fix: avoid issues when sigma_min is close to 0 (#1138)

* global bool

* reworked circular to global flag

* cleaner implementation of tiling support in sd cpp

* cleaned rope

* working simplified but still need wraps

* Further clean of rope

* resolve flux conflict

* switch to pad op circular only

* Set ggml to most recent

* Revert ggml temp

* Update ggml to most recent

* Revert unneded flux change

* move circular flag to the GGMLRunnerContext

* Pass through circular param in all places where conv is called

* fix of constant and minor cleanup

* Added back --circular option

* Conv2d circular in vae and various models

* Fix temporal padding for qwen image and other vaes

* Z Image circular tiling

* x and y axis seamless only

* First attempt at chroma seamless x and y

* refactor into pure x and y, almost there

* Fix crash on chroma

* Refactor into cleaner variable choices

* Removed redundant set_circular_enabled

* Sync ggml

* simplify circular parameter

* format code

* no need to perform circular pad on the clip

* simplify circular_axes setting

* unify function naming

* remove unnecessary member variables

* simplify rope

---------

Co-authored-by: Phylliida <phylliidadev@gmail.com>
Co-authored-by: leejet <leejet714@gmail.com>

.clang-tidy

@@ -0,0 +1,10 @@
+Checks: >
+  modernize-make-shared,
+  modernize-use-nullptr,
+  modernize-use-override,
+  modernize-pass-by-value,
+  modernize-return-braced-init-list,
+  modernize-deprecated-headers,
+HeaderFilterRegex: '^$'
+WarningsAsErrors: ''
+FormatStyle: none

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,73 @@
+name: 🐞 Bug Report
+description: Report a bug or unexpected behavior
+title: "[Bug] "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please use this template and include as many details as possible to help us reproduce and fix the issue.
+  - type: textarea
+    id: commit
+    attributes:
+      label: Git commit
+      description: Which commit are you trying to compile?
+      placeholder: |
+        $git rev-parse HEAD
+        40a6a8710ec15b1b5db6b5a098409f6bc8f654a4
+    validations:
+      required: true
+  - type: input
+    id: os
+    attributes:
+      label: Operating System & Version
+      placeholder: e.g. “Ubuntu 22.04”, “Windows 11 23H2”, “macOS 14.3”
+    validations:
+      required: true
+  - type: dropdown
+    id: backends
+    attributes:
+        label: GGML backends
+        description: Which GGML backends do you know to be affected?
+        options: [CPU, CUDA, HIP, Metal, Musa, SYCL, Vulkan, OpenCL]
+        multiple: true
+    validations:
+      required: true
+  - type: input
+    id: cmd_arguments
+    attributes:
+      label: Command-line arguments used
+      placeholder: The full command line you ran (with all flags)
+    validations:
+      required: true
+  - type: textarea
+    id: steps_to_reproduce
+    attributes:
+      label: Steps to reproduce
+      placeholder: A step-by-step list of what you did
+    validations:
+      required: true
+  - type: textarea
+    id: expected_behavior
+    attributes:
+      label: What you expected to happen
+      placeholder: Describe the expected behavior or result
+    validations:
+      required: true
+  - type: textarea
+    id: actual_behavior
+    attributes:
+      label: What actually happened
+      placeholder: Describe what you saw instead (errors, logs, crash, etc.)
+    validations:
+      required: true
+  - type: textarea
+    id: logs_and_errors
+    attributes:
+      label: Logs / error messages / stack trace
+      placeholder: Paste complete logs or error output
+  - type: textarea
+    id: additional_info
+    attributes:
+      label: Additional context / environment details
+      placeholder: e.g. CPU model, GPU, RAM, model file versions, quantization type, etc.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,33 @@
+name: 💡 Feature Request
+description: Suggest a new feature or improvement
+title: "[Feature] "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thank you for suggesting an improvement! Please fill in the fields below.
+  - type: input
+    id: summary
+    attributes:
+      label: Feature Summary
+      placeholder: A one-line summary of the feature you’d like
+    validations:
+      required: true
+  - type: textarea
+    id: description
+    attributes:
+      label: Detailed Description
+      placeholder: What problem does this solve? How do you expect it to work?
+    validations:
+      required: true
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives you considered
+      placeholder: Any alternative designs or workarounds you tried
+  - type: textarea
+    id: additional_context
+    attributes:
+      label: Additional context
+      placeholder: Any extra information (use cases, related functionalities, constraints)

.github/workflows/build.yml

@@ -4,17 +4,36 @@ on:
   workflow_dispatch: # allows manual triggering
     inputs:
       create_release:
-        description: 'Create new release'
+        description: "Create new release"
         required: true
         type: boolean
   push:
     branches:
       - master
       - ci
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths:
+      [
+        ".github/workflows/**",
+        "**/CMakeLists.txt",
+        "**/Makefile",
+        "**/*.h",
+        "**/*.hpp",
+        "**/*.c",
+        "**/*.cpp",
+        "**/*.cu",
+      ]
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths:
+      [
+        "**/CMakeLists.txt",
+        "**/Makefile",
+        "**/*.h",
+        "**/*.hpp",
+        "**/*.c",
+        "**/*.cpp",
+        "**/*.cu",
+      ]
 
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -46,7 +65,7 @@ jobs:
 
       - name: Get commit hash
         id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: pr-mpt/actions-commit-hash@v2
 
       - name: Fetch system info
@@ -67,12 +86,12 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
           path: |
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
 
-
   macOS-latest-cmake:
     runs-on: macos-latest
 
@@ -99,7 +118,7 @@ jobs:
 
       - name: Get commit hash
         id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: pr-mpt/actions-commit-hash@v2
 
       - name: Fetch system info
@@ -120,30 +139,33 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
           path: |
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
 
-
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025
+
+    env:
+      VULKAN_VERSION: 1.4.328.1
 
     strategy:
       matrix:
         include:
-          - build: 'noavx'
-            defines: '-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON'
-          - build: 'avx2'
-            defines: '-DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON'
-          - build: 'avx'
-            defines: '-DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON'
-          - build: 'avx512'
-            defines: '-DGGML_AVX512=ON -DSD_BUILD_SHARED_LIBS=ON'
-          - build: 'cuda12'
-            defines: '-DSD_CUBLAS=ON -DSD_BUILD_SHARED_LIBS=ON'
-          - build: 'rocm5.5'
-            defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
+          - 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: "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\"'"
+          - build: 'vulkan'
+            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
     steps:
       - name: Clone
         id: checkout
@@ -154,33 +176,32 @@ jobs:
       - name: Install cuda-toolkit
         id: cuda-toolkit
         if: ${{ matrix.build == 'cuda12' }}
-        uses: Jimver/cuda-toolkit@v0.2.11
+        uses: Jimver/cuda-toolkit@v0.2.22
         with:
-          cuda: '12.2.0'
-          method: 'network'
+          cuda: "12.8.1"
+          method: "network"
           sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
 
-      - name: Install rocm-toolkit
-        id: rocm-toolkit
-        if: ${{ matrix.build == 'rocm5.5' }}
-        uses: Cyberhan123/rocm-toolkit@v0.1.0
-        with:
-          rocm: '5.5.0'
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'vulkan' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
 
-      - name: Install Ninja
-        id: install-ninja
-        if: ${{ matrix.build == 'rocm5.5' }}
-        uses: urkle/action-get-ninja@v1
-        with:
-          version: 1.11.1
+      - name: Activate MSVC environment
+        id: msvc_dev_cmd
+        uses: ilammy/msvc-dev-cmd@v1
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release
+          cmake .. -DCMAKE_CXX_FLAGS='/bigobj' -G Ninja -DCMAKE_C_COMPILER=cl.exe -DCMAKE_CXX_COMPILER=cl.exe ${{ matrix.defines }}
+          cmake --build .
 
       - name: Check AVX512F support
         id: check_avx512f
@@ -222,7 +243,7 @@ jobs:
 
       - name: Copy and pack Cuda runtime
         id: pack_cuda_runtime
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' && matrix.build == 'cuda12' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
         run: |
           echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
           $dst='.\build\bin\cudart\'
@@ -230,19 +251,119 @@ jobs:
           7z a cudart-sd-bin-win-cu12-x64.zip $dst\*
 
       - name: Upload Cuda runtime
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' && matrix.build == 'cuda12' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
+        uses: actions/upload-artifact@v4
         with:
+          name: sd-cudart-sd-bin-win-cu12-x64.zip
           path: |
             cudart-sd-bin-win-cu12-x64.zip
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
           path: |
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
 
+  windows-latest-cmake-hip:
+    runs-on: windows-2022
+
+    env:
+      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      GPU_TARGETS: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+
+    steps:
+      - uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-x64
+          evict-old-files: 1d
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        run: |
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
+          # Set HIP_PATH environment variable for later steps
+          echo "HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)" >> $env:GITHUB_ENV
+
+      - name: Build
+        run: |
+          mkdir build
+          cd build
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake .. `
+            -G "Unix Makefiles" `
+            -DSD_HIPBLAS=ON `
+            -DSD_BUILD_SHARED_LIBS=ON `
+            -DGGML_NATIVE=OFF `
+            -DCMAKE_C_COMPILER=clang `
+            -DCMAKE_CXX_COMPILER=clang++ `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGPU_TARGETS="${{ env.GPU_TARGETS }}"
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          md "build\bin\rocblas\library\"
+          md "build\bin\hipblaslt\library"
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -252,11 +373,26 @@ jobs:
       - ubuntu-latest-cmake
       - macOS-latest-cmake
       - windows-latest-cmake
+      - windows-latest-cmake-hip
 
     steps:
+      - name: Clone
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+          pattern: sd-*
+          merge-multiple: true
+
+      - name: Get commit count
+        id: commit_count
+        run: |
+          echo "count=$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
 
       - name: Get commit hash
         id: commit
@@ -264,14 +400,16 @@ jobs:
 
       - name: Create release
         id: create_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
         uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
-          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+          tag_name: ${{ format('{0}-{1}-{2}', env.BRANCH_NAME, steps.commit_count.outputs.count, steps.commit.outputs.short) }}
 
       - name: Upload release
         id: upload_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
         uses: actions/github-script@v3
         with:
           github-token: ${{secrets.GITHUB_TOKEN}}

.gitignore

@@ -1,13 +1,15 @@
 build*/
+cmake-build-*/
 test/
 .vscode/
+.idea/
 .cache/
 *.swp
-.vscode/
 *.bat
 *.bin
 *.exe
 *.gguf
 output*.png
 models*
-*.log
+*.log
+preview.png

.gitmodules

@@ -1,3 +1,3 @@
 [submodule "ggml"]
     path = ggml
-	url = https://github.com/ggerganov/ggml.git
+	url = https://github.com/ggml-org/ggml.git

CMakeLists.txt

@@ -24,79 +24,168 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
-option(SD_CUBLAS                     "sd: cuda backend" OFF)
+option(SD_CUDA                       "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
-option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
+option(SD_VULKAN                     "sd: vulkan backend" OFF)
+option(SD_OPENCL                     "sd: opencl backend" OFF)
+option(SD_SYCL                       "sd: sycl backend" OFF)
+option(SD_MUSA                       "sd: musa backend" OFF)
 option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
+option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
+option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
-if(SD_CUBLAS)
-	message("Use CUBLAS as backend stable-diffusion")
-    set(GGML_CUBLAS ON)
-    add_definitions(-DSD_USE_CUBLAS)
+if(SD_CUDA)
+    message("-- Use CUDA as backend stable-diffusion")
+    set(GGML_CUDA ON)
+    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_METAL)
-	message("Use Metal as backend stable-diffusion")
+    message("-- Use Metal as backend stable-diffusion")
     set(GGML_METAL ON)
     add_definitions(-DSD_USE_METAL)
 endif()
 
+if (SD_VULKAN)
+    message("-- Use Vulkan as backend stable-diffusion")
+    set(GGML_VULKAN ON)
+    add_definitions(-DSD_USE_VULKAN)
+endif ()
+
+if (SD_OPENCL)
+    message("-- Use OpenCL as backend stable-diffusion")
+    set(GGML_OPENCL ON)
+    add_definitions(-DSD_USE_OPENCL)
+endif ()
+
 if (SD_HIPBLAS)
-    message("Use HIPBLAS as backend stable-diffusion")
-    set(GGML_HIPBLAS ON)
-    add_definitions(-DSD_USE_CUBLAS)
+    message("-- Use HIPBLAS as backend stable-diffusion")
+    set(GGML_HIP ON)
+    add_definitions(-DSD_USE_CUDA)
     if(SD_FAST_SOFTMAX)
         set(GGML_CUDA_FAST_SOFTMAX ON)
     endif()
 endif ()
 
-if(SD_FLASH_ATTN)
-    message("Use Flash Attention for memory optimization")
-    add_definitions(-DSD_USE_FLASH_ATTENTION)
+if(SD_MUSA)
+    message("-- Use MUSA as backend stable-diffusion")
+    set(GGML_MUSA ON)
+    add_definitions(-DSD_USE_CUDA)
+    if(SD_FAST_SOFTMAX)
+        set(GGML_CUDA_FAST_SOFTMAX ON)
+    endif()
 endif()
 
 set(SD_LIB stable-diffusion)
 
-file(GLOB SD_LIB_SOURCES 
+file(GLOB SD_LIB_SOURCES
     "*.h"
     "*.cpp"
     "*.hpp"
 )
 
-# we can get only one share lib
+find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
+if(GIT_EXE)
+    execute_process(COMMAND ${GIT_EXE} describe --tags --abbrev=7 --dirty=+
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE SDCPP_BUILD_VERSION
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE SDCPP_BUILD_COMMIT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+endif()
+
+if(NOT SDCPP_BUILD_VERSION)
+    set(SDCPP_BUILD_VERSION unknown)
+endif()
+message(STATUS "stable-diffusion.cpp version ${SDCPP_BUILD_VERSION}")
+
+if(NOT SDCPP_BUILD_COMMIT)
+    set(SDCPP_BUILD_COMMIT unknown)
+endif()
+message(STATUS "stable-diffusion.cpp commit ${SDCPP_BUILD_COMMIT}")
+
+set_property(
+  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp
+  APPEND PROPERTY COMPILE_DEFINITIONS
+  SDCPP_BUILD_COMMIT=${SDCPP_BUILD_COMMIT} SDCPP_BUILD_VERSION=${SDCPP_BUILD_VERSION}
+)
+
 if(SD_BUILD_SHARED_LIBS)
-    message("Build shared library")
-    set(BUILD_SHARED_LIBS OFF)
+    message("-- Build shared library")
     message(${SD_LIB_SOURCES})
+    if(NOT SD_BUILD_SHARED_GGML_LIB)
+        set(BUILD_SHARED_LIBS OFF)
+    endif()
     add_library(${SD_LIB} SHARED ${SD_LIB_SOURCES})
     add_definitions(-DSD_BUILD_SHARED_LIB)
     target_compile_definitions(${SD_LIB} PRIVATE -DSD_BUILD_DLL)
     set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 else()
-    message("Build static library")
+    message("-- Build static library")
+    if(NOT SD_BUILD_SHARED_GGML_LIB)
+        set(BUILD_SHARED_LIBS OFF)
+    endif()
     add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()
 
+if(SD_SYCL)
+    message("-- Use SYCL as backend stable-diffusion")
+    set(GGML_SYCL ON)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+    add_definitions(-DSD_USE_SYCL)
+    # disable fast-math on host, see:
+    # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
+    if (WIN32)
+        set(SYCL_COMPILE_OPTIONS /fp:precise)
+    else()
+        set(SYCL_COMPILE_OPTIONS -fp-model=precise)
+    endif()
+    message("-- Turn off fast-math for host in SYCL backend")
+    target_compile_options(${SD_LIB} PRIVATE ${SYCL_COMPILE_OPTIONS})
+endif()
 
 set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
-# see https://github.com/ggerganov/ggml/pull/682
-add_definitions(-DGGML_MAX_NAME=128)
+if (NOT SD_USE_SYSTEM_GGML)
+    # see https://github.com/ggerganov/ggml/pull/682
+    add_definitions(-DGGML_MAX_NAME=128)
+endif()
 
 # deps
-add_subdirectory(ggml)
+# Only add ggml if it hasn't been added yet
+if (NOT TARGET ggml)
+    if (SD_USE_SYSTEM_GGML)
+        find_package(ggml REQUIRED)
+        if (NOT ggml_FOUND)
+            message(FATAL_ERROR "System-installed GGML library not found.")
+        endif()
+        add_library(ggml ALIAS ggml::ggml)
+    else()
+        add_subdirectory(ggml)
+    endif()
+endif()
 
 add_subdirectory(thirdparty)
 
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
-target_compile_features(${SD_LIB} PUBLIC cxx_std_11)
+target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)
 
 
 if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
+set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)

Dockerfile

@@ -1,17 +1,22 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
-RUN apt-get update && apt-get install -y build-essential git cmake
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake
 
 WORKDIR /sd.cpp
 
 COPY . .
 
-RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+RUN cmake . -B ./build
+RUN cmake --build ./build --config Release --parallel
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM ubuntu:$UBUNTU_VERSION AS runtime
 
-COPY --from=build /sd.cpp/build/bin/sd /sd
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 && \
+    apt-get clean
 
-ENTRYPOINT [ "/sd" ]
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
+
+ENTRYPOINT [ "/sd-cli" ]

Dockerfile.musa

@@ -0,0 +1,23 @@
+ARG MUSA_VERSION=rc4.2.0
+ARG UBUNTU_VERSION=22.04
+
+FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build
+
+RUN apt-get update && apt-get install -y ccache cmake git
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+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 && \
+    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
+
+ENTRYPOINT [ "/sd-cli" ]

Dockerfile.sycl

@@ -0,0 +1,19 @@
+ARG SYCL_VERSION=2025.1.0-0
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
+
+RUN apt-get update && apt-get install -y cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release -j$(nproc)
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
+
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
+
+ENTRYPOINT [ "/sd-cli" ]

README.md

@@ -1,37 +1,86 @@
 <p align="center">
-  <img src="./assets/a%20lovely%20cat.png" width="256x">
+  <img src="./assets/logo.png" width="360x">
 </p>
 
 # stable-diffusion.cpp
 
-Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in pure C/C++
+<div align="center">
+<a href="https://trendshift.io/repositories/9714" target="_blank"><img src="https://trendshift.io/api/badge/repositories/9714" alt="leejet%2Fstable-diffusion.cpp | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</div>
+
+Diffusion model(SD,Flux,Wan,...) inference in pure C/C++
+
+***Note that this project is under active development. \
+API and command-line option may change frequently.***
+
+## 🔥Important News
+
+* **2025/12/01** 🚀 stable-diffusion.cpp now supports **Z-Image**  
+  👉 Details: [PR #1020](https://github.com/leejet/stable-diffusion.cpp/pull/1020)
+
+* **2025/11/30** 🚀 stable-diffusion.cpp now supports **FLUX.2-dev**  
+  👉 Details: [PR #1016](https://github.com/leejet/stable-diffusion.cpp/pull/1016)
+
+* **2025/10/13** 🚀 stable-diffusion.cpp now supports **Qwen-Image-Edit / Qwen-Image-Edit 2509**  
+  👉 Details: [PR #877](https://github.com/leejet/stable-diffusion.cpp/pull/877)
+
+* **2025/10/12** 🚀 stable-diffusion.cpp now supports **Qwen-Image**  
+  👉 Details: [PR #851](https://github.com/leejet/stable-diffusion.cpp/pull/851)
+
+* **2025/09/14** 🚀 stable-diffusion.cpp now supports **Wan2.1 Vace**  
+  👉 Details: [PR #819](https://github.com/leejet/stable-diffusion.cpp/pull/819)
+
+* **2025/09/06** 🚀 stable-diffusion.cpp now supports **Wan2.1 / Wan2.2**  
+  👉 Details: [PR #778](https://github.com/leejet/stable-diffusion.cpp/pull/778)
 
 ## Features
 
-- Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
+- Plain C/C++ implementation based on [ggml](https://github.com/ggml-org/ggml), working in the same way as [llama.cpp](https://github.com/ggml-org/llama.cpp)
 - Super lightweight and without external dependencies
-- SD1.x, SD2.x and SDXL support
-    - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
-
-- [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo) and [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo) support
-- 16-bit, 32-bit float support
-- 4-bit, 5-bit and 8-bit integer quantization support
-- Accelerated memory-efficient CPU inference
-    - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image, enabling Flash Attention just requires ~1.8GB.
-- AVX, AVX2 and AVX512 support for x86 architectures
-- Full CUDA and Metal backend for GPU acceleration.
-- Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
-    - No need to convert to `.ggml` or `.gguf` anymore!
-- Flash Attention for memory usage optimization (only cpu for now)
-- Original `txt2img` and `img2img` mode
+- 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)
+    - [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)
+    - [FLUX.2-dev](./docs/flux2.md)
+    - [Chroma](./docs/chroma.md)
+    - [Chroma1-Radiance](./docs/chroma_radiance.md)
+    - [Qwen Image](./docs/qwen_image.md)
+    - [Z-Image](./docs/z_image.md)
+    - [Ovis-Image](./docs/ovis_image.md)
+  - Image Edit Models
+    - [FLUX.1-Kontext-dev](./docs/kontext.md)
+    - [Qwen Image Edit series](./docs/qwen_image_edit.md)
+  - Video Models
+    - [Wan2.1/Wan2.2](./docs/wan.md)
+  - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) 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)
+- Supported backends
+  - CPU (AVX, AVX2 and AVX512 support for x86 architectures)
+  - CUDA
+  - Vulkan
+  - Metal
+  - OpenCL
+  - SYCL
+- Supported weight formats
+  - Pytorch checkpoint (`.ckpt` or `.pth`)
+  - Safetensors (`./safetensors`)
+  - GGUF (`.gguf`)
+- Supported platforms
+    - Linux
+    - Mac OS
+    - Windows
+    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
+- Flash Attention for memory usage optimization
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
-- 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)
 - VAE tiling processing for reduce memory usage
-- Control Net support with SD 1.5
 - Sampling method
     - `Euler A`
     - `Euler`
@@ -41,283 +90,84 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
     - [`DPM++ 2M v2`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457)
     - `DPM++ 2S a`
     - [`LCM`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13952)
-- Cross-platform reproducibility (`--rng cuda`, consistent with the `stable-diffusion-webui GPU RNG`)
+- Cross-platform reproducibility
+    - `--rng cuda`, default, consistent with the `stable-diffusion-webui GPU RNG`
+    - `--rng cpu`, consistent with the `comfyui RNG`
 - Embedds generation parameters into png output as webui-compatible text string
-- Supported platforms
-    - Linux
-    - Mac OS
-    - Windows
-    - Android (via Termux)
 
-### TODO
+## Quick Start
 
-- [ ] More sampling methods
-- [ ] Make inference faster
-    - The current implementation of ggml_conv_2d is slow and has high memory usage
-- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
-- [ ] Implement Inpainting support
-- [ ] k-quants support
+### Get the sd executable
 
-## Usage
+- Download pre-built binaries from the [releases page](https://github.com/leejet/stable-diffusion.cpp/releases)
+- Or build from source by following the [build guide](./docs/build.md)
 
-### Get the Code
+### Download model weights
 
-```
-git clone --recursive https://github.com/leejet/stable-diffusion.cpp
-cd stable-diffusion.cpp
-```
+- download weights(.ckpt or .safetensors or .gguf). For example
+    - Stable Diffusion v1.5 from https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 
 
-- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
-
-```
-cd stable-diffusion.cpp
-git pull origin master
-git submodule init
-git submodule update
-```
-
-### Download weights
-
-- download original weights(.ckpt or .safetensors). For example
-    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
-    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
-    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
-
-    ```shell
-    curl -L -O https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
-    # curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
-    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-nonema-pruned.safetensors
+    ```sh
+    curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
     ```
 
-### Build
-
-#### Build from scratch
-
-```shell
-mkdir build
-cd build
-cmake ..
-cmake --build . --config Release
-```
-
-##### Using OpenBLAS
-
-```
-cmake .. -DGGML_OPENBLAS=ON
-cmake --build . --config Release
-```
-
-##### Using CUBLAS
-
-This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
-
-```
-cmake .. -DSD_CUBLAS=ON
-cmake --build . --config Release
-```
-
-##### Using HipBLAS
-This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure to have the ROCm toolkit installed.
-
-Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
-
-```
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
-cmake --build . --config Release
-```
-
-
-##### Using Metal
-
-Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
-
-```
-cmake .. -DSD_METAL=ON
-cmake --build . --config Release
-```
-
-##### Using Flash Attention
-
-Enabling flash attention reduces memory usage by at least 400 MB. At the moment, it is not supported when CUBLAS is enabled because the kernel implementation is missing.
-
-```
-cmake .. -DSD_FLASH_ATTN=ON
-cmake --build . --config Release
-```
-
-### Run
-
-```
-usage: ./build/bin/sd [arguments]
-
-arguments:
-  -h, --help                         show this help message and exit
-  -M, --mode [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)
-  -t, --threads N                    number of threads to use during computation (default: -1).
-                                     If threads <= 0, then threads will be set to the number of CPU physical cores
-  -m, --model [MODEL]                path to model
-  --vae [VAE]                        path to vae
-  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
-  --control-net [CONTROL_PATH]       path to control net model
-  --embd-dir [EMBEDDING_PATH]        path to embeddings.
-  --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now.
-  --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)
-  --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0)
-                                     If not specified, the default is the type of the weight file.
-  --lora-model-dir [DIR]             lora model directory
-  -i, --init-img [IMAGE]             path to the input image, required by img2img
-  --control-image [IMAGE]            path to image condition, control net
-  -o, --output OUTPUT                path to write result image to (default: ./output.png)
-  -p, --prompt [PROMPT]              the prompt to render
-  -n, --negative-prompt PROMPT       the negative prompt (default: "")
-  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
-  --strength STRENGTH                strength for noising/unnoising (default: 0.75)
-  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
-                                     1.0 corresponds to full destruction of information in init image
-  -H, --height H                     image height, in pixel space (default: 512)
-  -W, --width W                      image width, in pixel space (default: 512)
-  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, lcm}
-                                     sampling method (default: "euler_a")
-  --steps  STEPS                     number of sample steps (default: 20)
-  --rng {std_default, cuda}          RNG (default: cuda)
-  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
-  -b, --batch-count COUNT            number of images to generate.
-  --schedule {discrete, karras}      Denoiser sigma schedule (default: discrete)
-  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
-                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
-  --vae-tiling                       process vae in tiles to reduce memory usage
-  --control-net-cpu                  keep controlnet in cpu (for low vram)
-  --canny                            apply canny preprocessor (edge detection)
-  -v, --verbose                      print extra info
-```
-
-#### Quantization
-
-You can specify the model weight type using the `--type` parameter. The weights are automatically converted when loading the model.
-
-- `f16` for 16-bit floating-point
-- `f32` for 32-bit floating-point
-- `q8_0` for 8-bit integer quantization
-- `q5_0` or `q5_1` for 5-bit integer quantization
-- `q4_0` or `q4_1` for 4-bit integer quantization
-
-#### Convert to GGUF
-
-You can also convert weights in the formats `ckpt/safetensors/diffusers` to gguf and perform quantization in advance, avoiding the need for quantization every time you load them.
-
-For example:
+### Generate an image with just one command
 
 ```sh
-./bin/sd -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
 ```
 
-#### txt2img example
+***For detailed command-line arguments, check out [cli doc](./examples/cli/README.md).***
 
-```sh
-./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
-# ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
-# ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
-```
+## Performance
 
-Using formats of different precisions will yield results of varying quality.
+If you want to improve performance or reduce VRAM/RAM usage, please refer to [performance guide](./docs/performance.md).
 
-| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
-| ----  |----  |----  |----  |----  |----  |----  |
-| ![](./assets/f32.png) |![](./assets/f16.png) |![](./assets/q8_0.png) |![](./assets/q5_0.png) |![](./assets/q5_1.png) |![](./assets/q4_0.png) |![](./assets/q4_1.png) |
+## More Guides
 
-#### img2img example
+- [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](./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)
+- [🔥Z-Image](./docs/z_image.md)
+- [Ovis-Image](./docs/ovis_image.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)
 
-- `./output.png` is the image generated from the above txt2img pipeline
+## Bindings
 
+These projects wrap `stable-diffusion.cpp` for easier use in other languages/frameworks.
 
-```
-./bin/sd --mode img2img -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
-```
+* Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
+* Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
+* C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
+* Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
+* Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)
+* Flutter/Dart: [rmatif/Local-Diffusion](https://github.com/rmatif/Local-Diffusion)
 
-<p align="center">
-  <img src="./assets/img2img_output.png" width="256x">
-</p>
+## UIs
 
-#### with LoRA
+These projects use `stable-diffusion.cpp` as a backend for their image generation.
 
-- You can specify the directory where the lora weights are stored via `--lora-model-dir`. If not specified, the default is the current working directory.
-
-- LoRA is specified via prompt, just like [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora).
-
-Here's a simple example:
-
-```
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
-```
-
-`../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
-
-#### LCM/LCM-LoRA
-
-- Download LCM-LoRA form https://huggingface.co/latent-consistency/lcm-lora-sdv1-5
-- Specify LCM-LoRA by adding `<lora:lcm-lora-sdv1-5:1>` to prompt
-- It's advisable to set `--cfg-scale` to `1.0` instead of the default `7.0`. For `--steps`, a range of `2-8` steps is recommended. For `--sampling-method`, `lcm`/`euler_a` is recommended.
-
-Here's a simple example:
-
-```
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
-```
-
-| without LCM-LoRA (--cfg-scale 7)  | with LCM-LoRA (--cfg-scale 1)  |
-| ----  |----    |
-| ![](./assets/without_lcm.png) |![](./assets/with_lcm.png)  |
-
-#### Using TAESD to faster decoding
-
-You can use TAESD to accelerate the decoding of latent images by following these steps:
-
-- Download the model [weights](https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors).
-
-Or curl
-
-```bash
-curl -L -O https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors
-```
-
-- Specify the model path using the `--taesd PATH` parameter. example:
-
-```bash
-sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
-```
-
-#### Using ESRGAN to upscale results
-
-You can use ESRGAN to upscale the generated images. At the moment, only the [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth) model is supported. Support for more models of this architecture will be added soon.
-
-- Specify the model path using the `--upscale-model PATH` parameter. example:
-
-```bash
-sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
-```
-
-### Docker
-
-#### Building using Docker
-
-```shell
-docker build -t sd .
-```
-
-#### Run
-
-```shell
-docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
-# For example
-# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
-```
-
-## Memory Requirements
-
-| precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
-| ----         | ----  |----  |----  |----  |----  |----  |----  |
-|  **Memory** (txt2img - 512 x 512) | ~2.8G | ~2.3G | ~2.1G | ~2.0G | ~2.0G | ~2.0G | ~2.0G |
-|  **Memory** (txt2img - 512 x 512) *with Flash Attention* | ~2.4G | ~1.9G | ~1.6G | ~1.5G | ~1.5G | ~1.5G | ~1.5G |
+- [Jellybox](https://jellybox.com)
+- [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
+- [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
+- [Local Diffusion](https://github.com/rmatif/Local-Diffusion)
+- [sd.cpp-webui](https://github.com/daniandtheweb/sd.cpp-webui)
+- [LocalAI](https://github.com/mudler/LocalAI)
+- [Neural-Pixel](https://github.com/Luiz-Alcantara/Neural-Pixel)
+- [KoboldCpp](https://github.com/LostRuins/koboldcpp)
 
 ## Contributors
 
@@ -325,13 +175,22 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 
 [![Contributors](https://contrib.rocks/image?repo=leejet/stable-diffusion.cpp)](https://github.com/leejet/stable-diffusion.cpp/graphs/contributors)
 
