fix: ignore tensors with the particular dim while loading (#233)

* apply pmid lora only once for multiple txt2img calls

* add better support for SDXL LoRA

* fix for some sdxl lora, like lcm-lora-xl

---------

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

.clang-format

@@ -3,11 +3,10 @@ UseTab: Never
 IndentWidth: 4
 TabWidth: 4
 AllowShortIfStatementsOnASingleLine: false
-IndentCaseLabels: false
 ColumnLimit: 0
 AccessModifierOffset: -4
 NamespaceIndentation: All
-FixNamespaceComments: false 
+FixNamespaceComments: false
 AlignAfterOpenBracket: true
 AlignConsecutiveAssignments: true
 IndentCaseLabels: true

.github/workflows/build.yml

@@ -30,7 +30,6 @@ jobs:
         with:
           submodules: recursive
 
-
       - name: Dependencies
         id: depends
         run: |
@@ -42,14 +41,37 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON
           cmake --build . --config Release
 
-      #- name: Test
-        #id: cmake_test
-        #run: |
-          #cd build
-          #ctest --verbose --timeout 900
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
+          echo "OS_NAME=`lsb_release -s -i`" >> "$GITHUB_OUTPUT"
+          echo "OS_VERSION=`lsb_release -s -r`" >> "$GITHUB_OUTPUT"
+          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+          zip -j sd-${{ 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 ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            sd-${{ 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
@@ -63,9 +85,8 @@ jobs:
 
       - name: Dependencies
         id: depends
-        continue-on-error: true
         run: |
-          brew update
+          brew install zip
 
       - name: Build
         id: cmake_build
@@ -73,14 +94,37 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DGGML_AVX2=ON -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" -DSD_BUILD_SHARED_LIBS=ON
           cmake --build . --config Release
 
-      #- name: Test
-        #id: cmake_test
-        #run: |
-          #cd build
-          #ctest --verbose --timeout 900
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
+          echo "OS_NAME=`sw_vers -productName`" >> "$GITHUB_OUTPUT"
+          echo "OS_VERSION=`sw_vers -productVersion`" >> "$GITHUB_OUTPUT"
+          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+          zip -j sd-${{ 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 ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            sd-${{ 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
@@ -89,14 +133,17 @@ jobs:
       matrix:
         include:
           - build: 'noavx'
-            defines: '-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
+            defines: '-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON'
           - build: 'avx2'
-            defines: '-DGGML_AVX2=ON'
+            defines: '-DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON'
           - build: 'avx'
-            defines: '-DGGML_AVX2=OFF'
+            defines: '-DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON'
           - build: 'avx512'
-            defines: '-DGGML_AVX512=ON'
-
+            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'
     steps:
       - name: Clone
         id: checkout
@@ -104,6 +151,29 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Install cuda-toolkit
+        id: cuda-toolkit
+        if: ${{ matrix.build == 'cuda12' }}
+        uses: Jimver/cuda-toolkit@v0.2.11
+        with:
+          cuda: '12.2.0'
+          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 Ninja
+        id: install-ninja
+        if: ${{ matrix.build == 'rocm5.5' }}
+        uses: urkle/action-get-ninja@v1
+        with:
+          version: 1.11.1
+
       - name: Build
         id: cmake_build
         run: |
@@ -125,12 +195,6 @@ jobs:
           & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
           .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
 
-      #- name: Test
-        #id: cmake_test
-        #run: |
-          #cd build
-          #ctest -C Release --verbose --timeout 900
-
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -140,9 +204,37 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          Copy-Item ggml/LICENSE .\build\bin\Release\ggml.txt
-          Copy-Item LICENSE .\build\bin\Release\stable-diffusion.cpp.txt
-          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          $filePath = ".\build\bin\Release\*"
+          if (Test-Path $filePath) {
+            echo "Exists at path $filePath"
+            Copy-Item ggml/LICENSE .\build\bin\Release\ggml.txt
+            Copy-Item LICENSE .\build\bin\Release\stable-diffusion.cpp.txt
+          } elseif (Test-Path ".\build\bin\stable-diffusion.dll") {
+          $filePath = ".\build\bin\*"
+            echo "Exists at path $filePath"
+            Copy-Item ggml/LICENSE .\build\bin\ggml.txt
+            Copy-Item LICENSE .\build\bin\stable-diffusion.cpp.txt
+          } else {
+            ls .\build\bin
+            throw "Can't find stable-diffusion.dll"
+          }
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip $filePath
+
+      - 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' }}
+        run: |
+          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          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
+        with:
+          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' }}

.gitignore

@@ -10,5 +10,4 @@ test/
 *.gguf
 output*.png
 models*
-!taesd-model.gguf
 *.log

.gitmodules

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

CMakeLists.txt

@@ -25,8 +25,11 @@ endif()
 #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_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(BUILD_SHARED_LIBS             "sd: build shared libs" 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_SERVER               "sd: build server example"                           ON)
 
 if(SD_CUBLAS)
@@ -35,21 +38,60 @@ if(SD_CUBLAS)
     add_definitions(-DSD_USE_CUBLAS)
 endif()
 
+if(SD_METAL)
+	message("Use Metal as backend stable-diffusion")
+    set(GGML_METAL ON)
+    add_definitions(-DSD_USE_METAL)
+endif()
+
+if (SD_HIPBLAS)
+    message("Use HIPBLAS as backend stable-diffusion")
+    set(GGML_HIPBLAS ON)
+    add_definitions(-DSD_USE_CUBLAS)
+    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)
 endif()
 
+set(SD_LIB stable-diffusion)
+
+file(GLOB SD_LIB_SOURCES 
+    "*.h"
+    "*.cpp"
+    "*.hpp"
+)
+
+# we can get only one share lib
+if(SD_BUILD_SHARED_LIBS)
+    message("Build shared library")
+    message(${SD_LIB_SOURCES})
+    set(BUILD_SHARED_LIBS OFF)
+    add_library(${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")
+    set(BUILD_SHARED_LIBS OFF)
+    add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
+endif()
+
 
 set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
+
+# see https://github.com/ggerganov/ggml/pull/682
+add_definitions(-DGGML_MAX_NAME=128)
+
 # deps
 add_subdirectory(ggml)
 
 add_subdirectory(thirdparty)
 
-set(SD_LIB stable-diffusion)
-
-add_library(${SD_LIB} stable-diffusion.h stable-diffusion.cpp model.h model.cpp util.h util.cpp)
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC cxx_std_11)

README.md

@@ -10,13 +10,17 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 
 - 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)
 - Super lightweight and without external dependencies
-- SD1.x and SD2.x support
+- 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
+- [PhotoMaker](https://github.com/TencentARC/PhotoMaker) 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 backend for GPU acceleration.
+- 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)
@@ -26,6 +30,9 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - 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`
@@ -48,14 +55,15 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - [ ] More sampling methods
 - [ ] Make inference faster
     - The current implementation of ggml_conv_2d is slow and has high memory usage
-    - Implement Winograd Convolution 2D for 3x3 kernel filtering
 - [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
-- [ ] Implement BPE Tokenizer
-- [ ] Implement [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN/tree/master) upscaler
+- [ ] Implement Inpainting support
 - [ ] k-quants support
 
 ## Usage
 
+For most users, you can download the built executable program from the latest [release](https://github.com/leejet/stable-diffusion.cpp/releases/latest).
+If the built product does not meet your requirements, you can choose to build it manually.
+
 ### Get the Code
 
 ```
@@ -112,7 +120,27 @@ cmake .. -DSD_CUBLAS=ON
 cmake --build . --config Release
 ```
 
-### Using Flash Attention
+##### 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.
 
@@ -124,25 +152,35 @@ cmake --build . --config Release
 ### Run
 
 ```
-usage: sd [arguments]
+usage: ./bin/sd [arguments]
 
 arguments:
   -h, --help                         show this help message and exit
-  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)
+  -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.
+  --stacked-id-embd-dir [DIR]        path to PHOTOMAKER stacked id embeddings.
+  --input-id-images-dir [DIR]        path to PHOTOMAKER input id images dir.
+  --normalize-input                  normalize PHOTOMAKER input id images
+  --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)
+  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%)
+  --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)
@@ -153,6 +191,11 @@ arguments:
   -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
 ```
 
@@ -166,11 +209,22 @@ You can specify the model weight type using the `--type` parameter. The weights
 - `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:
+
+```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
+```
+
 #### txt2img example
 
 ```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
 ```
 
 Using formats of different precisions will yield results of varying quality.
@@ -222,7 +276,7 @@ Here's a simple example:
 | ----  |----    |
 | ![](./assets/without_lcm.png) |![](./assets/with_lcm.png)  |
 
-## Using TAESD to faster decoding
+#### Using TAESD to faster decoding
 
 You can use TAESD to accelerate the decoding of latent images by following these steps:
 
@@ -240,6 +294,49 @@ curl -L -O https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_
 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
+```
+
+#### 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 `--stacked-id-embd-dir PATH` parameter.
+- Specify the input images path using the `--input-id-images-dir PATH` parameter.
+  - input images **must** have the same width and height for preprocessing (to be improved)
+
+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:
+
+- ```--style-ratio  (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 -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --stacked-id-embd-dir ../models/photomaker-v1.safetensors --input-id-images-dir ../assets/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 --style-ratio 10 --vae-on-cpu -o output.png
+```
+
 ### Docker
 
 #### Building using Docker
@@ -263,6 +360,19 @@ docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
 |  **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 |
 
+## Bindings
+
+These projects wrap `stable-diffusion.cpp` for easier use in other languages/frameworks.
+
+* Golang: [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
+* C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
+
+## UIs
+
+These projects use `stable-diffusion.cpp` as a backend for their image generation.
+
+- [Jellybox](https://jellybox.com)
+
 ## Contributors
 
 Thank you to all the people who have already contributed to stable-diffusion.cpp!
@@ -275,5 +385,8 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 - [stable-diffusion](https://github.com/CompVis/stable-diffusion)
 - [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)

common.hpp

@@ -0,0 +1,529 @@
+#ifndef __COMMON_HPP__
+#define __COMMON_HPP__
+
+#include "ggml_extend.hpp"
+
+class DownSampleBlock : public GGMLBlock {
+protected:
+    int channels;
+    int out_channels;
+    bool vae_downsample;
+
+public:
+    DownSampleBlock(int channels,
+                    int out_channels,
+                    bool vae_downsample = false)
+        : channels(channels),
+          out_channels(out_channels),
+          vae_downsample(vae_downsample) {
+        if (vae_downsample) {
+            blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {2, 2}, {0, 0}));
+        } else {
+            blocks["op"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {2, 2}, {1, 1}));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* 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 = conv->forward(ctx, x);
+        } else {
+            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["op"]);
+
+            x = conv->forward(ctx, x);
+        }
+        return x;  // [N, out_channels, h/2, w/2]
+    }
+};
+
+class UpSampleBlock : public GGMLBlock {
+protected:
+    int channels;
+    int out_channels;
+
+public:
+    UpSampleBlock(int channels,
+                  int out_channels)
+        : channels(channels),
+          out_channels(out_channels) {
+        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) {
+        // 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]
+        return x;
+    }
+};
+
+class ResBlock : public GGMLBlock {
+protected:
+    // network hparams
+    int64_t channels;      // model_channels * (1, 1, 1, 2, 2, 4, 4, 4)
+    int64_t emb_channels;  // time_embed_dim
+    int64_t out_channels;  // mult * model_channels
+    std::pair<int, int> kernel_size;
+    int dims;
+    bool skip_t_emb;
+    bool exchange_temb_dims;
+
+    std::shared_ptr<GGMLBlock> conv_nd(int dims,
+                                       int64_t in_channels,
+                                       int64_t out_channels,
+                                       std::pair<int, int> kernel_size,
+                                       std::pair<int, int> padding) {
+        GGML_ASSERT(dims == 2 || dims == 3);
+        if (dims == 3) {
+            return std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(in_channels, out_channels, kernel_size.first, 1, padding.first));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, {1, 1}, padding));
+        }
+    }
+
+public:
+    ResBlock(int64_t channels,
+             int64_t emb_channels,
+             int64_t out_channels,
+             std::pair<int, int> kernel_size = {3, 3},
+             int dims                        = 2,
+             bool exchange_temb_dims         = false,
+             bool skip_t_emb                 = false)
+        : channels(channels),
+          emb_channels(emb_channels),
+          out_channels(out_channels),
+          kernel_size(kernel_size),
+          dims(dims),
+          skip_t_emb(skip_t_emb),
+          exchange_temb_dims(exchange_temb_dims) {
+        std::pair<int, int> padding = {kernel_size.first / 2, kernel_size.second / 2};
+        blocks["in_layers.0"]       = std::shared_ptr<GGMLBlock>(new GroupNorm32(channels));
+        // in_layer_1 is nn.SILU()
+        blocks["in_layers.2"] = conv_nd(dims, channels, out_channels, kernel_size, padding);
+
+        if (!skip_t_emb) {
+            // emb_layer_0 is nn.SILU()
+            blocks["emb_layers.1"] = std::shared_ptr<GGMLBlock>(new Linear(emb_channels, out_channels));
+        }
+
+        blocks["out_layers.0"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(out_channels));
+        // out_layer_1 is nn.SILU()
+        // out_layer_2 is nn.Dropout(), skip for inference
+        blocks["out_layers.3"] = conv_nd(dims, out_channels, out_channels, kernel_size, padding);
+
+        if (out_channels != channels) {
+            blocks["skip_connection"] = conv_nd(dims, channels, out_channels, {1, 1}, {0, 0});
+        }
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = NULL) {
+        // 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]
+        // emb: [N, emb_channels] if dims == 2 else [N, t, emb_channels]
+        auto in_layers_0  = std::dynamic_pointer_cast<GroupNorm32>(blocks["in_layers.0"]);
+        auto in_layers_2  = std::dynamic_pointer_cast<UnaryBlock>(blocks["in_layers.2"]);
+        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) {
+            GGML_ASSERT(skip_t_emb);
+        }
+
+        // in_layers
+        auto h = in_layers_0->forward(ctx, x);
+        h      = ggml_silu_inplace(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);
+            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]
+            } 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]
+                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]
+                }
+            }
+
+            h = ggml_add(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);
+        // dropout, skip for inference
+        h = out_layers_3->forward(ctx, h);
+
+        // skip connection
+        if (out_channels != channels) {
+            auto skip_connection = std::dynamic_pointer_cast<UnaryBlock>(blocks["skip_connection"]);
+            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);
+        return h;  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+    }
+};
+
+class GEGLU : public GGMLBlock {
+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) {}
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // 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 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, ]
+
+        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_gelu_inplace(ctx, gate);
+
+        x = ggml_mul(ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
+
+        return x;
+    }
+};
+
+class FeedForward : public GGMLBlock {
+public:
+    FeedForward(int64_t dim,
+                int64_t dim_out,
+                int64_t mult = 4) {
+        int64_t inner_dim = dim * mult;
+
+        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));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* 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_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
+
+        x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
+        x = net_2->forward(ctx, x);  // [ne3, ne2, ne1, dim_out]
+        return x;
+    }
+};
+
+class CrossAttention : public GGMLBlock {
+protected:
+    int64_t query_dim;
+    int64_t context_dim;
+    int64_t n_head;
+    int64_t d_head;
+
+public:
+    CrossAttention(int64_t query_dim,
+                   int64_t context_dim,
+                   int64_t n_head,
+                   int64_t d_head)
+        : n_head(n_head),
+          d_head(d_head),
+          query_dim(query_dim),
+          context_dim(context_dim) {
+        int64_t inner_dim = d_head * n_head;
+
+        blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, false));
+        blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
+        blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
+
+        blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, query_dim));
+        // 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) {
+        // x: [N, n_token, query_dim]
+        // context: [N, n_context, context_dim]
+        // return: [N, n_token, query_dim]
+        auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+        auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+        auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+        auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+        int64_t n         = x->ne[2];
+        int64_t n_token   = x->ne[1];
+        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 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 = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
+        return x;
+    }
+};
+
+class BasicTransformerBlock : public GGMLBlock {
+protected:
+    int64_t n_head;
+    int64_t d_head;
+    bool ff_in;
+
+public:
+    BasicTransformerBlock(int64_t dim,
+                          int64_t n_head,
+                          int64_t d_head,
+                          int64_t context_dim,
+                          bool ff_in = false)
+        : n_head(n_head), d_head(d_head), ff_in(ff_in) {
+        // disable_self_attn is always False
+        // disable_temporal_crossattention is always False
+        // switch_temporal_ca_to_sa is always False
+        // inner_dim is always None or equal to dim
+        // gated_ff is always True
+        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head));
+        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head));
+        blocks["ff"]    = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
+        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm3"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+
+        if (ff_in) {
+            blocks["norm_in"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+            blocks["ff_in"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* 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]
+
+        auto attn1 = std::dynamic_pointer_cast<CrossAttention>(blocks["attn1"]);
+        auto attn2 = std::dynamic_pointer_cast<CrossAttention>(blocks["attn2"]);
+        auto ff    = std::dynamic_pointer_cast<FeedForward>(blocks["ff"]);
+        auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        auto norm3 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm3"]);
+
+        if (ff_in) {
+            auto norm_in = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_in"]);
+            auto ff_in   = std::dynamic_pointer_cast<FeedForward>(blocks["ff_in"]);
+
+            auto x_skip = x;
+            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);
+        }
+
+        auto r = x;
+        x      = norm1->forward(ctx, x);
+        x      = attn1->forward(ctx, x, x);  // self-attention
+        x      = ggml_add(ctx, x, r);
+        r      = x;
+        x      = norm2->forward(ctx, x);
+        x      = attn2->forward(ctx, x, context);  // cross-attention
+        x      = ggml_add(ctx, x, r);
+        r      = x;
+        x      = norm3->forward(ctx, x);
+        x      = ff->forward(ctx, x);
+        x      = ggml_add(ctx, x, r);
+
+        return x;
+    }
+};
+
+class SpatialTransformer : public GGMLBlock {
+protected:
+    int64_t in_channels;  // mult * model_channels
+    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
+
+public:
+    SpatialTransformer(int64_t in_channels,
+                       int64_t n_head,
+                       int64_t d_head,
+                       int64_t depth,
+                       int64_t context_dim)
+        : 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
+        // 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}));
+
+        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["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) {
+        // 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 x_in         = x;
+        int64_t n         = x->ne[3];
+        int64_t h         = x->ne[1];
+        int64_t w         = x->ne[0];
+        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]
+
+        for (int i = 0; i < depth; i++) {
+            std::string name       = "transformer_blocks." + std::to_string(i);
+            auto transformer_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[name]);
+
+            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]
+
+        // proj_out
+        x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
+
+        x = ggml_add(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);
+    }
+
+    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"]);
+        return sigmoid(alpha);
+    }
+
+public:
+    AlphaBlender() {
+        // merge_strategy is always learned_with_images
+        // for inference, we don't need to set alpha
+        // 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* 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));
+        return x;
+    }
+};
+
+class VideoResBlock : public ResBlock {
+public:
+    VideoResBlock(int channels,
+                  int emb_channels,
+                  int out_channels,
+                  std::pair<int, int> kernel_size = {3, 3},
+                  int64_t video_kernel_size       = 3,
+                  int dims                        = 2)  // always 2
+        : ResBlock(channels, emb_channels, out_channels, kernel_size, dims) {
+        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, emb_channels, out_channels, kernel_size, 3, true));
+        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* emb,
+                                int num_video_frames) {
+        // x: [N, channels, h, w] aka [b*t, channels, h, w]
+        // emb: [N, emb_channels] aka [b*t, emb_channels]
+        // image_only_indicator is always tensor([0.])
+        auto time_stack = std::dynamic_pointer_cast<ResBlock>(blocks["time_stack"]);
+        auto time_mixer = std::dynamic_pointer_cast<AlphaBlender>(blocks["time_mixer"]);
+
+        x = ResBlock::forward(ctx, x, emb);
+
+        int64_t T = num_video_frames;
+        int64_t B = x->ne[3] / T;
+        int64_t C = x->ne[2];
+        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)
+        auto x_mix = x;
+
+        emb = ggml_reshape_4d(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
+
+        return x;
+    }
+};
+
+#endif  // __COMMON_HPP__

