patch_embed(split_patch_embed),
+              bias(arch == LLMVisionArch::QWEN3_VL),
               patch_size(patch_size),
               temporal_patch_size(temporal_patch_size),
               in_channels(in_channels),
               embed_dim(embed_dim) {
-            bool bias = arch == LLMVisionArch::QWEN3_VL;
-            if (llama_cpp_style) {
+            if (split_patch_embed) {
                 blocks["proj.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels,
                                                                          embed_dim,
                                                                          {patch_size, patch_size},
                                                                          {patch_size, patch_size},
                                                                          {0, 0},
                                                                          {1, 1},
-                                                                         bias));
+                                                                         false));
                 blocks["proj.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels,
                                                                          embed_dim,
                                                                          {patch_size, patch_size},
                                                                          {patch_size, patch_size},
                                                                          {0, 0},
                                                                          {1, 1},
-                                                                         bias));
+                                                                         false));
             } else {
                 std::tuple<int, int, int> kernel_size = {(int)temporal_patch_size, (int)patch_size, (int)patch_size};
                 blocks["proj"]                        = std::shared_ptr<GGMLBlock>(new Conv3d(in_channels,
@@ -593,7 +641,7 @@ namespace LLM {
                                 temporal_patch_size,
                                 ggml_nelements(x) / (temporal_patch_size * patch_size * patch_size));
 
-            if (llama_cpp_style) {
+            if (split_patch_embed) {
                 auto proj_0 = std::dynamic_pointer_cast<Conv2d>(blocks["proj.0"]);
                 auto proj_1 = std::dynamic_pointer_cast<Conv2d>(blocks["proj.1"]);
 
@@ -606,6 +654,10 @@ namespace LLM {
                 x1      = proj_1->forward(ctx, x1);
 
                 x = ggml_add(ctx->ggml_ctx, x0, x1);
+                if (bias) {
+                    auto b = ggml_reshape_4d(ctx->ggml_ctx, params["bias"], 1, 1, embed_dim, 1);
+                    x      = ggml_add_inplace(ctx->ggml_ctx, x, b);
+                }
             } else {
                 auto proj = std::dynamic_pointer_cast<Conv3d>(blocks["proj"]);
 
@@ -798,7 +850,7 @@ namespace LLM {
               spatial_merge_size(vision_params.spatial_merge_size),
               num_grid_per_side(vision_params.num_position_embeddings > 0 ? static_cast<int>(std::sqrt(vision_params.num_position_embeddings)) : 0),
               fullatt_block_indexes(vision_params.fullatt_block_indexes) {
-            blocks["patch_embed"] = std::shared_ptr<GGMLBlock>(new VisionPatchEmbed(llama_cpp_style,
+            blocks["patch_embed"] = std::shared_ptr<GGMLBlock>(new VisionPatchEmbed(vision_params.split_patch_embed,
                                                                                     arch_,
                                                                                     vision_params.patch_size,
                                                                                     vision_params.temporal_patch_size,

src/model/te/t5.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_TE_T5_HPP__
+#ifndef __SD_MODEL_TE_T5_HPP__
 #define __SD_MODEL_TE_T5_HPP__
 
 #include <cfloat>
@@ -26,13 +26,66 @@ struct T5Config {
     static T5Config detect_from_weights(const String2TensorStorage& tensor_storage_map,
                                         const std::string& prefix,
                                         bool is_umt5 = false) {
-        (void)tensor_storage_map;
-        (void)prefix;
         T5Config config;
         if (is_umt5) {
             config.vocab_size         = 256384;
             config.relative_attention = false;
         }
+        auto find_tensor = [&](const std::string& suffix) -> const TensorStorage* {
+            auto it = tensor_storage_map.find(prefix + "." + suffix);
+            if (it != tensor_storage_map.end()) {
+                return &it->second;
+            }
+            it = tensor_storage_map.find(prefix + suffix);
+            if (it != tensor_storage_map.end()) {
+                return &it->second;
+            }
+            return nullptr;
+        };
+
+        if (const TensorStorage* shared = find_tensor("shared.weight")) {
+            if (shared->n_dims == 2) {
+                config.vocab_size = shared->ne[1];
+                config.model_dim  = shared->ne[0];
+            }
+        }
+        if (const TensorStorage* q = find_tensor("encoder.block.0.layer.0.SelfAttention.q.weight")) {
+            if (q->n_dims == 2) {
+                config.model_dim  = q->ne[0];
+                int64_t inner_dim = q->ne[1];
+                // Flan-T5/T5 uses d_kv=64 for common sizes.
+                if (inner_dim % 64 == 0) {
+                    config.num_heads = inner_dim / 64;
+                }
+            }
+        }
+        if (const TensorStorage* wi = find_tensor("encoder.block.0.layer.1.DenseReluDense.wi_0.weight")) {
+            if (wi->n_dims == 2) {
+                config.model_dim = wi->ne[0];
+                config.ff_dim    = wi->ne[1];
+            }
+        }
+        int64_t detected_layers = 0;
+        for (const auto& [name, _] : tensor_storage_map) {
+            std::string base = prefix;
+            if (!base.empty() && base.back() != '.') {
+                base += ".";
+            }
+            std::string layer_prefix = base + "encoder.block.";
+            if (!starts_with(name, layer_prefix)) {
+                continue;
+            }
+            size_t pos = layer_prefix.size();
+            size_t dot = name.find('.', pos);
+            if (dot == std::string::npos) {
+                continue;
+            }
+            int64_t layer   = atoi(name.substr(pos, dot - pos).c_str());
+            detected_layers = std::max(detected_layers, layer + 1);
+        }
+        if (detected_layers > 0) {
+            config.num_layers = detected_layers;
+        }
         return config;
     }
 };

src/model/upscaler/esrgan.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_UPSCALER_ESRGAN_HPP__
+#ifndef __SD_MODEL_UPSCALER_ESRGAN_HPP__
 #define __SD_MODEL_UPSCALER_ESRGAN_HPP__
 