+## 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/ggerganov/ggml)
+- [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)

cache_dit.hpp

@@ -0,0 +1,975 @@
+#ifndef __CACHE_DIT_HPP__
+#define __CACHE_DIT_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <limits>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "ggml_extend.hpp"
+
+struct DBCacheConfig {
+    bool enabled                        = false;
+    int Fn_compute_blocks               = 8;
+    int Bn_compute_blocks               = 0;
+    float residual_diff_threshold       = 0.08f;
+    int max_warmup_steps                = 8;
+    int max_cached_steps                = -1;
+    int max_continuous_cached_steps     = -1;
+    float max_accumulated_residual_diff = -1.0f;
+    std::vector<int> steps_computation_mask;
+    bool scm_policy_dynamic = true;
+};
+
+struct TaylorSeerConfig {
+    bool enabled            = false;
+    int n_derivatives       = 1;
+    int max_warmup_steps    = 2;
+    int skip_interval_steps = 1;
+};
+
+struct CacheDitConfig {
+    DBCacheConfig dbcache;
+    TaylorSeerConfig taylorseer;
+    int double_Fn_blocks = -1;
+    int double_Bn_blocks = -1;
+    int single_Fn_blocks = -1;
+    int single_Bn_blocks = -1;
+};
+
+struct TaylorSeerState {
+    int n_derivatives      = 1;
+    int current_step       = -1;
+    int last_computed_step = -1;
+    std::vector<std::vector<float>> dY_prev;
+    std::vector<std::vector<float>> dY_current;
+
+    void init(int n_deriv, size_t hidden_size) {
+        n_derivatives = n_deriv;
+        int order     = n_derivatives + 1;
+        dY_prev.resize(order);
+        dY_current.resize(order);
+        for (int i = 0; i < order; i++) {
+            dY_prev[i].clear();
+            dY_current[i].clear();
+        }
+        current_step       = -1;
+        last_computed_step = -1;
+    }
+
+    void reset() {
+        for (auto& v : dY_prev)
+            v.clear();
+        for (auto& v : dY_current)
+            v.clear();
+        current_step       = -1;
+        last_computed_step = -1;
+    }
+
+    bool can_approximate() const {
+        return last_computed_step >= n_derivatives && !dY_prev.empty() && !dY_prev[0].empty();
+    }
+
+    void update_derivatives(const float* Y, size_t size, int step) {
+        int order = n_derivatives + 1;
+        dY_prev   = dY_current;
+        dY_current[0].resize(size);
+        for (size_t i = 0; i < size; i++) {
+            dY_current[0][i] = Y[i];
+        }
+
+        int window = step - last_computed_step;
+        if (window <= 0)
+            window = 1;
+
+        for (int d = 0; d < n_derivatives; d++) {
+            if (!dY_prev[d].empty() && dY_prev[d].size() == size) {
+                dY_current[d + 1].resize(size);
+                for (size_t i = 0; i < size; i++) {
+                    dY_current[d + 1][i] = (dY_current[d][i] - dY_prev[d][i]) / static_cast<float>(window);
+                }
+            } else {
+                dY_current[d + 1].clear();
+            }
+        }
+
+        current_step       = step;
+        last_computed_step = step;
+    }
+
+    void approximate(float* output, size_t size, int target_step) const {
+        if (!can_approximate() || dY_prev[0].size() != size) {
+            return;
+        }
+
+        int elapsed = target_step - last_computed_step;
+        if (elapsed <= 0)
+            elapsed = 1;
+
+        std::fill(output, output + size, 0.0f);
+        float factorial = 1.0f;
+        int order       = static_cast<int>(dY_prev.size());
+
+        for (int o = 0; o < order; o++) {
+            if (dY_prev[o].empty() || dY_prev[o].size() != size)
+                continue;
+            if (o > 0)
+                factorial *= static_cast<float>(o);
+            float coeff = std::pow(static_cast<float>(elapsed), o) / factorial;
+            for (size_t i = 0; i < size; i++) {
+                output[i] += coeff * dY_prev[o][i];
+            }
+        }
+    }
+};
+
+struct BlockCacheEntry {
+    std::vector<float> residual_img;
+    std::vector<float> residual_txt;
+    std::vector<float> residual;
+    std::vector<float> prev_img;
+    std::vector<float> prev_txt;
+    std::vector<float> prev_output;
+    bool has_prev = false;
+};
+
+struct CacheDitState {
+    CacheDitConfig config;
+    bool initialized = false;
+
+    int total_double_blocks = 0;
+    int total_single_blocks = 0;
+    size_t hidden_size      = 0;
+
+    int current_step     = -1;
+    int total_steps      = 0;
+    int warmup_remaining = 0;
+    std::vector<int> cached_steps;
+    int continuous_cached_steps     = 0;
+    float accumulated_residual_diff = 0.0f;
+
+    std::vector<BlockCacheEntry> double_block_cache;
+    std::vector<BlockCacheEntry> single_block_cache;
+
+    std::vector<float> Fn_residual_img;
+    std::vector<float> Fn_residual_txt;
+    std::vector<float> prev_Fn_residual_img;
+    std::vector<float> prev_Fn_residual_txt;
+    bool has_prev_Fn_residual = false;
+
+    std::vector<float> Bn_buffer_img;
+    std::vector<float> Bn_buffer_txt;
+    std::vector<float> Bn_buffer;
+    bool has_Bn_buffer = false;
+
+    TaylorSeerState taylor_state;
+
+    bool can_cache_this_step  = false;
+    bool is_caching_this_step = false;
+
+    int total_blocks_computed = 0;
+    int total_blocks_cached   = 0;
+
+    void init(const CacheDitConfig& cfg, int num_double_blocks, int num_single_blocks, size_t h_size) {
+        config              = cfg;
+        total_double_blocks = num_double_blocks;
+        total_single_blocks = num_single_blocks;
+        hidden_size         = h_size;
+
+        initialized = cfg.dbcache.enabled || cfg.taylorseer.enabled;
+
+        if (!initialized)
+            return;
+
+        warmup_remaining = cfg.dbcache.max_warmup_steps;
+        double_block_cache.resize(total_double_blocks);
+        single_block_cache.resize(total_single_blocks);
+
+        if (cfg.taylorseer.enabled) {
+            taylor_state.init(cfg.taylorseer.n_derivatives, h_size);
+        }
+
+        reset_runtime();
+    }
+
+    void reset_runtime() {
+        current_step     = -1;
+        total_steps      = 0;
+        warmup_remaining = config.dbcache.max_warmup_steps;
+        cached_steps.clear();
+        continuous_cached_steps   = 0;
+        accumulated_residual_diff = 0.0f;
+
+        for (auto& entry : double_block_cache) {
+            entry.residual_img.clear();
+            entry.residual_txt.clear();
+            entry.prev_img.clear();
+            entry.prev_txt.clear();
+            entry.has_prev = false;
+        }
+
+        for (auto& entry : single_block_cache) {
+            entry.residual.clear();
+            entry.prev_output.clear();
+            entry.has_prev = false;
+        }
+
+        Fn_residual_img.clear();
+        Fn_residual_txt.clear();
+        prev_Fn_residual_img.clear();
+        prev_Fn_residual_txt.clear();
+        has_prev_Fn_residual = false;
+
+        Bn_buffer_img.clear();
+        Bn_buffer_txt.clear();
+        Bn_buffer.clear();
+        has_Bn_buffer = false;
+
+        taylor_state.reset();
+
+        can_cache_this_step  = false;
+        is_caching_this_step = false;
+
+        total_blocks_computed = 0;
+        total_blocks_cached   = 0;
+    }
+
+    bool enabled() const {
+        return initialized && (config.dbcache.enabled || config.taylorseer.enabled);
+    }
+
+    void begin_step(int step_index, float sigma = 0.0f) {
+        if (!enabled())
+            return;
+        if (step_index == current_step)
+            return;
+
+        current_step = step_index;
+        total_steps++;
+
+        bool in_warmup = warmup_remaining > 0;
+        if (in_warmup) {
+            warmup_remaining--;
+        }
+
+        bool scm_allows_cache = true;
+        if (!config.dbcache.steps_computation_mask.empty()) {
+            if (step_index < static_cast<int>(config.dbcache.steps_computation_mask.size())) {
+                scm_allows_cache = (config.dbcache.steps_computation_mask[step_index] == 0);
+                if (!config.dbcache.scm_policy_dynamic && scm_allows_cache) {
+                    can_cache_this_step  = true;
+                    is_caching_this_step = false;
+                    return;
+                }
+            }
+        }
+
+        bool max_cached_ok = (config.dbcache.max_cached_steps < 0) ||
+                             (static_cast<int>(cached_steps.size()) < config.dbcache.max_cached_steps);
+
+        bool max_cont_ok = (config.dbcache.max_continuous_cached_steps < 0) ||
+                           (continuous_cached_steps < config.dbcache.max_continuous_cached_steps);
+
+        bool accum_ok = (config.dbcache.max_accumulated_residual_diff < 0.0f) ||
+                        (accumulated_residual_diff < config.dbcache.max_accumulated_residual_diff);
+
+        can_cache_this_step  = !in_warmup && scm_allows_cache && max_cached_ok && max_cont_ok && accum_ok && has_prev_Fn_residual;
+        is_caching_this_step = false;
+    }
+
+    void end_step(bool was_cached) {
+        if (was_cached) {
+            cached_steps.push_back(current_step);
+            continuous_cached_steps++;
+        } else {
+            continuous_cached_steps = 0;
+        }
+    }
+
+    static float calculate_residual_diff(const float* prev, const float* curr, size_t size) {
+        if (size == 0)
+            return 0.0f;
+
+        float sum_diff = 0.0f;
+        float sum_abs  = 0.0f;
+
+        for (size_t i = 0; i < size; i++) {
+            sum_diff += std::fabs(prev[i] - curr[i]);
+            sum_abs += std::fabs(prev[i]);
+        }
+
+        return sum_diff / (sum_abs + 1e-6f);
+    }
+
+    static float calculate_residual_diff(const std::vector<float>& prev, const std::vector<float>& curr) {
+        if (prev.size() != curr.size() || prev.empty())
+            return 1.0f;
+        return calculate_residual_diff(prev.data(), curr.data(), prev.size());
+    }
+
+    int get_double_Fn_blocks() const {
+        return (config.double_Fn_blocks >= 0) ? config.double_Fn_blocks : config.dbcache.Fn_compute_blocks;
+    }
+
+    int get_double_Bn_blocks() const {
+        return (config.double_Bn_blocks >= 0) ? config.double_Bn_blocks : config.dbcache.Bn_compute_blocks;
+    }
+
+    int get_single_Fn_blocks() const {
+        return (config.single_Fn_blocks >= 0) ? config.single_Fn_blocks : config.dbcache.Fn_compute_blocks;
+    }
+
+    int get_single_Bn_blocks() const {
+        return (config.single_Bn_blocks >= 0) ? config.single_Bn_blocks : config.dbcache.Bn_compute_blocks;
+    }
+
+    bool is_Fn_double_block(int block_idx) const {
+        return block_idx < get_double_Fn_blocks();
+    }
+
+    bool is_Bn_double_block(int block_idx) const {
+        int Bn = get_double_Bn_blocks();
+        return Bn > 0 && block_idx >= (total_double_blocks - Bn);
+    }
+
+    bool is_Mn_double_block(int block_idx) const {
+        return !is_Fn_double_block(block_idx) && !is_Bn_double_block(block_idx);
+    }
+
+    bool is_Fn_single_block(int block_idx) const {
+        return block_idx < get_single_Fn_blocks();
+    }
+
+    bool is_Bn_single_block(int block_idx) const {
+        int Bn = get_single_Bn_blocks();
+        return Bn > 0 && block_idx >= (total_single_blocks - Bn);
+    }
+
+    bool is_Mn_single_block(int block_idx) const {
+        return !is_Fn_single_block(block_idx) && !is_Bn_single_block(block_idx);
+    }
+
+    void store_Fn_residual(const float* img, const float* txt, size_t img_size, size_t txt_size, const float* input_img, const float* input_txt) {
+        Fn_residual_img.resize(img_size);
+        Fn_residual_txt.resize(txt_size);
+
+        for (size_t i = 0; i < img_size; i++) {
+            Fn_residual_img[i] = img[i] - input_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            Fn_residual_txt[i] = txt[i] - input_txt[i];
+        }
+    }
+
+    bool check_cache_decision() {
+        if (!can_cache_this_step) {
+            is_caching_this_step = false;
+            return false;
+        }
+
+        if (!has_prev_Fn_residual || prev_Fn_residual_img.empty()) {
+            is_caching_this_step = false;
+            return false;
+        }
+
+        float diff_img = calculate_residual_diff(prev_Fn_residual_img, Fn_residual_img);
+        float diff_txt = calculate_residual_diff(prev_Fn_residual_txt, Fn_residual_txt);
+        float diff     = (diff_img + diff_txt) / 2.0f;
+
+        if (diff < config.dbcache.residual_diff_threshold) {
+            is_caching_this_step = true;
+            accumulated_residual_diff += diff;
+            return true;
+        }
+
+        is_caching_this_step = false;
+        return false;
+    }
+
+    void update_prev_Fn_residual() {
+        prev_Fn_residual_img = Fn_residual_img;
+        prev_Fn_residual_txt = Fn_residual_txt;
+        has_prev_Fn_residual = !prev_Fn_residual_img.empty();
+    }
+
+    void store_double_block_residual(int block_idx, const float* img, const float* txt, size_t img_size, size_t txt_size, const float* prev_img, const float* prev_txt) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(double_block_cache.size()))
+            return;
+
+        BlockCacheEntry& entry = double_block_cache[block_idx];
+
+        entry.residual_img.resize(img_size);
+        entry.residual_txt.resize(txt_size);
+        for (size_t i = 0; i < img_size; i++) {
+            entry.residual_img[i] = img[i] - prev_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            entry.residual_txt[i] = txt[i] - prev_txt[i];
+        }
+
+        entry.prev_img.resize(img_size);
+        entry.prev_txt.resize(txt_size);
+        for (size_t i = 0; i < img_size; i++) {
+            entry.prev_img[i] = img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            entry.prev_txt[i] = txt[i];
+        }
+        entry.has_prev = true;
+    }
+
+    void apply_double_block_cache(int block_idx, float* img, float* txt, size_t img_size, size_t txt_size) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(double_block_cache.size()))
+            return;
+
+        const BlockCacheEntry& entry = double_block_cache[block_idx];
+        if (entry.residual_img.size() != img_size || entry.residual_txt.size() != txt_size)
+            return;
+
+        for (size_t i = 0; i < img_size; i++) {
+            img[i] += entry.residual_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            txt[i] += entry.residual_txt[i];
+        }
+
+        total_blocks_cached++;
+    }
+
+    void store_single_block_residual(int block_idx, const float* output, size_t size, const float* input) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(single_block_cache.size()))
+            return;
+
+        BlockCacheEntry& entry = single_block_cache[block_idx];
+
+        entry.residual.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.residual[i] = output[i] - input[i];
+        }
+
+        entry.prev_output.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.prev_output[i] = output[i];
+        }
+        entry.has_prev = true;
+    }
+
+    void apply_single_block_cache(int block_idx, float* output, size_t size) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(single_block_cache.size()))
+            return;
+
+        const BlockCacheEntry& entry = single_block_cache[block_idx];
+        if (entry.residual.size() != size)
+            return;
+
+        for (size_t i = 0; i < size; i++) {
+            output[i] += entry.residual[i];
+        }
+
+        total_blocks_cached++;
+    }
+
+    void store_Bn_buffer(const float* img, const float* txt, size_t img_size, size_t txt_size, const float* Bn_start_img, const float* Bn_start_txt) {
+        Bn_buffer_img.resize(img_size);
+        Bn_buffer_txt.resize(txt_size);
+
+        for (size_t i = 0; i < img_size; i++) {
+            Bn_buffer_img[i] = img[i] - Bn_start_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            Bn_buffer_txt[i] = txt[i] - Bn_start_txt[i];
+        }
+        has_Bn_buffer = true;
+    }
+
+    void apply_Bn_buffer(float* img, float* txt, size_t img_size, size_t txt_size) {
+        if (!has_Bn_buffer)