control.hpp

@@ -0,0 +1,458 @@
+#ifndef __CONTROL_HPP__
+#define __CONTROL_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+#include "model.h"
+
+#define CONTROL_NET_GRAPH_SIZE 1536
+
+/*
+    =================================== ControlNet ===================================
+    Reference: https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cldm/cldm.py
+
+*/
+class ControlNetBlock : public GGMLBlock {
+protected:
+    SDVersion version = VERSION_1_x;
+    // network hparams
+    int in_channels                        = 4;
+    int out_channels                       = 4;
+    int hint_channels                      = 3;
+    int num_res_blocks                     = 2;
+    std::vector<int> attention_resolutions = {4, 2, 1};
+    std::vector<int> channel_mult          = {1, 2, 4, 4};
+    std::vector<int> transformer_depth     = {1, 1, 1, 1};
+    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
+
+public:
+    int model_channels  = 320;
+    int adm_in_channels = 2816;  // only for VERSION_XL
+
+    ControlNetBlock(SDVersion version = VERSION_1_x)
+        : version(version) {
+        if (version == VERSION_2_x) {
+            context_dim       = 1024;
+            num_head_channels = 64;
+            num_heads         = -1;
+        } else if (version == VERSION_XL) {
+            context_dim           = 2048;
+            attention_resolutions = {4, 2};
+            channel_mult          = {1, 2, 4};
+            transformer_depth     = {1, 2, 10};
+            num_head_channels     = 64;
+            num_heads             = -1;
+        } else if (version == VERSION_SVD) {
+            in_channels       = 8;
+            out_channels      = 4;
+            context_dim       = 1024;
+            adm_in_channels   = 768;
+            num_head_channels = 64;
+            num_heads         = -1;
+        }
+
+        blocks["time_embed.0"] = std::shared_ptr<GGMLBlock>(new Linear(model_channels, time_embed_dim));
+        // 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) {
+            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));
+        }
+
+        // input_blocks
+        blocks["input_blocks.0.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, model_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        std::vector<int> input_block_chans;
+        input_block_chans.push_back(model_channels);
+        int ch              = model_channels;
+        int input_block_idx = 0;
+        int ds              = 1;
+
+        auto get_resblock = [&](int64_t channels, int64_t emb_channels, int64_t out_channels) -> ResBlock* {
+            return new ResBlock(channels, emb_channels, out_channels);
+        };
+
+        auto get_attention_layer = [&](int64_t in_channels,
+                                       int64_t n_head,
+                                       int64_t d_head,
+                                       int64_t depth,
+                                       int64_t context_dim) -> SpatialTransformer* {
+            return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
+        };
+
+        auto make_zero_conv = [&](int64_t channels) {
+            return new Conv2d(channels, channels, {1, 1});
+        };
+
+        blocks["zero_convs.0.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(model_channels));
+
+        blocks["input_hint_block.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(hint_channels, 16, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(16, 16, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(16, 32, {3, 3}, {2, 2}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.6"] = std::shared_ptr<GGMLBlock>(new Conv2d(32, 32, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.8"] = std::shared_ptr<GGMLBlock>(new Conv2d(32, 96, {3, 3}, {2, 2}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.10"] = std::shared_ptr<GGMLBlock>(new Conv2d(96, 96, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.12"] = std::shared_ptr<GGMLBlock>(new Conv2d(96, 256, {3, 3}, {2, 2}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.14"] = std::shared_ptr<GGMLBlock>(new Conv2d(256, model_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        size_t len_mults = channel_mult.size();
+        for (int i = 0; i < len_mults; i++) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, mult * model_channels));
+
+                ch = mult * model_channels;
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    int n_head = num_heads;
+                    int d_head = ch / num_heads;
+                    if (num_head_channels != -1) {
+                        d_head = num_head_channels;
+                        n_head = ch / d_head;
+                    }
+                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
+                                                                                      n_head,
+                                                                                      d_head,
+                                                                                      transformer_depth[i],
+                                                                                      context_dim));
+                }
+                blocks["zero_convs." + std::to_string(input_block_idx) + ".0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
+                input_block_chans.push_back(ch);
+            }
+            if (i != len_mults - 1) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(ch, ch));
+
+                blocks["zero_convs." + std::to_string(input_block_idx) + ".0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
+
+                input_block_chans.push_back(ch);
+                ds *= 2;
+            }
+        }
+
+        // middle blocks
+        int n_head = num_heads;
+        int d_head = ch / num_heads;
+        if (num_head_channels != -1) {
+            d_head = num_head_channels;
+            n_head = ch / d_head;
+        }
+        blocks["middle_block.0"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
+        blocks["middle_block.1"] = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
+                                                                                  n_head,
+                                                                                  d_head,
+                                                                                  transformer_depth[transformer_depth.size() - 1],
+                                                                                  context_dim));
+        blocks["middle_block.2"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
+
+        // middle_block_out
+        blocks["middle_block_out.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
+    }
+
+    struct ggml_tensor* resblock_forward(std::string name,
+                                         struct ggml_context* ctx,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* emb) {
+        auto block = std::dynamic_pointer_cast<ResBlock>(blocks[name]);
+        return block->forward(ctx, x, emb);
+    }
+
+    struct ggml_tensor* attention_layer_forward(std::string name,
+                                                struct ggml_context* 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* hint,
+                                                 struct ggml_tensor* emb,
+                                                 struct ggml_tensor* context) {
+        int num_input_blocks = 15;
+        auto h               = hint;
+        for (int i = 0; i < num_input_blocks; i++) {
+            if (i % 2 == 0) {
+                auto block = std::dynamic_pointer_cast<Conv2d>(blocks["input_hint_block." + std::to_string(i)]);
+
+                h = block->forward(ctx, h);
+            } else {
+                h = ggml_silu_inplace(ctx, h);
+            }
+        }
+        return h;
+    }
+
+    std::vector<struct ggml_tensor*> forward(struct ggml_context* 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) {
+        // 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->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]));
+            }
+        }
+
+        if (y != NULL) {
+            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]));
+            }
+        }
+
+        auto time_embed_0     = std::dynamic_pointer_cast<Linear>(blocks["time_embed.0"]);
+        auto time_embed_2     = std::dynamic_pointer_cast<Linear>(blocks["time_embed.2"]);
+        auto input_blocks_0_0 = std::dynamic_pointer_cast<Conv2d>(blocks["input_blocks.0.0"]);
+        auto zero_convs_0     = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs.0.0"]);
+
+        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 emb = time_embed_0->forward(ctx, t_emb);
+        emb      = ggml_silu_inplace(ctx, emb);
+        emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]
+
+        // SDXL/SVD
+        if (y != NULL) {
+            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      = label_embed_2->forward(ctx, label_emb);  // [N, time_embed_dim]
+
+            emb = ggml_add(ctx, emb, label_emb);  // [N, time_embed_dim]
+        }
+
+        std::vector<struct ggml_tensor*> outs;
+
+        if (guided_hint == NULL) {
+            guided_hint = input_hint_block_forward(ctx, hint, emb, context);
+        }
+        outs.push_back(guided_hint);
+
+        // input_blocks
+
+        // input block 0
+        auto h = input_blocks_0_0->forward(ctx, x);
+        h      = ggml_add(ctx, h, guided_hint);
+        outs.push_back(zero_convs_0->forward(ctx, h));
+
+        // input block 1-11
+        size_t len_mults    = channel_mult.size();
+        int input_block_idx = 0;
+        int ds              = 1;
+        for (int i = 0; i < len_mults; i++) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                h                = resblock_forward(name, ctx, h, emb);  // [N, mult*model_channels, h, w]
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
+                    h                = attention_layer_forward(name, ctx, h, context);  // [N, mult*model_channels, h, w]
+                }
+
+                auto zero_conv = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs." + std::to_string(input_block_idx) + ".0"]);
+
+                outs.push_back(zero_conv->forward(ctx, h));
+            }
+            if (i != len_mults - 1) {
+                ds *= 2;
+                input_block_idx += 1;
+
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                auto block       = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
+
+                h = block->forward(ctx, h);  // [N, mult*model_channels, h/(2^(i+1)), w/(2^(i+1))]
+
+                auto zero_conv = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs." + std::to_string(input_block_idx) + ".0"]);
+
+                outs.push_back(zero_conv->forward(ctx, h));
+            }
+        }
+        // [N, 4*model_channels, h/8, w/8]
+
+        // middle_block
+        h = resblock_forward("middle_block.0", ctx, h, emb);             // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, h, context);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb);             // [N, 4*model_channels, h/8, w/8]
+
+        // out
+        outs.push_back(middle_block_out->forward(ctx, h));
+        return outs;
+    }
+};
+
+struct ControlNet : public GGMLModule {
+    SDVersion version = VERSION_1_x;
+    ControlNetBlock control_net;
+
+    ggml_backend_buffer_t control_buffer = NULL;  // keep control output tensors in backend memory
+    ggml_context* control_ctx            = NULL;
+    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
+    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);
+    }
+
+    ~ControlNet() {
+        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.no_alloc   = true;
+        control_ctx       = ggml_init(params);
+
+        controls.resize(outs.size() - 1);
+
+        size_t control_buffer_size = 0;
+
+        guided_hint = ggml_dup_tensor(control_ctx, outs[0]);
+        control_buffer_size += ggml_nbytes(guided_hint);
+
+        for (int i = 0; i < outs.size() - 1; i++) {
+            controls[i] = ggml_dup_tensor(control_ctx, outs[i + 1]);
+            control_buffer_size += ggml_nbytes(controls[i]);
+        }
+
+        control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, backend);
+
+        LOG_DEBUG("control buffer size %.2fMB", control_buffer_size * 1.f / 1024.f / 1024.f);
+    }
+
+    void free_control_ctx() {
+        if (control_buffer != NULL) {
+            ggml_backend_buffer_free(control_buffer);
+            control_buffer = NULL;
+        }
+        if (control_ctx != NULL) {
+            ggml_free(control_ctx);
+            control_ctx = NULL;
+        }
+        guided_hint        = NULL;
+        guided_hint_cached = false;
+        controls.clear();
+    }
+
+    std::string get_desc() {
+        return "control_net";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        control_net.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                    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);
+
+        x = to_backend(x);
+        if (guided_hint_cached) {
+            hint = NULL;
+        } else {
+            hint = to_backend(hint);
+        }
+        context   = to_backend(context);
+        y         = to_backend(y);
+        timesteps = to_backend(timesteps);
+
+        auto outs = control_net.forward(compute_ctx,
+                                        x,
+                                        hint,
+                                        guided_hint_cached ? guided_hint : NULL,
+                                        timesteps,
+                                        context,
+                                        y);
+
+        if (control_ctx == NULL) {
+            alloc_control_ctx(outs);
+        }
+
+        ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[0], guided_hint));
+        for (int i = 0; i < outs.size() - 1; i++) {
+            ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[i + 1], controls[i]));
+        }
+
+        return gf;
+    }
+
+    void 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) {
+        // x: [N, in_channels, h, w]
+        // timesteps: [N, ]
+        // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
+        // y: [N, adm_in_channels] or [1, adm_in_channels]
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(x, hint, timesteps, context, y);
+        };
+
+        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
+        guided_hint_cached = true;
+    }
+
+    bool load_from_file(const std::string& file_path) {
+        LOG_INFO("loading control net from '%s'", file_path.c_str());
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> tensors;
+        control_net.get_param_tensors(tensors);
+        std::set<std::string> ignore_tensors;
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(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);
+
+        if (!success) {
+            LOG_ERROR("load control net tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("control net model loaded");
+        return success;
+    }
+};
+
+#endif  // __CONTROL_HPP__

denoiser.hpp

@@ -0,0 +1,125 @@
+#ifndef __DENOISER_HPP__
+#define __DENOISER_HPP__
+
+#include "ggml_extend.hpp"
+
+/*================================================= CompVisDenoiser ==================================================*/
+
+// Ref: https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/external.py
+
+#define TIMESTEPS 1000
+
+struct SigmaSchedule {
+    float alphas_cumprod[TIMESTEPS];
+    float sigmas[TIMESTEPS];
+    float log_sigmas[TIMESTEPS];
+
+    virtual std::vector<float> get_sigmas(uint32_t n) = 0;
+
+    float sigma_to_t(float sigma) {
+        float log_sigma = std::log(sigma);
+        std::vector<float> dists;
+        dists.reserve(TIMESTEPS);
+        for (float log_sigma_val : log_sigmas) {
+            dists.push_back(log_sigma - log_sigma_val);
+        }
+
+        int low_idx = 0;
+        for (size_t i = 0; i < TIMESTEPS; i++) {
+            if (dists[i] >= 0) {
+                low_idx++;
+            }
+        }
+        low_idx      = std::min(std::max(low_idx - 1, 0), TIMESTEPS - 2);
+        int high_idx = low_idx + 1;
+
+        float low  = log_sigmas[low_idx];
+        float high = log_sigmas[high_idx];
+        float w    = (low - log_sigma) / (low - high);
+        w          = std::max(0.f, std::min(1.f, w));
+        float t    = (1.0f - w) * low_idx + w * high_idx;
+
+        return t;
+    }
+
+    float t_to_sigma(float t) {
+        int low_idx     = static_cast<int>(std::floor(t));
+        int high_idx    = static_cast<int>(std::ceil(t));
+        float w         = t - static_cast<float>(low_idx);
+        float log_sigma = (1.0f - w) * log_sigmas[low_idx] + w * log_sigmas[high_idx];
+        return std::exp(log_sigma);
+    }
+};
+
+struct DiscreteSchedule : SigmaSchedule {
+    std::vector<float> get_sigmas(uint32_t n) {
+        std::vector<float> result;
+
+        int t_max = TIMESTEPS - 1;
+
+        if (n == 0) {
+            return result;
+        } else if (n == 1) {
+            result.push_back(t_to_sigma((float)t_max));
+            result.push_back(0);
+            return result;
+        }
+
+        float step = static_cast<float>(t_max) / static_cast<float>(n - 1);
+        for (uint32_t i = 0; i < n; ++i) {
+            float t = t_max - step * i;
+            result.push_back(t_to_sigma(t));
+        }
+        result.push_back(0);
+        return result;
+    }
+};
+
+struct KarrasSchedule : SigmaSchedule {
+    std::vector<float> get_sigmas(uint32_t n) {
+        // These *COULD* be function arguments here,
+        // but does anybody ever bother to touch them?
+        float sigma_min = 0.1f;
+        float sigma_max = 10.f;
+        float rho       = 7.f;
+
+        std::vector<float> result(n + 1);
+
+        float min_inv_rho = pow(sigma_min, (1.f / rho));
+        float max_inv_rho = pow(sigma_max, (1.f / rho));
+        for (uint32_t i = 0; i < n; i++) {
+            // Eq. (5) from Karras et al 2022
+            result[i] = pow(max_inv_rho + (float)i / ((float)n - 1.f) * (min_inv_rho - max_inv_rho), rho);
+        }
+        result[n] = 0.;
+        return result;
+    }
+};
+
+struct Denoiser {
+    std::shared_ptr<SigmaSchedule> schedule              = std::make_shared<DiscreteSchedule>();
+    virtual std::vector<float> get_scalings(float sigma) = 0;
+};
+
+struct CompVisDenoiser : public Denoiser {
+    float sigma_data = 1.0f;
+
+    std::vector<float> get_scalings(float sigma) {
+        float c_out = -sigma;
+        float c_in  = 1.0f / std::sqrt(sigma * sigma + sigma_data * sigma_data);
+        return {c_out, c_in};
+    }
+};
+
+struct CompVisVDenoiser : public Denoiser {
+    float sigma_data = 1.0f;
+
+    std::vector<float> get_scalings(float sigma) {
+        float c_skip = sigma_data * sigma_data / (sigma * sigma + sigma_data * sigma_data);
+        float c_out  = -sigma * sigma_data / std::sqrt(sigma * sigma + sigma_data * sigma_data);
+        float c_in   = 1.0f / std::sqrt(sigma * sigma + sigma_data * sigma_data);
+        return {c_skip, c_out, c_in};
+    }
+};
+
+#endif  // __DENOISER_HPP__

docs/hipBLAS_on_Windows.md

@@ -0,0 +1,85 @@
+# Using hipBLAS on Windows
+
+To get hipBLAS in `stable-diffusion.cpp` working on Windows, go through this guide section by section.
+
+## Build Tools for Visual Studio 2022
+
+Skip this step if you already have Build Tools installed.
+
+To install Build Tools, go to [Visual Studio Downloads](https://visualstudio.microsoft.com/vs/), download `Visual Studio 2022 and other Products` and run the installer.
+
+## CMake
+
+Skip this step if you already have CMake installed: running `cmake --version` should output `cmake version x.y.z`.
+
+Download latest `Windows x64 Installer` from [Download | CMake](https://cmake.org/download/) and run it.
+
+## ROCm
+
+Skip this step if you already have Build Tools installed.
+
+The [validation tools](https://rocm.docs.amd.com/en/latest/reference/validation_tools.html) not support on Windows. So you should confirm the Version of `ROCM` by yourself.
+
+Fortunately, `AMD` provides complete help documentation, you can use the help documentation to install [ROCM](https://rocm.docs.amd.com/en/latest/deploy/windows/quick_start.html)
+
+>**If you encounter an error, if it is [AMD ROCm Windows Installation Error 215](https://github.com/RadeonOpenCompute/ROCm/issues/2363), don't worry about this error. ROCM has been installed correctly, but the vs studio plugin installation failed, we can ignore it.**
+
+Then we must set `ROCM` as environment variables before running cmake.
+
+Usually if you install according to the official tutorial and do not modify the ROCM path, then there is a high probability that it is here `C:\Program Files\AMD\ROCm\5.5\bin`
+
+This is what I use to set the clang:
+```Commandline
+set CC=C:\Program Files\AMD\ROCm\5.5\bin\clang.exe
+set CXX=C:\Program Files\AMD\ROCm\5.5\bin\clang++.exe
+```
+
+## Ninja
+
+Skip this step if you already have Ninja installed: running `ninja --version` should output `1.11.1`.
+
+Download latest `ninja-win.zip` from [GitHub Releases Page](https://github.com/ninja-build/ninja/releases/tag/v1.11.1) and unzip. Then set as environment variables. I unzipped it in `C:\Program Files\ninja`, so I set it like this:
+
+```Commandline
+set ninja=C:\Program Files\ninja\ninja.exe
+```
+## Building stable-diffusion.cpp
+
+The thing different from the regular CPU build is `-DSD_HIPBLAS=ON` ,
+`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1100`
+
+>**Notice**: check the `clang` and `clang++` information:
+```Commandline
+clang --version
+clang++ --version
+```
+
+If you see like this, we can continue:
+```
+clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
+Target: x86_64-pc-windows-msvc
+Thread model: posix
+InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+```
+
+```
+clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
+Target: x86_64-pc-windows-msvc
+Thread model: posix
+InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+```
+
+>**Notice** that the `gfx1100` is the GPU architecture of my GPU, you can change it to your GPU architecture. Click here to see your architecture [LLVM Target](https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus)
+
+My GPU is AMD Radeon™ RX 7900 XTX Graphics, so I set it to `gfx1100`.
+
+option:
+
+```commandline
+mkdir build
+cd build
+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
+```
+
+If everything went OK, `build\bin\sd.exe` file should appear.

esrgan.hpp

@@ -0,0 +1,198 @@
+#ifndef __ESRGAN_HPP__
+#define __ESRGAN_HPP__
+
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*
+    ===================================    ESRGAN  ===================================
+    References:
+    https://github.com/xinntao/Real-ESRGAN/blob/master/inference_realesrgan.py
+    https://github.com/XPixelGroup/BasicSR/blob/v1.4.2/basicsr/archs/rrdbnet_arch.py
+
+*/
+
+class ResidualDenseBlock : public GGMLBlock {
+protected:
+    int num_feat;
+    int num_grow_ch;
+
+public:
+    ResidualDenseBlock(int num_feat = 64, int num_grow_ch = 32)
+        : num_feat(num_feat), num_grow_ch(num_grow_ch) {
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv3"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv4"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv5"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 4 * num_grow_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d>(blocks["conv3"]);
+        auto conv4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv4"]);
+        auto conv5 = std::dynamic_pointer_cast<Conv2d>(blocks["conv5"]);
+
+        auto x1    = lrelu(ctx, conv1->forward(ctx, x));
+        auto x_cat = ggml_concat(ctx, x, x1);
+        auto x2    = lrelu(ctx, conv2->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x2);
+        auto x3    = lrelu(ctx, conv3->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x3);
+        auto x4    = lrelu(ctx, conv4->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x4);
+        auto x5    = conv5->forward(ctx, x_cat);
+
+        x5 = ggml_add(ctx, ggml_scale(ctx, x5, 0.2f), x);
+        return x5;
+    }
+};
+
+class RRDB : public GGMLBlock {
+public:
+    RRDB(int num_feat, int num_grow_ch = 32) {
+        blocks["rdb1"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb2"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto rdb1 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb1"]);
+        auto rdb2 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb2"]);
+        auto rdb3 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb3"]);
+
+        auto out = rdb1->forward(ctx, x);
+        out      = rdb2->forward(ctx, out);
+        out      = rdb3->forward(ctx, out);
+
+        out = ggml_add(ctx, ggml_scale(ctx, out, 0.2f), x);
+        return out;
+    }
+};
+
+class RRDBNet : public GGMLBlock {
+protected:
+    int scale       = 4;  // default RealESRGAN_x4plus_anime_6B
+    int num_block   = 6;  // default RealESRGAN_x4plus_anime_6B
+    int num_in_ch   = 3;
+    int num_out_ch  = 3;
+    int num_feat    = 64;  // default RealESRGAN_x4plus_anime_6B
+    int num_grow_ch = 32;  // default RealESRGAN_x4plus_anime_6B
+
+public:
+    RRDBNet() {
+        blocks["conv_first"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_in_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        for (int i = 0; i < num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new RRDB(num_feat, num_grow_ch));
+        }
+        blocks["conv_body"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        // upsample
+        blocks["conv_up1"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_up2"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_hr"]   = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_last"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_out_ch, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_in_ch, h, w]
+        // return: [n, num_out_ch, h*4, w*4]
+        auto conv_first = std::dynamic_pointer_cast<Conv2d>(blocks["conv_first"]);
+        auto conv_body  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_body"]);
+        auto conv_up1   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
+        auto conv_up2   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
+        auto conv_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
+        auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
+
+        auto feat      = conv_first->forward(ctx, x);
+        auto body_feat = feat;
+        for (int i = 0; i < num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
+
+            body_feat = block->forward(ctx, body_feat);
+        }
+        body_feat = conv_body->forward(ctx, body_feat);
+        feat      = ggml_add(ctx, feat, body_feat);
+        // upsample
+        feat     = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2)));
+        feat     = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2)));
+        auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        return out;
+    }
+};
+
+struct ESRGAN : public GGMLModule {
+    RRDBNet rrdb_net;
+    int scale     = 4;
+    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
+
+    ESRGAN(ggml_backend_t backend,
+           ggml_type wtype)
+        : GGMLModule(backend, wtype) {
+        rrdb_net.init(params_ctx, wtype);
+    }
+
+    std::string get_desc() {
+        return "esrgan";
+    }
+
+    bool load_from_file(const std::string& file_path) {
+        LOG_INFO("loading esrgan from '%s'", file_path.c_str());
+
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> esrgan_tensors;
+        rrdb_net.get_param_tensors(esrgan_tensors);
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init esrgan model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool success = model_loader.load_tensors(esrgan_tensors, backend);
+
+        if (!success) {
+            LOG_ERROR("load esrgan tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("esrgan model loaded");
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
+        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        x                       = to_backend(x);
+        struct ggml_tensor* out = rrdb_net.forward(compute_ctx, x);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* x,
+                 ggml_tensor** output,
+                 ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(x);
+        };
+        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+};
+
+#endif  // __ESRGAN_HPP__

examples/cli/main.cpp

@@ -1,21 +1,25 @@
 #include <stdio.h>
-#include <ctime>
+#include <string.h>
+#include <time.h>
+#include <iostream>
 #include <random>
-#include "ggml/ggml.h"
+#include <string>
+#include <vector>
+
+// #include "preprocessing.hpp"
 #include "stable-diffusion.h"
-#include "util.h"
 
 #define STB_IMAGE_IMPLEMENTATION
+#define STB_IMAGE_STATIC
 #include "stb_image.h"
 
 #define STB_IMAGE_WRITE_IMPLEMENTATION
 #define STB_IMAGE_WRITE_STATIC
 #include "stb_image_write.h"
 
-#include <cstring>
-#include <iostream>
-#include <string>
-#include <vector>
+#define STB_IMAGE_RESIZE_IMPLEMENTATION
+#define STB_IMAGE_RESIZE_STATIC
+#include "stb_image_resize.h"
 
 const char* rng_type_to_str[] = {
     "std_default",
@@ -44,11 +48,15 @@ const char* schedule_str[] = {
 const char* modes_str[] = {
     "txt2img",
     "img2img",
+    "img2vid",
+    "convert",
 };
 
 enum SDMode {
     TXT2IMG,
     IMG2IMG,
+    IMG2VID,
+    CONVERT,
     MODE_COUNT
 };
 
@@ -59,25 +67,48 @@ struct SDParams {
     std::string model_path;
     std::string vae_path;
     std::string taesd_path;
-    ggml_type wtype = GGML_TYPE_COUNT;
+    std::string esrgan_path;
+    std::string controlnet_path;
+    std::string embeddings_path;
+    std::string stacked_id_embeddings_path;
+    std::string input_id_images_path;
+    sd_type_t wtype = SD_TYPE_COUNT;
     std::string lora_model_dir;
     std::string output_path = "output.png";
     std::string input_path;
+    std::string control_image_path;
 
     std::string prompt;
     std::string negative_prompt;
-    float cfg_scale = 7.0f;
-    int width       = 512;
-    int height      = 512;
-    int batch_count = 1;
+    float min_cfg     = 1.0f;
+    float cfg_scale   = 7.0f;
+    float style_ratio = 20.f;
+    int clip_skip     = -1;  // <= 0 represents unspecified
+    int width         = 512;
+    int height        = 512;
+    int batch_count   = 1;
 
-    SampleMethod sample_method = EULER_A;
-    Schedule schedule          = DEFAULT;
-    int sample_steps           = 20;
-    float strength             = 0.75f;
-    RNGType rng_type           = CUDA_RNG;
-    int64_t seed               = 42;
-    bool verbose               = false;
+    int video_frames         = 6;
+    int motion_bucket_id     = 127;
+    int fps                  = 6;
+    float augmentation_level = 0.f;
+
+    sample_method_t sample_method = EULER_A;
+    schedule_t schedule           = DEFAULT;
+    int sample_steps              = 20;
+    float strength                = 0.75f;
+    float control_strength        = 0.9f;
+    rng_type_t rng_type           = CUDA_RNG;
+    int64_t seed                  = 42;
+    bool verbose                  = false;
+    bool vae_tiling               = false;
+    bool control_net_cpu          = false;
+    bool normalize_input          = false;
+    bool clip_on_cpu              = false;
+    bool vae_on_cpu               = false;
+    bool canny_preprocess         = false;
+    bool color                    = false;
+    int upscale_repeats           = 1;
 };
 
 void print_params(SDParams params) {
@@ -85,14 +116,28 @@ void print_params(SDParams params) {
     printf("    n_threads:         %d\n", params.n_threads);
     printf("    mode:              %s\n", modes_str[params.mode]);
     printf("    model_path:        %s\n", params.model_path.c_str());
-    printf("    wtype:             %s\n", params.wtype < GGML_TYPE_COUNT ? ggml_type_name(params.wtype) : "unspecified");
+    printf("    wtype:             %s\n", params.wtype < SD_TYPE_COUNT ? sd_type_name(params.wtype) : "unspecified");
     printf("    vae_path:          %s\n", params.vae_path.c_str());
     printf("    taesd_path:        %s\n", params.taesd_path.c_str());
+    printf("    esrgan_path:       %s\n", params.esrgan_path.c_str());
+    printf("    controlnet_path:   %s\n", params.controlnet_path.c_str());
+    printf("    embeddings_path:   %s\n", params.embeddings_path.c_str());
+    printf("    stacked_id_embeddings_path:   %s\n", params.stacked_id_embeddings_path.c_str());
+    printf("    input_id_images_path:   %s\n", params.input_id_images_path.c_str());
+    printf("    style ratio:       %.2f\n", params.style_ratio);
+    printf("    normzalize input image :  %s\n", params.normalize_input ? "true" : "false");
     printf("    output_path:       %s\n", params.output_path.c_str());
     printf("    init_img:          %s\n", params.input_path.c_str());
+    printf("    control_image:     %s\n", params.control_image_path.c_str());
+    printf("    clip on cpu:       %s\n", params.clip_on_cpu ? "true" : "false");
+    printf("    controlnet cpu:    %s\n", params.control_net_cpu ? "true" : "false");
+    printf("    vae decoder on cpu:%s\n", params.vae_on_cpu ? "true" : "false");
+    printf("    strength(control): %.2f\n", params.control_strength);
     printf("    prompt:            %s\n", params.prompt.c_str());
     printf("    negative_prompt:   %s\n", params.negative_prompt.c_str());
+    printf("    min_cfg:           %.2f\n", params.min_cfg);
     printf("    cfg_scale:         %.2f\n", params.cfg_scale);
+    printf("    clip_skip:         %d\n", params.clip_skip);
     printf("    width:             %d\n", params.width);
     printf("    height:            %d\n", params.height);
     printf("    sample_method:     %s\n", sample_method_str[params.sample_method]);
@@ -102,6 +147,8 @@ void print_params(SDParams params) {
     printf("    rng:               %s\n", rng_type_to_str[params.rng_type]);
     printf("    seed:              %ld\n", params.seed);
     printf("    batch_count:       %d\n", params.batch_count);
+    printf("    vae_tiling:        %s\n", params.vae_tiling ? "true" : "false");
+    printf("    upscale_repeats:   %d\n", params.upscale_repeats);
 }
 
 void print_usage(int argc, const char* argv[]) {
@@ -109,21 +156,31 @@ void print_usage(int argc, const char* argv[]) {
     printf("\n");
     printf("arguments:\n");
     printf("  -h, --help                         show this help message and exit\n");
-    printf("  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)\n");
+    printf("  -M, --mode [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)\n");
     printf("  -t, --threads N                    number of threads to use during computation (default: -1).\n");
     printf("                                     If threads <= 0, then threads will be set to the number of CPU physical cores\n");
     printf("  -m, --model [MODEL]                path to model\n");
     printf("  --vae [VAE]                        path to vae\n");
     printf("  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)\n");
+    printf("  --control-net [CONTROL_PATH]       path to control net model\n");
+    printf("  --embd-dir [EMBEDDING_PATH]        path to embeddings.\n");
+    printf("  --stacked-id-embd-dir [DIR]        path to PHOTOMAKER stacked id embeddings.\n");
+    printf("  --input-id-images-dir [DIR]        path to PHOTOMAKER input id images dir.\n");
+    printf("  --normalize-input                  normalize PHOTOMAKER input id images\n");
+    printf("  --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now.\n");
+    printf("  --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)\n");
     printf("  --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0)\n");
     printf("                                     If not specified, the default is the type of the weight file.\n");
     printf("  --lora-model-dir [DIR]             lora model directory\n");
     printf("  -i, --init-img [IMAGE]             path to the input image, required by img2img\n");
+    printf("  --control-image [IMAGE]            path to image condition, control net\n");
     printf("  -o, --output OUTPUT                path to write result image to (default: ./output.png)\n");
     printf("  -p, --prompt [PROMPT]              the prompt to render\n");
     printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
     printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
     printf("  --strength STRENGTH                strength for noising/unnoising (default: 0.75)\n");
+    printf("  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%%)\n");
+    printf("  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)\n");
     printf("                                     1.0 corresponds to full destruction of information in init image\n");
     printf("  -H, --height H                     image height, in pixel space (default: 512)\n");
     printf("  -W, --width W                      image width, in pixel space (default: 512)\n");
@@ -134,6 +191,11 @@ void print_usage(int argc, const char* argv[]) {
     printf("  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)\n");
     printf("  -b, --batch-count COUNT            number of images to generate.\n");
     printf("  --schedule {discrete, karras}      Denoiser sigma schedule (default: discrete)\n");
+    printf("  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)\n");
+    printf("                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x\n");
+    printf("  --vae-tiling                       process vae in tiles to reduce memory usage\n");
+    printf("  --control-net-cpu                  keep controlnet in cpu (for low vram)\n");
+    printf("  --canny                            apply canny preprocessor (edge detection)\n");
     printf("  -v, --verbose                      print extra info\n");
 }
 