 #include <algorithm>

src/model/upscaler/ltx_latent_upscaler.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
+#ifndef __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
 #define __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
 
 #include <algorithm>

src/model/vae/auto_encoder_kl.hpp

@@ -682,7 +682,7 @@ struct AutoEncoderKL : public VAE {
         } else if (sd_version_is_sd3(version)) {
             scale_factor = 1.5305f;
             shift_factor = 0.0609f;
-        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_longcat(version)) {
+        } else if (sd_version_uses_flux_vae(version)) {
             scale_factor = 0.3611f;
             shift_factor = 0.1159f;
         } else if (sd_version_uses_flux2_vae(version)) {
@@ -816,12 +816,13 @@ struct AutoEncoderKL : public VAE {
     }
 
     sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        auto latents_ = sd_version_is_sefi_image(version) ? sd::ops::slice(latents, 2, 16, 144) : latents;
         if (sd_version_uses_flux2_vae(version)) {
             int channel_dim                = 2;
-            auto [mean_tensor, std_tensor] = get_latents_mean_std(latents, channel_dim);
-            return (latents * std_tensor) / scale_factor + mean_tensor;
+            auto [mean_tensor, std_tensor] = get_latents_mean_std(latents_, channel_dim);
+            return (latents_ * std_tensor) / scale_factor + mean_tensor;
         }
-        return (latents / scale_factor) + shift_factor;
+        return (latents_ / scale_factor) + shift_factor;
     }
 
     sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {

src/model/vae/ltx_audio_vae.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
+#ifndef __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
 #define __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
 
 #include <cmath>
@@ -214,7 +214,7 @@ namespace LTXV {
 
         auto x = ggml_reshape_3d(ctx, waveform, time, 1, channels * batch);
         if (left_pad > 0) {
-            x = ggml_pad_ext(ctx, x, static_cast<int>(left_pad), 0, 0, 0, 0, 0, 0, 0);
+            x = ggml_ext_pad_ext(ctx, runner_ctx->backend, x, static_cast<int>(left_pad), 0, 0, 0, 0, 0, 0, 0);
         }
 
         auto frames = ggml_conv_1d(ctx, forward_basis, x, hop_length, 0, 1);
@@ -451,6 +451,7 @@ namespace LTXV {
             int pad_h = kernel_size.first - 1;
             int pad_w = kernel_size.second - 1;
             x         = ggml_ext_pad_ext(ctx->ggml_ctx,
+                                         ctx->backend,
                                          x,
                                          pad_w / 2,
                                          pad_w - pad_w / 2,

src/model/vae/tae.hpp

@@ -408,7 +408,7 @@ public:
             h = conv->forward(ctx, h);
             for (int j = 0; j < num_blocks; j++) {
                 auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
-                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                auto mem   = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, h, 0, 0, 0, 0, 0, 0, 1, 0);
                 mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
                 h          = block->forward(ctx, h, mem);
             }
@@ -479,7 +479,7 @@ public:
         int index = 3;
         for (int i = 0; i < num_layers; i++) {
             for (int j = 0; j < num_blocks; j++) {
-                auto mem = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                auto mem = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, h, 0, 0, 0, 0, 0, 0, 1, 0);
                 mem      = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
                 if (is_wide) {
                     auto block = std::dynamic_pointer_cast<WideMemBlock>(blocks[std::to_string(index++)]);
@@ -548,7 +548,7 @@ public:
         }
         auto result = decoder->forward(ctx, z);
         if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
-            // (W, H, C, T) -> (W, H, T, C)
+            // (W, H, T, C) -> (W, H, C, T)
             result = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, result, 0, 1, 3, 2));
         }
         return result;
@@ -556,8 +556,10 @@ public:
 
     ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
         auto encoder = std::dynamic_pointer_cast<TinyVideoEncoder>(blocks["encoder"]);
-        // (W, H, T, C) -> (W, H, C, T)
-        x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
+            // (W, H, T, C) -> (W, H, C, T)
+            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        }
         int64_t num_frames = x->ne[3];
         if (num_frames % encoder->t_downscale) {
             // pad to multiple of encoder->t_downscale at the end
@@ -567,7 +569,10 @@ public:
             }
         }
         x = encoder->forward(ctx, x);
-        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
+            // (W, H, C, T) -> (W, H, T, C)
+            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        }
         return x;
     }
 };

src/model/vae/vae.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_VAE_VAE_HPP__
+#ifndef __SD_MODEL_VAE_VAE_HPP__
 #define __SD_MODEL_VAE_VAE_HPP__
 
 #include "core/tensor_ggml.hpp"
@@ -78,7 +78,7 @@ public:
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
-        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1) {
+        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version)) {
             scale_factor = 1;
         }
         return scale_factor;
@@ -133,7 +133,11 @@ public:
             int64_t H              = input.shape()[1] / scale_factor;
             float tile_overlap;
             int tile_size_x, tile_size_y;
-            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, 1.30539f);
+            // Image VAE encode is more sensitive to tile boundary context than decode.
+            // Keep the smaller legacy factor for video VAEs, but default image encode
+            // tiles to 64 latent pixels so a 512px SD image is encoded as one tile.
+            const float encode_tile_factor = (sd_version_is_wan(version) || sd_version_is_ltxav(version)) ? 1.30539f : 2.0f;
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, encode_tile_factor);
             LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
             output = tiled_compute(input,
                                    n_threads,