+            return;
+        if (Bn_buffer_img.size() != img_size || Bn_buffer_txt.size() != txt_size)
+            return;
+
+        for (size_t i = 0; i < img_size; i++) {
+            img[i] += Bn_buffer_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            txt[i] += Bn_buffer_txt[i];
+        }
+    }
+
+    void taylor_update(const float* hidden_state, size_t size) {
+        if (!config.taylorseer.enabled)
+            return;
+        taylor_state.update_derivatives(hidden_state, size, current_step);
+    }
+
+    bool taylor_can_approximate() const {
+        return config.taylorseer.enabled && taylor_state.can_approximate();
+    }
+
+    void taylor_approximate(float* output, size_t size) {
+        if (!config.taylorseer.enabled)
+            return;
+        taylor_state.approximate(output, size, current_step);
+    }
+
+    bool should_use_taylor_this_step() const {
+        if (!config.taylorseer.enabled)
+            return false;
+        if (current_step < config.taylorseer.max_warmup_steps)
+            return false;
+
+        int interval = config.taylorseer.skip_interval_steps;
+        if (interval <= 0)
+            interval = 1;
+
+        return (current_step % (interval + 1)) != 0;
+    }
+
+    void log_metrics() const {
+        if (!enabled())
+            return;
+
+        int total_blocks  = total_blocks_computed + total_blocks_cached;
+        float cache_ratio = (total_blocks > 0) ? (static_cast<float>(total_blocks_cached) / total_blocks * 100.0f) : 0.0f;
+
+        float step_cache_ratio = (total_steps > 0) ? (static_cast<float>(cached_steps.size()) / total_steps * 100.0f) : 0.0f;
+
+        LOG_INFO("CacheDIT: steps_cached=%zu/%d (%.1f%%), blocks_cached=%d/%d (%.1f%%), accum_diff=%.4f",
+                 cached_steps.size(), total_steps, step_cache_ratio,
+                 total_blocks_cached, total_blocks, cache_ratio,
+                 accumulated_residual_diff);
+    }
+
+    std::string get_summary() const {
+        char buf[256];
+        snprintf(buf, sizeof(buf),
+                 "CacheDIT[thresh=%.2f]: cached %zu/%d steps, %d/%d blocks",
+                 config.dbcache.residual_diff_threshold,
+                 cached_steps.size(), total_steps,
+                 total_blocks_cached, total_blocks_computed + total_blocks_cached);
+        return std::string(buf);
+    }
+};
+
+inline std::vector<int> parse_scm_mask(const std::string& mask_str) {
+    std::vector<int> mask;
+    if (mask_str.empty())
+        return mask;
+
+    size_t pos   = 0;
+    size_t start = 0;
+    while ((pos = mask_str.find(',', start)) != std::string::npos) {
+        std::string token = mask_str.substr(start, pos - start);
+        mask.push_back(std::stoi(token));
+        start = pos + 1;
+    }
+    if (start < mask_str.length()) {
+        mask.push_back(std::stoi(mask_str.substr(start)));
+    }
+
+    return mask;
+}
+
+inline std::vector<int> generate_scm_mask(
+    const std::vector<int>& compute_bins,
+    const std::vector<int>& cache_bins,
+    int total_steps) {
+    std::vector<int> mask;
+    size_t c_idx = 0, cache_idx = 0;
+
+    while (static_cast<int>(mask.size()) < total_steps) {
+        if (c_idx < compute_bins.size()) {
+            for (int i = 0; i < compute_bins[c_idx] && static_cast<int>(mask.size()) < total_steps; i++) {
+                mask.push_back(1);
+            }
+            c_idx++;
+        }
+        if (cache_idx < cache_bins.size()) {
+            for (int i = 0; i < cache_bins[cache_idx] && static_cast<int>(mask.size()) < total_steps; i++) {
+                mask.push_back(0);
+            }
+            cache_idx++;
+        }
+        if (c_idx >= compute_bins.size() && cache_idx >= cache_bins.size())
+            break;
+    }
+
+    if (!mask.empty()) {
+        mask.back() = 1;
+    }
+
+    return mask;
+}
+
+inline std::vector<int> get_scm_preset(const std::string& preset, int total_steps) {
+    struct Preset {
+        std::vector<int> compute_bins;
+        std::vector<int> cache_bins;
+    };
+
+    Preset slow   = {{8, 3, 3, 2, 1, 1}, {1, 2, 2, 2, 3}};
+    Preset medium = {{6, 2, 2, 2, 2, 1}, {1, 3, 3, 3, 3}};
+    Preset fast   = {{6, 1, 1, 1, 1, 1}, {1, 3, 4, 5, 4}};
+    Preset ultra  = {{4, 1, 1, 1, 1}, {2, 5, 6, 7}};
+
+    Preset* p = nullptr;
+    if (preset == "slow" || preset == "s" || preset == "S")
+        p = &slow;
+    else if (preset == "medium" || preset == "m" || preset == "M")
+        p = &medium;
+    else if (preset == "fast" || preset == "f" || preset == "F")
+        p = &fast;
+    else if (preset == "ultra" || preset == "u" || preset == "U")
+        p = &ultra;
+    else
+        return {};
+
+    if (total_steps != 28 && total_steps > 0) {
+        float scale = static_cast<float>(total_steps) / 28.0f;
+        std::vector<int> scaled_compute, scaled_cache;
+
+        for (int v : p->compute_bins) {
+            scaled_compute.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
+        }
+        for (int v : p->cache_bins) {
+            scaled_cache.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
+        }
+
+        return generate_scm_mask(scaled_compute, scaled_cache, total_steps);
+    }
+
+    return generate_scm_mask(p->compute_bins, p->cache_bins, total_steps);
+}
+
+inline float get_preset_threshold(const std::string& preset) {
+    if (preset == "slow" || preset == "s" || preset == "S")
+        return 0.20f;
+    if (preset == "medium" || preset == "m" || preset == "M")
+        return 0.25f;
+    if (preset == "fast" || preset == "f" || preset == "F")
+        return 0.30f;
+    if (preset == "ultra" || preset == "u" || preset == "U")
+        return 0.34f;
+    return 0.08f;
+}
+
+inline int get_preset_warmup(const std::string& preset) {
+    if (preset == "slow" || preset == "s" || preset == "S")
+        return 8;
+    if (preset == "medium" || preset == "m" || preset == "M")
+        return 6;
+    if (preset == "fast" || preset == "f" || preset == "F")
+        return 6;
+    if (preset == "ultra" || preset == "u" || preset == "U")
+        return 4;
+    return 8;
+}
+
+inline int get_preset_Fn(const std::string& preset) {
+    if (preset == "slow" || preset == "s" || preset == "S")
+        return 8;
+    if (preset == "medium" || preset == "m" || preset == "M")
+        return 8;
+    if (preset == "fast" || preset == "f" || preset == "F")
+        return 6;
+    if (preset == "ultra" || preset == "u" || preset == "U")
+        return 4;
+    return 8;
+}
+
+inline int get_preset_Bn(const std::string& preset) {
+    (void)preset;
+    return 0;
+}
+
+inline void parse_dbcache_options(const std::string& opts, DBCacheConfig& cfg) {
+    if (opts.empty())
+        return;
+
+    int Fn = 8, Bn = 0, warmup = 8, max_cached = -1, max_cont = -1;
+    float thresh = 0.08f;
+
+    sscanf(opts.c_str(), "%d,%d,%f,%d,%d,%d",
+           &Fn, &Bn, &thresh, &warmup, &max_cached, &max_cont);
+
+    cfg.Fn_compute_blocks           = Fn;
+    cfg.Bn_compute_blocks           = Bn;
+    cfg.residual_diff_threshold     = thresh;
+    cfg.max_warmup_steps            = warmup;
+    cfg.max_cached_steps            = max_cached;
+    cfg.max_continuous_cached_steps = max_cont;
+}
+
+inline void parse_taylorseer_options(const std::string& opts, TaylorSeerConfig& cfg) {
+    if (opts.empty())
+        return;
+
+    int n_deriv = 1, warmup = 2, interval = 1;
+    sscanf(opts.c_str(), "%d,%d,%d", &n_deriv, &warmup, &interval);
+
+    cfg.n_derivatives       = n_deriv;
+    cfg.max_warmup_steps    = warmup;
+    cfg.skip_interval_steps = interval;
+}
+
+struct CacheDitConditionState {
+    DBCacheConfig config;
+    TaylorSeerConfig taylor_config;
+    bool initialized = false;
+
+    int current_step_index = -1;
+    bool step_active       = false;
+    bool skip_current_step = false;
+    bool initial_step      = true;
+    int warmup_remaining   = 0;
+    std::vector<int> cached_steps;
+    int continuous_cached_steps     = 0;
+    float accumulated_residual_diff = 0.0f;
+    int total_steps_skipped         = 0;
+
+    const void* anchor_condition = nullptr;
+
+    struct CacheEntry {
+        std::vector<float> diff;
+        std::vector<float> prev_input;
+        std::vector<float> prev_output;
+        bool has_prev = false;
+    };
+    std::unordered_map<const void*, CacheEntry> cache_diffs;
+
+    TaylorSeerState taylor_state;
+
+    float start_sigma = std::numeric_limits<float>::max();
+    float end_sigma   = 0.0f;
+
+    void reset_runtime() {
+        current_step_index = -1;
+        step_active        = false;
+        skip_current_step  = false;
+        initial_step       = true;
+        warmup_remaining   = config.max_warmup_steps;
+        cached_steps.clear();
+        continuous_cached_steps   = 0;
+        accumulated_residual_diff = 0.0f;
+        total_steps_skipped       = 0;
+        anchor_condition          = nullptr;
+        cache_diffs.clear();
+        taylor_state.reset();
+    }
+
+    void init(const DBCacheConfig& dbcfg, const TaylorSeerConfig& tcfg) {
+        config        = dbcfg;
+        taylor_config = tcfg;
+        initialized   = dbcfg.enabled || tcfg.enabled;
+        reset_runtime();
+
+        if (taylor_config.enabled) {
+            taylor_state.init(taylor_config.n_derivatives, 0);
+        }
+    }
+
+    void set_sigmas(const std::vector<float>& sigmas) {
+        if (!initialized || sigmas.size() < 2)
+            return;
+
+        float start_percent = 0.15f;
+        float end_percent   = 0.95f;
+
+        size_t n_steps    = sigmas.size() - 1;
+        size_t start_step = static_cast<size_t>(start_percent * n_steps);
+        size_t end_step   = static_cast<size_t>(end_percent * n_steps);
+
+        if (start_step >= n_steps)
+            start_step = n_steps - 1;
+        if (end_step >= n_steps)
+            end_step = n_steps - 1;
+
+        start_sigma = sigmas[start_step];
+        end_sigma   = sigmas[end_step];
+
+        if (start_sigma < end_sigma) {
+            std::swap(start_sigma, end_sigma);
+        }
+    }
+
+    bool enabled() const {
+        return initialized && (config.enabled || taylor_config.enabled);
+    }
+
+    void begin_step(int step_index, float sigma) {
+        if (!enabled())
+            return;
+        if (step_index == current_step_index)
+            return;
+
+        current_step_index = step_index;
+        skip_current_step  = false;
+        step_active        = false;
+
+        if (sigma > start_sigma)
+            return;
+        if (!(sigma > end_sigma))
+            return;
+
+        step_active = true;
+
+        if (warmup_remaining > 0) {
+            warmup_remaining--;
+            return;
+        }
+
+        if (!config.steps_computation_mask.empty()) {
+            if (step_index < static_cast<int>(config.steps_computation_mask.size())) {
+                if (config.steps_computation_mask[step_index] == 1) {
+                    return;
+                }
+            }
+        }
+
+        if (config.max_cached_steps >= 0 &&
+            static_cast<int>(cached_steps.size()) >= config.max_cached_steps) {
+            return;
+        }
+
+        if (config.max_continuous_cached_steps >= 0 &&
+            continuous_cached_steps >= config.max_continuous_cached_steps) {
+            return;
+        }
+    }
+
+    bool step_is_active() const {
+        return enabled() && step_active;
+    }
+
+    bool is_step_skipped() const {
+        return enabled() && step_active && skip_current_step;
+    }
+
+    bool has_cache(const void* cond) const {
+        auto it = cache_diffs.find(cond);
+        return it != cache_diffs.end() && !it->second.diff.empty();
+    }
+
+    void update_cache(const void* cond, const float* input, const float* output, size_t size) {
+        CacheEntry& entry = cache_diffs[cond];
+        entry.diff.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.diff[i] = output[i] - input[i];
+        }
+
+        entry.prev_input.resize(size);
+        entry.prev_output.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.prev_input[i]  = input[i];
+            entry.prev_output[i] = output[i];
+        }
+        entry.has_prev = true;
+    }
+
+    void apply_cache(const void* cond, const float* input, float* output, size_t size) {
+        auto it = cache_diffs.find(cond);
+        if (it == cache_diffs.end() || it->second.diff.empty())
+            return;
+        if (it->second.diff.size() != size)
+            return;
+
+        for (size_t i = 0; i < size; i++) {
+            output[i] = input[i] + it->second.diff[i];
+        }
+    }
+
+    bool before_condition(const void* cond, struct ggml_tensor* input, struct ggml_tensor* output, float sigma, int step_index) {
+        if (!enabled() || step_index < 0)
+            return false;
+
+        if (step_index != current_step_index) {
+            begin_step(step_index, sigma);
+        }
+
+        if (!step_active)
+            return false;
+
+        if (initial_step) {
+            anchor_condition = cond;
+            initial_step     = false;
+        }
+
+        bool is_anchor = (cond == anchor_condition);
+
+        if (skip_current_step) {
+            if (has_cache(cond)) {
+                apply_cache(cond, (float*)input->data, (float*)output->data,
+                            static_cast<size_t>(ggml_nelements(output)));
+                return true;
+            }
+            return false;
+        }
+
+        if (!is_anchor)
+            return false;
+
+        auto it = cache_diffs.find(cond);
+        if (it == cache_diffs.end() || !it->second.has_prev)
+            return false;
+
+        size_t ne = static_cast<size_t>(ggml_nelements(input));
+        if (it->second.prev_input.size() != ne)
+            return false;
+
+        float* input_data = (float*)input->data;
+        float diff        = CacheDitState::calculate_residual_diff(
+                   it->second.prev_input.data(), input_data, ne);
+
+        float effective_threshold = config.residual_diff_threshold;
+        if (config.Fn_compute_blocks > 0) {
+            float fn_confidence = 1.0f + 0.02f * (config.Fn_compute_blocks - 8);
+            fn_confidence       = std::max(0.5f, std::min(2.0f, fn_confidence));
+            effective_threshold *= fn_confidence;
+        }
+        if (config.Bn_compute_blocks > 0) {
+            float bn_quality = 1.0f - 0.03f * config.Bn_compute_blocks;
+            bn_quality       = std::max(0.5f, std::min(1.0f, bn_quality));
+            effective_threshold *= bn_quality;
+        }
+
+        if (diff < effective_threshold) {
+            skip_current_step = true;
+            total_steps_skipped++;
+            cached_steps.push_back(current_step_index);
+            continuous_cached_steps++;
+            accumulated_residual_diff += diff;
+            apply_cache(cond, input_data, (float*)output->data, ne);
+            return true;
+        }
+
+        continuous_cached_steps = 0;
+        return false;
+    }
+
+    void after_condition(const void* cond, struct ggml_tensor* input, struct ggml_tensor* output) {
+        if (!step_is_active())
+            return;
+
+        size_t ne = static_cast<size_t>(ggml_nelements(output));
+        update_cache(cond, (float*)input->data, (float*)output->data, ne);
+
+        if (cond == anchor_condition && taylor_config.enabled) {
+            taylor_state.update_derivatives((float*)output->data, ne, current_step_index);
+        }
+    }
+
+    void log_metrics() const {
+        if (!enabled())
+            return;
+
+        LOG_INFO("CacheDIT: steps_skipped=%d/%d (%.1f%%), accum_residual_diff=%.4f",
+                 total_steps_skipped,
+                 current_step_index + 1,
+                 (current_step_index > 0) ? (100.0f * total_steps_skipped / (current_step_index + 1)) : 0.0f,
+                 accumulated_residual_diff);
+    }
+};
+
+#endif

common.hpp

@@ -23,12 +23,12 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
         // x: [N, channels, h, w]
         if (vae_downsample) {
             auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
 