@@ -162,7 +224,8 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 }
             }
             if (mode_found == -1) {
-                fprintf(stderr, "error: invalid mode %s, must be one of [txt2img, img2img]\n",
+                fprintf(stderr,
+                        "error: invalid mode %s, must be one of [txt2img, img2img, img2vid, convert]\n",
                         mode_selected);
                 exit(1);
             }
@@ -185,6 +248,36 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.taesd_path = argv[i];
+        } else if (arg == "--control-net") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.controlnet_path = argv[i];
+        } else if (arg == "--upscale-model") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.esrgan_path = argv[i];
+        } else if (arg == "--embd-dir") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.embeddings_path = argv[i];
+        } else if (arg == "--stacked-id-embd-dir") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.stacked_id_embeddings_path = argv[i];
+        } else if (arg == "--input-id-images-dir") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.input_id_images_path = argv[i];
         } else if (arg == "--type") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -192,19 +285,19 @@ void parse_args(int argc, const char** argv, SDParams& params) {
             }
             std::string type = argv[i];
             if (type == "f32") {
-                params.wtype = GGML_TYPE_F32;
+                params.wtype = SD_TYPE_F32;
             } else if (type == "f16") {
-                params.wtype = GGML_TYPE_F16;
+                params.wtype = SD_TYPE_F16;
             } else if (type == "q4_0") {
-                params.wtype = GGML_TYPE_Q4_0;
+                params.wtype = SD_TYPE_Q4_0;
             } else if (type == "q4_1") {
-                params.wtype = GGML_TYPE_Q4_1;
+                params.wtype = SD_TYPE_Q4_1;
             } else if (type == "q5_0") {
-                params.wtype = GGML_TYPE_Q5_0;
+                params.wtype = SD_TYPE_Q5_0;
             } else if (type == "q5_1") {
-                params.wtype = GGML_TYPE_Q5_1;
+                params.wtype = SD_TYPE_Q5_1;
             } else if (type == "q8_0") {
-                params.wtype = GGML_TYPE_Q8_0;
+                params.wtype = SD_TYPE_Q8_0;
             } else {
                 fprintf(stderr, "error: invalid weight format %s, must be one of [f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0]\n",
                         type.c_str());
@@ -222,6 +315,12 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.input_path = argv[i];
+        } else if (arg == "--control-image") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.control_image_path = argv[i];
         } else if (arg == "-o" || arg == "--output") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -234,6 +333,16 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.prompt = argv[i];
+        } else if (arg == "--upscale-repeats") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.upscale_repeats = std::stoi(argv[i]);
+            if (params.upscale_repeats < 1) {
+                fprintf(stderr, "error: upscale multiplier must be at least 1\n");
+                exit(1);
+            }
         } else if (arg == "-n" || arg == "--negative-prompt") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -252,6 +361,18 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.strength = std::stof(argv[i]);
+        } else if (arg == "--style-ratio") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.style_ratio = std::stof(argv[i]);
+        } else if (arg == "--control-strength") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.control_strength = std::stof(argv[i]);
         } else if (arg == "-H" || arg == "--height") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -270,6 +391,24 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.sample_steps = std::stoi(argv[i]);
+        } else if (arg == "--clip-skip") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.clip_skip = std::stoi(argv[i]);
+        } else if (arg == "--vae-tiling") {
+            params.vae_tiling = true;
+        } else if (arg == "--control-net-cpu") {
+            params.control_net_cpu = true;
+        } else if (arg == "--normalize-input") {
+            params.normalize_input = true;
+        } else if (arg == "--clip-on-cpu") {
+            params.clip_on_cpu = true;  // will slow down get_learned_condiotion but necessary for low MEM GPUs
+        } else if (arg == "--vae-on-cpu") {
+            params.vae_on_cpu = true;  // will slow down latent decoding but necessary for low MEM GPUs
+        } else if (arg == "--canny") {
+            params.canny_preprocess = true;
         } else if (arg == "-b" || arg == "--batch-count") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -306,7 +445,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 invalid_arg = true;
                 break;
             }
-            params.schedule = (Schedule)schedule_found;
+            params.schedule = (schedule_t)schedule_found;
         } else if (arg == "-s" || arg == "--seed") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -329,12 +468,14 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 invalid_arg = true;
                 break;
             }
-            params.sample_method = (SampleMethod)sample_method_found;
+            params.sample_method = (sample_method_t)sample_method_found;
         } else if (arg == "-h" || arg == "--help") {
             print_usage(argc, argv);
             exit(0);
         } else if (arg == "-v" || arg == "--verbose") {
             params.verbose = true;
+        } else if (arg == "--color") {
+            params.color = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             print_usage(argc, argv);
@@ -350,7 +491,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         params.n_threads = get_num_physical_cores();
     }
 
-    if (params.prompt.length() == 0) {
+    if (params.mode != CONVERT && params.mode != IMG2VID && params.prompt.length() == 0) {
         fprintf(stderr, "error: the following arguments are required: prompt\n");
         print_usage(argc, argv);
         exit(1);
@@ -362,7 +503,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         exit(1);
     }
 
-    if (params.mode == IMG2IMG && params.input_path.length() == 0) {
+    if ((params.mode == IMG2IMG || params.mode == IMG2VID) && params.input_path.length() == 0) {
         fprintf(stderr, "error: when using the img2img mode, the following arguments are required: init-img\n");
         print_usage(argc, argv);
         exit(1);
@@ -398,6 +539,24 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         srand((int)time(NULL));
         params.seed = rand();
     }
+
+    if (params.mode == CONVERT) {
+        if (params.output_path == "output.png") {
+            params.output_path = "output.gguf";
+        }
+    }
+}
+
+static std::string sd_basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
 }
 
 std::string get_image_params(SDParams params, int64_t seed) {
@@ -409,7 +568,7 @@ std::string get_image_params(SDParams params, int64_t seed) {
     parameter_string += "CFG scale: " + std::to_string(params.cfg_scale) + ", ";
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(params.width) + "x" + std::to_string(params.height) + ", ";
-    parameter_string += "Model: " + basename(params.model_path) + ", ";
+    parameter_string += "Model: " + sd_basename(params.model_path) + ", ";
     parameter_string += "RNG: " + std::string(rng_type_to_str[params.rng_type]) + ", ";
     parameter_string += "Sampler: " + std::string(sample_method_str[params.sample_method]);
     if (params.schedule == KARRAS) {
@@ -420,86 +579,318 @@ std::string get_image_params(SDParams params, int64_t seed) {
     return parameter_string;
 }
 
+/* Enables Printing the log level tag in color using ANSI escape codes */
+void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
+    SDParams* params = (SDParams*)data;
+    int tag_color;
+    const char* level_str;
+    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
+
+    if (!log || (!params->verbose && level <= SD_LOG_DEBUG)) {
+        return;
+    }
+
+    switch (level) {
+        case SD_LOG_DEBUG:
+            tag_color = 37;
+            level_str = "DEBUG";
+            break;
+        case SD_LOG_INFO:
+            tag_color = 34;
+            level_str = "INFO";
+            break;
+        case SD_LOG_WARN:
+            tag_color = 35;
+            level_str = "WARN";
+            break;
+        case SD_LOG_ERROR:
+            tag_color = 31;
+            level_str = "ERROR";
+            break;
+        default: /* Potential future-proofing */
+            tag_color = 33;
+            level_str = "?????";
+            break;
+    }
+
+    if (params->color == true) {
+        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
+    } else {
+        fprintf(out_stream, "[%-5s] ", level_str);
+    }
+    fputs(log, out_stream);
+    fflush(out_stream);
+}
+
 int main(int argc, const char* argv[]) {
     SDParams params;
     parse_args(argc, argv, params);
 
+    sd_set_log_callback(sd_log_cb, (void*)&params);
+
     if (params.verbose) {
         print_params(params);
-        printf("%s", sd_get_system_info().c_str());
-        set_sd_log_level(SDLogLevel::DEBUG);
+        printf("%s", sd_get_system_info());
     }
 
-    bool vae_decode_only        = true;
-    uint8_t* input_image_buffer = NULL;
-    if (params.mode == IMG2IMG) {
+    if (params.mode == CONVERT) {
+        bool success = convert(params.model_path.c_str(), params.vae_path.c_str(), params.output_path.c_str(), params.wtype);
+        if (!success) {
+            fprintf(stderr,
+                    "convert '%s'/'%s' to '%s' failed\n",
+                    params.model_path.c_str(),
+                    params.vae_path.c_str(),
+                    params.output_path.c_str());
+            return 1;
+        } else {
+            printf("convert '%s'/'%s' to '%s' success\n",
+                   params.model_path.c_str(),
+                   params.vae_path.c_str(),
+                   params.output_path.c_str());
+            return 0;
+        }
+    }
+
+    if (params.mode == IMG2VID) {
+        fprintf(stderr, "SVD support is broken, do not use it!!!\n");
+        return 1;
+    }
+
+    bool vae_decode_only          = true;
+    uint8_t* input_image_buffer   = NULL;
+    uint8_t* control_image_buffer = NULL;
+    if (params.mode == IMG2IMG || params.mode == IMG2VID) {
         vae_decode_only = false;
 
         int c              = 0;
-        input_image_buffer = stbi_load(params.input_path.c_str(), &params.width, &params.height, &c, 3);
+        int width          = 0;
+        int height         = 0;
+        input_image_buffer = stbi_load(params.input_path.c_str(), &width, &height, &c, 3);
         if (input_image_buffer == NULL) {
             fprintf(stderr, "load image from '%s' failed\n", params.input_path.c_str());
             return 1;
         }
-        if (c != 3) {
-            fprintf(stderr, "input image must be a 3 channels RGB image, but got %d channels\n", c);
+        if (c < 3) {
+            fprintf(stderr, "the number of channels for the input image must be >= 3, but got %d channels\n", c);
             free(input_image_buffer);
             return 1;
         }
-        if (params.width <= 0 || params.width % 64 != 0) {
-            fprintf(stderr, "error: the width of image must be a multiple of 64\n");
+        if (width <= 0) {
+            fprintf(stderr, "error: the width of image must be greater than 0\n");
             free(input_image_buffer);
             return 1;
         }
-        if (params.height <= 0 || params.height % 64 != 0) {
-            fprintf(stderr, "error: the height of image must be a multiple of 64\n");
+        if (height <= 0) {
+            fprintf(stderr, "error: the height of image must be greater than 0\n");
             free(input_image_buffer);
             return 1;
         }
+
+        // Resize input image ...
+        if (params.height != height || params.width != width) {
+            printf("resize input image from %dx%d to %dx%d\n", width, height, params.width, params.height);
+            int resized_height = params.height;
+            int resized_width  = params.width;
+
+            uint8_t* resized_image_buffer = (uint8_t*)malloc(resized_height * resized_width * 3);
+            if (resized_image_buffer == NULL) {
+                fprintf(stderr, "error: allocate memory for resize input image\n");
+                free(input_image_buffer);
+                return 1;
+            }
+            stbir_resize(input_image_buffer, width, height, 0,
+                         resized_image_buffer, resized_width, resized_height, 0, STBIR_TYPE_UINT8,
+                         3 /*RGB channel*/, STBIR_ALPHA_CHANNEL_NONE, 0,
+                         STBIR_EDGE_CLAMP, STBIR_EDGE_CLAMP,
+                         STBIR_FILTER_BOX, STBIR_FILTER_BOX,
+                         STBIR_COLORSPACE_SRGB, nullptr);
+
+            // Save resized result
+            free(input_image_buffer);
+            input_image_buffer = resized_image_buffer;
+        }
     }
 
-    StableDiffusion sd(params.n_threads, vae_decode_only, params.taesd_path, true, params.lora_model_dir, params.rng_type);
+    sd_ctx_t* sd_ctx = new_sd_ctx(params.model_path.c_str(),
+                                  params.vae_path.c_str(),
+                                  params.taesd_path.c_str(),
+                                  params.controlnet_path.c_str(),
+                                  params.lora_model_dir.c_str(),
+                                  params.embeddings_path.c_str(),
+                                  params.stacked_id_embeddings_path.c_str(),
+                                  vae_decode_only,
+                                  params.vae_tiling,
+                                  true,
+                                  params.n_threads,
+                                  params.wtype,
+                                  params.rng_type,
+                                  params.schedule,
+                                  params.clip_on_cpu,
+                                  params.control_net_cpu,
+                                  params.vae_on_cpu);
 
-    if (!sd.load_from_file(params.model_path, params.vae_path, params.wtype, params.schedule)) {
+    if (sd_ctx == NULL) {
+        printf("new_sd_ctx_t failed\n");
         return 1;
     }
 
-    std::vector<uint8_t*> results;
+    sd_image_t* control_image = NULL;
+    if (params.controlnet_path.size() > 0 && params.control_image_path.size() > 0) {
+        int c                = 0;
+        control_image_buffer = stbi_load(params.control_image_path.c_str(), &params.width, &params.height, &c, 3);
+        if (control_image_buffer == NULL) {
+            fprintf(stderr, "load image from '%s' failed\n", params.control_image_path.c_str());
+            return 1;
+        }
+        control_image = new sd_image_t{(uint32_t)params.width,
+                                       (uint32_t)params.height,
+                                       3,
+                                       control_image_buffer};
+        if (params.canny_preprocess) {  // apply preprocessor
+            control_image->data = preprocess_canny(control_image->data,
+                                                   control_image->width,
+                                                   control_image->height,
+                                                   0.08f,
+                                                   0.08f,
+                                                   0.8f,
+                                                   1.0f,
+                                                   false);
+        }
+    }
+
+    sd_image_t* results;
     if (params.mode == TXT2IMG) {
-        results = sd.txt2img(params.prompt,
-                             params.negative_prompt,
-                             params.cfg_scale,
-                             params.width,
-                             params.height,
-                             params.sample_method,
-                             params.sample_steps,
-                             params.seed,
-                             params.batch_count);
+        results = txt2img(sd_ctx,
+                          params.prompt.c_str(),
+                          params.negative_prompt.c_str(),
+                          params.clip_skip,
+                          params.cfg_scale,
+                          params.width,
+                          params.height,
+                          params.sample_method,
+                          params.sample_steps,
+                          params.seed,
+                          params.batch_count,
+                          control_image,
+                          params.control_strength,
+                          params.style_ratio,
+                          params.normalize_input,
+                          params.input_id_images_path.c_str());
     } else {
-        results = sd.img2img(input_image_buffer,
-                             params.prompt,
-                             params.negative_prompt,
-                             params.cfg_scale,
-                             params.width,
-                             params.height,
-                             params.sample_method,
-                             params.sample_steps,
-                             params.strength,
-                             params.seed);
+        sd_image_t input_image = {(uint32_t)params.width,
+                                  (uint32_t)params.height,
+                                  3,
+                                  input_image_buffer};
+
+        if (params.mode == IMG2VID) {
+            results = img2vid(sd_ctx,
+                              input_image,
+                              params.width,
+                              params.height,
+                              params.video_frames,
+                              params.motion_bucket_id,
+                              params.fps,
+                              params.augmentation_level,
+                              params.min_cfg,
+                              params.cfg_scale,
+                              params.sample_method,
+                              params.sample_steps,
+                              params.strength,
+                              params.seed);
+            if (results == NULL) {
+                printf("generate failed\n");
+                free_sd_ctx(sd_ctx);
+                return 1;
+            }
+            size_t last            = params.output_path.find_last_of(".");
+            std::string dummy_name = last != std::string::npos ? params.output_path.substr(0, last) : params.output_path;
+            for (int i = 0; i < params.video_frames; i++) {
+                if (results[i].data == NULL) {
+                    continue;
+                }
+                std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ".png" : dummy_name + ".png";
+                stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
+                               results[i].data, 0, get_image_params(params, params.seed + i).c_str());
+                printf("save result image to '%s'\n", final_image_path.c_str());
+                free(results[i].data);
+                results[i].data = NULL;
+            }
+            free(results);
+            free_sd_ctx(sd_ctx);
+            return 0;
+        } else {
+            results = img2img(sd_ctx,
+                              input_image,
+                              params.prompt.c_str(),
+                              params.negative_prompt.c_str(),
+                              params.clip_skip,
+                              params.cfg_scale,
+                              params.width,
+                              params.height,
+                              params.sample_method,
+                              params.sample_steps,
+                              params.strength,
+                              params.seed,
+                              params.batch_count,
+                              control_image,
+                              params.control_strength,
+                              params.style_ratio,
+                              params.normalize_input,
+                              params.input_id_images_path.c_str());
+        }
     }
 
-    if (results.size() == 0 || results.size() != params.batch_count) {
-        LOG_ERROR("generate failed");
+    if (results == NULL) {
+        printf("generate failed\n");
+        free_sd_ctx(sd_ctx);
         return 1;
     }
 
+    int upscale_factor = 4;  // unused for RealESRGAN_x4plus_anime_6B.pth
+    if (params.esrgan_path.size() > 0 && params.upscale_repeats > 0) {
+        upscaler_ctx_t* upscaler_ctx = new_upscaler_ctx(params.esrgan_path.c_str(),
+                                                        params.n_threads,
+                                                        params.wtype);
+
+        if (upscaler_ctx == NULL) {
+            printf("new_upscaler_ctx failed\n");
+        } else {
+            for (int i = 0; i < params.batch_count; i++) {
+                if (results[i].data == NULL) {
+                    continue;
+                }
+                sd_image_t current_image = results[i];
+                for (int u = 0; u < params.upscale_repeats; ++u) {
+                    sd_image_t upscaled_image = upscale(upscaler_ctx, current_image, upscale_factor);
+                    if (upscaled_image.data == NULL) {
+                        printf("upscale failed\n");
+                        break;
+                    }
+                    free(current_image.data);
+                    current_image = upscaled_image;
+                }
+                results[i] = current_image;  // Set the final upscaled image as the result
+            }
+        }
+    }
+
     size_t last            = params.output_path.find_last_of(".");
     std::string dummy_name = last != std::string::npos ? params.output_path.substr(0, last) : params.output_path;
     for (int i = 0; i < params.batch_count; i++) {
+        if (results[i].data == NULL) {
+            continue;
+        }
         std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ".png" : dummy_name + ".png";
-        stbi_write_png(final_image_path.c_str(), params.width, params.height, 3, results[i], 0, get_image_params(params, params.seed + i).c_str());
-        LOG_INFO("save result image to '%s'", final_image_path.c_str());
+        stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
+                       results[i].data, 0, get_image_params(params, params.seed + i).c_str());
+        printf("save result image to '%s'\n", final_image_path.c_str());
+        free(results[i].data);
+        results[i].data = NULL;
     }
+    free(results);
+    free_sd_ctx(sd_ctx);
+    free(control_image_buffer);
+    free(input_image_buffer);
 
     return 0;
 }

format-code.sh

@@ -1,2 +1,2 @@
-clang-format -style=file -i *.cpp *.h
+clang-format -style=file -i *.cpp *.h *.hpp
 clang-format -style=file -i examples/cli/*.cpp

lora.hpp

@@ -0,0 +1,189 @@
+#ifndef __LORA_HPP__
+#define __LORA_HPP__
+
+#include "ggml_extend.hpp"
+
+#define LORA_GRAPH_SIZE 10240
+
+struct LoraModel : public GGMLModule {
+    float multiplier = 1.0f;
+    std::map<std::string, struct ggml_tensor*> lora_tensors;
+    std::string file_path;
+    ModelLoader model_loader;
+    bool load_failed = false;
+    bool applied     = false;
+
+    LoraModel(ggml_backend_t backend,
+              ggml_type wtype,
+              const std::string& file_path = "",
+              const std::string& prefix    = "")
+        : file_path(file_path), GGMLModule(backend, wtype) {
+        if (!model_loader.init_from_file(file_path, prefix)) {
+            load_failed = true;
+        }
+    }
+
+    std::string get_desc() {
+        return "lora";
+    }
+
+    bool load_from_file(bool filter_tensor = false) {
+        LOG_INFO("loading LoRA from '%s'", file_path.c_str());
+
+        if (load_failed) {
+            LOG_ERROR("init lora model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool dry_run          = true;
+        auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+            const std::string& name = tensor_storage.name;
+
+            if (filter_tensor && !contains(name, "lora")) {
+                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
+                return true;
+            }
+
+            if (dry_run) {
+                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                           tensor_storage.type,
+                                                           tensor_storage.n_dims,
+                                                           tensor_storage.ne);
+                lora_tensors[name]       = real;
+            } else {
+                auto real   = lora_tensors[name];
+                *dst_tensor = real;
+            }
+
+            return true;
+        };
+
+        model_loader.load_tensors(on_new_tensor_cb, backend);
+        alloc_params_buffer();
+
+        dry_run = false;
+        model_loader.load_tensors(on_new_tensor_cb, backend);
+
+        LOG_DEBUG("finished loaded lora");
+        return true;
+    }
+
+    struct ggml_cgraph* build_lora_graph(std::map<std::string, struct ggml_tensor*> model_tensors) {
+        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, LORA_GRAPH_SIZE, false);
+
+        std::set<std::string> applied_lora_tensors;
+        for (auto it : model_tensors) {
+            std::string k_tensor       = it.first;
+            struct ggml_tensor* weight = model_tensors[it.first];
+
+            size_t k_pos = k_tensor.find(".weight");
+            if (k_pos == std::string::npos) {
+                continue;
+            }
+            k_tensor = k_tensor.substr(0, k_pos);
+            replace_all_chars(k_tensor, '.', '_');
+            // LOG_DEBUG("k_tensor %s", k_tensor.c_str());
+            std::string lora_up_name = "lora." + k_tensor + ".lora_up.weight";
+            if (lora_tensors.find(lora_up_name) == lora_tensors.end()) {
+                if (k_tensor == "model_diffusion_model_output_blocks_2_2_conv") {
+                    // fix for some sdxl lora, like lcm-lora-xl
+                    k_tensor     = "model_diffusion_model_output_blocks_2_1_conv";
+                    lora_up_name = "lora." + k_tensor + ".lora_up.weight";
+                }
+            }
+
+            std::string lora_down_name = "lora." + k_tensor + ".lora_down.weight";
+            std::string alpha_name     = "lora." + k_tensor + ".alpha";
+            std::string scale_name     = "lora." + k_tensor + ".scale";
+
+            ggml_tensor* lora_up   = NULL;
+            ggml_tensor* lora_down = NULL;
+
+            if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
+                lora_up = lora_tensors[lora_up_name];
+            }
+
+            if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
+                lora_down = lora_tensors[lora_down_name];
+            }
+
+            if (lora_up == NULL || lora_down == NULL) {
+                continue;
+            }
+
+            applied_lora_tensors.insert(lora_up_name);
+            applied_lora_tensors.insert(lora_down_name);
+            applied_lora_tensors.insert(alpha_name);
+            applied_lora_tensors.insert(scale_name);
+
+            // calc_cale
+            int64_t dim       = lora_down->ne[ggml_n_dims(lora_down) - 1];
+            float scale_value = 1.0f;
+            if (lora_tensors.find(scale_name) != lora_tensors.end()) {
+                scale_value = ggml_backend_tensor_get_f32(lora_tensors[scale_name]);
+            } else if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                scale_value = alpha / dim;
+            }
+            scale_value *= multiplier;
+
+            // flat lora tensors to multiply it
+            int64_t lora_up_rows   = lora_up->ne[ggml_n_dims(lora_up) - 1];
+            lora_up                = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
+            int64_t lora_down_rows = lora_down->ne[ggml_n_dims(lora_down) - 1];
+            lora_down              = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
+
+            // ggml_mul_mat requires tensor b transposed
+            lora_down                  = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, lora_down));
+            struct ggml_tensor* updown = ggml_mul_mat(compute_ctx, lora_up, lora_down);
+            updown                     = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, updown));
+            updown                     = ggml_reshape(compute_ctx, updown, weight);
+            GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(weight));
+            updown = ggml_scale_inplace(compute_ctx, updown, scale_value);
+            ggml_tensor* final_weight;
+            // if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
+            //     final_weight = ggml_new_tensor(compute_ctx, GGML_TYPE_F32, weight->n_dims, weight->ne);
+            //     final_weight = ggml_cpy_inplace(compute_ctx, weight, final_weight);
+            //     final_weight = ggml_add_inplace(compute_ctx, final_weight, updown);
+            //     final_weight = ggml_cpy_inplace(compute_ctx, final_weight, weight);
+            // } else {
+            //     final_weight = ggml_add_inplace(compute_ctx, weight, updown);
+            // }
+            final_weight = ggml_add_inplace(compute_ctx, weight, updown);  // apply directly
+            ggml_build_forward_expand(gf, final_weight);
+        }
+
+        size_t total_lora_tensors_count   = 0;
+        size_t applied_lora_tensors_count = 0;
+
+        for (auto& kv : lora_tensors) {
+            total_lora_tensors_count++;
+            if (applied_lora_tensors.find(kv.first) == applied_lora_tensors.end()) {
+                LOG_WARN("unused lora tensor %s", kv.first.c_str());
+            } else {
+                applied_lora_tensors_count++;
+            }
+        }
+        /* Don't worry if this message shows up twice in the logs per LoRA,
+         * this function is called once to calculate the required buffer size
+         * and then again to actually generate a graph to be used */
+        if (applied_lora_tensors_count != total_lora_tensors_count) {
+            LOG_WARN("Only (%lu / %lu) LoRA tensors have been applied",
+                     applied_lora_tensors_count, total_lora_tensors_count);
+        } else {
+            LOG_DEBUG("(%lu / %lu) LoRA tensors applied successfully",
+                      applied_lora_tensors_count, total_lora_tensors_count);
+        }
+
+        return gf;
+    }
+
+    void apply(std::map<std::string, struct ggml_tensor*> model_tensors, int n_threads) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_lora_graph(model_tensors);
+        };
+        GGMLModule::compute(get_graph, n_threads, true);
+    }
+};
+
+#endif  // __LORA_HPP__

model.cpp

@@ -1,6 +1,7 @@
 #include <stdarg.h>
 #include <fstream>
 #include <regex>
+#include <set>
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -14,6 +15,12 @@
 #include "ggml/ggml-backend.h"
 #include "ggml/ggml.h"
 