-            x = ggml_pad(ctx, x, 1, 1, 0, 0);
+            x = ggml_ext_pad(ctx->ggml_ctx, x, 1, 1, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
             x = conv->forward(ctx, x);
         } else {
             auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["op"]);
@@ -52,12 +52,12 @@ public:
         blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
         // x: [N, channels, h, w]
         auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
 
-        x = ggml_upscale(ctx, x, 2);  // [N, channels, h*2, w*2]
-        x = conv->forward(ctx, x);    // [N, out_channels, h*2, w*2]
+        x = ggml_upscale(ctx->ggml_ctx, x, 2, GGML_SCALE_MODE_NEAREST);  // [N, channels, h*2, w*2]
+        x = conv->forward(ctx, x);                                       // [N, out_channels, h*2, w*2]
         return x;
     }
 };
@@ -121,7 +121,7 @@ public:
         }
     }
 
-    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = NULL) {
+    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = nullptr) {
         // For dims==3, we reduce dimension from 5d to 4d by merging h and w, in order not to change ggml
         // [N, c, t, h, w] => [N, c, t, h * w]
         // x: [N, channels, h, w] if dims == 2 else [N, channels, t, h, w]
@@ -131,38 +131,38 @@ public:
         auto out_layers_0 = std::dynamic_pointer_cast<GroupNorm32>(blocks["out_layers.0"]);
         auto out_layers_3 = std::dynamic_pointer_cast<UnaryBlock>(blocks["out_layers.3"]);
 
-        if (emb == NULL) {
+        if (emb == nullptr) {
             GGML_ASSERT(skip_t_emb);
         }
 
         // in_layers
         auto h = in_layers_0->forward(ctx, x);
-        h      = ggml_silu_inplace(ctx, h);
+        h      = ggml_silu_inplace(ctx->ggml_ctx, h);
         h      = in_layers_2->forward(ctx, h);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
 
         // emb_layers
         if (!skip_t_emb) {
             auto emb_layer_1 = std::dynamic_pointer_cast<Linear>(blocks["emb_layers.1"]);
 
-            auto emb_out = ggml_silu(ctx, emb);
+            auto emb_out = ggml_silu(ctx->ggml_ctx, emb);
             emb_out      = emb_layer_1->forward(ctx, emb_out);  // [N, out_channels] if dims == 2 else [N, t, out_channels]
 
             if (dims == 2) {
-                emb_out = ggml_reshape_4d(ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]);  // [N, out_channels, 1, 1]
+                emb_out = ggml_reshape_4d(ctx->ggml_ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]);  // [N, out_channels, 1, 1]
             } else {
-                emb_out = ggml_reshape_4d(ctx, emb_out, 1, emb_out->ne[0], emb_out->ne[1], emb_out->ne[2]);  // [N, t, out_channels, 1]
+                emb_out = ggml_reshape_4d(ctx->ggml_ctx, emb_out, 1, emb_out->ne[0], emb_out->ne[1], emb_out->ne[2]);  // [N, t, out_channels, 1]
                 if (exchange_temb_dims) {
                     // emb_out = rearrange(emb_out, "b t c ... -> b c t ...")
-                    emb_out = ggml_cont(ctx, ggml_permute(ctx, emb_out, 0, 2, 1, 3));  // [N, out_channels, t, 1]
+                    emb_out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, emb_out, 0, 2, 1, 3));  // [N, out_channels, t, 1]
                 }
             }
 
-            h = ggml_add(ctx, h, emb_out);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+            h = ggml_add(ctx->ggml_ctx, h, emb_out);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
         }
 
         // out_layers
         h = out_layers_0->forward(ctx, h);
-        h = ggml_silu_inplace(ctx, h);
+        h = ggml_silu_inplace(ctx->ggml_ctx, h);
         // dropout, skip for inference
         h = out_layers_3->forward(ctx, h);
 
@@ -172,65 +172,97 @@ public:
             x                    = skip_connection->forward(ctx, x);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
         }
 
-        h = ggml_add(ctx, h, x);
+        h = ggml_add(ctx->ggml_ctx, h, x);
         return h;  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
     }
 };
 
-class GEGLU : public GGMLBlock {
+class GEGLU : public UnaryBlock {
 protected:
     int64_t dim_in;
     int64_t dim_out;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["proj.weight"] = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
-        params["proj.bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim_out * 2);
-    }
-
 public:
     GEGLU(int64_t dim_in, int64_t dim_out)
-        : dim_in(dim_in), dim_out(dim_out) {}
+        : dim_in(dim_in), dim_out(dim_out) {
+        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out * 2));
+    }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
         // x: [ne3, ne2, ne1, dim_in]
         // return: [ne3, ne2, ne1, dim_out]
-        struct ggml_tensor* w = params["proj.weight"];
-        struct ggml_tensor* b = params["proj.bias"];
+        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
 
-        auto x_w    = ggml_view_2d(ctx, w, w->ne[0], w->ne[1] / 2, w->nb[1], 0);                        // [dim_out, dim_in]
-        auto x_b    = ggml_view_1d(ctx, b, b->ne[0] / 2, 0);                                            // [dim_out, dim_in]
-        auto gate_w = ggml_view_2d(ctx, w, w->ne[0], w->ne[1] / 2, w->nb[1], w->nb[1] * w->ne[1] / 2);  // [dim_out, ]
-        auto gate_b = ggml_view_1d(ctx, b, b->ne[0] / 2, b->nb[0] * b->ne[0] / 2);                      // [dim_out, ]
+        x          = proj->forward(ctx, x);  // [ne3, ne2, ne1, dim_out*2]
+        auto x_vec = ggml_ext_chunk(ctx->ggml_ctx, x, 2, 0, false);
+        x          = x_vec[0];  // [ne3, ne2, ne1, dim_out]
+        auto gate  = x_vec[1];  // [ne3, ne2, ne1, dim_out]
 
-        auto x_in = x;
-        x         = ggml_nn_linear(ctx, x_in, x_w, x_b);        // [ne3, ne2, ne1, dim_out]
-        auto gate = ggml_nn_linear(ctx, x_in, gate_w, gate_b);  // [ne3, ne2, ne1, dim_out]
+        gate = ggml_cont(ctx->ggml_ctx, gate);
 
-        gate = ggml_gelu_inplace(ctx, gate);
+        gate = ggml_gelu_inplace(ctx->ggml_ctx, gate);
 
-        x = ggml_mul(ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
+        x = ggml_mul(ctx->ggml_ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
 
         return x;
     }
 };
 
+class GELU : public UnaryBlock {
+public:
+    GELU(int64_t dim_in, int64_t dim_out, bool bias = true) {
+        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
+    }
+
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+        // x: [ne3, ne2, ne1, dim_in]
+        // return: [ne3, ne2, ne1, dim_out]
+        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+
+        x = proj->forward(ctx, x);
+        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        return x;
+    }
+};
+
 class FeedForward : public GGMLBlock {
 public:
+    enum class Activation {
+        GEGLU,
+        GELU
+    };
     FeedForward(int64_t dim,
                 int64_t dim_out,
-                int64_t mult = 4) {
+                int64_t mult          = 4,
+                Activation activation = Activation::GEGLU,
+                bool precision_fix    = false) {
         int64_t inner_dim = dim * mult;
+        if (activation == Activation::GELU) {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
+        } else {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
+        }
 
-        blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
         // net_1 is nn.Dropout(), skip for inference
-        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out));
+        bool force_prec_f32 = false;
+        float scale         = 1.f;
+        if (precision_fix) {
+            scale = 1.f / 128.f;
+#ifdef SD_USE_VULKAN
+            force_prec_f32 = true;
+#endif
+        }
+        // The purpose of the scale here is to prevent NaN issues in certain situations.
+        // For example, when using Vulkan without enabling force_prec_f32,
+        // or when using CUDA but the weights are k-quants.
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, force_prec_f32, scale));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
         // x: [ne3, ne2, ne1, dim]
         // return: [ne3, ne2, ne1, dim_out]
 
-        auto net_0 = std::dynamic_pointer_cast<GEGLU>(blocks["net.0"]);
+        auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
         auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
 
         x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
@@ -265,7 +297,9 @@ public:
         // to_out_1 is nn.Dropout(), skip for inference
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* context) {
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
         // x: [N, n_token, query_dim]
         // context: [N, n_context, context_dim]
         // return: [N, n_token, query_dim]
@@ -279,26 +313,11 @@ public:
         int64_t n_context = context->ne[1];
         int64_t inner_dim = d_head * n_head;
 
-        auto q = to_q->forward(ctx, x);                                 // [N, n_token, inner_dim]
-        q      = ggml_reshape_4d(ctx, q, d_head, n_head, n_token, n);   // [N, n_token, n_head, d_head]
-        q      = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));      // [N, n_head, n_token, d_head]
-        q      = ggml_reshape_3d(ctx, q, d_head, n_token, n_head * n);  // [N * n_head, n_token, d_head]
+        auto q = to_q->forward(ctx, x);        // [N, n_token, inner_dim]
+        auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
+        auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
 
-        auto k = to_k->forward(ctx, context);                             // [N, n_context, inner_dim]
-        k      = ggml_reshape_4d(ctx, k, d_head, n_head, n_context, n);   // [N, n_context, n_head, d_head]
-        k      = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3));        // [N, n_head, n_context, d_head]
-        k      = ggml_reshape_3d(ctx, k, d_head, n_context, n_head * n);  // [N * n_head, n_context, d_head]
-
-        auto v = to_v->forward(ctx, context);                             // [N, n_context, inner_dim]
-        v      = ggml_reshape_4d(ctx, v, d_head, n_head, n_context, n);   // [N, n_context, n_head, d_head]
-        v      = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));        // [N, n_head, d_head, n_context]
-        v      = ggml_reshape_3d(ctx, v, n_context, d_head, n_head * n);  // [N * n_head, d_head, n_context]
-
-        auto kqv = ggml_nn_attention(ctx, q, k, v, false);  // [N * n_head, n_token, d_head]
-        kqv      = ggml_reshape_4d(ctx, kqv, d_head, n_token, n_head, n);
-        kqv      = ggml_cont(ctx, ggml_permute(ctx, kqv, 0, 2, 1, 3));  // [N, n_token, n_head, d_head]
-
-        x = ggml_reshape_3d(ctx, kqv, d_head * n_head, n_token, n);  // [N, n_token, inner_dim]
+        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_token, inner_dim]
 
         x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
         return x;
@@ -336,7 +355,9 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* context) {
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
         // x: [N, n_token, query_dim]
         // context: [N, n_context, context_dim]
         // return: [N, n_token, query_dim]
@@ -356,21 +377,21 @@ public:
             x           = norm_in->forward(ctx, x);
             x           = ff_in->forward(ctx, x);
             // self.is_res is always True
-            x = ggml_add(ctx, x, x_skip);
+            x = ggml_add(ctx->ggml_ctx, x, x_skip);
         }
 
         auto r = x;
         x      = norm1->forward(ctx, x);
         x      = attn1->forward(ctx, x, x);  // self-attention
-        x      = ggml_add(ctx, x, r);
+        x      = ggml_add(ctx->ggml_ctx, x, r);
         r      = x;
         x      = norm2->forward(ctx, x);
         x      = attn2->forward(ctx, x, context);  // cross-attention
-        x      = ggml_add(ctx, x, r);
+        x      = ggml_add(ctx->ggml_ctx, x, r);
         r      = x;
         x      = norm3->forward(ctx, x);
         x      = ff->forward(ctx, x);
-        x      = ggml_add(ctx, x, r);
+        x      = ggml_add(ctx->ggml_ctx, x, r);
 
         return x;
     }
@@ -382,39 +403,67 @@ protected:
     int64_t n_head;
     int64_t d_head;
     int64_t depth       = 1;    // 1
-    int64_t context_dim = 768;  // hidden_size, 1024 for VERSION_2_x
+    int64_t context_dim = 768;  // hidden_size, 1024 for VERSION_SD2
+    bool use_linear     = false;
+
+    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
+        auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
+        if (iter != tensor_storage_map.end()) {
+            int64_t inner_dim = n_head * d_head;
+            if (iter->second.n_dims == 4 && use_linear) {
+                use_linear         = false;
+                blocks["proj_in"]  = std::make_shared<Conv2d>(in_channels, inner_dim, std::pair{1, 1});
+                blocks["proj_out"] = std::make_shared<Conv2d>(inner_dim, in_channels, std::pair{1, 1});
+            } else if (iter->second.n_dims == 2 && !use_linear) {
+                use_linear         = true;
+                blocks["proj_in"]  = std::make_shared<Linear>(in_channels, inner_dim);
+                blocks["proj_out"] = std::make_shared<Linear>(inner_dim, in_channels);
+            }
+        }
+    }
 
 public:
     SpatialTransformer(int64_t in_channels,
                        int64_t n_head,
                        int64_t d_head,
                        int64_t depth,
-                       int64_t context_dim)
+                       int64_t context_dim,
+                       bool use_linear)
         : in_channels(in_channels),
           n_head(n_head),
           d_head(d_head),
           depth(depth),
-          context_dim(context_dim) {
-        // We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False
+          context_dim(context_dim),
+          use_linear(use_linear) {
         // disable_self_attn is always False
         int64_t inner_dim = n_head * d_head;  // in_channels
         blocks["norm"]    = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
-        blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, inner_dim, {1, 1}));
+        if (use_linear) {
+            blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, inner_dim));
+        } else {
+            blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, inner_dim, {1, 1}));
+        }
 
         for (int i = 0; i < depth; i++) {
             std::string name = "transformer_blocks." + std::to_string(i);
-            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim));
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim, false));
         }
 
-        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
+        if (use_linear) {
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, in_channels));
+        } else {
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
+        }
     }
 
-    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* context) {
+    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                        struct ggml_tensor* x,
+                                        struct ggml_tensor* context) {
         // x: [N, in_channels, h, w]
         // context: [N, max_position(aka n_token), hidden_size(aka context_dim)]
         auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
-        auto proj_in  = std::dynamic_pointer_cast<Conv2d>(blocks["proj_in"]);
-        auto proj_out = std::dynamic_pointer_cast<Conv2d>(blocks["proj_out"]);
+        auto proj_in  = std::dynamic_pointer_cast<UnaryBlock>(blocks["proj_in"]);
+        auto proj_out = std::dynamic_pointer_cast<UnaryBlock>(blocks["proj_out"]);
 
         auto x_in         = x;
         int64_t n         = x->ne[3];
@@ -423,10 +472,15 @@ public:
         int64_t inner_dim = n_head * d_head;
 
         x = norm->forward(ctx, x);
-        x = proj_in->forward(ctx, x);  // [N, inner_dim, h, w]
-
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
-        x = ggml_reshape_3d(ctx, x, inner_dim, w * h, n);      // [N, h * w, inner_dim]
+        if (use_linear) {
+            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
+            x = ggml_reshape_3d(ctx->ggml_ctx, x, inner_dim, w * h, n);                // [N, h * w, inner_dim]
+            x = proj_in->forward(ctx, x);                                              // [N, inner_dim, h, w]
+        } else {
+            x = proj_in->forward(ctx, x);                                              // [N, inner_dim, h, w]
+            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
+            x = ggml_reshape_3d(ctx->ggml_ctx, x, inner_dim, w * h, n);                // [N, h * w, inner_dim]
+        }
 
         for (int i = 0; i < depth; i++) {
             std::string name       = "transformer_blocks." + std::to_string(i);
@@ -435,27 +489,37 @@ public:
             x = transformer_block->forward(ctx, x, context);
         }
 