+#include "stable-diffusion.h"
+
+#ifdef SD_USE_METAL
+#include "ggml-metal.h"
+#endif
+
 #define ST_HEADER_SIZE_LEN 8
 
 uint64_t read_u64(uint8_t* buffer) {
@@ -54,7 +61,6 @@ std::string self_attn_names[] = {
     "self_attn.q_proj.weight",
     "self_attn.k_proj.weight",
     "self_attn.v_proj.weight",
-
     "self_attn.q_proj.bias",
     "self_attn.k_proj.bias",
     "self_attn.v_proj.bias",
@@ -75,13 +81,16 @@ const char* unused_tensors[] = {
     "cond_stage_model.transformer.text_model.embeddings.position_ids",
     "cond_stage_model.model.logit_scale",
     "cond_stage_model.model.text_projection",
+    "conditioner.embedders.0.transformer.text_model.embeddings.position_ids",
+    "conditioner.embedders.0.model.logit_scale",
+    "conditioner.embedders.1.model.logit_scale",
     "model.diffusion_model.time_embedding.cond_proj.weight",
     "unet.time_embedding.cond_proj.weight",
     "model_ema.decay",
     "model_ema.num_updates",
     "model_ema.diffusion_model",
-    "control_model",
     "embedding_manager",
+    "denoiser.sigmas",
 };
 
 bool is_unused_tensor(std::string name) {
@@ -94,11 +103,19 @@ bool is_unused_tensor(std::string name) {
 }
 
 std::unordered_map<std::string, std::string> open_clip_to_hf_clip_model = {
-    {"cond_stage_model.model.ln_final.bias", "cond_stage_model.transformer.text_model.final_layer_norm.bias"},
-    {"cond_stage_model.model.ln_final.weight", "cond_stage_model.transformer.text_model.final_layer_norm.weight"},
-    {"cond_stage_model.model.positional_embedding", "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight"},
-    {"cond_stage_model.model.token_embedding.weight", "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight"},
-
+    {"model.ln_final.bias", "transformer.text_model.final_layer_norm.bias"},
+    {"model.ln_final.weight", "transformer.text_model.final_layer_norm.weight"},
+    {"model.positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"},
+    {"model.token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"},
+    {"model.text_projection", "transformer.text_model.text_projection"},
+    {"model.visual.class_embedding", "transformer.vision_model.embeddings.class_embedding"},
+    {"model.visual.conv1.weight", "transformer.vision_model.embeddings.patch_embedding.weight"},
+    {"model.visual.ln_post.bias", "transformer.vision_model.post_layernorm.bias"},
+    {"model.visual.ln_post.weight", "transformer.vision_model.post_layernorm.weight"},
+    {"model.visual.ln_pre.bias", "transformer.vision_model.pre_layernorm.bias"},
+    {"model.visual.ln_pre.weight", "transformer.vision_model.pre_layernorm.weight"},
+    {"model.visual.positional_embedding", "transformer.vision_model.embeddings.position_embedding.weight"},
+    {"model.visual.proj", "transformer.visual_projection.weight"},
 };
 
 std::unordered_map<std::string, std::string> open_clip_to_hk_clip_resblock = {
@@ -126,28 +143,58 @@ std::unordered_map<std::string, std::string> vae_decoder_name_map = {
 };
 
 std::string convert_open_clip_to_hf_clip(const std::string& name) {
-    std::string new_name                  = name;
-    std::string open_clip_resblock_prefix = "cond_stage_model.model.transformer.resblocks.";
-    std::string hf_clip_resblock_prefix   = "cond_stage_model.transformer.text_model.encoder.layers.";
-
-    if (open_clip_to_hf_clip_model.find(name) != open_clip_to_hf_clip_model.end()) {
-        new_name = open_clip_to_hf_clip_model[name];
+    std::string new_name = name;
+    std::string prefix;
+    if (starts_with(new_name, "conditioner.embedders.0.open_clip.")) {
+        prefix   = "cond_stage_model.";
+        new_name = new_name.substr(strlen("conditioner.embedders.0.open_clip."));
+    } else if (starts_with(new_name, "conditioner.embedders.0.")) {
+        prefix   = "cond_stage_model.";
+        new_name = new_name.substr(strlen("conditioner.embedders.0."));
+    } else if (starts_with(new_name, "conditioner.embedders.1.")) {
+        prefix   = "cond_stage_model.1.";
+        new_name = new_name.substr(strlen("conditioner.embedders.0."));
+    } else if (starts_with(new_name, "cond_stage_model.")) {
+        prefix   = "cond_stage_model.";
+        new_name = new_name.substr(strlen("cond_stage_model."));
+    } else if (ends_with(new_name, "vision_model.visual_projection.weight")) {
+        prefix   = new_name.substr(0, new_name.size() - strlen("vision_model.visual_projection.weight"));
+        new_name = prefix + "visual_projection.weight";
+        return new_name;
+    } else {
+        return new_name;
     }
 
-    if (name.find(open_clip_resblock_prefix) == 0) {
-        std::string remain = name.substr(open_clip_resblock_prefix.length());
-        std::string idx    = remain.substr(0, remain.find("."));
-        std::string suffix = remain.substr(idx.length() + 1);
+    if (open_clip_to_hf_clip_model.find(new_name) != open_clip_to_hf_clip_model.end()) {
+        new_name = open_clip_to_hf_clip_model[new_name];
+    }
 
-        if (suffix == "attn.in_proj_weight" || suffix == "attn.in_proj_bias") {
-            new_name = hf_clip_resblock_prefix + idx + "." + suffix;
-        } else if (open_clip_to_hk_clip_resblock.find(suffix) != open_clip_to_hk_clip_resblock.end()) {
-            std::string new_suffix = open_clip_to_hk_clip_resblock[suffix];
-            new_name               = hf_clip_resblock_prefix + idx + "." + new_suffix;
+    std::string open_clip_resblock_prefix = "model.transformer.resblocks.";
+    std::string hf_clip_resblock_prefix   = "transformer.text_model.encoder.layers.";
+
+    auto replace_suffix = [&]() {
+        if (new_name.find(open_clip_resblock_prefix) == 0) {
+            std::string remain = new_name.substr(open_clip_resblock_prefix.length());
+            std::string idx    = remain.substr(0, remain.find("."));
+            std::string suffix = remain.substr(idx.length() + 1);
+
+            if (suffix == "attn.in_proj_weight" || suffix == "attn.in_proj_bias") {
+                new_name = hf_clip_resblock_prefix + idx + "." + suffix;
+            } else if (open_clip_to_hk_clip_resblock.find(suffix) != open_clip_to_hk_clip_resblock.end()) {
+                std::string new_suffix = open_clip_to_hk_clip_resblock[suffix];
+                new_name               = hf_clip_resblock_prefix + idx + "." + new_suffix;
+            }
         }
-    }
+    };
 
-    return new_name;
+    replace_suffix();
+
+    open_clip_resblock_prefix = "model.visual.transformer.resblocks.";
+    hf_clip_resblock_prefix   = "transformer.vision_model.encoder.layers.";
+
+    replace_suffix();
+
+    return prefix + new_name;
 }
 
 std::string convert_vae_decoder_name(const std::string& name) {
@@ -157,6 +204,25 @@ std::string convert_vae_decoder_name(const std::string& name) {
     return name;
 }
 
+/* If not a SDXL LoRA the unet" prefix will have already been replaced by this
+ * point and "te2" and "te1" don't seem to appear in non-SDXL only "te_" */
+std::string convert_sdxl_lora_name(std::string tensor_name) {
+    const std::pair<std::string, std::string> sdxl_lora_name_lookup[] = {
+        {"unet", "model_diffusion_model"},
+        {"te2", "cond_stage_model_1_transformer"},
+        {"te1", "cond_stage_model_transformer"},
+        {"text_encoder_2", "cond_stage_model_1_transformer"},
+        {"text_encoder", "cond_stage_model_transformer"},
+    };
+    for (auto& pair_i : sdxl_lora_name_lookup) {
+        if (tensor_name.compare(0, pair_i.first.length(), pair_i.first) == 0) {
+            tensor_name = std::regex_replace(tensor_name, std::regex(pair_i.first), pair_i.second);
+            break;
+        }
+    }
+    return tensor_name;
+}
+
 std::unordered_map<std::string, std::unordered_map<std::string, std::string>> suffix_conversion_underline = {
     {
         "attentions",
@@ -205,7 +271,7 @@ std::unordered_map<std::string, std::unordered_map<std::string, std::string>> su
     },
 };
 
-std::string convert_diffusers_name_to_compvis(const std::string& key, char seq) {
+std::string convert_diffusers_name_to_compvis(std::string key, char seq) {
     std::vector<std::string> m;
 
     auto match = [](std::vector<std::string>& match_list, const std::regex& regex, const std::string& key) {
@@ -239,6 +305,11 @@ std::string convert_diffusers_name_to_compvis(const std::string& key, char seq)
         return inner_key;
     };
 
+    // convert attn to out
+    if (ends_with(key, "to_out")) {
+        key += format("%c0", seq);
+    }
+
     // unet
     if (match(m, std::regex(format("unet%cconv_in(.*)", seq)), key)) {
         return format("model%cdiffusion_model%cinput_blocks%c0%c0", seq, seq, seq, seq) + m[0];
@@ -348,18 +419,27 @@ std::string convert_diffusers_name_to_compvis(const std::string& key, char seq)
 }
 
 std::string convert_tensor_name(const std::string& name) {
-    std::string new_name;
-    if (starts_with(name, "cond_stage_model.model")) {
+    std::string new_name = name;
+    if (starts_with(name, "cond_stage_model.") || starts_with(name, "conditioner.embedders.") || ends_with(name, ".vision_model.visual_projection.weight")) {
         new_name = convert_open_clip_to_hf_clip(name);
     } else if (starts_with(name, "first_stage_model.decoder")) {
         new_name = convert_vae_decoder_name(name);
+    } else if (starts_with(name, "control_model.")) {  // for controlnet pth models
+        size_t pos = name.find('.');
+        if (pos != std::string::npos) {
+            new_name = name.substr(pos + 1);
+        }
     } else if (starts_with(name, "lora_")) {  // for lora
         size_t pos = name.find('.');
         if (pos != std::string::npos) {
             std::string name_without_network_parts = name.substr(5, pos - 5);
             std::string network_part               = name.substr(pos + 1);
+
             // LOG_DEBUG("%s %s", name_without_network_parts.c_str(), network_part.c_str());
             std::string new_key = convert_diffusers_name_to_compvis(name_without_network_parts, '_');
+            /* For dealing with the new SDXL LoRA tensor naming convention */
+            new_key = convert_sdxl_lora_name(new_key);
+
             if (new_key.empty()) {
                 new_name = name;
             } else {
@@ -368,6 +448,33 @@ std::string convert_tensor_name(const std::string& name) {
         } else {
             new_name = name;
         }
+    } else if (contains(name, "lora_up") || contains(name, "lora_down") ||
+               contains(name, "lora.up") || contains(name, "lora.down") ||
+               contains(name, "lora_linear")) {
+        size_t pos = new_name.find(".processor");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, strlen(".processor"), "");
+        }
+        pos = new_name.rfind("lora");
+        if (pos != std::string::npos) {
+            std::string name_without_network_parts = new_name.substr(0, pos - 1);
+            std::string network_part               = new_name.substr(pos);
+            // LOG_DEBUG("%s %s", name_without_network_parts.c_str(), network_part.c_str());
+            std::string new_key = convert_diffusers_name_to_compvis(name_without_network_parts, '.');
+            new_key             = convert_sdxl_lora_name(new_key);
+            replace_all_chars(new_key, '.', '_');
+            size_t npos = network_part.rfind("_linear_layer");
+            if (npos != std::string::npos) {
+                network_part.replace(npos, strlen("_linear_layer"), "");
+            }
+            if (starts_with(network_part, "lora.")) {
+                network_part = "lora_" + network_part.substr(5);
+            }
+            if (new_key.size() > 0) {
+                new_name = "lora." + new_key + "." + network_part;
+            }
+            // LOG_DEBUG("new name: %s", new_name.c_str());
+        }
     } else if (starts_with(name, "unet") || starts_with(name, "vae") || starts_with(name, "te")) {  // for diffuser
         size_t pos = name.find_last_of('.');
         if (pos != std::string::npos) {
@@ -410,7 +517,7 @@ void preprocess_tensor(TensorStorage tensor_storage,
 
     tensor_storage.name = new_name;
 
-    if (starts_with(new_name, "cond_stage_model.transformer.text_model.encoder.layers.") &&
+    if (new_name.find("cond_stage_model") != std::string::npos &&
         ends_with(new_name, "attn.in_proj_weight")) {
         size_t prefix_size = new_name.find("attn.in_proj_weight");
         std::string prefix = new_name.substr(0, prefix_size);
@@ -422,7 +529,7 @@ void preprocess_tensor(TensorStorage tensor_storage,
 
         processed_tensor_storages.insert(processed_tensor_storages.end(), chunks.begin(), chunks.end());
 
-    } else if (starts_with(new_name, "cond_stage_model.transformer.text_model.encoder.layers.") &&
+    } else if (new_name.find("cond_stage_model") != std::string::npos &&
                ends_with(new_name, "attn.in_proj_bias")) {
         size_t prefix_size = new_name.find("attn.in_proj_bias");
         std::string prefix = new_name.substr(0, prefix_size);
@@ -450,7 +557,13 @@ void bf16_to_f32_vec(uint16_t* src, float* dst, int64_t n) {
     }
 }
 
-void convert_tensor(void* src, ggml_type src_type, void* dst, ggml_type dst_type, int n) {
+void convert_tensor(void* src,
+                    ggml_type src_type,
+                    void* dst,
+                    ggml_type dst_type,
+                    int nrows,
+                    int n_per_row) {
+    int n = nrows * n_per_row;
     if (src_type == dst_type) {
         size_t nbytes = n * ggml_type_size(src_type) / ggml_blck_size(src_type);
         memcpy(((char*)dst), ((char*)src), nbytes);
@@ -459,7 +572,9 @@ void convert_tensor(void* src, ggml_type src_type, void* dst, ggml_type dst_type
             ggml_fp32_to_fp16_row((float*)src, (ggml_fp16_t*)dst, n);
         } else {
             int64_t hist[16];
-            ggml_quantize_chunk(dst_type, (float*)src, dst, 0, n, hist);
+            std::vector<float> imatrix(n_per_row, 1.0f);  // dummy importance matrix
+            const float* im = imatrix.data();
+            ggml_quantize_chunk(dst_type, (float*)src, dst, 0, nrows, n_per_row, hist, im);
         }
     } else if (dst_type == GGML_TYPE_F32) {
         if (src_type == GGML_TYPE_F16) {
@@ -488,7 +603,9 @@ void convert_tensor(void* src, ggml_type src_type, void* dst, ggml_type dst_type
             ggml_fp32_to_fp16_row((float*)src_data_f32, (ggml_fp16_t*)dst, n);
         } else {
             int64_t hist[16];
-            ggml_quantize_chunk(dst_type, (float*)src_data_f32, dst, 0, n, hist);
+            std::vector<float> imatrix(n_per_row, 1.0f);  // dummy importance matrix
+            const float* im = imatrix.data();
+            ggml_quantize_chunk(dst_type, (float*)src_data_f32, dst, 0, nrows, n_per_row, hist, im);
         }
     }
 }
@@ -588,7 +705,7 @@ bool is_safetensors_file(const std::string& file_path) {
     }
 
     size_t header_size_ = read_u64(header_size_buf);
-    if (header_size_ >= file_size_) {
+    if (header_size_ >= file_size_ || header_size_ <= 2) {
         return false;
     }
 
@@ -652,7 +769,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
 
         // LOG_DEBUG("%s", name.c_str());
 
-        TensorStorage tensor_storage(prefix + name, dummy->type, dummy->ne, dummy->n_dims, file_index, offset);
+        TensorStorage tensor_storage(prefix + name, dummy->type, dummy->ne, ggml_n_dims(dummy), file_index, offset);
 
         GGML_ASSERT(ggml_nbytes(dummy) == tensor_storage.nbytes());
 
@@ -751,19 +868,32 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
             return false;
         }
 
-        if (shape.size() > 4) {
+        if (shape.size() > SD_MAX_DIMS) {
             LOG_ERROR("invalid tensor '%s'", name.c_str());
             return false;
         }
 
-        int n_dims    = (int)shape.size();
-        int64_t ne[4] = {1, 1, 1, 1};
+        int n_dims              = (int)shape.size();
+        int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
         for (int i = 0; i < n_dims; i++) {
             ne[i] = shape[i].get<int64_t>();
         }
 
-        TensorStorage tensor_storage(prefix + name, type, ne, n_dims, file_index, ST_HEADER_SIZE_LEN + header_size_ + begin);
+        if (n_dims == 5) {
+            if (ne[3] == 1 && ne[4] == 1) {
+                n_dims = 4;
+            } else {
+                LOG_ERROR("invalid tensor '%s'", name.c_str());
+                return false;
+            }
+        }
 
+        // ggml/src/ggml.c:2745
+        if (n_dims < 1 || n_dims > GGML_MAX_DIMS) {
+            continue;
+        }
+
+        TensorStorage tensor_storage(prefix + name, type, ne, n_dims, file_index, ST_HEADER_SIZE_LEN + header_size_ + begin);
         tensor_storage.reverse_ne();
 
         size_t tensor_data_size = end - begin;
@@ -776,6 +906,8 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
         }
 
         tensor_storages.push_back(tensor_storage);
+
+        // LOG_DEBUG("%s %s", tensor_storage.to_string().c_str(), dtype.c_str());
     }
 
     return true;
@@ -919,7 +1051,7 @@ struct PickleTensorReader {
                 phase = READ_NAME;
             }
         } else if (phase == READ_DIMENS) {
-            if (tensor_storage.n_dims + 1 > 4) {  // too many dimens
+            if (tensor_storage.n_dims + 1 > SD_MAX_DIMS) {  // too many dimens
                 phase                 = READ_NAME;
                 tensor_storage.n_dims = 0;
             }
@@ -1100,7 +1232,9 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
                     if (reader.phase == PickleTensorReader::READ_DIMENS) {
                         reader.tensor_storage.reverse_ne();
                         reader.tensor_storage.file_index = file_index;
-                        reader.tensor_storage.name       = prefix + reader.tensor_storage.name;
+                        // if(strcmp(prefix.c_str(), "scarlett") == 0)
+                        // printf(" got tensor %s \n ", reader.tensor_storage.name.c_str());
+                        reader.tensor_storage.name = prefix + reader.tensor_storage.name;
                         tensor_storages.push_back(reader.tensor_storage);
                         // LOG_DEBUG("%s", reader.tensor_storage.name.c_str());
                         // reset
@@ -1156,12 +1290,24 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
 SDVersion ModelLoader::get_sd_version() {
     TensorStorage token_embedding_weight;
     for (auto& tensor_storage : tensor_storages) {
+        if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos) {
+            return VERSION_XL;
+        }
+        if (tensor_storage.name.find("cond_stage_model.1") != std::string::npos) {
+            return VERSION_XL;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.input_blocks.8.0.time_mixer.mix_factor") != std::string::npos) {
+            return VERSION_SVD;
+        }
+
         if (tensor_storage.name == "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight" ||
             tensor_storage.name == "cond_stage_model.model.token_embedding.weight" ||
             tensor_storage.name == "text_model.embeddings.token_embedding.weight" ||
-            tensor_storage.name == "te.text_model.embeddings.token_embedding.weight") {
+            tensor_storage.name == "te.text_model.embeddings.token_embedding.weight" ||
+            tensor_storage.name == "conditioner.embedders.0.model.token_embedding.weight" ||
+            tensor_storage.name == "conditioner.embedders.0.transformer.text_model.embeddings.token_embedding.weight") {
             token_embedding_weight = tensor_storage;
-            break;
+            // break;
         }
     }
     if (token_embedding_weight.ne[0] == 768) {
@@ -1186,23 +1332,46 @@ ggml_type ModelLoader::get_sd_wtype() {
     return GGML_TYPE_COUNT;
 }
 
-bool ModelLoader::load_vocab(on_new_token_cb_t on_new_token_cb) {
-    char* vocab_buffer          = reinterpret_cast<char*>(vocab_json);
-    nlohmann::json vocab        = nlohmann::json::parse(vocab_buffer);
-    std::map<char, int> decoder = unicode_to_byte();
-    for (auto& it : vocab.items()) {
-        int token_id          = it.value();
-        std::string token_str = it.key();
-        std::string token     = "";
-        for (char c : token_str) {
-            token += decoder[c];
-        }
-        on_new_token_cb(token, token_id);
-    }
-    return true;
+std::string ModelLoader::load_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(merges_utf8_c_str), sizeof(merges_utf8_c_str));
+    return merges_utf8_str;
 }
 
-bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
+std::vector<TensorStorage> remove_duplicates(const std::vector<TensorStorage>& vec) {
+    std::vector<TensorStorage> res;
+    std::unordered_map<std::string, size_t> name_to_index_map;
+
+    for (size_t i = 0; i < vec.size(); ++i) {
+        const std::string& current_name = vec[i].name;
+        auto it                         = name_to_index_map.find(current_name);
+
+        if (it != name_to_index_map.end()) {
+            res[it->second] = vec[i];
+        } else {
+            name_to_index_map[current_name] = i;
+            res.push_back(vec[i]);
+        }
+    }
+
+    // vec.resize(name_to_index_map.size());
+
+    return res;
+}
+
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend) {
+    std::vector<TensorStorage> processed_tensor_storages;
+    for (auto& tensor_storage : tensor_storages) {
+        // LOG_DEBUG("%s", name.c_str());
+
+        if (is_unused_tensor(tensor_storage.name)) {
+            continue;
+        }
+
+        preprocess_tensor(tensor_storage, processed_tensor_storages);
+    }
+    std::vector<TensorStorage> dedup = remove_duplicates(processed_tensor_storages);
+    processed_tensor_storages        = dedup;
+
     bool success = true;
     for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
         std::string file_path = file_paths_[file_index];
@@ -1260,24 +1429,10 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
             return true;
         };
 
-        std::vector<TensorStorage> processed_tensor_storages;
-        for (auto& tensor_storage : tensor_storages) {
+        for (auto& tensor_storage : processed_tensor_storages) {
             if (tensor_storage.file_index != file_index) {
                 continue;
             }
-
-            // LOG_DEBUG("%s", name.c_str());
-
-            if (is_unused_tensor(tensor_storage.name)) {
-                continue;
-            }
-
-            preprocess_tensor(tensor_storage, processed_tensor_storages);
-        }
-
-        for (auto& tensor_storage : processed_tensor_storages) {
-            // LOG_DEBUG("%s", name.c_str());
-
             ggml_tensor* dst_tensor = NULL;
 
             success = on_new_tensor_cb(tensor_storage, &dst_tensor);
@@ -1290,11 +1445,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
                 continue;
             }
 
-            ggml_backend_t backend = ggml_get_backend(dst_tensor);
-
             size_t nbytes_to_read = tensor_storage.nbytes_to_read();
 
-            if (backend == NULL || ggml_backend_is_cpu(backend)) {
+            if (dst_tensor->buffer == NULL || ggml_backend_buffer_is_host(dst_tensor->buffer)) {
                 // for the CPU and Metal backend, we can copy directly into the tensor
                 if (tensor_storage.type == dst_tensor->type) {
                     GGML_ASSERT(ggml_nbytes(dst_tensor) == tensor_storage.nbytes());
@@ -1314,7 +1467,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
                     }
 
                     convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
-                                   dst_tensor->type, (int)tensor_storage.nelements());
+                                   dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
                 }
             } else {
                 read_buffer.resize(tensor_storage.nbytes());
@@ -1333,7 +1486,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
                     convert_buffer.resize(ggml_nbytes(dst_tensor));
                     convert_tensor((void*)read_buffer.data(), tensor_storage.type,
                                    (void*)convert_buffer.data(), dst_tensor->type,
-                                   (int)tensor_storage.nelements());
+                                   (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
                     ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
                 }
             }
@@ -1350,7 +1503,134 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
     return success;
 }
 