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
-        x = ggml_reshape_4d(ctx, x, w, h, inner_dim, n);       // [N, inner_dim, h, w]
+        if (use_linear) {
+            // proj_out
+            x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
 
-        // proj_out
-        x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
+            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
+            x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, inner_dim, n);                 // [N, inner_dim, h, w]
+        } else {
+            x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
+            x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, inner_dim, n);                 // [N, inner_dim, h, w]
 
-        x = ggml_add(ctx, x, x_in);
+            // proj_out
+            x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
+        }
+
+        x = ggml_add(ctx->ggml_ctx, x, x_in);
         return x;
     }
 };
 
 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["mix_factor"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
+        // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
+        enum ggml_type wtype = GGML_TYPE_F32;
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 
     float get_alpha() {
         // image_only_indicator is always tensor([0.]) and since mix_factor.shape is [1,]
         // so learned_with_images is same as learned
-        float alpha = ggml_backend_tensor_get_f32(params["mix_factor"]);
+        float alpha = ggml_ext_backend_tensor_get_f32(params["mix_factor"]);
         return sigmoid(alpha);
     }
 
@@ -466,14 +530,14 @@ public:
         // since mix_factor.shape is [1,], we don't need rearrange using  rearrange_pattern
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx,
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                 struct ggml_tensor* x_spatial,
                                 struct ggml_tensor* x_temporal) {
         // image_only_indicator is always tensor([0.])
         float alpha = get_alpha();
-        auto x      = ggml_add(ctx,
-                               ggml_scale(ctx, x_spatial, alpha),
-                               ggml_scale(ctx, x_temporal, 1.0f - alpha));
+        auto x      = ggml_add(ctx->ggml_ctx,
+                               ggml_scale(ctx->ggml_ctx, x_spatial, alpha),
+                               ggml_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
         return x;
     }
 };
@@ -491,7 +555,7 @@ public:
         blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx,
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* emb,
                                 int num_video_frames) {
@@ -509,21 +573,21 @@ public:
         int64_t H = x->ne[1];
         int64_t W = x->ne[0];
 
-        x          = ggml_reshape_4d(ctx, x, W * H, C, T, B);           // (b t) c h w -> b t c (h w)
-        x          = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
+        x          = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, C, T, B);                     // (b t) c h w -> b t c (h w)
+        x          = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
         auto x_mix = x;
 
-        emb = ggml_reshape_4d(ctx, emb, emb->ne[0], T, B, emb->ne[3]);  // (b t) ... -> b t ...
+        emb = ggml_reshape_4d(ctx->ggml_ctx, emb, emb->ne[0], T, B, emb->ne[3]);  // (b t) ... -> b t ...
 
         x = time_stack->forward(ctx, x, emb);  // b t c (h w)
 
         x = time_mixer->forward(ctx, x_mix, x);  // b t c (h w)
 
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
-        x = ggml_reshape_4d(ctx, x, W, H, C, T * B);           // b t c (h w) -> (b t) c h w
+        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
+        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
 
         return x;
     }
 };
 
-#endif  // __COMMON_HPP__
+#endif  // __COMMON_HPP__

control.hpp

@@ -14,7 +14,7 @@
 */
 class ControlNetBlock : public GGMLBlock {
 protected:
-    SDVersion version = VERSION_1_x;
+    SDVersion version = VERSION_SD1;
     // network hparams
     int in_channels                        = 4;
     int out_channels                       = 4;
@@ -26,19 +26,20 @@ protected:
     int time_embed_dim                     = 1280;  // model_channels*4
     int num_heads                          = 8;
     int num_head_channels                  = -1;   // channels // num_heads
-    int context_dim                        = 768;  // 1024 for VERSION_2_x, 2048 for VERSION_XL
+    int context_dim                        = 768;  // 1024 for VERSION_SD2, 2048 for VERSION_SDXL
+    bool use_linear_projection             = false;
 
 public:
     int model_channels  = 320;
-    int adm_in_channels = 2816;  // only for VERSION_XL
+    int adm_in_channels = 2816;  // only for VERSION_SDXL
 
-    ControlNetBlock(SDVersion version = VERSION_1_x)
+    ControlNetBlock(SDVersion version = VERSION_SD1)
         : version(version) {
-        if (version == VERSION_2_x) {
+        if (sd_version_is_sd2(version)) {
             context_dim       = 1024;
             num_head_channels = 64;
             num_heads         = -1;
-        } else if (version == VERSION_XL) {
+        } else if (sd_version_is_sdxl(version)) {
             context_dim           = 2048;
             attention_resolutions = {4, 2};
             channel_mult          = {1, 2, 4};
@@ -58,7 +59,7 @@ public:
         // time_embed_1 is nn.SiLU()
         blocks["time_embed.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
 
-        if (version == VERSION_XL || version == VERSION_SVD) {
+        if (sd_version_is_sdxl(version) || version == VERSION_SVD) {
             blocks["label_emb.0.0"] = std::shared_ptr<GGMLBlock>(new Linear(adm_in_channels, time_embed_dim));
             // label_emb_1 is nn.SiLU()
             blocks["label_emb.0.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
@@ -82,7 +83,7 @@ public:
                                        int64_t d_head,
                                        int64_t depth,
                                        int64_t context_dim) -> SpatialTransformer* {
-            return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
+            return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, use_linear_projection);
         };
 
         auto make_zero_conv = [&](int64_t channels) {
@@ -165,7 +166,7 @@ public:
     }
 
     struct ggml_tensor* resblock_forward(std::string name,
-                                         struct ggml_context* ctx,
+                                         GGMLRunnerContext* ctx,
                                          struct ggml_tensor* x,
                                          struct ggml_tensor* emb) {
         auto block = std::dynamic_pointer_cast<ResBlock>(blocks[name]);
@@ -173,14 +174,14 @@ public:
     }
 
     struct ggml_tensor* attention_layer_forward(std::string name,
-                                                struct ggml_context* ctx,
+                                                GGMLRunnerContext* ctx,
                                                 struct ggml_tensor* x,
                                                 struct ggml_tensor* context) {
         auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
         return block->forward(ctx, x, context);
     }
 
-    struct ggml_tensor* input_hint_block_forward(struct ggml_context* ctx,
+    struct ggml_tensor* input_hint_block_forward(GGMLRunnerContext* ctx,
                                                  struct ggml_tensor* hint,
                                                  struct ggml_tensor* emb,
                                                  struct ggml_tensor* context) {
@@ -192,32 +193,32 @@ public:
 
                 h = block->forward(ctx, h);
             } else {
-                h = ggml_silu_inplace(ctx, h);
+                h = ggml_silu_inplace(ctx->ggml_ctx, h);
             }
         }
         return h;
     }
 
-    std::vector<struct ggml_tensor*> forward(struct ggml_context* ctx,
+    std::vector<struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
                                              struct ggml_tensor* x,
                                              struct ggml_tensor* hint,
                                              struct ggml_tensor* guided_hint,
                                              struct ggml_tensor* timesteps,
                                              struct ggml_tensor* context,
-                                             struct ggml_tensor* y = NULL) {
+                                             struct ggml_tensor* y = nullptr) {
         // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
         // timesteps: [N,]
         // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
-        if (context != NULL) {
+        if (context != nullptr) {
             if (context->ne[2] != x->ne[3]) {
-                context = ggml_repeat(ctx, context, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
+                context = ggml_repeat(ctx->ggml_ctx, context, ggml_new_tensor_3d(ctx->ggml_ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
             }
         }
 
-        if (y != NULL) {
+        if (y != nullptr) {
             if (y->ne[1] != x->ne[3]) {
-                y = ggml_repeat(ctx, y, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
+                y = ggml_repeat(ctx->ggml_ctx, y, ggml_new_tensor_2d(ctx->ggml_ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
             }
         }
 
@@ -228,27 +229,27 @@ public:
 
         auto middle_block_out = std::dynamic_pointer_cast<Conv2d>(blocks["middle_block_out.0"]);
 
-        auto t_emb = ggml_nn_timestep_embedding(ctx, timesteps, model_channels);  // [N, model_channels]
+        auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, model_channels);  // [N, model_channels]
 
         auto emb = time_embed_0->forward(ctx, t_emb);
-        emb      = ggml_silu_inplace(ctx, emb);
+        emb      = ggml_silu_inplace(ctx->ggml_ctx, emb);
         emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]
 
         // SDXL/SVD
-        if (y != NULL) {
+        if (y != nullptr) {
             auto label_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.0"]);
             auto label_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.2"]);
 
             auto label_emb = label_embed_0->forward(ctx, y);
-            label_emb      = ggml_silu_inplace(ctx, label_emb);
+            label_emb      = ggml_silu_inplace(ctx->ggml_ctx, label_emb);
             label_emb      = label_embed_2->forward(ctx, label_emb);  // [N, time_embed_dim]
 
-            emb = ggml_add(ctx, emb, label_emb);  // [N, time_embed_dim]
+            emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
         }
 
         std::vector<struct ggml_tensor*> outs;
 
-        if (guided_hint == NULL) {
+        if (guided_hint == nullptr) {
             guided_hint = input_hint_block_forward(ctx, hint, emb, context);
         }
         outs.push_back(guided_hint);
@@ -257,7 +258,7 @@ public:
 
         // input block 0
         auto h = input_blocks_0_0->forward(ctx, x);
-        h      = ggml_add(ctx, h, guided_hint);
+        h      = ggml_add(ctx->ggml_ctx, h, guided_hint);
         outs.push_back(zero_convs_0->forward(ctx, h));
 
         // input block 1-11
@@ -306,31 +307,32 @@ public:
     }
 };
 
-struct ControlNet : public GGMLModule {
-    SDVersion version = VERSION_1_x;
+struct ControlNet : public GGMLRunner {
+    SDVersion version = VERSION_SD1;
     ControlNetBlock control_net;
 
-    ggml_backend_buffer_t control_buffer = NULL;  // keep control output tensors in backend memory
-    ggml_context* control_ctx            = NULL;
+    ggml_backend_buffer_t control_buffer = nullptr;  // keep control output tensors in backend memory
+    ggml_context* control_ctx            = nullptr;
     std::vector<struct ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
-    struct ggml_tensor* guided_hint = NULL;     // guided_hint cache, for faster inference
+    struct ggml_tensor* guided_hint = nullptr;  // guided_hint cache, for faster inference
     bool guided_hint_cached         = false;
 
     ControlNet(ggml_backend_t backend,
-               ggml_type wtype,
-               SDVersion version = VERSION_1_x)
-        : GGMLModule(backend, wtype), control_net(version) {
-        control_net.init(params_ctx, wtype);
+               bool offload_params_to_cpu,
+               const String2TensorStorage& tensor_storage_map = {},
+               SDVersion version                              = VERSION_SD1)
+        : GGMLRunner(backend, offload_params_to_cpu), control_net(version) {
+        control_net.init(params_ctx, tensor_storage_map, "");
     }
 
-    ~ControlNet() {
+    ~ControlNet() override {
         free_control_ctx();
     }
 
     void alloc_control_ctx(std::vector<struct ggml_tensor*> outs) {
         struct ggml_init_params params;
         params.mem_size   = static_cast<size_t>(outs.size() * ggml_tensor_overhead()) + 1024 * 1024;
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
         params.no_alloc   = true;
         control_ctx       = ggml_init(params);
 
@@ -346,37 +348,29 @@ struct ControlNet : public GGMLModule {
             control_buffer_size += ggml_nbytes(controls[i]);
         }
 
-        control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, backend);
+        control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, runtime_backend);
 
         LOG_DEBUG("control buffer size %.2fMB", control_buffer_size * 1.f / 1024.f / 1024.f);
     }
 
     void free_control_ctx() {
-        if (control_buffer != NULL) {
+        if (control_buffer != nullptr) {
             ggml_backend_buffer_free(control_buffer);
-            control_buffer = NULL;
+            control_buffer = nullptr;
         }
-        if (control_ctx != NULL) {
+        if (control_ctx != nullptr) {
             ggml_free(control_ctx);
-            control_ctx = NULL;
+            control_ctx = nullptr;
         }
-        guided_hint        = NULL;
+        guided_hint        = nullptr;
         guided_hint_cached = false;
         controls.clear();
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "control_net";
     }
 
-    size_t get_params_mem_size() {
-        return control_net.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return control_net.get_params_num();
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         control_net.get_param_tensors(tensors, prefix);
     }
@@ -385,12 +379,12 @@ struct ControlNet : public GGMLModule {
                                     struct ggml_tensor* hint,
                                     struct ggml_tensor* timesteps,
                                     struct ggml_tensor* context,
-                                    struct ggml_tensor* y = NULL) {
-        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, CONTROL_NET_GRAPH_SIZE, false);
+                                    struct ggml_tensor* y = nullptr) {
+        struct ggml_cgraph* gf = new_graph_custom(CONTROL_NET_GRAPH_SIZE);
 
         x = to_backend(x);
         if (guided_hint_cached) {
-            hint = NULL;
+            hint = nullptr;
         } else {
             hint = to_backend(hint);
         }
@@ -398,15 +392,17 @@ struct ControlNet : public GGMLModule {
         y         = to_backend(y);
         timesteps = to_backend(timesteps);
 
-        auto outs = control_net.forward(compute_ctx,
+        auto runner_ctx = get_context();
+
+        auto outs = control_net.forward(&runner_ctx,
                                         x,
                                         hint,
-                                        guided_hint_cached ? guided_hint : NULL,
+                                        guided_hint_cached ? guided_hint : nullptr,
                                         timesteps,
                                         context,
                                         y);
 
-        if (control_ctx == NULL) {
+        if (control_ctx == nullptr) {
             alloc_control_ctx(outs);
         }
 
@@ -418,14 +414,14 @@ struct ControlNet : public GGMLModule {
         return gf;
     }
 
-    void compute(int n_threads,
+    bool compute(int n_threads,
                  struct ggml_tensor* x,
                  struct ggml_tensor* hint,
                  struct ggml_tensor* timesteps,
                  struct ggml_tensor* context,
                  struct ggml_tensor* y,
-                 struct ggml_tensor** output     = NULL,
-                 struct ggml_context* output_ctx = NULL) {
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
@@ -434,11 +430,15 @@ struct ControlNet : public GGMLModule {
             return build_graph(x, hint, timesteps, context, y);
         };
 
-        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
-        guided_hint_cached = true;
+        bool res = GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        if (res) {
+            // cache guided_hint
+            guided_hint_cached = true;
+        }
+        return res;
     }
 
-    bool load_from_file(const std::string& file_path) {
+    bool load_from_file(const std::string& file_path, int n_threads) {
         LOG_INFO("loading control net from '%s'", file_path.c_str());
         alloc_params_buffer();
         std::map<std::string, ggml_tensor*> tensors;
@@ -446,12 +446,12 @@ struct ControlNet : public GGMLModule {
         std::set<std::string> ignore_tensors;
 
         ModelLoader model_loader;
-        if (!model_loader.init_from_file(file_path)) {
+        if (!model_loader.init_from_file_and_convert_name(file_path)) {
             LOG_ERROR("init control net model loader from file failed: '%s'", file_path.c_str());
             return false;
         }
 