-int64_t ModelLoader::cal_mem_size(ggml_backend_t backend) {
+bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+                               ggml_backend_t backend,
+                               std::set<std::string> ignore_tensors) {
+    std::set<std::string> tensor_names_in_file;
+    auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+        const std::string& name = tensor_storage.name;
+        // LOG_DEBUG("%s", tensor_storage.to_string().c_str());
+        tensor_names_in_file.insert(name);
+
+        struct ggml_tensor* real;
+        if (tensors.find(name) != tensors.end()) {
+            real = tensors[name];
+        } else {
+            for (auto& ignore_tensor : ignore_tensors) {
+                if (starts_with(name, ignore_tensor)) {
+                    return true;
+                }
+            }
+            LOG_INFO("unknown tensor '%s' in model file", tensor_storage.to_string().c_str());
+            return true;
+        }
+
+        if (
+            real->ne[0] != tensor_storage.ne[0] ||
+            real->ne[1] != tensor_storage.ne[1] ||
+            real->ne[2] != tensor_storage.ne[2] ||
+            real->ne[3] != tensor_storage.ne[3]) {
+            LOG_ERROR(
+                "tensor '%s' has wrong shape in model file: "
+                "got [%d, %d, %d, %d], expected [%d, %d, %d, %d]",
+                name.c_str(),
+                (int)tensor_storage.ne[0], (int)tensor_storage.ne[1], (int)tensor_storage.ne[2], (int)tensor_storage.ne[3],
+                (int)real->ne[0], (int)real->ne[1], (int)real->ne[2], (int)real->ne[3]);
+            return false;
+        }
+
+        *dst_tensor = real;
+
+        return true;
+    };
+
+    bool success = load_tensors(on_new_tensor_cb, backend);
+    if (!success) {
+        LOG_ERROR("load tensors from file failed");
+        return false;
+    }
+
+    bool some_tensor_not_init = false;
+
+    for (auto pair : tensors) {
+        if (pair.first.find("cond_stage_model.transformer.text_model.encoder.layers.23") != std::string::npos) {
+            continue;
+        }
+        if (pair.first.find("alphas_cumprod") != std::string::npos) {
+            continue;
+        }
+
+        if (pair.first.find("alphas_cumprod") != std::string::npos) {
+            continue;
+        }
+
+        if (tensor_names_in_file.find(pair.first) == tensor_names_in_file.end()) {
+            LOG_ERROR("tensor '%s' not in model file", pair.first.c_str());
+            some_tensor_not_init = true;
+        }
+    }
+
+    if (some_tensor_not_init) {
+        return false;
+    }
+    return true;
+}
+
+bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type) {
+    auto backend    = ggml_backend_cpu_init();
+    size_t mem_size = 1 * 1024 * 1024;  // for padding
+    mem_size += tensor_storages.size() * ggml_tensor_overhead();
+    mem_size += get_params_mem_size(backend, type);
+    LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
+    ggml_context* ggml_ctx = ggml_init({mem_size, NULL, false});
+
+    gguf_context* gguf_ctx = gguf_init_empty();
+
+    auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+        const std::string& name = tensor_storage.name;
+
+        ggml_type tensor_type = tensor_storage.type;
+        if (type != GGML_TYPE_COUNT) {
+            if (ggml_is_quantized(type) && tensor_storage.ne[0] % 32 != 0) {
+                tensor_type = GGML_TYPE_F16;
+            } else {
+                tensor_type = type;
+            }
+        }
+
+        ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
+        if (tensor == NULL) {
+            LOG_ERROR("ggml_new_tensor failed");
+            return false;
+        }
+        ggml_set_name(tensor, name.c_str());
+
+        // LOG_DEBUG("%s %d %s %d[%d %d %d %d] %d[%d %d %d %d]", name.c_str(),
+        // ggml_nbytes(tensor), ggml_type_name(tensor_type),
+        // tensor_storage.n_dims,
+        // tensor_storage.ne[0], tensor_storage.ne[1], tensor_storage.ne[2], tensor_storage.ne[3],
+        // tensor->n_dims, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+
+        *dst_tensor = tensor;
+
+        gguf_add_tensor(gguf_ctx, tensor);
+
+        return true;
+    };
+
+    bool success = load_tensors(on_new_tensor_cb, backend);
+    ggml_backend_free(backend);
+    LOG_INFO("load tensors done");
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    if (success) {
+        gguf_write_to_file(gguf_ctx, file_path.c_str(), false);
+    }
+    ggml_free(ggml_ctx);
+    gguf_free(gguf_ctx);
+    return success;
+}
+
+int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type) {
     size_t alignment = 128;
     if (backend != NULL) {
         alignment = ggml_backend_get_alignment(backend);
@@ -1365,8 +1645,35 @@ int64_t ModelLoader::cal_mem_size(ggml_backend_t backend) {
     }
 
     for (auto& tensor_storage : processed_tensor_storages) {
+        ggml_type tensor_type = tensor_storage.type;
+        if (type != GGML_TYPE_COUNT) {
+            if (ggml_is_quantized(type) && tensor_storage.ne[0] % 32 != 0) {
+                tensor_type = GGML_TYPE_F16;
+            } else {
+                tensor_type = type;
+            }
+        }
+        tensor_storage.type = tensor_type;
         mem_size += tensor_storage.nbytes() + alignment;
     }
 
     return mem_size;
 }
+
+bool convert(const char* input_path, const char* vae_path, const char* output_path, sd_type_t output_type) {
+    ModelLoader model_loader;
+
+    if (!model_loader.init_from_file(input_path)) {
+        LOG_ERROR("init model loader from file failed: '%s'", input_path);
+        return false;
+    }
+
+    if (vae_path != NULL && strlen(vae_path) > 0) {
+        if (!model_loader.init_from_file(vae_path, "vae.")) {
+            LOG_ERROR("init model loader from file failed: '%s'", vae_path);
+            return false;
+        }
+    }
+    bool success = model_loader.save_to_gguf_file(output_path, (ggml_type)output_type);
+    return success;
+}

model.h

@@ -4,7 +4,10 @@
 #include <functional>
 #include <map>
 #include <memory>
+#include <set>
+#include <sstream>
 #include <string>
+#include <tuple>
 #include <vector>
 
 #include "ggml/ggml-backend.h"
@@ -12,19 +15,22 @@
 #include "json.hpp"
 #include "zip.h"
 
+#define SD_MAX_DIMS 5
+
 enum SDVersion {
     VERSION_1_x,
     VERSION_2_x,
     VERSION_XL,
+    VERSION_SVD,
     VERSION_COUNT,
 };
 
 struct TensorStorage {
     std::string name;
-    ggml_type type = GGML_TYPE_F32;
-    bool is_bf16   = false;
-    int64_t ne[4]  = {1, 1, 1, 1};
-    int n_dims     = 0;
+    ggml_type type          = GGML_TYPE_F32;
+    bool is_bf16            = false;
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
 
     size_t file_index = 0;
     int index_in_zip  = -1;  // >= means stored in a zip file
@@ -40,7 +46,11 @@ struct TensorStorage {
     }
 
     int64_t nelements() const {
-        return ne[0] * ne[1] * ne[2] * ne[3];
+        int64_t n = 1;
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            n *= ne[i];
+        }
+        return n;
     }
 
     int64_t nbytes() const {
@@ -68,6 +78,7 @@ struct TensorStorage {
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
         size_t chunk_size = nbytes_to_read() / n;
+        // printf("%d/%d\n", chunk_size, nbytes_to_read());
         reverse_ne();
         for (int i = 0; i < n; i++) {
             TensorStorage chunk_i = *this;
@@ -81,7 +92,7 @@ struct TensorStorage {
     }
 
     void reverse_ne() {
-        int64_t new_ne[4] = {1, 1, 1, 1};
+        int64_t new_ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
         for (int i = 0; i < n_dims; i++) {
             new_ne[i] = ne[n_dims - 1 - i];
         }
@@ -89,10 +100,27 @@ struct TensorStorage {
             ne[i] = new_ne[i];
         }
     }
+
+    std::string to_string() const {
+        std::stringstream ss;
+        const char* type_name = ggml_type_name(type);
+        if (is_bf16) {
+            type_name = "bf16";
+        }
+        ss << name << " | " << type_name << " | ";
+        ss << n_dims << " [";
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            ss << ne[i];
+            if (i != SD_MAX_DIMS - 1) {
+                ss << ", ";
+            }
+        }
+        ss << "]";
+        return ss.str();
+    }
 };
 
 typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
-typedef std::function<void(const std::string&, int32_t)> on_new_token_cb_t;
 
 class ModelLoader {
 protected:
@@ -115,9 +143,13 @@ public:
     bool init_from_file(const std::string& file_path, const std::string& prefix = "");
     SDVersion get_sd_version();
     ggml_type get_sd_wtype();
-    bool load_vocab(on_new_token_cb_t on_new_token_cb);
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb);
-    int64_t cal_mem_size(ggml_backend_t backend);
+    std::string load_merges();
+    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend);
+    bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+                      ggml_backend_t backend,
+                      std::set<std::string> ignore_tensors = {});
+    bool save_to_gguf_file(const std::string& file_path, ggml_type type);
+    int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;
 };
 #endif  // __MODEL_H__

pmid.hpp

@@ -0,0 +1,295 @@
+#ifndef __PMI_HPP__
+#define __PMI_HPP__
+
+#include "ggml_extend.hpp"
+
+#include "clip.hpp"
+#include "lora.hpp"
+
+struct FuseBlock : public GGMLBlock {
+    // network hparams
+    int in_dim;
+    int out_dim;
+    int hidden_dim;
+    bool use_residue;
+
+public:
+    FuseBlock(int i_d, int o_d, int h_d, bool use_residue = true)
+        : in_dim(i_d), out_dim(o_d), hidden_dim(h_d), use_residue(use_residue) {
+        blocks["fc1"]       = std::shared_ptr<GGMLBlock>(new Linear(in_dim, hidden_dim, true));
+        blocks["fc2"]       = std::shared_ptr<GGMLBlock>(new Linear(hidden_dim, out_dim, true));
+        blocks["layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(in_dim));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, channels, h, w]
+
+        auto fc1        = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
+        auto fc2        = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
+        auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layernorm"]);
+
+        struct ggml_tensor* r = x;
+        // x = ggml_nn_layer_norm(ctx, x, ln_w, ln_b);
+        x = layer_norm->forward(ctx, x);
+        // x = ggml_add(ctx, ggml_mul_mat(ctx, fc1_w, x),  fc1_b);
+        x = fc1->forward(ctx, x);
+        x = ggml_gelu_inplace(ctx, x);
+        x = fc2->forward(ctx, x);
+        // x = ggml_add(ctx, ggml_mul_mat(ctx, fc2_w, x),  fc2_b);
+        if (use_residue)
+            x = ggml_add(ctx, x, r);
+        return x;
+    }
+};
+
+struct FuseModule : public GGMLBlock {
+    // network hparams
+    int embed_dim;
+
+public:
+    FuseModule(int imb_d)
+        : embed_dim(imb_d) {
+        blocks["mlp1"]       = std::shared_ptr<GGMLBlock>(new FuseBlock(imb_d * 2, imb_d, imb_d, false));
+        blocks["mlp2"]       = std::shared_ptr<GGMLBlock>(new FuseBlock(imb_d, imb_d, imb_d, true));
+        blocks["layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(embed_dim));
+    }
+
+    struct ggml_tensor* fuse_fn(struct ggml_context* ctx,
+                                struct ggml_tensor* prompt_embeds,
+                                struct ggml_tensor* id_embeds) {
+        auto mlp1       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp1"]);
+        auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
+        auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
+
+        auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
+        auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
+        // concat is along dim 2
+        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0);
+        stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
+
+        // stacked_id_embeds = mlp1.forward(ctx, stacked_id_embeds);
+        // stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
+        // stacked_id_embeds = mlp2.forward(ctx, stacked_id_embeds);
+        // stacked_id_embeds = ggml_nn_layer_norm(ctx, stacked_id_embeds, ln_w, ln_b);
+
+        stacked_id_embeds = mlp1->forward(ctx, stacked_id_embeds);
+        stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
+        stacked_id_embeds = mlp2->forward(ctx, stacked_id_embeds);
+        stacked_id_embeds = layer_norm->forward(ctx, stacked_id_embeds);
+
+        return stacked_id_embeds;
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* prompt_embeds,
+                                struct ggml_tensor* id_embeds,
+                                struct ggml_tensor* class_tokens_mask,
+                                struct ggml_tensor* class_tokens_mask_pos,
+                                struct ggml_tensor* left,
+                                struct ggml_tensor* right) {
+        // x: [N, channels, h, w]
+
+        struct ggml_tensor* valid_id_embeds = id_embeds;
+        // # slice out the image token embeddings
+        // print_ggml_tensor(class_tokens_mask_pos, false);
+        ggml_set_name(class_tokens_mask_pos, "class_tokens_mask_pos");
+        ggml_set_name(prompt_embeds, "prompt_embeds");
+        // print_ggml_tensor(valid_id_embeds, true, "valid_id_embeds");
+        // print_ggml_tensor(class_tokens_mask_pos, true, "class_tokens_mask_pos");
+        struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx, prompt_embeds, class_tokens_mask_pos);
+        ggml_set_name(image_token_embeds, "image_token_embeds");
+        struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
+
+        stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        if (left && right) {
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds);
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right);
+        } else if (left) {
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds);
+        } else if (right) {
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right);
+        }
+        stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        class_tokens_mask                         = ggml_cont(ctx, ggml_transpose(ctx, class_tokens_mask));
+        class_tokens_mask                         = ggml_repeat(ctx, class_tokens_mask, prompt_embeds);
+        prompt_embeds                             = ggml_mul(ctx, prompt_embeds, class_tokens_mask);
+        struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx, prompt_embeds, stacked_id_embeds);
+        ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
+        return updated_prompt_embeds;
+    }
+};
+
+struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
+    PhotoMakerIDEncoderBlock()
+        : CLIPVisionModelProjection(OPENAI_CLIP_VIT_L_14) {
+        blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
+        blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* id_pixel_values,
+                                struct ggml_tensor* prompt_embeds,
+                                struct ggml_tensor* class_tokens_mask,
+                                struct ggml_tensor* class_tokens_mask_pos,
+                                struct ggml_tensor* left,
+                                struct ggml_tensor* right) {
+        // x: [N, channels, h, w]
+        auto vision_model        = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
+        auto visual_projection   = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
+        auto visual_projection_2 = std::dynamic_pointer_cast<Linear>(blocks["visual_projection_2"]);
+        auto fuse_module         = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
+
+        struct ggml_tensor* shared_id_embeds = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        struct ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);    // [N, proj_dim(768)]
+        struct ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);  // [N, 1280]
+
+        id_embeds   = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
+        id_embeds_2 = ggml_cont(ctx, ggml_permute(ctx, id_embeds_2, 2, 0, 1, 3));
+
+        id_embeds = ggml_concat(ctx, id_embeds, id_embeds_2);  // [batch_size, seq_length, 1, 2048] check whether concat at dim 2 is right
+        id_embeds = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 1, 2, 0, 3));
+
+        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+                                                                         prompt_embeds,
+                                                                         id_embeds,
+                                                                         class_tokens_mask,
+                                                                         class_tokens_mask_pos,
+                                                                         left, right);
+        return updated_prompt_embeds;
+    }
+};
+
+struct PhotoMakerIDEncoder : public GGMLModule {
+public:
+    SDVersion version = VERSION_XL;
+    PhotoMakerIDEncoderBlock id_encoder;
+    float style_strength;
+
+    std::vector<float> ctm;
+    std::vector<ggml_fp16_t> ctmf16;
+    std::vector<int> ctmpos;
+
+    std::vector<ggml_fp16_t> zeros_left_16;
+    std::vector<float> zeros_left;
+    std::vector<ggml_fp16_t> zeros_right_16;
+    std::vector<float> zeros_right;
+
+public:
+    PhotoMakerIDEncoder(ggml_backend_t backend, ggml_type wtype, SDVersion version = VERSION_XL, float sty = 20.f)
+        : GGMLModule(backend, wtype),
+          version(version),
+          style_strength(sty) {
+        id_encoder.init(params_ctx, wtype);
+    }
+
+    std::string get_desc() {
+        return "pmid";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        id_encoder.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_cgraph* build_graph(  // struct ggml_allocr* allocr,
+        struct ggml_tensor* id_pixel_values,
+        struct ggml_tensor* prompt_embeds,
+        std::vector<bool>& class_tokens_mask) {
+        ctm.clear();
+        ctmf16.clear();
+        ctmpos.clear();
+        zeros_left.clear();
+        zeros_left_16.clear();
+        zeros_right.clear();
+        zeros_right_16.clear();
+
+        ggml_context* ctx0 = compute_ctx;
+
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        int64_t hidden_size = prompt_embeds->ne[0];
+        int64_t seq_length  = prompt_embeds->ne[1];
+        ggml_type type      = GGML_TYPE_F32;
+
+        struct ggml_tensor* class_tokens_mask_d = ggml_new_tensor_1d(ctx0, type, class_tokens_mask.size());
+
+        struct ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
+        struct ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
+
+        struct ggml_tensor* left  = NULL;
+        struct ggml_tensor* right = NULL;
+        for (int i = 0; i < class_tokens_mask.size(); i++) {
+            if (class_tokens_mask[i]) {
+                ctm.push_back(0.f);                        // here use 0.f instead of 1.f to make a scale mask
+                ctmf16.push_back(ggml_fp32_to_fp16(0.f));  // here use 0.f instead of 1.f to make a scale mask
+                ctmpos.push_back(i);
+            } else {
+                ctm.push_back(1.f);                        // here use 1.f instead of 0.f to make a scale mask
+                ctmf16.push_back(ggml_fp32_to_fp16(1.f));  // here use 0.f instead of 1.f to make a scale mask
+            }
+        }
+        if (ctmpos[0] > 0) {
+            left = ggml_new_tensor_3d(ctx0, type, hidden_size, 1, ctmpos[0]);
+        }
+        if (ctmpos[ctmpos.size() - 1] < seq_length - 1) {
+            right = ggml_new_tensor_3d(ctx0, type,
+                                       hidden_size, 1, seq_length - ctmpos[ctmpos.size() - 1] - 1);
+        }
+        struct ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ctmpos.size());
+
+        {
+            if (type == GGML_TYPE_F16)
+                set_backend_tensor_data(class_tokens_mask_d, ctmf16.data());
+            else
+                set_backend_tensor_data(class_tokens_mask_d, ctm.data());
+            set_backend_tensor_data(class_tokens_mask_pos, ctmpos.data());
+            if (left) {
+                if (type == GGML_TYPE_F16) {
+                    for (int i = 0; i < ggml_nelements(left); ++i)
+                        zeros_left_16.push_back(ggml_fp32_to_fp16(0.f));
+                    set_backend_tensor_data(left, zeros_left_16.data());
+                } else {
+                    for (int i = 0; i < ggml_nelements(left); ++i)
+                        zeros_left.push_back(0.f);
+                    set_backend_tensor_data(left, zeros_left.data());
+                }
+            }
+            if (right) {
+                if (type == GGML_TYPE_F16) {
+                    for (int i = 0; i < ggml_nelements(right); ++i)
+                        zeros_right_16.push_back(ggml_fp32_to_fp16(0.f));
+                    set_backend_tensor_data(right, zeros_right_16.data());
+                } else {
+                    for (int i = 0; i < ggml_nelements(right); ++i)
+                        zeros_right.push_back(0.f);
+                    set_backend_tensor_data(right, zeros_right.data());
+                }
+            }
+        }
+        struct ggml_tensor* updated_prompt_embeds = id_encoder.forward(ctx0,
+                                                                       id_pixel_values_d,
+                                                                       prompt_embeds_d,
+                                                                       class_tokens_mask_d,
+                                                                       class_tokens_mask_pos,
+                                                                       left, right);
+        ggml_build_forward_expand(gf, updated_prompt_embeds);
+
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* id_pixel_values,
+                 struct ggml_tensor* prompt_embeds,
+                 std::vector<bool>& class_tokens_mask,
+                 struct ggml_tensor** updated_prompt_embeds,
+                 ggml_context* output_ctx) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            // return build_graph(compute_allocr, id_pixel_values, prompt_embeds, class_tokens_mask);
+            return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask);
+        };
+
+        // GGMLModule::compute(get_graph, n_threads, updated_prompt_embeds);
+        GGMLModule::compute(get_graph, n_threads, true, updated_prompt_embeds, output_ctx);
+    }
+};
+
+#endif  // __PMI_HPP__

preprocessing.hpp

@@ -0,0 +1,227 @@
+#ifndef __PREPROCESSING_HPP__
+#define __PREPROCESSING_HPP__
+
+#include "ggml_extend.hpp"
+#define M_PI_ 3.14159265358979323846
+
+void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml_tensor* kernel, int padding) {
+    struct ggml_init_params params;
+    params.mem_size                 = 20 * 1024 * 1024;  // 10
+    params.mem_buffer               = NULL;
+    params.no_alloc                 = false;
+    struct ggml_context* ctx0       = ggml_init(params);
+    struct ggml_tensor* kernel_fp16 = ggml_new_tensor_4d(ctx0, GGML_TYPE_F16, kernel->ne[0], kernel->ne[1], 1, 1);
+    ggml_fp32_to_fp16_row((float*)kernel->data, (ggml_fp16_t*)kernel_fp16->data, ggml_nelements(kernel));
+    ggml_tensor* h  = ggml_conv_2d(ctx0, kernel_fp16, input, 1, 1, padding, padding, 1, 1);
+    ggml_cgraph* gf = ggml_new_graph(ctx0);
+    ggml_build_forward_expand(gf, ggml_cpy(ctx0, h, output));
+    ggml_graph_compute_with_ctx(ctx0, gf, 1);
+    ggml_free(ctx0);
+}
+
+void gaussian_kernel(struct ggml_tensor* kernel) {
+    int ks_mid   = kernel->ne[0] / 2;
+    float sigma  = 1.4f;
+    float normal = 1.f / (2.0f * M_PI_ * powf(sigma, 2.0f));
+    for (int y = 0; y < kernel->ne[0]; y++) {
+        float gx = -ks_mid + y;
+        for (int x = 0; x < kernel->ne[1]; x++) {
+            float gy = -ks_mid + x;
+            float k_ = expf(-((gx * gx + gy * gy) / (2.0f * powf(sigma, 2.0f)))) * normal;
+            ggml_tensor_set_f32(kernel, k_, x, y);
+        }
+    }
+}
+
+void grayscale(struct ggml_tensor* rgb_img, struct ggml_tensor* grayscale) {
+    for (int iy = 0; iy < rgb_img->ne[1]; iy++) {
+        for (int ix = 0; ix < rgb_img->ne[0]; ix++) {
+            float r    = ggml_tensor_get_f32(rgb_img, ix, iy);
+            float g    = ggml_tensor_get_f32(rgb_img, ix, iy, 1);
+            float b    = ggml_tensor_get_f32(rgb_img, ix, iy, 2);
+            float gray = 0.2989f * r + 0.5870f * g + 0.1140f * b;
+            ggml_tensor_set_f32(grayscale, gray, ix, iy);
+        }
+    }
+}
+
+void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) {
+    int n_elements = ggml_nelements(h);
+    float* dx      = (float*)x->data;
+    float* dy      = (float*)y->data;
+    float* dh      = (float*)h->data;
+    for (int i = 0; i < n_elements; i++) {
+        dh[i] = sqrtf(dx[i] * dx[i] + dy[i] * dy[i]);
+    }
+}
+
+void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) {
+    int n_elements = ggml_nelements(h);
+    float* dx      = (float*)x->data;
+    float* dy      = (float*)y->data;
+    float* dh      = (float*)h->data;
+    for (int i = 0; i < n_elements; i++) {
+        dh[i] = atan2f(dy[i], dx[i]);
+    }
+}
+
+void normalize_tensor(struct ggml_tensor* g) {
+    int n_elements = ggml_nelements(g);
+    float* dg      = (float*)g->data;
+    float max      = -INFINITY;
+    for (int i = 0; i < n_elements; i++) {
+        max = dg[i] > max ? dg[i] : max;
+    }
+    max = 1.0f / max;
+    for (int i = 0; i < n_elements; i++) {
+        dg[i] *= max;
+    }
+}
+
+void non_max_supression(struct ggml_tensor* result, struct ggml_tensor* G, struct ggml_tensor* D) {
+    for (int iy = 1; iy < result->ne[1] - 1; iy++) {
+        for (int ix = 1; ix < result->ne[0] - 1; ix++) {
+            float angle = ggml_tensor_get_f32(D, ix, iy) * 180.0f / M_PI_;
+            angle       = angle < 0.0f ? angle += 180.0f : angle;
+            float q     = 1.0f;
+            float r     = 1.0f;
+
+            // angle 0
+            if ((0 >= angle && angle < 22.5f) || (157.5f >= angle && angle <= 180)) {
+                q = ggml_tensor_get_f32(G, ix, iy + 1);
+                r = ggml_tensor_get_f32(G, ix, iy - 1);
+            }
+            // angle 45
+            else if (22.5f >= angle && angle < 67.5f) {
+                q = ggml_tensor_get_f32(G, ix + 1, iy - 1);
+                r = ggml_tensor_get_f32(G, ix - 1, iy + 1);
+            }
+            // angle 90
+            else if (67.5f >= angle && angle < 112.5) {
+                q = ggml_tensor_get_f32(G, ix + 1, iy);
+                r = ggml_tensor_get_f32(G, ix - 1, iy);
+            }
+            // angle 135
+            else if (112.5 >= angle && angle < 157.5f) {
+                q = ggml_tensor_get_f32(G, ix - 1, iy - 1);
+                r = ggml_tensor_get_f32(G, ix + 1, iy + 1);
+            }
+
+            float cur = ggml_tensor_get_f32(G, ix, iy);
+            if ((cur >= q) && (cur >= r)) {
+                ggml_tensor_set_f32(result, cur, ix, iy);
+            } else {
+                ggml_tensor_set_f32(result, 0.0f, ix, iy);
+            }
+        }
+    }
+}
+
+void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float low_threshold, float weak, float strong) {
+    int n_elements = ggml_nelements(img);
+    float* imd     = (float*)img->data;
+    float max      = -INFINITY;
+    for (int i = 0; i < n_elements; i++) {
+        max = imd[i] > max ? imd[i] : max;
+    }
+    float ht = max * high_threshold;
+    float lt = ht * low_threshold;
+    for (int i = 0; i < n_elements; i++) {
+        float img_v = imd[i];
+        if (img_v >= ht) {  // strong pixel
+            imd[i] = strong;
+        } else if (img_v <= ht && img_v >= lt) {  // strong pixel
+            imd[i] = weak;
+        }
+    }
+
+    for (int iy = 0; iy < img->ne[1]; iy++) {
+        for (int ix = 0; ix < img->ne[0]; ix++) {
+            if (ix >= 3 && ix <= img->ne[0] - 3 && iy >= 3 && iy <= img->ne[1] - 3) {
+                ggml_tensor_set_f32(img, ggml_tensor_get_f32(img, ix, iy), ix, iy);
+            } else {
+                ggml_tensor_set_f32(img, 0.0f, ix, iy);
+            }
+        }
+    }
+
+    // hysteresis
+    for (int iy = 1; iy < img->ne[1] - 1; iy++) {
+        for (int ix = 1; ix < img->ne[0] - 1; ix++) {
+            float imd_v = ggml_tensor_get_f32(img, ix, iy);
+            if (imd_v == weak) {
+                if (ggml_tensor_get_f32(img, ix + 1, iy - 1) == strong || ggml_tensor_get_f32(img, ix + 1, iy) == strong ||
+                    ggml_tensor_get_f32(img, ix, iy - 1) == strong || ggml_tensor_get_f32(img, ix, iy + 1) == strong ||
+                    ggml_tensor_get_f32(img, ix - 1, iy - 1) == strong || ggml_tensor_get_f32(img, ix - 1, iy) == strong) {
+                    ggml_tensor_set_f32(img, strong, ix, iy);
+                } else {
+                    ggml_tensor_set_f32(img, 0.0f, ix, iy);
+                }
+            }
+        }
+    }
+}
+
+uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
+    struct ggml_init_params params;
+    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10
+    params.mem_buffer             = NULL;
+    params.no_alloc               = false;
+    struct ggml_context* work_ctx = ggml_init(params);
+
+    if (!work_ctx) {
+        LOG_ERROR("ggml_init() failed");
+        return NULL;
+    }
+
+    float kX[9] = {
+        -1, 0, 1,
+        -2, 0, 2,
+        -1, 0, 1};
+
+    float kY[9] = {
+        1, 2, 1,
+        0, 0, 0,
+        -1, -2, -1};
+
+    // generate kernel
+    int kernel_size             = 5;
+    struct ggml_tensor* gkernel = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, kernel_size, kernel_size, 1, 1);
+    struct ggml_tensor* sf_kx   = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
+    memcpy(sf_kx->data, kX, ggml_nbytes(sf_kx));
+    struct ggml_tensor* sf_ky = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
+    memcpy(sf_ky->data, kY, ggml_nbytes(sf_ky));
+    gaussian_kernel(gkernel);
+    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 3, 1);
+    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 1, 1);
+    struct ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
+    struct ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
+    struct ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
+    struct ggml_tensor* tetha      = ggml_dup_tensor(work_ctx, image_gray);
+    sd_image_to_tensor(img, image);
+    grayscale(image, image_gray);
+    convolve(image_gray, image_gray, gkernel, 2);
+    convolve(image_gray, iX, sf_kx, 1);
+    convolve(image_gray, iY, sf_ky, 1);
+    prop_hypot(iX, iY, G);
+    normalize_tensor(G);
+    prop_arctan2(iX, iY, tetha);
+    non_max_supression(image_gray, G, tetha);
+    threshold_hystersis(image_gray, high_threshold, low_threshold, weak, strong);
+    // to RGB channels
+    for (int iy = 0; iy < height; iy++) {
+        for (int ix = 0; ix < width; ix++) {
+            float gray = ggml_tensor_get_f32(image_gray, ix, iy);
+            gray       = inverse ? 1.0f - gray : gray;
+            ggml_tensor_set_f32(image, gray, ix, iy);
+            ggml_tensor_set_f32(image, gray, ix, iy, 1);
+            ggml_tensor_set_f32(image, gray, ix, iy, 2);
+        }
+    }
+    free(img);
+    uint8_t* output = sd_tensor_to_image(image);
+    ggml_free(work_ctx);
+    return output;
+}
+
+#endif  // __PREPROCESSING_HPP__

rng_philox.hpp

@@ -4,7 +4,7 @@
 #include <cmath>
 #include <vector>
 
-#include "rng.h"
+#include "rng.hpp"
 
 // RNG imitiating torch cuda randn on CPU.
 // Port from: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/5ef669de080814067961f28357256e8fe27544f4/modules/rng_philox.py

stable-diffusion.h

@@ -1,16 +1,39 @@
 #ifndef __STABLE_DIFFUSION_H__
 #define __STABLE_DIFFUSION_H__
 
-#include <memory>
-#include <string>
-#include <vector>
+#if defined(_WIN32) || defined(__CYGWIN__)
+#ifndef SD_BUILD_SHARED_LIB
+#define SD_API
+#else
+#ifdef SD_BUILD_DLL
+#define SD_API __declspec(dllexport)
+#else
+#define SD_API __declspec(dllimport)
+#endif
+#endif
+#else
+#if __GNUC__ >= 4
+#define SD_API __attribute__((visibility("default")))
+#else
+#define SD_API
+#endif
+#endif
 
-enum RNGType {
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+enum rng_type_t {
     STD_DEFAULT_RNG,
     CUDA_RNG
 };
 
-enum SampleMethod {
+enum sample_method_t {
     EULER_A,
     EULER,
     HEUN,
@@ -22,54 +45,165 @@ enum SampleMethod {
     N_SAMPLE_METHODS
 };
 
-enum Schedule {
+enum schedule_t {
     DEFAULT,
     DISCRETE,
     KARRAS,
     N_SCHEDULES
 };
 
-class StableDiffusionGGML;
-
-class StableDiffusion {
-private:
-    std::shared_ptr<StableDiffusionGGML> sd;
-
-public:
-    StableDiffusion(int n_threads                = -1,
-                    bool vae_decode_only         = false,
-                    std::string taesd_path       = "",
-                    bool free_params_immediately = false,
-                    std::string lora_model_dir   = "",
-                    RNGType rng_type             = STD_DEFAULT_RNG);
-    bool load_from_file(const std::string& model_path,
-                        const std::string& vae_path,
-                        ggml_type wtype,
-                        Schedule d = DEFAULT);
-    std::vector<uint8_t*> txt2img(
-        std::string prompt,
-        std::string negative_prompt,
-        float cfg_scale,
-        int width,
-        int height,
-        SampleMethod sample_method,
-        int sample_steps,
-        int64_t seed,
-        int batch_count);
-
-    std::vector<uint8_t*> img2img(
-        const uint8_t* init_img_data,
-        std::string prompt,
-        std::string negative_prompt,
-        float cfg_scale,
-        int width,
-        int height,
-        SampleMethod sample_method,
-        int sample_steps,
-        float strength,
-        int64_t seed);
+// same as enum ggml_type
+enum sd_type_t {
+    SD_TYPE_F32  = 0,
+    SD_TYPE_F16  = 1,
+    SD_TYPE_Q4_0 = 2,
+    SD_TYPE_Q4_1 = 3,
+    // SD_TYPE_Q4_2 = 4, support has been removed
+    // SD_TYPE_Q4_3 (5) support has been removed
+    SD_TYPE_Q5_0 = 6,
+    SD_TYPE_Q5_1 = 7,
+    SD_TYPE_Q8_0 = 8,
+    SD_TYPE_Q8_1 = 9,
+    // k-quantizations
+    SD_TYPE_Q2_K    = 10,
+    SD_TYPE_Q3_K    = 11,
+    SD_TYPE_Q4_K    = 12,
+    SD_TYPE_Q5_K    = 13,
+    SD_TYPE_Q6_K    = 14,
+    SD_TYPE_Q8_K    = 15,
+    SD_TYPE_IQ2_XXS = 16,
+    SD_TYPE_IQ2_XS  = 17,
+    SD_TYPE_IQ3_XXS = 18,
+    SD_TYPE_IQ1_S   = 19,
+    SD_TYPE_IQ4_NL  = 20,
+    SD_TYPE_IQ3_S   = 21,
+    SD_TYPE_IQ2_S   = 22,
+    SD_TYPE_IQ4_XS  = 23,
+    SD_TYPE_I8,
+    SD_TYPE_I16,
+    SD_TYPE_I32,
+    SD_TYPE_COUNT,
 };
 
-std::string sd_get_system_info();
+SD_API const char* sd_type_name(enum sd_type_t type);
 
-#endif  // __STABLE_DIFFUSION_H__
+enum sd_log_level_t {
+    SD_LOG_DEBUG,
+    SD_LOG_INFO,
+    SD_LOG_WARN,
+    SD_LOG_ERROR
+};
+
+typedef void (*sd_log_cb_t)(enum sd_log_level_t level, const char* text, void* data);
+typedef void (*sd_progress_cb_t)(int step, int steps, float time, void* data);
+
+SD_API void sd_set_log_callback(sd_log_cb_t sd_log_cb, void* data);
+SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
+SD_API int32_t get_num_physical_cores();
+SD_API const char* sd_get_system_info();
+
+typedef struct {
+    uint32_t width;
+    uint32_t height;
+    uint32_t channel;
+    uint8_t* data;
+} sd_image_t;
+
+typedef struct sd_ctx_t sd_ctx_t;
+
+SD_API sd_ctx_t* new_sd_ctx(const char* model_path,
+                            const char* vae_path,
+                            const char* taesd_path,
+                            const char* control_net_path_c_str,
+                            const char* lora_model_dir,
+                            const char* embed_dir_c_str,
+                            const char* stacked_id_embed_dir_c_str,
+                            bool vae_decode_only,
+                            bool vae_tiling,
+                            bool free_params_immediately,
+                            int n_threads,
+                            enum sd_type_t wtype,
+                            enum rng_type_t rng_type,
+                            enum schedule_t s,
+                            bool keep_clip_on_cpu,
+                            bool keep_control_net_cpu,
+                            bool keep_vae_on_cpu);
+
+SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
+
+SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
+                           const char* prompt,
+                           const char* negative_prompt,
+                           int clip_skip,
+                           float cfg_scale,
+                           int width,
+                           int height,
+                           enum sample_method_t sample_method,
+                           int sample_steps,
+                           int64_t seed,
+                           int batch_count,
+                           const sd_image_t* control_cond,
+                           float control_strength,
+                           float style_strength,
+                           bool normalize_input,
+                           const char* input_id_images_path);
+
+SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
+                           sd_image_t init_image,
+                           const char* prompt,
+                           const char* negative_prompt,
+                           int clip_skip,
+                           float cfg_scale,
+                           int width,
+                           int height,
+                           enum sample_method_t sample_method,
+                           int sample_steps,
+                           float strength,
+                           int64_t seed,
+                           int batch_count,
+                           const sd_image_t* control_cond,
+                           float control_strength,
+                           float style_strength,
+                           bool normalize_input,
+                           const char* input_id_images_path);
+
+SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,
+                           sd_image_t init_image,
+                           int width,
+                           int height,
+                           int video_frames,
+                           int motion_bucket_id,
+                           int fps,
+                           float augmentation_level,
+                           float min_cfg,
+                           float cfg_scale,
+                           enum sample_method_t sample_method,
+                           int sample_steps,
+                           float strength,
+                           int64_t seed);
+
+typedef struct upscaler_ctx_t upscaler_ctx_t;
+
+SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
+                                        int n_threads,
+                                        enum sd_type_t wtype);
+SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
+
+SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor);
+
+SD_API bool convert(const char* input_path, const char* vae_path, const char* output_path, sd_type_t output_type);
+
+SD_API uint8_t* preprocess_canny(uint8_t* img,
+                                 int width,
+                                 int height,
+                                 float high_threshold,
+                                 float low_threshold,
+                                 float weak,
+                                 float strong,
+                                 bool inverse);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // __STABLE_DIFFUSION_H__

tae.hpp

@@ -0,0 +1,251 @@
+#ifndef __TAE_HPP__
+#define __TAE_HPP__
+
+#include "ggml_extend.hpp"
+
+#include "model.h"
+
+/*
+    ===================================    TinyAutoEncoder  ===================================
+    References:
+    https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/autoencoders/vae.py
+    https://github.com/madebyollin/taesd/blob/main/taesd.py
+
+*/
+
+class TAEBlock : public UnaryBlock {
+protected:
+    int n_in;
+    int n_out;
+
+public:
+    TAEBlock(int n_in, int n_out)
+        : n_in(n_in), n_out(n_out) {
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
+        if (n_in != n_out) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {1, 1}, {1, 1}, {1, 1}, {1, 1}, false));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, n_in, h, w]
+        // return: [n, n_out, h, w]
+
+        auto conv_0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv_2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
+        auto conv_4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+
+        auto h = conv_0->forward(ctx, x);
+        h      = ggml_relu_inplace(ctx, h);
+        h      = conv_2->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx, h);
+        h      = conv_4->forward(ctx, h);
+
+        if (n_in != n_out) {
+            auto skip = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            LOG_DEBUG("skip");
+            x = skip->forward(ctx, x);
+        }
+
+        h = ggml_add(ctx, h, x);
+        h = ggml_relu_inplace(ctx, h);
+        return h;
+    }
+};
+
+class TinyEncoder : public UnaryBlock {
+    int in_channels = 3;
+    int channels    = 64;
+    int z_channels  = 4;
+    int num_blocks  = 3;
+
+public:
+    TinyEncoder() {
+        int index                       = 0;
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, in_channels, h, w]
+        // return: [n, z_channels, h/8, w/8]
+
+        for (int i = 0; i < num_blocks * 3 + 6; i++) {
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
+
+            x = block->forward(ctx, x);
+        }
+
+        return x;
+    }
+};
+
+class TinyDecoder : public UnaryBlock {
+    int z_channels   = 4;
+    int channels     = 64;
+    int out_channels = 3;
+    int num_blocks   = 3;
+
+public:
+    TinyDecoder(int index = 0) {
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        blocks[std::to_string(index++)] = 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* z) {
+        // z: [n, z_channels, h, w]
+        // return: [n, out_channels, h*8, w*8]
+
+        auto h = ggml_scale(ctx, z, 1.0f / 3.0f);
+        h      = ggml_tanh_inplace(ctx, h);
+        h      = ggml_scale(ctx, h, 3.0f);
+
+        for (int i = 0; i < num_blocks * 3 + 10; i++) {
+            if (blocks.find(std::to_string(i)) == blocks.end()) {
+                if (i == 1) {
+                    h = ggml_relu_inplace(ctx, h);
+                } else {
+                    h = ggml_upscale(ctx, h, 2);
+                }
+                continue;
+            }
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
+
+            h = block->forward(ctx, h);
+        }
+
+        return h;
+    }
+};
+
+class TAESD : public GGMLBlock {
+protected:
+    bool decode_only;
+
+public:
+    TAESD(bool decode_only = true)
+        : decode_only(decode_only) {
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder());
+
+        if (!decode_only) {
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder());
+        }
+    }
+
+    struct ggml_tensor* decode(struct ggml_context* ctx, struct ggml_tensor* z) {
+        auto decoder = std::dynamic_pointer_cast<TinyDecoder>(blocks["decoder.layers"]);
+        return decoder->forward(ctx, z);
+    }
+
+    struct ggml_tensor* encode(struct ggml_context* ctx, struct ggml_tensor* x) {
+        auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
+        return encoder->forward(ctx, x);
+    }
+};
+
+struct TinyAutoEncoder : public GGMLModule {
+    TAESD taesd;
+    bool decode_only = false;
+
+    TinyAutoEncoder(ggml_backend_t backend,
+                    ggml_type wtype,
+                    bool decoder_only = true)
+        : decode_only(decoder_only),
+          taesd(decode_only),
+          GGMLModule(backend, wtype) {
+        taesd.init(params_ctx, wtype);
+    }
+
+    std::string get_desc() {
+        return "taesd";
+    }
+
+    bool load_from_file(const std::string& file_path) {
+        LOG_INFO("loading taesd from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> taesd_tensors;
+        taesd.get_param_tensors(taesd_tensors);
+        std::set<std::string> ignore_tensors;
+        if (decode_only) {
+            ignore_tensors.insert("encoder.");
+        }
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init taesd model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool success = model_loader.load_tensors(taesd_tensors, backend, ignore_tensors);
+
+        if (!success) {
+            LOG_ERROR("load tae tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("taesd model loaded");
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        z                       = to_backend(z);
+        struct ggml_tensor* out = decode_graph ? taesd.decode(compute_ctx, z) : taesd.encode(compute_ctx, z);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* z,
+                 bool decode_graph,
+                 struct ggml_tensor** output,
+                 struct ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(z, decode_graph);
+        };
+
+        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+};
+
+#endif  // __TAE_HPP__