-        bool success = model_loader.load_tensors(tensors, backend, ignore_tensors);
+        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads);
 
         if (!success) {
             LOG_ERROR("load control net tensors from model loader failed");

diffusion_model.hpp

@@ -0,0 +1,450 @@
+#ifndef __DIFFUSION_MODEL_H__
+#define __DIFFUSION_MODEL_H__
+
+#include "flux.hpp"
+#include "mmdit.hpp"
+#include "qwen_image.hpp"
+#include "unet.hpp"
+#include "wan.hpp"
+#include "z_image.hpp"
+
+struct DiffusionParams {
+    struct ggml_tensor* x                     = nullptr;
+    struct ggml_tensor* timesteps             = nullptr;
+    struct ggml_tensor* context               = nullptr;
+    struct ggml_tensor* c_concat              = nullptr;
+    struct ggml_tensor* y                     = nullptr;
+    struct ggml_tensor* guidance              = nullptr;
+    std::vector<ggml_tensor*> ref_latents     = {};
+    bool increase_ref_index                   = false;
+    int num_video_frames                      = -1;
+    std::vector<struct ggml_tensor*> controls = {};
+    float control_strength                    = 0.f;
+    struct ggml_tensor* vace_context          = nullptr;
+    float vace_strength                       = 1.f;
+    std::vector<int> skip_layers              = {};
+};
+
+struct DiffusionModel {
+    virtual std::string get_desc()                                                      = 0;
+    virtual bool compute(int n_threads,
+                         DiffusionParams diffusion_params,
+                         struct ggml_tensor** output     = nullptr,
+                         struct ggml_context* output_ctx = nullptr)                     = 0;
+    virtual void alloc_params_buffer()                                                  = 0;
+    virtual void free_params_buffer()                                                   = 0;
+    virtual void free_compute_buffer()                                                  = 0;
+    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) = 0;
+    virtual size_t get_params_buffer_size()                                             = 0;
+    virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
+    virtual int64_t get_adm_in_channels()                            = 0;
+    virtual void set_flash_attn_enabled(bool enabled)                = 0;
+    virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
+};
+
+struct UNetModel : public DiffusionModel {
+    UNetModelRunner unet;
+
+    UNetModel(ggml_backend_t backend,
+              bool offload_params_to_cpu,
+              const String2TensorStorage& tensor_storage_map = {},
+              SDVersion version                              = VERSION_SD1)
+        : unet(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model", version) {
+    }
+
+    std::string get_desc() override {
+        return unet.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        unet.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        unet.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        unet.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        unet.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() override {
+        return unet.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        unet.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return unet.unet.adm_in_channels;
+    }
+
+    void set_flash_attn_enabled(bool enabled) {
+        unet.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        unet.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return unet.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.num_video_frames,
+                            diffusion_params.controls,
+                            diffusion_params.control_strength, output, output_ctx);
+    }
+};
+
+struct MMDiTModel : public DiffusionModel {
+    MMDiTRunner mmdit;
+
+    MMDiTModel(ggml_backend_t backend,
+               bool offload_params_to_cpu,
+               const String2TensorStorage& tensor_storage_map = {})
+        : mmdit(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model") {
+    }
+
+    std::string get_desc() override {
+        return mmdit.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        mmdit.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        mmdit.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        mmdit.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        mmdit.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() override {
+        return mmdit.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        mmdit.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768 + 1280;
+    }
+
+    void set_flash_attn_enabled(bool enabled) {
+        mmdit.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        mmdit.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return mmdit.compute(n_threads,
+                             diffusion_params.x,
+                             diffusion_params.timesteps,
+                             diffusion_params.context,
+                             diffusion_params.y,
+                             output,
+                             output_ctx,
+                             diffusion_params.skip_layers);
+    }
+};
+
+struct FluxModel : public DiffusionModel {
+    Flux::FluxRunner flux;
+
+    FluxModel(ggml_backend_t backend,
+              bool offload_params_to_cpu,
+              const String2TensorStorage& tensor_storage_map = {},
+              SDVersion version                              = VERSION_FLUX,
+              bool use_mask                                  = false)
+        : flux(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model", version, use_mask) {
+    }
+
+    std::string get_desc() override {
+        return flux.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        flux.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        flux.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        flux.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        flux.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() override {
+        return flux.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        flux.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attn_enabled(bool enabled) {
+        flux.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        flux.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return flux.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.guidance,
+                            diffusion_params.ref_latents,
+                            diffusion_params.increase_ref_index,
+                            output,
+                            output_ctx,
+                            diffusion_params.skip_layers);
+    }
+};
+
+struct WanModel : public DiffusionModel {
+    std::string prefix;
+    WAN::WanRunner wan;
+
+    WanModel(ggml_backend_t backend,
+             bool offload_params_to_cpu,
+             const String2TensorStorage& tensor_storage_map = {},
+             const std::string prefix                       = "model.diffusion_model",
+             SDVersion version                              = VERSION_WAN2)
+        : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_storage_map, prefix, version) {
+    }
+
+    std::string get_desc() override {
+        return wan.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        wan.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        wan.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        wan.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        wan.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return wan.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        wan.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attn_enabled(bool enabled) {
+        wan.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        wan.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return wan.compute(n_threads,
+                           diffusion_params.x,
+                           diffusion_params.timesteps,
+                           diffusion_params.context,
+                           diffusion_params.y,
+                           diffusion_params.c_concat,
+                           nullptr,
+                           diffusion_params.vace_context,
+                           diffusion_params.vace_strength,
+                           output,
+                           output_ctx);
+    }
+};
+
+struct QwenImageModel : public DiffusionModel {
+    std::string prefix;
+    Qwen::QwenImageRunner qwen_image;
+
+    QwenImageModel(ggml_backend_t backend,
+                   bool offload_params_to_cpu,
+                   const String2TensorStorage& tensor_storage_map = {},
+                   const std::string prefix                       = "model.diffusion_model",
+                   SDVersion version                              = VERSION_QWEN_IMAGE,
+                   bool zero_cond_t                               = false)
+        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_storage_map, prefix, version, zero_cond_t) {
+    }
+
+    std::string get_desc() override {
+        return qwen_image.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        qwen_image.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        qwen_image.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        qwen_image.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        qwen_image.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return qwen_image.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        qwen_image.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attn_enabled(bool enabled) {
+        qwen_image.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        qwen_image.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return qwen_image.compute(n_threads,
+                                  diffusion_params.x,
+                                  diffusion_params.timesteps,
+                                  diffusion_params.context,
+                                  diffusion_params.ref_latents,
+                                  true,  // increase_ref_index
+                                  output,
+                                  output_ctx);
+    }
+};
+
+struct ZImageModel : public DiffusionModel {
+    std::string prefix;
+    ZImage::ZImageRunner z_image;
+
+    ZImageModel(ggml_backend_t backend,
+                bool offload_params_to_cpu,
+                const String2TensorStorage& tensor_storage_map = {},
+                const std::string prefix                       = "model.diffusion_model",
+                SDVersion version                              = VERSION_Z_IMAGE)
+        : prefix(prefix), z_image(backend, offload_params_to_cpu, tensor_storage_map, prefix, version) {
+    }
+
+    std::string get_desc() override {
+        return z_image.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        z_image.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        z_image.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        z_image.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        z_image.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return z_image.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        z_image.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attn_enabled(bool enabled) {
+        z_image.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        z_image.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return z_image.compute(n_threads,
+                               diffusion_params.x,
+                               diffusion_params.timesteps,
+                               diffusion_params.context,
+                               diffusion_params.ref_latents,
+                               true,  // increase_ref_index
+                               output,
+                               output_ctx);
+    }
+};
+
+#endif

docs/build.md

@@ -0,0 +1,173 @@
+# Build from scratch
+
+## Get the Code
+
+```
+git clone --recursive https://github.com/leejet/stable-diffusion.cpp
+cd stable-diffusion.cpp
+```
+
+- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
+
+```
+cd stable-diffusion.cpp
+git pull origin master
+git submodule init
+git submodule update
+```
+
+## Build (CPU only)
+
+If you don't have a GPU or CUDA installed, you can build a CPU-only version.
+
+```shell
+mkdir build && cd build
+cmake ..
+cmake --build . --config Release
+```
+
+## Build with OpenBLAS
+
+```shell
+mkdir build && cd build
+cmake .. -DGGML_OPENBLAS=ON
+cmake --build . --config Release
+```
+
+## Build with CUDA
+
+This provides GPU acceleration using NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_CUDA=ON
+cmake --build . --config Release
+```
+
+## Build with HipBLAS
+
+This provides GPU acceleration using AMD GPU. Make sure to have the ROCm toolkit installed.
+To build for another GPU architecture than installed in your system, set `$GFX_NAME` manually to the desired architecture (replace first command). This is also necessary if your GPU is not officially supported by ROCm, for example you have to set `$GFX_NAME` manually to `gfx1030` for consumer RDNA2 cards.
+
+Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
+
+```shell
+mkdir build && cd build
+if command -v rocminfo; then export GFX_NAME=$(rocminfo | awk '/ *Name: +gfx[1-9]/ {print $2; exit}'); else echo "rocminfo missing!"; fi
+if [ -z "${GFX_NAME}" ]; then echo "Error: Couldn't detect GPU!"; else echo "Building for GPU: ${GFX_NAME}"; fi
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DAMDGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+cmake --build . --config Release
+```
+
+## Build with MUSA
+
+This provides GPU acceleration using Moore Threads GPU. Make sure to have the MUSA toolkit installed.
+
+```shell
+mkdir build && cd build
+cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build . --config Release
+```
+
+## Build with Metal
+
+Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_METAL=ON
+cmake --build . --config Release
+```
+
+## Build with Vulkan
+
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_VULKAN=ON
+cmake --build . --config Release
+```
+
+## Build with OpenCL (for Adreno GPU)
+
+Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
+
+To build for Windows ARM please refers to [Windows 11 Arm64](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
+
+Building for Android:
+
+  Android NDK:
+       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
+
+Setup OpenCL Dependencies for NDK:
+
+You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
+
+*   OpenCL Headers:
+    ```bash
+    # In a temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-Headers
+    cd OpenCL-Headers
+    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
+    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    cd ..
+    ```
+
+*   OpenCL ICD Loader:
+    ```shell
+    # In the same temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+    cd OpenCL-ICD-Loader
+    mkdir build_ndk && cd build_ndk
+
+    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
+    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=24 \
+      -DANDROID_STL=c++_shared
+
+    ninja
+    # Replace <YOUR_NDK_PATH>
+    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    cd ../..
+    ```
+
+Build `stable-diffusion.cpp` for Android with OpenCL:
+
+```shell
+mkdir build-android && cd build-android
+
+# Replace <YOUR_NDK_PATH> with your actual NDK installation path
+# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
+cmake .. -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DGGML_OPENMP=OFF \
+  -DSD_OPENCL=ON
+
+ninja
+```
+*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
+
+## Build with SYCL
+
+Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
+
+```shell
+# Export relevant ENV variables
+source /opt/intel/oneapi/setvars.sh
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Option 2: Use FP16
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+
+cmake --build . --config Release
+```

docs/caching.md

@@ -0,0 +1,126 @@
+## Caching
+
+Caching methods accelerate diffusion inference by reusing intermediate computations when changes between steps are small.
+
+### Cache Modes
+
+| Mode | Target | Description |
+|------|--------|-------------|
+| `ucache` | UNET models | Condition-level caching with error tracking |
+| `easycache` | DiT models | Condition-level cache |
+| `dbcache` | DiT models | Block-level L1 residual threshold |
+| `taylorseer` | DiT models | Taylor series approximation |
+| `cache-dit` | DiT models | Combined DBCache + TaylorSeer |
+
+### UCache (UNET Models)
+
+UCache caches the residual difference (output - input) and reuses it when input changes are below threshold.
+
+```bash
+sd-cli -m model.safetensors -p "a cat" --cache-mode ucache --cache-option "threshold=1.5"
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `threshold` | Error threshold for reuse decision | 1.0 |
+| `start` | Start caching at this percent of steps | 0.15 |
+| `end` | Stop caching at this percent of steps | 0.95 |
+| `decay` | Error decay rate (0-1) | 1.0 |
+| `relative` | Scale threshold by output norm (0/1) | 1 |
+| `reset` | Reset error after computing (0/1) | 1 |
+
+#### Reset Parameter
+
+The `reset` parameter controls error accumulation behavior:
+
+- `reset=1` (default): Resets accumulated error after each computed step. More aggressive caching, works well with most samplers.
+- `reset=0`: Keeps error accumulated. More conservative, recommended for `euler_a` sampler.
+
+### EasyCache (DiT Models)
+
+Condition-level caching for DiT models. Caches and reuses outputs when input changes are below threshold.
+
+```bash
+--cache-mode easycache --cache-option "threshold=0.3"
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `threshold` | Input change threshold for reuse | 0.2 |
+| `start` | Start caching at this percent of steps | 0.15 |
+| `end` | Stop caching at this percent of steps | 0.95 |
+
+### Cache-DIT (DiT Models)
+
+For DiT models like FLUX and QWEN, use block-level caching modes.
+
+#### DBCache
+
+Caches blocks based on L1 residual difference threshold:
+
+```bash
+--cache-mode dbcache --cache-option "threshold=0.25,warmup=4"
+```
+
+#### TaylorSeer
+
+Uses Taylor series approximation to predict block outputs:
+
+```bash
+--cache-mode taylorseer
+```
+
+#### Cache-DIT (Combined)
+
+Combines DBCache and TaylorSeer:
+
+```bash
+--cache-mode cache-dit --cache-preset fast
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `Fn` | Front blocks to always compute | 8 |
+| `Bn` | Back blocks to always compute | 0 |
+| `threshold` | L1 residual difference threshold | 0.08 |
+| `warmup` | Steps before caching starts | 8 |
+
+#### Presets
+
+Available presets: `slow`, `medium`, `fast`, `ultra` (or `s`, `m`, `f`, `u`).
+
+```bash
+--cache-mode cache-dit --cache-preset fast
+```
+
+#### SCM Options
+
+Steps Computation Mask controls which steps can be cached:
+
+```bash
+--scm-mask "1,1,1,1,0,0,1,0,0,0,1,0,0,0,1,0,0,0,1,1"
+```
+
+Mask values: `1` = compute, `0` = can cache.
+
+| Policy | Description |
+|--------|-------------|
+| `dynamic` | Check threshold before caching |
+| `static` | Always cache on cacheable steps |
+
+```bash
+--scm-policy dynamic
+```
+
+### Performance Tips
+
+- Start with default thresholds and adjust based on output quality
+- Lower threshold = better quality, less speedup
+- Higher threshold = more speedup, potential quality loss
+- More steps generally means more caching opportunities

docs/chroma.md

@@ -0,0 +1,33 @@
+# How to Use
+
+You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download Chroma
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF)
+    - Otherwise, download chroma's safetensors from [lodestones/Chroma](https://huggingface.co/lodestones/Chroma)
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert Chroma weights
+
+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
+```
+
+## Run
+
+### Example
+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
+```
+
+![](../assets/flux/chroma_v40.png)
+
+
+

docs/chroma_radiance.md

@@ -0,0 +1,21 @@
+# How to Use
+
+## Download weights
+
+- Download Chroma1-Radiance
+    - safetensors: https://huggingface.co/lodestones/Chroma1-Radiance/tree/main
+    - gguf: https://huggingface.co/silveroxides/Chroma1-Radiance-GGUF/tree/main
+
+- Download t5xxl
+    - safetensors: https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## 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
+```
+
+<img alt="Chroma1-Radiance" src="../assets/flux/chroma1-radiance.png" />
+
+
+