unet.hpp

@@ -0,0 +1,653 @@
+#ifndef __UNET_HPP__
+#define __UNET_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*==================================================== UnetModel =====================================================*/
+
+#define UNET_GRAPH_SIZE 10240
+
+class SpatialVideoTransformer : public SpatialTransformer {
+protected:
+    int64_t time_depth;
+    int64_t max_time_embed_period;
+
+public:
+    SpatialVideoTransformer(int64_t in_channels,
+                            int64_t n_head,
+                            int64_t d_head,
+                            int64_t depth,
+                            int64_t context_dim,
+                            int64_t time_depth            = 1,
+                            int64_t max_time_embed_period = 10000)
+        : SpatialTransformer(in_channels, n_head, d_head, depth, context_dim),
+          max_time_embed_period(max_time_embed_period) {
+        // We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False
+        // use_spatial_context is always True
+        // merge_strategy is always learned_with_images
+        // merge_factor is loaded from weights
+        // time_context_dim is always None
+        // ff_in is always True
+        // disable_self_attn is always False
+        // disable_temporal_crossattention is always False
+
+        int64_t inner_dim = n_head * d_head;
+
+        GGML_ASSERT(depth == time_depth);
+        GGML_ASSERT(in_channels == inner_dim);
+
+        int64_t time_mix_d_head    = d_head;
+        int64_t n_time_mix_heads   = n_head;
+        int64_t time_mix_inner_dim = time_mix_d_head * n_time_mix_heads;  // equal to inner_dim
+        int64_t time_context_dim   = context_dim;
+
+        for (int i = 0; i < time_depth; i++) {
+            std::string name = "time_stack." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim,
+                                                                                    n_time_mix_heads,
+                                                                                    time_mix_d_head,
+                                                                                    time_context_dim,
+                                                                                    true));
+        }
+
+        int64_t time_embed_dim     = in_channels * 4;
+        blocks["time_pos_embed.0"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, time_embed_dim));
+        // time_pos_embed.1 is nn.SiLU()
+        blocks["time_pos_embed.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, in_channels));
+
+        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context,
+                                int timesteps) {
+        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w], t == timesteps
+        // context: [N, max_position(aka n_context), hidden_size(aka context_dim)] aka [b*t, n_context, context_dim], t == timesteps
+        // t_emb: [N, in_channels] aka [b*t, in_channels]
+        // timesteps is num_frames
+        // time_context is always None
+        // image_only_indicator is always tensor([0.])
+        // transformer_options is not used
+        // GGML_ASSERT(ggml_n_dims(context) == 3);
+
+        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 time_pos_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["time_pos_embed.0"]);
+        auto time_pos_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["time_pos_embed.2"]);
+        auto time_mixer       = std::dynamic_pointer_cast<AlphaBlender>(blocks["time_mixer"]);
+
+        auto x_in         = x;
+        int64_t n         = x->ne[3];
+        int64_t h         = x->ne[1];
+        int64_t w         = x->ne[0];
+        int64_t inner_dim = n_head * d_head;
+
+        GGML_ASSERT(n == timesteps);  // We compute cond and uncond separately, so batch_size==1
+
+        auto time_context    = context;  // [b*t, n_context, context_dim]
+        auto spatial_context = context;
+        // time_context_first_timestep = time_context[::timesteps]
+        auto time_context_first_timestep = ggml_view_3d(ctx,
+                                                        time_context,
+                                                        time_context->ne[0],
+                                                        time_context->ne[1],
+                                                        1,
+                                                        time_context->nb[1],
+                                                        time_context->nb[2],
+                                                        0);  // [b, n_context, context_dim]
+        time_context                     = ggml_new_tensor_3d(ctx, GGML_TYPE_F32,
+                                                              time_context_first_timestep->ne[0],
+                                                              time_context_first_timestep->ne[1],
+                                                              time_context_first_timestep->ne[2] * h * w);
+        time_context                     = ggml_repeat(ctx, time_context_first_timestep, time_context);  // [b*h*w, n_context, context_dim]
+
+        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]
+
+        auto num_frames = ggml_arange(ctx, 0, timesteps, 1);
+        // since b is 1, no need to do repeat
+        auto t_emb = ggml_nn_timestep_embedding(ctx, num_frames, in_channels, max_time_embed_period);  // [N, in_channels]
+
+        auto emb = time_pos_embed_0->forward(ctx, t_emb);
+        emb      = ggml_silu_inplace(ctx, emb);
+        emb      = time_pos_embed_2->forward(ctx, emb);                   // [N, in_channels]
+        emb      = ggml_reshape_3d(ctx, emb, emb->ne[0], 1, emb->ne[1]);  // [N, 1, in_channels]
+
+        for (int i = 0; i < depth; i++) {
+            std::string transformer_name = "transformer_blocks." + std::to_string(i);
+            std::string time_stack_name  = "time_stack." + std::to_string(i);
+
+            auto block     = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[transformer_name]);
+            auto mix_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[time_stack_name]);
+
+            x = block->forward(ctx, x, spatial_context);  // [N, h * w, inner_dim]
+
+            // in_channels == inner_dim
+            auto x_mix = x;
+            x_mix      = ggml_add(ctx, x_mix, emb);  // [N, h * w, inner_dim]
+
+            int64_t N = x_mix->ne[2];
+            int64_t T = timesteps;
+            int64_t B = N / T;
+            int64_t S = x_mix->ne[1];
+            int64_t C = x_mix->ne[0];
+
+            x_mix = ggml_reshape_4d(ctx, x_mix, C, S, T, B);               // (b t) s c -> b t s c
+            x_mix = ggml_cont(ctx, ggml_permute(ctx, x_mix, 0, 2, 1, 3));  // b t s c -> b s t c
+            x_mix = ggml_reshape_3d(ctx, x_mix, C, T, S * B);              // b s t c -> (b s) t c
+
+            x_mix = mix_block->forward(ctx, x_mix, time_context);  // [B * h * w, T, inner_dim]
+
+            x_mix = ggml_reshape_4d(ctx, x_mix, C, T, S, B);               // (b s) t c -> b s t c
+            x_mix = ggml_cont(ctx, ggml_permute(ctx, x_mix, 0, 2, 1, 3));  // b s t c -> b t s c
+            x_mix = ggml_reshape_3d(ctx, x_mix, C, S, T * B);              // b t s c -> (b t) s c
+
+            x = time_mixer->forward(ctx, x, x_mix);  // [N, h * w, inner_dim]
+        }
+
+        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]
+
+        // proj_out
+        x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
+
+        x = ggml_add(ctx, x, x_in);
+        return x;
+    }
+};
+
+// ldm.modules.diffusionmodules.openaimodel.UNetModel
+class UnetModelBlock : public GGMLBlock {
+protected:
+    SDVersion version = VERSION_1_x;
+    // network hparams
+    int in_channels                        = 4;
+    int out_channels                       = 4;
+    int num_res_blocks                     = 2;
+    std::vector<int> attention_resolutions = {4, 2, 1};
+    std::vector<int> channel_mult          = {1, 2, 4, 4};
+    std::vector<int> transformer_depth     = {1, 1, 1, 1};
+    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
+
+public:
+    int model_channels  = 320;
+    int adm_in_channels = 2816;  // only for VERSION_XL/SVD
+
+    UnetModelBlock(SDVersion version = VERSION_1_x)
+        : version(version) {
+        if (version == VERSION_2_x) {
+            context_dim       = 1024;
+            num_head_channels = 64;
+            num_heads         = -1;
+        } else if (version == VERSION_XL) {
+            context_dim           = 2048;
+            attention_resolutions = {4, 2};
+            channel_mult          = {1, 2, 4};
+            transformer_depth     = {1, 2, 10};
+            num_head_channels     = 64;
+            num_heads             = -1;
+        } else if (version == VERSION_SVD) {
+            in_channels       = 8;
+            out_channels      = 4;
+            context_dim       = 1024;
+            adm_in_channels   = 768;
+            num_head_channels = 64;
+            num_heads         = -1;
+        }
+        // dims is always 2
+        // use_temporal_attention is always True for SVD
+
+        blocks["time_embed.0"] = std::shared_ptr<GGMLBlock>(new Linear(model_channels, time_embed_dim));
+        // 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) {
+            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));
+        }
+
+        // input_blocks
+        blocks["input_blocks.0.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, model_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        std::vector<int> input_block_chans;
+        input_block_chans.push_back(model_channels);
+        int ch              = model_channels;
+        int input_block_idx = 0;
+        int ds              = 1;
+
+        auto get_resblock = [&](int64_t channels, int64_t emb_channels, int64_t out_channels) -> ResBlock* {
+            if (version == VERSION_SVD) {
+                return new VideoResBlock(channels, emb_channels, out_channels);
+            } else {
+                return new ResBlock(channels, emb_channels, out_channels);
+            }
+        };
+
+        auto get_attention_layer = [&](int64_t in_channels,
+                                       int64_t n_head,
+                                       int64_t d_head,
+                                       int64_t depth,
+                                       int64_t context_dim) -> SpatialTransformer* {
+            if (version == VERSION_SVD) {
+                return new SpatialVideoTransformer(in_channels, n_head, d_head, depth, context_dim);
+            } else {
+                return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
+            }
+        };
+
+        size_t len_mults = channel_mult.size();
+        for (int i = 0; i < len_mults; i++) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, mult * model_channels));
+
+                ch = mult * model_channels;
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    int n_head = num_heads;
+                    int d_head = ch / num_heads;
+                    if (num_head_channels != -1) {
+                        d_head = num_head_channels;
+                        n_head = ch / d_head;
+                    }
+                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
+                                                                                      n_head,
+                                                                                      d_head,
+                                                                                      transformer_depth[i],
+                                                                                      context_dim));
+                }
+                input_block_chans.push_back(ch);
+            }
+            if (i != len_mults - 1) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(ch, ch));
+
+                input_block_chans.push_back(ch);
+                ds *= 2;
+            }
+        }
+
+        // middle blocks
+        int n_head = num_heads;
+        int d_head = ch / num_heads;
+        if (num_head_channels != -1) {
+            d_head = num_head_channels;
+            n_head = ch / d_head;
+        }
+        blocks["middle_block.0"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
+        blocks["middle_block.1"] = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
+                                                                                  n_head,
+                                                                                  d_head,
+                                                                                  transformer_depth[transformer_depth.size() - 1],
+                                                                                  context_dim));
+        blocks["middle_block.2"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
+
+        // output_blocks
+        int output_block_idx = 0;
+        for (int i = (int)len_mults - 1; i >= 0; i--) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                int ich = input_block_chans.back();
+                input_block_chans.pop_back();
+
+                std::string name = "output_blocks." + std::to_string(output_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(get_resblock(ch + ich, time_embed_dim, mult * model_channels));
+
+                ch                = mult * model_channels;
+                int up_sample_idx = 1;
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    int n_head = num_heads;
+                    int d_head = ch / num_heads;
+                    if (num_head_channels != -1) {
+                        d_head = num_head_channels;
+                        n_head = ch / d_head;
+                    }
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + ".1";
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(get_attention_layer(ch, n_head, d_head, transformer_depth[i], context_dim));
+
+                    up_sample_idx++;
+                }
+
+                if (i > 0 && j == num_res_blocks) {
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + "." + std::to_string(up_sample_idx);
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(new UpSampleBlock(ch, ch));
+
+                    ds /= 2;
+                }
+
+                output_block_idx += 1;
+            }
+        }
+
+        // out
+        blocks["out.0"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(ch));  // ch == model_channels
+        // out_1 is nn.SiLU()
+        blocks["out.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(model_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* resblock_forward(std::string name,
+                                         struct ggml_context* ctx,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* emb,
+                                         int num_video_frames) {
+        if (version == VERSION_SVD) {
+            auto block = std::dynamic_pointer_cast<VideoResBlock>(blocks[name]);
+
+            return block->forward(ctx, x, emb, num_video_frames);
+        } else {
+            auto block = std::dynamic_pointer_cast<ResBlock>(blocks[name]);
+
+            return block->forward(ctx, x, emb);
+        }
+    }
+
+    struct ggml_tensor* attention_layer_forward(std::string name,
+                                                struct ggml_context* ctx,
+                                                struct ggml_tensor* x,
+                                                struct ggml_tensor* context,
+                                                int timesteps) {
+        if (version == VERSION_SVD) {
+            auto block = std::dynamic_pointer_cast<SpatialVideoTransformer>(blocks[name]);
+
+            return block->forward(ctx, x, context, timesteps);
+        } else {
+            auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
+
+            return block->forward(ctx, x, context);
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* timesteps,
+                                struct ggml_tensor* context,
+                                struct ggml_tensor* c_concat              = NULL,
+                                struct ggml_tensor* y                     = NULL,
+                                int num_video_frames                      = -1,
+                                std::vector<struct ggml_tensor*> controls = {},
+                                float control_strength                    = 0.f) {
+        // 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]
+        // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
+        // y: [N, adm_in_channels] or [1, adm_in_channels]
+        // return: [N, out_channels, h, w]
+        if (context != NULL) {
+            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]));
+            }
+        }
+
+        if (c_concat != NULL) {
+            if (c_concat->ne[3] != x->ne[3]) {
+                c_concat = ggml_repeat(ctx, c_concat, x);
+            }
+            x = ggml_concat(ctx, x, c_concat);
+        }
+
+        if (y != NULL) {
+            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]));
+            }
+        }
+
+        auto time_embed_0     = std::dynamic_pointer_cast<Linear>(blocks["time_embed.0"]);
+        auto time_embed_2     = std::dynamic_pointer_cast<Linear>(blocks["time_embed.2"]);
+        auto input_blocks_0_0 = std::dynamic_pointer_cast<Conv2d>(blocks["input_blocks.0.0"]);
+
+        auto out_0 = std::dynamic_pointer_cast<GroupNorm32>(blocks["out.0"]);
+        auto out_2 = std::dynamic_pointer_cast<Conv2d>(blocks["out.2"]);
+
+        auto t_emb = ggml_nn_timestep_embedding(ctx, timesteps, model_channels);  // [N, model_channels]
+
+        auto emb = time_embed_0->forward(ctx, t_emb);
+        emb      = ggml_silu_inplace(ctx, emb);
+        emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]
+
+        // SDXL/SVD
+        if (y != NULL) {
+            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      = label_embed_2->forward(ctx, label_emb);  // [N, time_embed_dim]
+
+            emb = ggml_add(ctx, emb, label_emb);  // [N, time_embed_dim]
+        }
+
+        // input_blocks
+        std::vector<struct ggml_tensor*> hs;
+
+        // input block 0
+        auto h = input_blocks_0_0->forward(ctx, x);
+
+        ggml_set_name(h, "bench-start");
+        hs.push_back(h);
+        // input block 1-11
+        size_t len_mults    = channel_mult.size();
+        int input_block_idx = 0;
+        int ds              = 1;
+        for (int i = 0; i < len_mults; i++) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                h                = resblock_forward(name, ctx, h, emb, num_video_frames);  // [N, mult*model_channels, h, w]
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
+                    h                = attention_layer_forward(name, ctx, h, context, num_video_frames);  // [N, mult*model_channels, h, w]
+                }
+                hs.push_back(h);
+            }
+            if (i != len_mults - 1) {
+                ds *= 2;
+                input_block_idx += 1;
+
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                auto block       = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
+
+                h = block->forward(ctx, h);  // [N, mult*model_channels, h/(2^(i+1)), w/(2^(i+1))]
+                hs.push_back(h);
+            }
+        }
+        // [N, 4*model_channels, h/8, w/8]
+
+        // middle_block
+        h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, h, context, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
+
+        if (controls.size() > 0) {
+            auto cs = ggml_scale_inplace(ctx, controls[controls.size() - 1], control_strength);
+            h       = ggml_add(ctx, h, cs);  // middle control
+        }
+        int control_offset = controls.size() - 2;
+
+        // output_blocks
+        int output_block_idx = 0;
+        for (int i = (int)len_mults - 1; i >= 0; i--) {
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                auto h_skip = hs.back();
+                hs.pop_back();
+
+                if (controls.size() > 0) {
+                    auto cs = ggml_scale_inplace(ctx, controls[control_offset], control_strength);
+                    h_skip  = ggml_add(ctx, h_skip, cs);  // control net condition
+                    control_offset--;
+                }
+
+                h = ggml_concat(ctx, h, h_skip);
+
+                std::string name = "output_blocks." + std::to_string(output_block_idx) + ".0";
+
+                h = resblock_forward(name, ctx, h, emb, num_video_frames);
+
+                int up_sample_idx = 1;
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + ".1";
+
+                    h = attention_layer_forward(name, ctx, h, context, num_video_frames);
+
+                    up_sample_idx++;
+                }
+
+                if (i > 0 && j == num_res_blocks) {
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + "." + std::to_string(up_sample_idx);
+                    auto block       = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
+
+                    h = block->forward(ctx, h);
+
+                    ds /= 2;
+                }
+
+                output_block_idx += 1;
+            }
+        }
+
+        // out
+        h = out_0->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);
+        h = out_2->forward(ctx, h);
+        ggml_set_name(h, "bench-end");
+        return h;  // [N, out_channels, h, w]
+    }
+};
+
+struct UNetModel : public GGMLModule {
+    SDVersion version = VERSION_1_x;
+    UnetModelBlock unet;
+
+    UNetModel(ggml_backend_t backend,
+              ggml_type wtype,
+              SDVersion version = VERSION_1_x)
+        : GGMLModule(backend, wtype), unet(version) {
+        unet.init(params_ctx, wtype);
+    }
+
+    std::string get_desc() {
+        return "unet";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        unet.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                    struct ggml_tensor* timesteps,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* c_concat              = NULL,
+                                    struct ggml_tensor* y                     = NULL,
+                                    int num_video_frames                      = -1,
+                                    std::vector<struct ggml_tensor*> controls = {},
+                                    float control_strength                    = 0.f) {
+        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, UNET_GRAPH_SIZE, false);
+
+        if (num_video_frames == -1) {
+            num_video_frames = x->ne[3];
+        }
+
+        x         = to_backend(x);
+        context   = to_backend(context);
+        y         = to_backend(y);
+        timesteps = to_backend(timesteps);
+
+        for (int i = 0; i < controls.size(); i++) {
+            controls[i] = to_backend(controls[i]);
+        }
+
+        struct ggml_tensor* out = unet.forward(compute_ctx,
+                                               x,
+                                               timesteps,
+                                               context,
+                                               c_concat,
+                                               y,
+                                               num_video_frames,
+                                               controls,
+                                               control_strength);
+
+        ggml_build_forward_expand(gf, out);
+
+        return gf;
+    }
+
+    void compute(int n_threads,
+                 struct ggml_tensor* x,
+                 struct ggml_tensor* timesteps,
+                 struct ggml_tensor* context,
+                 struct ggml_tensor* c_concat,
+                 struct ggml_tensor* y,
+                 int num_video_frames                      = -1,
+                 std::vector<struct ggml_tensor*> controls = {},
+                 float control_strength                    = 0.f,
+                 struct ggml_tensor** output               = NULL,
+                 struct ggml_context* output_ctx           = NULL) {
+        // x: [N, in_channels, h, w]
+        // timesteps: [N, ]
+        // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
+        // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
+        // y: [N, adm_in_channels] or [1, adm_in_channels]
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength);
+        };
+
+        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+
+    void test() {
+        struct ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        struct ggml_context* work_ctx = ggml_init(params);
+        GGML_ASSERT(work_ctx != NULL);
+
+        {
+            // CPU, num_video_frames = 1, x{num_video_frames, 8, 8, 8}: Pass
+            // CUDA, num_video_frames = 1, x{num_video_frames, 8, 8, 8}: Pass
+            // CPU, num_video_frames = 3, x{num_video_frames, 8, 8, 8}: Wrong result
+            // CUDA, num_video_frames = 3, x{num_video_frames, 8, 8, 8}: nan
+            int num_video_frames = 3;
+
+            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 8, num_video_frames);
+            std::vector<float> timesteps_vec(num_video_frames, 999.f);
+            auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+            ggml_set_f32(x, 0.5f);
+            // print_ggml_tensor(x);
+
+            auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 1024, 1, num_video_frames);
+            ggml_set_f32(context, 0.5f);
+            // print_ggml_tensor(context);
+
+            auto y = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 768, num_video_frames);
+            ggml_set_f32(y, 0.5f);
+            // print_ggml_tensor(y);
+
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            compute(8, x, timesteps, context, NULL, y, num_video_frames, {}, 0.f, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("unet test done in %dms", t1 - t0);
+        }
+    };
+};
+
+#endif  // __UNET_HPP__

upscaler.cpp

@@ -0,0 +1,123 @@
+#include "esrgan.hpp"
+#include "ggml_extend.hpp"
+#include "model.h"
+#include "stable-diffusion.h"
+
+struct UpscalerGGML {
+    ggml_backend_t backend    = NULL;  // general backend
+    ggml_type model_data_type = GGML_TYPE_F16;
+    std::shared_ptr<ESRGAN> esrgan_upscaler;
+    std::string esrgan_path;
+    int n_threads;
+
+    UpscalerGGML(int n_threads)
+        : n_threads(n_threads) {
+    }
+
+    bool load_from_file(const std::string& esrgan_path) {
+#ifdef SD_USE_CUBLAS
+        LOG_DEBUG("Using CUDA backend");
+        backend = ggml_backend_cuda_init(0);
+#endif
+#ifdef SD_USE_METAL
+        LOG_DEBUG("Using Metal backend");
+        ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
+        backend = ggml_backend_metal_init();
+#endif
+
+        if (!backend) {
+            LOG_DEBUG("Using CPU backend");
+            backend = ggml_backend_cpu_init();
+        }
+        LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_data_type);
+        if (!esrgan_upscaler->load_from_file(esrgan_path)) {
+            return false;
+        }
+        return true;
+    }
+
+    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
+        // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
+        sd_image_t upscaled_image = {0, 0, 0, NULL};
+        int output_width          = (int)input_image.width * esrgan_upscaler->scale;
+        int output_height         = (int)input_image.height * esrgan_upscaler->scale;
+        LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
+                 input_image.width, input_image.height, output_width, output_height);
+
+        struct ggml_init_params params;
+        params.mem_size = output_width * output_height * 3 * sizeof(float) * 2;
+        params.mem_size += 2 * ggml_tensor_overhead();
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        // draft context
+        struct ggml_context* upscale_ctx = ggml_init(params);
+        if (!upscale_ctx) {
+            LOG_ERROR("ggml_init() failed");
+            return upscaled_image;
+        }
+        LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
+        ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
+        sd_image_to_tensor(input_image.data, input_image_tensor);
+
+        ggml_tensor* upscaled = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, output_width, output_height, 3, 1);
+        auto on_tiling        = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+            esrgan_upscaler->compute(n_threads, in, &out);
+        };
+        int64_t t0 = ggml_time_ms();
+        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);
+        esrgan_upscaler->free_compute_buffer();
+        ggml_tensor_clamp(upscaled, 0.f, 1.f);
+        uint8_t* upscaled_data = sd_tensor_to_image(upscaled);
+        ggml_free(upscale_ctx);
+        int64_t t3 = ggml_time_ms();
+        LOG_INFO("input_image_tensor upscaled, taking %.2fs", (t3 - t0) / 1000.0f);
+        upscaled_image = {
+            (uint32_t)output_width,
+            (uint32_t)output_height,
+            3,
+            upscaled_data,
+        };
+        return upscaled_image;
+    }
+};
+
+struct upscaler_ctx_t {
+    UpscalerGGML* upscaler = NULL;
+};
+
+upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
+                                 int n_threads,
+                                 enum sd_type_t wtype) {
+    upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
+    if (upscaler_ctx == NULL) {
+        return NULL;
+    }
+    std::string esrgan_path(esrgan_path_c_str);
+
+    upscaler_ctx->upscaler = new UpscalerGGML(n_threads);
+    if (upscaler_ctx->upscaler == NULL) {
+        return NULL;
+    }
+
+    if (!upscaler_ctx->upscaler->load_from_file(esrgan_path)) {
+        delete upscaler_ctx->upscaler;
+        upscaler_ctx->upscaler = NULL;
+        free(upscaler_ctx);
+        return NULL;
+    }
+    return upscaler_ctx;
+}
+
+sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor) {
+    return upscaler_ctx->upscaler->upscale(input_image, upscale_factor);
+}
+
+void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx) {
+    if (upscaler_ctx->upscaler != NULL) {
+        delete upscaler_ctx->upscaler;
+        upscaler_ctx->upscaler = NULL;
+    }
+    free(upscaler_ctx);
+}

util.cpp

@@ -1,10 +1,16 @@
 #include "util.h"
-
 #include <stdarg.h>
+#include <algorithm>
+#include <cmath>
+#include <codecvt>
 #include <fstream>
+#include <locale>
+#include <sstream>
+#include <string>
 #include <thread>
 #include <unordered_set>
 #include <vector>
+#include "preprocessing.hpp"
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/sysctl.h>
@@ -16,6 +22,12 @@
 #include <unistd.h>
 #endif
 
+#include "ggml/ggml.h"
+#include "stable-diffusion.h"
+
+#define STB_IMAGE_RESIZE_IMPLEMENTATION
+#include "stb_image_resize.h"
+
 bool ends_with(const std::string& str, const std::string& ending) {
     if (str.length() >= ending.length()) {
         return (str.compare(str.length() - ending.length(), ending.length(), ending) == 0);
@@ -31,6 +43,13 @@ bool starts_with(const std::string& str, const std::string& start) {
     return false;
 }
 
+bool contains(const std::string& str, const std::string& substr) {
+    if (str.find(substr) != std::string::npos) {
+        return true;
+    }
+    return false;
+}
+
 void replace_all_chars(std::string& str, char target, char replacement) {
     for (size_t i = 0; i < str.length(); ++i) {
         if (str[i] == target) {
@@ -65,6 +84,57 @@ bool is_directory(const std::string& path) {
     return (attributes != INVALID_FILE_ATTRIBUTES && (attributes & FILE_ATTRIBUTE_DIRECTORY));
 }
 
+std::string get_full_path(const std::string& dir, const std::string& filename) {
+    std::string full_path = dir + "\\" + filename;
+
+    WIN32_FIND_DATA find_file_data;
+    HANDLE hFind = FindFirstFile(full_path.c_str(), &find_file_data);
+
+    if (hFind != INVALID_HANDLE_VALUE) {
+        FindClose(hFind);
+        return full_path;
+    } else {
+        return "";
+    }
+}
+
+std::vector<std::string> get_files_from_dir(const std::string& dir) {
+    std::vector<std::string> files;
+
+    WIN32_FIND_DATA findFileData;
+    HANDLE hFind;
+
+    char currentDirectory[MAX_PATH];
+    GetCurrentDirectory(MAX_PATH, currentDirectory);
+
+    char directoryPath[MAX_PATH];  // this is absolute path
+    sprintf(directoryPath, "%s\\%s\\*", currentDirectory, dir.c_str());
+
+    // Find the first file in the directory
+    hFind = FindFirstFile(directoryPath, &findFileData);
+
+    // Check if the directory was found
+    if (hFind == INVALID_HANDLE_VALUE) {
+        printf("Unable to find directory.\n");
+        return files;
+    }
+
+    // Loop through all files in the directory
+    do {
+        // Check if the found file is a regular file (not a directory)
+        if (!(findFileData.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY)) {
+            files.push_back(std::string(currentDirectory) + "\\" + dir + "\\" + std::string(findFileData.cFileName));
+        }
+    } while (FindNextFile(hFind, &findFileData) != 0);
+
+    // Close the handle
+    FindClose(hFind);
+
+    sort(files.begin(), files.end());
+
+    return files;
+}
+
 #else  // Unix
 #include <dirent.h>
 #include <sys/stat.h>
@@ -79,6 +149,47 @@ bool is_directory(const std::string& path) {
     return (stat(path.c_str(), &buffer) == 0 && S_ISDIR(buffer.st_mode));
 }
 
+// TODO: add windows version
+std::string get_full_path(const std::string& dir, const std::string& filename) {
+    DIR* dp = opendir(dir.c_str());
+
+    if (dp != nullptr) {
+        struct dirent* entry;
+
+        while ((entry = readdir(dp)) != nullptr) {
+            if (strcasecmp(entry->d_name, filename.c_str()) == 0) {
+                closedir(dp);
+                return dir + "/" + entry->d_name;
+            }
+        }
+
+        closedir(dp);
+    }
+
+    return "";
+}
+
+std::vector<std::string> get_files_from_dir(const std::string& dir) {
+    std::vector<std::string> files;
+
+    DIR* dp = opendir(dir.c_str());
+
+    if (dp != nullptr) {
+        struct dirent* entry;
+
+        while ((entry = readdir(dp)) != nullptr) {
+            std::string fname = dir + "/" + entry->d_name;
+            if (!is_directory(fname))
+                files.push_back(fname);
+        }
+        closedir(dp);
+    }
+
+    sort(files.begin(), files.end());
+
+    return files;
+}
+
 #endif
 
 // get_num_physical_cores is copy from
@@ -119,7 +230,25 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-std::string basename(const std::string& path) {
+static sd_progress_cb_t sd_progress_cb = NULL;
+void* sd_progress_cb_data              = NULL;
+
+std::u32string utf8_to_utf32(const std::string& utf8_str) {
+    std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+    return converter.from_bytes(utf8_str);
+}
+
+std::string utf32_to_utf8(const std::u32string& utf32_str) {
+    std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+    return converter.to_bytes(utf32_str);
+}
+
+std::u32string unicode_value_to_utf32(int unicode_value) {
+    std::u32string utf32_string = {static_cast<char32_t>(unicode_value)};
+    return utf32_string;
+}
+
+static std::string sd_basename(const std::string& path) {
     size_t pos = path.find_last_of('/');
     if (pos != std::string::npos) {
         return path.substr(pos + 1);
@@ -147,40 +276,290 @@ std::string path_join(const std::string& p1, const std::string& p2) {
     return p1 + "/" + p2;
 }
 
-static SDLogLevel log_level = SDLogLevel::INFO;
+sd_image_t* preprocess_id_image(sd_image_t* img) {
+    int shortest_edge   = 224;
+    int size            = shortest_edge;
+    sd_image_t* resized = NULL;
+    uint32_t w          = img->width;
+    uint32_t h          = img->height;
+    uint32_t c          = img->channel;
 
-void set_sd_log_level(SDLogLevel level) {
-    log_level = level;
+    // 1. do resize using stb_resize functions
+
+    unsigned char* buf = (unsigned char*)malloc(sizeof(unsigned char) * 3 * size * size);
+    if (!stbir_resize_uint8(img->data, w, h, 0,
+                            buf, size, size, 0,
+                            c)) {
+        fprintf(stderr, "%s: resize operation failed \n ", __func__);
+        return resized;
+    }
+
+    // 2. do center crop (likely unnecessary due to step 1)
+
+    // 3. do rescale
+
+    // 4. do normalize
+
+    // 3 and 4 will need to be done in float format.
+
+    resized = new sd_image_t{(uint32_t)shortest_edge,
+                             (uint32_t)shortest_edge,
+                             3,
+                             buf};
+    return resized;
 }
 
-void log_printf(SDLogLevel level, const char* file, int line, const char* format, ...) {
-    if (level < log_level) {
+void pretty_progress(int step, int steps, float time) {
+    if (sd_progress_cb) {
+        sd_progress_cb(step, steps, time, sd_progress_cb_data);
         return;
     }
+    if (step == 0) {
+        return;
+    }
+    std::string progress = "  |";
+    int max_progress     = 50;
+    int32_t current      = (int32_t)(step * 1.f * max_progress / steps);
+    for (int i = 0; i < 50; i++) {
+        if (i > current) {
+            progress += " ";
+        } else if (i == current && i != max_progress - 1) {
+            progress += ">";
+        } else {
+            progress += "=";
+        }
+    }
+    progress += "|";
+    printf(time > 1.0f ? "\r%s %i/%i - %.2fs/it" : "\r%s %i/%i - %.2fit/s",
+           progress.c_str(), step, steps,
+           time > 1.0f || time == 0 ? time : (1.0f / time));
+    fflush(stdout);  // for linux
+    if (step == steps) {
+        printf("\n");
+    }
+}
+
+std::string ltrim(const std::string& s) {
+    auto it = std::find_if(s.begin(), s.end(), [](int ch) {
+        return !std::isspace(ch);
+    });
+    return std::string(it, s.end());
+}
+
+std::string rtrim(const std::string& s) {
+    auto it = std::find_if(s.rbegin(), s.rend(), [](int ch) {
+        return !std::isspace(ch);
+    });
+    return std::string(s.begin(), it.base());
+}
+
+std::string trim(const std::string& s) {
+    return rtrim(ltrim(s));
+}
+
+static sd_log_cb_t sd_log_cb = NULL;
+void* sd_log_cb_data         = NULL;
+
+#define LOG_BUFFER_SIZE 1024
+
+void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
     va_list args;
     va_start(args, format);
 
-    if (level == SDLogLevel::DEBUG) {
-        printf("[DEBUG] %s:%-4d - ", basename(file).c_str(), line);
-        vprintf(format, args);
-        printf("\n");
-        fflush(stdout);
-    } else if (level == SDLogLevel::INFO) {
-        printf("[INFO]  %s:%-4d - ", basename(file).c_str(), line);
-        vprintf(format, args);
-        printf("\n");
-        fflush(stdout);
-    } else if (level == SDLogLevel::WARN) {
-        fprintf(stdout, "[WARN]  %s:%-4d - ", basename(file).c_str(), line);
-        vfprintf(stdout, format, args);
-        fprintf(stdout, "\n");
-        fflush(stdout);
-    } else {
-        fprintf(stderr, "[ERROR] %s:%-4d - ", basename(file).c_str(), line);
-        vfprintf(stderr, format, args);
-        fprintf(stderr, "\n");
-        fflush(stderr);
+    static char log_buffer[LOG_BUFFER_SIZE + 1];
+    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
+
+    if (written >= 0 && written < LOG_BUFFER_SIZE) {
+        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
+    }
+    strncat(log_buffer, "\n", LOG_BUFFER_SIZE - strlen(log_buffer));
+
+    if (sd_log_cb) {
+        sd_log_cb(level, log_buffer, sd_log_cb_data);
     }
 
     va_end(args);
 }
+
+void sd_set_log_callback(sd_log_cb_t cb, void* data) {
+    sd_log_cb      = cb;
+    sd_log_cb_data = data;
+}
+void sd_set_progress_callback(sd_progress_cb_t cb, void* data) {
+    sd_progress_cb      = cb;
+    sd_progress_cb_data = data;
+}
+const char* sd_get_system_info() {
+    static char buffer[1024];
+    std::stringstream ss;
+    ss << "System Info: \n";
+    ss << "    BLAS = " << ggml_cpu_has_blas() << std::endl;
+    ss << "    SSE3 = " << ggml_cpu_has_sse3() << std::endl;
+    ss << "    AVX = " << ggml_cpu_has_avx() << std::endl;
+    ss << "    AVX2 = " << ggml_cpu_has_avx2() << std::endl;
+    ss << "    AVX512 = " << ggml_cpu_has_avx512() << std::endl;
+    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << std::endl;
+    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << std::endl;
+    ss << "    FMA = " << ggml_cpu_has_fma() << std::endl;
+    ss << "    NEON = " << ggml_cpu_has_neon() << std::endl;
+    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << std::endl;
+    ss << "    F16C = " << ggml_cpu_has_f16c() << std::endl;
+    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << std::endl;
+    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << std::endl;
+    ss << "    VSX = " << ggml_cpu_has_vsx() << std::endl;
+    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
+    return buffer;
+}
+
+const char* sd_type_name(enum sd_type_t type) {
+    return ggml_type_name((ggml_type)type);
+}
+
+sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image) {
+    sd_image_f32_t converted_image;
+    converted_image.width   = image.width;
+    converted_image.height  = image.height;
+    converted_image.channel = image.channel;
+
+    // Allocate memory for float data
+    converted_image.data = (float*)malloc(image.width * image.height * image.channel * sizeof(float));
+
+    for (int i = 0; i < image.width * image.height * image.channel; i++) {
+        // Convert uint8_t to float
+        converted_image.data[i] = (float)image.data[i];
+    }
+
+    return converted_image;
+}
+
+// Function to perform double linear interpolation
+float interpolate(float v1, float v2, float v3, float v4, float x_ratio, float y_ratio) {
+    return v1 * (1 - x_ratio) * (1 - y_ratio) + v2 * x_ratio * (1 - y_ratio) + v3 * (1 - x_ratio) * y_ratio + v4 * x_ratio * y_ratio;
+}
+
+sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height) {
+    sd_image_f32_t resized_image;
+    resized_image.width   = target_width;
+    resized_image.height  = target_height;
+    resized_image.channel = image.channel;
+
+    // Allocate memory for resized float data
+    resized_image.data = (float*)malloc(target_width * target_height * image.channel * sizeof(float));
+
+    for (int y = 0; y < target_height; y++) {
+        for (int x = 0; x < target_width; x++) {
+            float original_x = (float)x * image.width / target_width;
+            float original_y = (float)y * image.height / target_height;
+
+            int x1 = (int)original_x;
+            int y1 = (int)original_y;
+            int x2 = x1 + 1;
+            int y2 = y1 + 1;
+
+            for (int k = 0; k < image.channel; k++) {
+                float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
+                float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
+                float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
+                float v4 = *(image.data + y2 * image.width * image.channel + x2 * image.channel + k);
+
+                float x_ratio = original_x - x1;
+                float y_ratio = original_y - y1;
+
+                float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
+
+                *(resized_image.data + y * target_width * image.channel + x * image.channel + k) = value;
+            }
+        }
+    }
+
+    return resized_image;
+}
+
+void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]) {
+    for (int y = 0; y < image.height; y++) {
+        for (int x = 0; x < image.width; x++) {
+            for (int k = 0; k < image.channel; k++) {
+                int index         = (y * image.width + x) * image.channel + k;
+                image.data[index] = (image.data[index] - means[k]) / stds[k];
+            }
+        }
+    }
+}
+
+// Constants for means and std
+float means[3] = {0.48145466, 0.4578275, 0.40821073};
+float stds[3]  = {0.26862954, 0.26130258, 0.27577711};
+
+// Function to clip and preprocess sd_image_f32_t
+sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
+    float scale = (float)size / fmin(image.width, image.height);
+
+    // Interpolation
+    int new_width       = (int)(scale * image.width);
+    int new_height      = (int)(scale * image.height);
+    float* resized_data = (float*)malloc(new_width * new_height * image.channel * sizeof(float));
+
+    for (int y = 0; y < new_height; y++) {
+        for (int x = 0; x < new_width; x++) {
+            float original_x = (float)x * image.width / new_width;
+            float original_y = (float)y * image.height / new_height;
+
+            int x1 = (int)original_x;
+            int y1 = (int)original_y;
+            int x2 = x1 + 1;
+            int y2 = y1 + 1;
+
+            for (int k = 0; k < image.channel; k++) {
+                float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
+                float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
+                float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
+                float v4 = *(image.data + y2 * image.width * image.channel + x2 * image.channel + k);
+
+                float x_ratio = original_x - x1;
+                float y_ratio = original_y - y1;
+
+                float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
+
+                *(resized_data + y * new_width * image.channel + x * image.channel + k) = value;
+            }
+        }
+    }
+
+    // Clip and preprocess
+    int h = (new_height - size) / 2;
+    int w = (new_width - size) / 2;
+
+    sd_image_f32_t result;
+    result.width   = size;
+    result.height  = size;
+    result.channel = image.channel;
+    result.data    = (float*)malloc(size * size * image.channel * sizeof(float));
+
+    for (int k = 0; k < image.channel; k++) {
+        for (int i = 0; i < size; i++) {
+            for (int j = 0; j < size; j++) {
+                *(result.data + i * size * image.channel + j * image.channel + k) =
+                    fmin(fmax(*(resized_data + (i + h) * new_width * image.channel + (j + w) * image.channel + k), 0.0f), 255.0f) / 255.0f;
+            }
+        }
+    }
+
+    // Free allocated memory
+    free(resized_data);
+
+    // Normalize
+    for (int k = 0; k < image.channel; k++) {
+        for (int i = 0; i < size; i++) {
+            for (int j = 0; j < size; j++) {
+                // *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f;
+                int offset  = i * size * image.channel + j * image.channel + k;
+                float value = *(result.data + offset);
+                value       = (value - means[k]) / stds[k];
+                // value = 0.5f;
+                *(result.data + offset) = value;
+            }
+        }
+    }
+
+    return result;
+}

util.h

@@ -1,10 +1,15 @@
 #ifndef __UTIL_H__
 #define __UTIL_H__
 
+#include <cstdint>
 #include <string>
+#include <vector>
+
+#include "stable-diffusion.h"
 
 bool ends_with(const std::string& str, const std::string& ending);
 bool starts_with(const std::string& str, const std::string& start);
+bool contains(const std::string& str, const std::string& substr);
 
 std::string format(const char* fmt, ...);
 
@@ -12,26 +17,43 @@ void replace_all_chars(std::string& str, char target, char replacement);
 
 bool file_exists(const std::string& filename);
 bool is_directory(const std::string& path);
+std::string get_full_path(const std::string& dir, const std::string& filename);
 
-std::string basename(const std::string& path);
+std::vector<std::string> get_files_from_dir(const std::string& dir);
+
+std::u32string utf8_to_utf32(const std::string& utf8_str);
+std::string utf32_to_utf8(const std::u32string& utf32_str);
+std::u32string unicode_value_to_utf32(int unicode_value);
+
+sd_image_t* preprocess_id_image(sd_image_t* img);
+
+// std::string sd_basename(const std::string& path);
+
+typedef struct {
+    uint32_t width;
+    uint32_t height;
+    uint32_t channel;
+    float* data;
+} sd_image_f32_t;
+
+void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]);
+
+sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image);
+
+sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height);
+
+sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size);
 
 std::string path_join(const std::string& p1, const std::string& p2);
 
-int32_t get_num_physical_cores();
+void pretty_progress(int step, int steps, float time);
 
-enum SDLogLevel {
-    DEBUG,
-    INFO,
-    WARN,
-    ERROR
-};
+void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);
 
-void set_sd_log_level(SDLogLevel level);
+std::string trim(const std::string& s);
 
-void log_printf(SDLogLevel level, const char* file, int line, const char* format, ...);
-
-#define LOG_DEBUG(format, ...) log_printf(SDLogLevel::DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_INFO(format, ...) log_printf(SDLogLevel::INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_WARN(format, ...) log_printf(SDLogLevel::WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_ERROR(format, ...) log_printf(SDLogLevel::ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#endif  // __UTIL_H__
+#define LOG_DEBUG(format, ...) log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_INFO(format, ...) log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_WARN(format, ...) log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_ERROR(format, ...) log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#endif  // __UTIL_H__

vae.hpp

@@ -0,0 +1,605 @@
+#ifndef __VAE_HPP__
+#define __VAE_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+
+/*================================================== AutoEncoderKL ===================================================*/
+
+#define VAE_GRAPH_SIZE 20480
+
+class ResnetBlock : public UnaryBlock {
+protected:
+    int64_t in_channels;
+    int64_t out_channels;
+
+public:
+    ResnetBlock(int64_t in_channels,
+                int64_t out_channels)
+        : in_channels(in_channels),
+          out_channels(out_channels) {
+        // temb_channels is always 0
+        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(out_channels));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        if (out_channels != in_channels) {
+            blocks["nin_shortcut"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {1, 1}));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        // t_emb is always None
+        auto norm1 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm1"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto norm2 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+
+        auto h = x;
+        h      = norm1->forward(ctx, h);
+        h      = ggml_silu_inplace(ctx, h);  // swish
+        h      = conv1->forward(ctx, h);
+        // return h;
+
+        h = norm2->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);  // swish
+        // dropout, skip for inference
+        h = conv2->forward(ctx, h);
+
+        // skip connection
+        if (out_channels != in_channels) {
+            auto nin_shortcut = std::dynamic_pointer_cast<Conv2d>(blocks["nin_shortcut"]);
+
+            x = nin_shortcut->forward(ctx, x);  // [N, out_channels, h, w]
+        }
+
+        h = ggml_add(ctx, h, x);
+        return h;  // [N, out_channels, h, w]
+    }
+};
+
+class AttnBlock : public UnaryBlock {
+protected:
+    int64_t in_channels;
+
+public:
+    AttnBlock(int64_t in_channels)
+        : in_channels(in_channels) {
+        blocks["norm"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        blocks["q"]    = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+        blocks["k"]    = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+        blocks["v"]    = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
+        auto q_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["q"]);
+        auto k_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["k"]);
+        auto v_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["v"]);
+        auto proj_out = std::dynamic_pointer_cast<Conv2d>(blocks["proj_out"]);
+
+        auto h_ = norm->forward(ctx, x);
+
+        const int64_t n = h_->ne[3];
+        const int64_t c = h_->ne[2];
+        const int64_t h = h_->ne[1];
+        const int64_t w = h_->ne[0];
+
+        auto q = q_proj->forward(ctx, h_);                          // [N, in_channels, h, w]
+        q      = ggml_cont(ctx, ggml_permute(ctx, q, 1, 2, 0, 3));  // [N, h, w, in_channels]
+        q      = ggml_reshape_3d(ctx, q, c, h * w, n);              // [N, h * w, in_channels]
+
+        auto k = k_proj->forward(ctx, h_);                          // [N, in_channels, h, w]
+        k      = ggml_cont(ctx, ggml_permute(ctx, k, 1, 2, 0, 3));  // [N, h, w, in_channels]
+        k      = ggml_reshape_3d(ctx, k, c, h * w, n);              // [N, h * w, in_channels]
+
+        auto v = v_proj->forward(ctx, h_);              // [N, in_channels, h, w]
+        v      = ggml_reshape_3d(ctx, v, h * w, c, n);  // [N, in_channels, h * w]
+
+        h_ = ggml_nn_attention(ctx, q, k, v, false);  // [N, h * w, in_channels]
+
+        h_ = ggml_cont(ctx, ggml_permute(ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
+        h_ = ggml_reshape_4d(ctx, h_, w, h, c, n);               // [N, in_channels, h, w]
+
+        h_ = proj_out->forward(ctx, h_);  // [N, in_channels, h, w]
+
+        h_ = ggml_add(ctx, h_, x);
+        return h_;
+    }
+};
+
+class AE3DConv : public Conv2d {
+public:
+    AE3DConv(int64_t in_channels,
+             int64_t out_channels,
+             std::pair<int, int> kernel_size,
+             int64_t video_kernel_size    = 3,
+             std::pair<int, int> stride   = {1, 1},
+             std::pair<int, int> padding  = {0, 0},
+             std::pair<int, int> dilation = {1, 1},
+             bool bias                    = true)
+        : Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias) {
+        int64_t kernel_padding  = video_kernel_size / 2;
+        blocks["time_mix_conv"] = std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(out_channels,
+                                                                             out_channels,
+                                                                             video_kernel_size,
+                                                                             1,
+                                                                             kernel_padding));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x) {
+        // timesteps always None
+        // skip_video always False
+        // x: [N, IC, IH, IW]
+        // result: [N, OC, OH, OW]
+        auto time_mix_conv = std::dynamic_pointer_cast<Conv3dnx1x1>(blocks["time_mix_conv"]);
+
+        x = Conv2d::forward(ctx, x);
+        // timesteps = x.shape[0]
+        // x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+        // x = conv3d(x)
+        // return rearrange(x, "b c t h w -> (b t) c h w")
+        int64_t T = x->ne[3];
+        int64_t B = x->ne[3] / T;
+        int64_t C = x->ne[2];
+        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 = time_mix_conv->forward(ctx, x);                    // [B, OC, T, OH * OW]
+        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
+        return x;                                              // [B*T, OC, OH, OW]
+    }
+};
+
+class VideoResnetBlock : public ResnetBlock {
+protected:
+    void init_params(struct ggml_context* ctx, ggml_type wtype) {
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+    }
+
+    float get_alpha() {
+        float alpha = ggml_backend_tensor_get_f32(params["mix_factor"]);
+        return sigmoid(alpha);
+    }
+
+public:
+    VideoResnetBlock(int64_t in_channels,
+                     int64_t out_channels,
+                     int video_kernel_size = 3)
+        : ResnetBlock(in_channels, out_channels) {
+        // merge_strategy is always learned
+        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, 0, out_channels, {video_kernel_size, 1}, 3, false, true));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w]
+        // return: [N, out_channels, h, w] aka [b*t, out_channels, h, w]
+        // t_emb is always None
+        // skip_video is always False
+        // timesteps is always None
+        auto time_stack = std::dynamic_pointer_cast<ResBlock>(blocks["time_stack"]);
+
+        x = ResnetBlock::forward(ctx, x);  // [N, out_channels, h, w]
+        // return x;
+
+        int64_t T = x->ne[3];
+        int64_t B = x->ne[3] / T;
+        int64_t C = x->ne[2];
+        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)
+        auto x_mix = x;
+
+        x = time_stack->forward(ctx, x);  // b t c (h w)
+
+        float alpha = get_alpha();
+        x           = ggml_add(ctx,
+                               ggml_scale(ctx, x, alpha),
+                               ggml_scale(ctx, x_mix, 1.0f - alpha));
+
+        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
+
+        return x;
+    }
+};
+
+// ldm.modules.diffusionmodules.model.Encoder
+class Encoder : public GGMLBlock {
+protected:
+    int ch                   = 128;
+    std::vector<int> ch_mult = {1, 2, 4, 4};
+    int num_res_blocks       = 2;
+    int in_channels          = 3;
+    int z_channels           = 4;
+    bool double_z            = true;
+
+public:
+    Encoder(int ch,
+            std::vector<int> ch_mult,
+            int num_res_blocks,
+            int in_channels,
+            int z_channels,
+            bool double_z = true)
+        : ch(ch),
+          ch_mult(ch_mult),
+          num_res_blocks(num_res_blocks),
+          in_channels(in_channels),
+          z_channels(z_channels),
+          double_z(double_z) {
+        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, ch, {3, 3}, {1, 1}, {1, 1}));
+
+        size_t num_resolutions = ch_mult.size();
+
+        int block_in = 1;
+        for (int i = 0; i < num_resolutions; i++) {
+            if (i == 0) {
+                block_in = ch;
+            } else {
+                block_in = ch * ch_mult[i - 1];
+            }
+            int block_out = ch * ch_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_out));
+                block_in         = block_out;
+            }
+            if (i != num_resolutions - 1) {
+                std::string name = "down." + std::to_string(i) + ".downsample";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(block_in, block_in, true));
+            }
+        }
+
+        blocks["mid.block_1"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
+        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in));
+        blocks["mid.block_2"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
+
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
+        blocks["conv_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(block_in, double_z ? z_channels * 2 : z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+
+        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
+        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
+        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
+        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
+        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
+        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
+
+        auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
+
+        // downsampling
+        size_t num_resolutions = ch_mult.size();
+        for (int i = 0; i < num_resolutions; i++) {
+            for (int j = 0; j < num_res_blocks; j++) {
+                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
+                auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
+
+                h = down_block->forward(ctx, h);
+            }
+            if (i != num_resolutions - 1) {
+                std::string name = "down." + std::to_string(i) + ".downsample";
+                auto down_sample = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
+
+                h = down_sample->forward(ctx, h);
+            }
+        }
+
+        // middle
+        h = mid_block_1->forward(ctx, h);
+        h = mid_attn_1->forward(ctx, h);
+        h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+
+        // end
+        h = norm_out->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);  // nonlinearity/swish
+        h = conv_out->forward(ctx, h);  // [N, z_channels*2, h, w]
+        return h;
+    }
+};
+
+// ldm.modules.diffusionmodules.model.Decoder
+class Decoder : public GGMLBlock {
+protected:
+    int ch                   = 128;
+    int out_ch               = 3;
+    std::vector<int> ch_mult = {1, 2, 4, 4};
+    int num_res_blocks       = 2;
+    int z_channels           = 4;
+    bool video_decoder       = false;
+    int video_kernel_size    = 3;
+
+    virtual std::shared_ptr<GGMLBlock> get_conv_out(int64_t in_channels,
+                                                    int64_t out_channels,
+                                                    std::pair<int, int> kernel_size,
+                                                    std::pair<int, int> stride  = {1, 1},
+                                                    std::pair<int, int> padding = {0, 0}) {
+        if (video_decoder) {
+            return std::shared_ptr<GGMLBlock>(new AE3DConv(in_channels, out_channels, kernel_size, video_kernel_size, stride, padding));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, stride, padding));
+        }
+    }
+
+    virtual std::shared_ptr<GGMLBlock> get_resnet_block(int64_t in_channels,
+                                                        int64_t out_channels) {
+        if (video_decoder) {
+            return std::shared_ptr<GGMLBlock>(new VideoResnetBlock(in_channels, out_channels, video_kernel_size));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new ResnetBlock(in_channels, out_channels));
+        }
+    }
+
+public:
+    Decoder(int ch,
+            int out_ch,
+            std::vector<int> ch_mult,
+            int num_res_blocks,
+            int z_channels,
+            bool video_decoder    = false,
+            int video_kernel_size = 3)
+        : ch(ch),
+          out_ch(out_ch),
+          ch_mult(ch_mult),
+          num_res_blocks(num_res_blocks),
+          z_channels(z_channels),
+          video_decoder(video_decoder),
+          video_kernel_size(video_kernel_size) {
+        size_t num_resolutions = ch_mult.size();
+        int block_in           = ch * ch_mult[num_resolutions - 1];
+
+        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, block_in, {3, 3}, {1, 1}, {1, 1}));
+
+        blocks["mid.block_1"] = get_resnet_block(block_in, block_in);
+        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in));
+        blocks["mid.block_2"] = get_resnet_block(block_in, block_in);
+
+        for (int i = num_resolutions - 1; i >= 0; i--) {
+            int mult      = ch_mult[i];
+            int block_out = ch * mult;
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
+                blocks[name]     = get_resnet_block(block_in, block_out);
+
+                block_in = block_out;
+            }
+            if (i != 0) {
+                std::string name = "up." + std::to_string(i) + ".upsample";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new UpSampleBlock(block_in, block_in));
+            }
+        }
+
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
+        blocks["conv_out"] = get_conv_out(block_in, out_ch, {3, 3}, {1, 1}, {1, 1});
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* z) {
+        // z: [N, z_channels, h, w]
+        // alpha is always 0
+        // merge_strategy is always learned
+        // time_mode is always conv-only, so we need to replace conv_out_op/resnet_op to AE3DConv/VideoResBlock
+        // AttnVideoBlock will not be used
+        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
+        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
+        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
+        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
+        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
+        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
+
+        // conv_in
+        auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
+
+        // middle
+        h = mid_block_1->forward(ctx, h);
+        // return h;
+
+        h = mid_attn_1->forward(ctx, h);
+        h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+
+        // upsampling
+        size_t num_resolutions = ch_mult.size();
+        for (int i = num_resolutions - 1; i >= 0; i--) {
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
+                auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
+
+                h = up_block->forward(ctx, h);
+            }
+            if (i != 0) {
+                std::string name = "up." + std::to_string(i) + ".upsample";
+                auto up_sample   = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
+
+                h = up_sample->forward(ctx, h);
+            }
+        }
+
+        h = norm_out->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);  // nonlinearity/swish
+        h = conv_out->forward(ctx, h);  // [N, out_ch, h*8, w*8]
+        return h;
+    }
+};
+
+// ldm.models.autoencoder.AutoencoderKL
+class AutoencodingEngine : public GGMLBlock {
+protected:
+    bool decode_only       = true;
+    bool use_video_decoder = false;
+    int embed_dim          = 4;
+    struct {
+        int z_channels           = 4;
+        int resolution           = 256;
+        int in_channels          = 3;
+        int out_ch               = 3;
+        int ch                   = 128;
+        std::vector<int> ch_mult = {1, 2, 4, 4};
+        int num_res_blocks       = 2;
+        bool double_z            = true;
+    } dd_config;
+
+public:
+    AutoencodingEngine(bool decode_only       = true,
+                       bool use_video_decoder = false)
+        : decode_only(decode_only), use_video_decoder(use_video_decoder) {
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder(dd_config.ch,
+                                                                   dd_config.out_ch,
+                                                                   dd_config.ch_mult,
+                                                                   dd_config.num_res_blocks,
+                                                                   dd_config.z_channels,
+                                                                   use_video_decoder));
+        if (!use_video_decoder) {
+            blocks["post_quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(dd_config.z_channels,
+                                                                              embed_dim,
+                                                                              {1, 1}));
+        }
+        if (!decode_only) {
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder(dd_config.ch,
+                                                                       dd_config.ch_mult,
+                                                                       dd_config.num_res_blocks,
+                                                                       dd_config.in_channels,
+                                                                       dd_config.z_channels,
+                                                                       dd_config.double_z));
+            if (!use_video_decoder) {
+                int factor = dd_config.double_z ? 2 : 1;
+
+                blocks["quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(embed_dim * factor,
+                                                                             dd_config.z_channels * factor,
+                                                                             {1, 1}));
+            }
+        }
+    }
+
+    struct ggml_tensor* decode(struct ggml_context* ctx, struct ggml_tensor* z) {
+        // z: [N, z_channels, h, w]
+        if (!use_video_decoder) {
+            auto post_quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["post_quant_conv"]);
+            z                    = post_quant_conv->forward(ctx, z);  // [N, z_channels, h, w]
+        }
+        auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
+
+        ggml_set_name(z, "bench-start");
+        auto h = decoder->forward(ctx, z);
+        ggml_set_name(h, "bench-end");
+        return h;
+    }
+
+    struct ggml_tensor* encode(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        auto encoder = std::dynamic_pointer_cast<Encoder>(blocks["encoder"]);
+
+        auto h = encoder->forward(ctx, x);  // [N, 2*z_channels, h/8, w/8]
+        if (!use_video_decoder) {
+            auto quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["quant_conv"]);
+            h               = quant_conv->forward(ctx, h);  // [N, 2*embed_dim, h/8, w/8]
+        }
+        return h;
+    }
+};
+
+struct AutoEncoderKL : public GGMLModule {
+    bool decode_only = true;
+    AutoencodingEngine ae;
+
+    AutoEncoderKL(ggml_backend_t backend,
+                  ggml_type wtype,
+                  bool decode_only       = false,
+                  bool use_video_decoder = false)
+        : decode_only(decode_only), ae(decode_only, use_video_decoder), GGMLModule(backend, wtype) {
+        ae.init(params_ctx, wtype);
+    }
+
+    std::string get_desc() {
+        return "vae";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        ae.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        z = to_backend(z);
+
+        struct ggml_tensor* out = decode_graph ? ae.decode(compute_ctx, z) : ae.encode(compute_ctx, z);
+
+        ggml_build_forward_expand(gf, out);
+
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* z,
+                 bool decode_graph,
+                 struct ggml_tensor** output,
+                 struct ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(z, decode_graph);
+        };
+        // ggml_set_f32(z, 0.5f);
+        // print_ggml_tensor(z);
+        GGMLModule::compute(get_graph, n_threads, true, output, output_ctx);
+    }
+
+    void test() {
+        struct ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        struct ggml_context* work_ctx = ggml_init(params);
+        GGML_ASSERT(work_ctx != NULL);
+
+        {
+            // CPU, x{1, 3, 64, 64}: Pass
+            // CUDA, x{1, 3, 64, 64}: Pass, but sill get wrong result for some image, may be due to interlnal nan
+            // CPU, x{2, 3, 64, 64}: Wrong result
+            // CUDA, x{2, 3, 64, 64}: Wrong result, and different from CPU result
+            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 64, 64, 3, 2);
+            ggml_set_f32(x, 0.5f);
+            print_ggml_tensor(x);
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            compute(8, x, false, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("encode test done in %dms", t1 - t0);
+        }
+
+        if (false) {
+            // CPU, z{1, 4, 8, 8}: Pass
+            // CUDA, z{1, 4, 8, 8}: Pass
+            // CPU, z{3, 4, 8, 8}: Wrong result
+            // CUDA, z{3, 4, 8, 8}: Wrong result, and different from CPU result
+            auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
+            ggml_set_f32(z, 0.5f);
+            print_ggml_tensor(z);
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            compute(8, z, true, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("decode test done in %dms", t1 - t0);
+        }
+    };
+};
+
+#endif