docs/distilled_sd.md

@@ -0,0 +1,99 @@
+# Running distilled models: SSD1B and SDx.x with tiny U-Nets
+
+## Preface 
+
+These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
+Similarly, SD1.x- and SD2.x-style models with a tiny U-Net consist of only 6 U-Net blocks, leading to very small files and time savings of up to 50%. For more information, see the paper: https://arxiv.org/pdf/2305.15798.pdf.
+
+## SSD1B
+
+Note that not all of these models follow the standard parameter naming conventions. However, several useful SSD-1B models are available online, such as:
+
+ * https://huggingface.co/segmind/SSD-1B/resolve/main/SSD-1B-A1111.safetensors
+ * https://huggingface.co/hassenhamdi/SSD-1B-fp8_e4m3fn/resolve/main/SSD-1B_fp8_e4m3fn.safetensors
+
+Useful LoRAs are also available:
+
+ * https://huggingface.co/seungminh/lora-swarovski-SSD-1B/resolve/main/pytorch_lora_weights.safetensors
+ * https://huggingface.co/kylielee505/mylcmlorassd/resolve/main/pytorch_lora_weights.safetensors
+
+These files can be used out-of-the-box, unlike the models described in the next section.
+
+
+## SD1.x, SD2.x with tiny U-Nets
+
+These models require conversion before use. You will need a Python script provided by the diffusers team, available on GitHub:
+
+ * https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/scripts/convert_diffusers_to_original_stable_diffusion.py
+
+### SD2.x
+
+NotaAI provides the following model online:
+
+* https://huggingface.co/nota-ai/bk-sdm-v2-tiny
+
+Creating a .safetensors file involves two steps. First, run this short Python script to download the model from Hugging Face:
+
+```python
+from diffusers import StableDiffusionPipeline
+pipe = StableDiffusionPipeline.from_pretrained("nota-ai/bk-sdm-v2-tiny",cache_dir="./")
+```
+
+Second, create the .safetensors file by running:
+
+```bash
+python convert_diffusers_to_original_stable_diffusion.py \
+      --model_path  models--nota-ai--bk-sdm-v2-tiny/snapshots/68277af553777858cd47e133f92e4db47321bc74 \
+      --checkpoint_path bk-sdm-v2-tiny.safetensors --half --use_safetensors
+```
+
+This will generate the **file bk-sdm-v2-tiny.safetensors**, which is now ready for use with sd.cpp.
+
+### SD1.x
+
+Several Tiny SD 1.x models are available online, such as:
+
+ * https://huggingface.co/segmind/tiny-sd
+ * https://huggingface.co/segmind/portrait-finetuned
+ * https://huggingface.co/nota-ai/bk-sdm-tiny
+
+These models also require conversion, partly because some tensors are stored in a non-contiguous manner. To create a usable checkpoint file, follow these simple steps:
+Download and prepare the model using Python: 
+
+##### Download the model using Python on your computer, for example this way:
+
+```python
+import torch
+from diffusers import StableDiffusionPipeline
+pipe = StableDiffusionPipeline.from_pretrained("segmind/tiny-sd")
+unet=pipe.unet
+for param in unet.parameters():
+    param.data = param.data.contiguous()     # <- important here
+pipe.save_pretrained("segmindtiny-sd", safe_serialization=True)
+```
+
+##### Run the conversion script:
+
+```bash
+python convert_diffusers_to_original_stable_diffusion.py \
+      --model_path  ./segmindtiny-sd \
+      --checkpoint_path ./segmind_tiny-sd.ckpt --half
+```
+
+The file segmind_tiny-sd.ckpt will be generated and is now ready for use with sd.cpp. You can follow a similar process for the other models mentioned above.
+
+
+### Another available .ckpt file:
+
+ * https://huggingface.co/ClashSAN/small-sd/resolve/main/tinySDdistilled.ckpt
+
+To use this file, you must first adjust its non-contiguous tensors:
+
+```python
+import torch
+ckpt = torch.load("tinySDdistilled.ckpt", map_location=torch.device('cpu'))
+for key, value in ckpt['state_dict'].items():
+    if isinstance(value, torch.Tensor):
+        ckpt['state_dict'][key] = value.contiguous()
+torch.save(ckpt, "tinySDdistilled_fixed.ckpt")
+```

docs/docker.md

@@ -0,0 +1,15 @@
+## Docker
+
+### Building using Docker
+
+```shell
+docker build -t sd .
+```
+
+### Run
+
+```shell
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd-cli [args...]
+# For example
+# docker run -v ./models:/models -v ./build:/output sd-cli -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
+```

docs/esrgan.md

@@ -0,0 +1,9 @@
+## Using ESRGAN to upscale results
+
+You can use ESRGAN to upscale the generated images. At the moment, only the [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth) model is supported. Support for more models of this architecture will be added soon.
+
+- Specify the model path using the `--upscale-model PATH` parameter. example:
+
+```bash
+sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
+```

docs/flux.md

@@ -0,0 +1,66 @@
+# How to Use
+
+You can run Flux using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download flux
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf)
+    - Otherwise, download flux-dev from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors or flux-schnell from https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/flux1-schnell.safetensors
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download clip_l from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert flux weights
+
+You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf), this way you don't have to do the conversion yourself.
+
+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
+```
+
+## Run
+
+- `--cfg-scale` is recommended to be set to 1. 
+
+### Flux-dev
+For example:
+
+```
+ .\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| Type | q8_0  | q4_0  | q4_k  | q3_k  | q2_k |
+|---- | ----  |----  |----  |----  |----  |
+| **Memory** | 12068.09 MB  | 6394.53 MB | 6395.17 MB | 4888.16 MB  | 3735.73 MB |
+| **Result** | ![](../assets/flux/flux1-dev-q8_0.png) |![](../assets/flux/flux1-dev-q4_0.png) |![](../assets/flux/flux1-dev-q4_k.png) |![](../assets/flux/flux1-dev-q3_k.png) |![](../assets/flux/flux1-dev-q2_k.png)|
+
+
+
+### Flux-schnell
+
+
+```
+ .\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
+```
+
+| q8_0  |
+| ----  |
+|![](../assets/flux/flux1-schnell-q8_0.png) |
+
+## Run with LoRA
+
+Since many flux LoRA training libraries have used various LoRA naming formats, it is possible that not all flux LoRA naming formats are supported. It is recommended to use LoRA with naming formats compatible with ComfyUI.
+
+### Flux-dev q8_0 with LoRA
+
+- LoRA model from https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main (using comfy converted version!!!)
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
+```
+
+![output](../assets/flux/flux1-dev-q8_0%20with%20lora.png)

docs/flux2.md

@@ -0,0 +1,21 @@
+# How to Use
+
+## Download weights
+
+- Download FLUX.2-dev
+    - gguf: https://huggingface.co/city96/FLUX.2-dev-gguf/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Mistral-Small-3.2-24B-Instruct-2506-GGUF
+    - gguf: https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main
+
+## 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
+```
+
+<img alt="flux2 example" src="../assets/flux2/example.png" />
+
+
+

docs/hipBLAS_on_Windows.md

@@ -82,4 +82,4 @@ cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_H
 cmake --build . --config Release
 ```
 
-If everything went OK, `build\bin\sd.exe` file should appear.
+If everything went OK, `build\bin\sd-cli.exe` file should appear.

docs/kontext.md

@@ -0,0 +1,39 @@
+# How to Use
+
+You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download Kontext
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF)
+    - Otherwise, download FLUX.1-Kontext-dev from https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev/blob/main/flux1-kontext-dev.safetensors
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download clip_l from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert Kontext weights
+
+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
+```
+
+## Run
+
+- `--cfg-scale` is recommended to be set to 1. 
+
+### Example
+For example:
+
+```
+ .\bin\Release\sd-cli.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+```
+
+
+| ref_image | prompt  | output  |
+| ---- | ----  |----  |
+| ![](../assets/flux/flux1-dev-q8_0.png) | change 'flux.cpp' to 'kontext.cpp' |![](../assets/flux/kontext1_dev_output.png) |
+
+
+

docs/lcm.md

@@ -0,0 +1,15 @@
+## LCM/LCM-LoRA
+
+- Download LCM-LoRA form https://huggingface.co/latent-consistency/lcm-lora-sdv1-5
+- Specify LCM-LoRA by adding `<lora:lcm-lora-sdv1-5:1>` to prompt
+- It's advisable to set `--cfg-scale` to `1.0` instead of the default `7.0`. For `--steps`, a range of `2-8` steps is recommended. For `--sampling-method`, `lcm`/`euler_a` is recommended.
+
+Here's a simple example:
+
+```
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
+```
+
+| without LCM-LoRA (--cfg-scale 7)  | with LCM-LoRA (--cfg-scale 1)  |
+| ----  |----    |
+| ![](../assets/without_lcm.png) |![](../assets/with_lcm.png)  |

docs/lora.md

@@ -0,0 +1,26 @@
+## LoRA
+
+- You can specify the directory where the lora weights are stored via `--lora-model-dir`. If not specified, the default is the current working directory.
+
+- LoRA is specified via prompt, just like [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora).
+
+Here's a simple example:
+
+```
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
+```
+
+`../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
+
+# Lora Apply Mode
+
+There are two ways to apply LoRA: **immediately** and **at_runtime**. You can specify it using the `--lora-apply-mode` parameter.
+
+By default, the mode is selected automatically:
+
+* If the model weights contain any quantized parameters, the **at_runtime** mode is used;
+* Otherwise, the **immediately** mode is used.
+
+The **immediately** mode may have precision and compatibility issues with quantized parameters, but it usually offers faster inference speed and, in some cases, lower memory usage.
+In contrast, the **at_runtime** mode provides better compatibility and higher precision, but inference may be slower and memory usage may be higher in some cases.
+

docs/ovis_image.md

@@ -0,0 +1,19 @@
+# How to Use
+
+## Download weights
+
+- Download Ovis-Image-7B
+    - safetensors: https://huggingface.co/Comfy-Org/Ovis-Image/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/leejet/Ovis-Image-7B-GGUF
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
+- Download Ovis 2.5
+    - safetensors: https://huggingface.co/Comfy-Org/Ovis-Image/tree/main/split_files/text_encoders
+
+## 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
+```
+
+<img alt="ovis image example" src="../assets/ovis_image/example.png" />

docs/performance.md

@@ -0,0 +1,26 @@
+## Use Flash Attention to save memory and improve speed.
+
+Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
+eg.:
+ - flux 768x768 ~600mb
+ - SD2 768x768 ~1400mb
+
+For most backends, it slows things down, but for cuda it generally speeds it up too.
+At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
+
+Run by adding `--diffusion-fa` to the arguments and watch for:
+```
+[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
+```
+and the compute buffer shrink in the debug log:
+```
+[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
+```
+
+## Offload weights to the CPU to save VRAM without reducing generation speed.
+
+Using `--offload-to-cpu` allows you to offload weights to the CPU, saving VRAM without reducing generation speed.
+
+## Use quantization to reduce memory usage.
+
+[quantization](./quantization_and_gguf.md)

docs/photo_maker.md

@@ -0,0 +1,53 @@
+## Using PhotoMaker to personalize image generation
+
+You can use [PhotoMaker](https://github.com/TencentARC/PhotoMaker) to personalize generated images with your own ID.
+
+**NOTE**, currently PhotoMaker **ONLY** works with **SDXL** (any SDXL model files will work).
+
+Download PhotoMaker model file (in safetensor format) [here](https://huggingface.co/bssrdf/PhotoMaker). The official release of the model file (in .bin format) does not work with ```stablediffusion.cpp```.
+
+- Specify the PhotoMaker model path using the `--photo-maker PATH` parameter.
+- Specify the input images path using the `--pm-id-images-dir PATH` parameter.
+
+In prompt, make sure you have a class word followed by the trigger word ```"img"``` (hard-coded for now). The class word could be one of ```"man, woman, girl, boy"```. If input ID images contain asian faces, add ```Asian``` before the class
+word.
+
+Another PhotoMaker specific parameter:
+
+- ```--pm-style-strength  (0-100)%```: default is 20 and 10-20 typically gets good results. Lower ratio means more faithfully following input ID (not necessarily better quality).
+
+Other parameters recommended for running Photomaker:
+
+- ```--cfg-scale 5.0```
+- ```-H 1024```
+- ```-W 1024```
+
+If on low memory GPUs (<= 8GB), recommend running with ```--vae-on-cpu``` option to get artifact free images.
+
+Example:
+
+```bash
+bin/sd-cli -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
+```
+
+## PhotoMaker Version 2
+
+[PhotoMaker Version 2 (PMV2)](https://github.com/TencentARC/PhotoMaker/blob/main/README_pmv2.md) has some key improvements. Unfortunately it has a very heavy dependency which makes running it a bit involved in ```SD.cpp```. 
+
+Running PMV2 is now a two-step process:
+
+- Run a python script ```face_detect.py``` to obtain **id_embeds** for the given input images
+```
+python face_detect.py input_image_dir
+```
+An ```id_embeds.bin``` file will be generated in ```input_images_dir```
+
+**Note: this step is only needed to run once; the same ```id_embeds``` can be reused**
+
+- Run the same command as in version 1 but replacing ```photomaker-v1.safetensors``` with ```photomaker-v2.safetensors```.
+
+  You can download ```photomaker-v2.safetensors``` from [here](https://huggingface.co/bssrdf/PhotoMakerV2)
+
+- All the command line parameters from Version 1 remain the same for Version 2 plus one extra pointing to a valid ```id_embeds``` file:  --pm-id-embed-path [path_to__id_embeds.bin] 
+
+

docs/quantization_and_gguf.md

@@ -0,0 +1,27 @@
+## Quantization
+
+You can specify the model weight type using the `--type` parameter. The weights are automatically converted when loading the model.
+
+- `f16` for 16-bit floating-point
+- `f32` for 32-bit floating-point
+- `q8_0` for 8-bit integer quantization
+- `q5_0` or `q5_1` for 5-bit integer quantization
+- `q4_0` or `q4_1` for 4-bit integer quantization
+
+
+### Memory Requirements of Stable Diffusion 1.x
+
+| precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----         | ----  |----  |----  |----  |----  |----  |----  |
+|  **Memory** (txt2img - 512 x 512) | ~2.8G | ~2.3G | ~2.1G | ~2.0G | ~2.0G | ~2.0G | ~2.0G |
+|  **Memory** (txt2img - 512 x 512) *with Flash Attention* | ~2.4G | ~1.9G | ~1.6G | ~1.5G | ~1.5G | ~1.5G | ~1.5G |
+
+## Convert to GGUF
+
+You can also convert weights in the formats `ckpt/safetensors/diffusers` to gguf and perform quantization in advance, avoiding the need for quantization every time you load them.
+
+For example:
+
+```sh
+./bin/sd-cli -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
+```

docs/qwen_image.md

@@ -0,0 +1,23 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/QuantStack/Qwen-Image-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## 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
+```
+
+<img alt="qwen example" src="../assets/qwen/example.png" />
+
+
+

docs/qwen_image_edit.md

@@ -0,0 +1,48 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - Qwen Image Edit
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-GGUF/tree/main
+    - Qwen Image Edit 2509
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/tree/main
+    - Qwen Image Edit 2511
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## Examples
+
+### 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
+```
+
+<img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
+
+
+### 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'"
+```
+
+<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.
+
+```
+.\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
+```
+
+<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2511.png" />

docs/sd.md

@@ -0,0 +1,37 @@
+## Download weights
+
+- download original weights(.ckpt or .safetensors). For example
+    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
+    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
+    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
+    - Stable Diffusion 3 2B from https://huggingface.co/stabilityai/stable-diffusion-3-medium
+
+### txt2img example
+
+```sh
+./bin/sd-cli -m ../models/sd-v1-4.ckpt -p "a lovely cat"
+# ./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+# ./bin/sd-cli -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
+# ./bin/sd-cli -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd-cli --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd-cli -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----  |----  |----  |----  |----  |----  |----  |
+| ![](../assets/f32.png) |![](../assets/f16.png) |![](../assets/q8_0.png) |![](../assets/q5_0.png) |![](../assets/q5_1.png) |![](../assets/q4_0.png) |![](../assets/q4_1.png) |
+
+### img2img example
+
+- `./output.png` is the image generated from the above txt2img pipeline
+
+
+```
+./bin/sd-cli -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+```
+
+<p align="center">
+  <img src="../assets/img2img_output.png" width="256x">
+</p>

docs/sd3.md

@@ -0,0 +1,20 @@
+# How to Use
+
+## Download weights
+
+- Download sd3.5_large from https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/sd3.5_large.safetensors
+- Download clip_g from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_g.safetensors
+- Download clip_l from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp16.safetensors
+
+
+## Run
+
+### SD3.5 Large
+For example:
+
+```
+.\bin\Release\sd-cli.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+```
+
+![](../assets/sd3.5_large.png)