fix: avoid double free and fix sdxl lora naming conversion

* Fixed a double free issue when running multiple backends on the CPU, eg: CLIP
and the primary backend, as this would result in the *_backend pointers both
pointing to the same thing resulting in a segfault when calling the
StableDiffusionGGML destructor.

* Improve logging to allow for a color switch on the command line interface.
Changed the base log_printf function to not bake the log level directly into
the log buffer as that information is already passed the logging function via
the level parameter and it's easier to add in there than strip it out.

* Added a fix for certain SDXL LoRAs that don't seem to follow the expected
naming convention, converts over the tensor name during the LoRA model
loading. Added some logging of useful LoRA loading information. Had to
increase the base size of the GGML graph as the existing size results in an
insufficient graph memory error when using SDXL LoRAs.

* small fixes

---------

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

.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' }}

CMakeLists.txt

@@ -28,7 +28,8 @@ 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)
@@ -59,16 +60,25 @@ endif()
 
 set(SD_LIB stable-diffusion)
 
-add_library(${SD_LIB} stable-diffusion.h stable-diffusion.cpp model.h model.cpp util.h util.cpp upscaler.cpp
-             ggml_extend.hpp clip.hpp common.hpp unet.hpp tae.hpp esrgan.hpp lora.hpp denoiser.hpp rng.hpp rng_philox.hpp)
+file(GLOB SD_LIB_SOURCES 
+    "*.h"
+    "*.cpp"
+    "*.hpp"
+)
 
-if(BUILD_SHARED_LIBS)
+# 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()
 
 

README.md

@@ -14,6 +14,7 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
     - !!!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
@@ -60,6 +61,9 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 
 ## 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
 
 ```
@@ -160,7 +164,11 @@ arguments:
   --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
@@ -171,6 +179,7 @@ arguments:
   -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)
@@ -186,6 +195,7 @@ arguments:
                                      <= 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
 ```
 
@@ -294,6 +304,39 @@ You can use ESRGAN to upscale the generated images. At the moment, only the [Rea
 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
@@ -317,6 +360,13 @@ 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)
+
 ## Contributors
 
 Thank you to all the people who have already contributed to stable-diffusion.cpp!
@@ -329,6 +379,8 @@ 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

@@ -3,541 +3,527 @@
 
 #include "ggml_extend.hpp"
 
-struct DownSample {
-    // hparams
+class DownSampleBlock : public GGMLBlock {
+protected:
     int channels;
     int out_channels;
+    bool vae_downsample;
 
-    // conv2d params
-    struct ggml_tensor* op_w;  // [out_channels, channels, 3, 3]
-    struct ggml_tensor* op_b;  // [out_channels,]
-
-    bool vae_downsample = false;
-
-    size_t calculate_mem_size(ggml_type wtype) {
-        size_t mem_size = 0;
-        mem_size += ggml_row_size(GGML_TYPE_F16, out_channels * channels * 3 * 3);  // op_w
-        mem_size += ggml_row_size(GGML_TYPE_F32, out_channels);                     // op_b
-        return mem_size;
-    }
-
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        op_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, out_channels);
-        op_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-    }
-
-    void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+public:
+    DownSampleBlock(int channels,
+                    int out_channels,
+                    bool vae_downsample = false)
+        : channels(channels),
+          out_channels(out_channels),
+          vae_downsample(vae_downsample) {
         if (vae_downsample) {
-            tensors[prefix + "conv.weight"] = op_w;
-            tensors[prefix + "conv.bias"]   = op_b;
+            blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {2, 2}, {0, 0}));
         } else {
-            tensors[prefix + "op.weight"] = op_w;
-            tensors[prefix + "op.bias"]   = op_b;
+            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]
-        struct ggml_tensor* c = NULL;
         if (vae_downsample) {
-            c = ggml_pad(ctx, x, 1, 1, 0, 0);
-            c = ggml_nn_conv_2d(ctx, c, op_w, op_b, 2, 2, 0, 0);
+            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
+
+            x = ggml_pad(ctx, x, 1, 1, 0, 0);
+            x = conv->forward(ctx, x);
         } else {
-            c = ggml_nn_conv_2d(ctx, x, op_w, op_b, 2, 2, 1, 1);
+            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["op"]);
+
+            x = conv->forward(ctx, x);
         }
-        return c;  // [N, out_channels, h/2, w/2]
+        return x;  // [N, out_channels, h/2, w/2]
     }
 };
 
-struct UpSample {
-    // hparams
+class UpSampleBlock : public GGMLBlock {
+protected:
     int channels;
     int out_channels;
 
-    // conv2d params
-    struct ggml_tensor* conv_w;  // [out_channels, channels, 3, 3]
-    struct ggml_tensor* conv_b;  // [out_channels,]
-
-    size_t calculate_mem_size(ggml_type wtype) {
-        size_t mem_size = 0;
-        mem_size += ggml_row_size(GGML_TYPE_F16, out_channels * channels * 3 * 3);  // op_w
-        mem_size += ggml_row_size(GGML_TYPE_F32, out_channels);                     // op_b
-        return mem_size;
-    }
-
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        conv_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, out_channels);
-        conv_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-    }
-
-    void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
-        tensors[prefix + "conv.weight"] = conv_w;
-        tensors[prefix + "conv.bias"]   = conv_b;
+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]
-        x = ggml_upscale(ctx, x, 2);                              // [N, channels, h*2, w*2]
-        x = ggml_nn_conv_2d(ctx, x, conv_w, conv_b, 1, 1, 1, 1);  // [N, out_channels, h*2, w*2]
+        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;
     }
 };
 
-struct ResBlock {
+class ResBlock : public GGMLBlock {
+protected:
     // network hparams
-    int channels;      // model_channels * (1, 1, 1, 2, 2, 4, 4, 4)
-    int emb_channels;  // time_embed_dim
-    int out_channels;  // mult * model_channels
+    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;
 
-    // network params
-    // in_layers
-    struct ggml_tensor* in_layer_0_w;  // [channels, ]
-    struct ggml_tensor* in_layer_0_b;  // [channels, ]
-    // in_layer_1 is nn.SILU()
-    struct ggml_tensor* in_layer_2_w;  // [out_channels, channels, 3, 3]
-    struct ggml_tensor* in_layer_2_b;  // [out_channels, ]
-
-    // emb_layers
-    // emb_layer_0 is nn.SILU()
-    struct ggml_tensor* emb_layer_1_w;  // [out_channels, emb_channels]
-    struct ggml_tensor* emb_layer_1_b;  // [out_channels, ]
-
-    // out_layers
-    struct ggml_tensor* out_layer_0_w;  // [out_channels, ]
-    struct ggml_tensor* out_layer_0_b;  // [out_channels, ]
-    // out_layer_1 is nn.SILU()
-    // out_layer_2 is nn.Dropout(), p = 0 for inference
-    struct ggml_tensor* out_layer_3_w;  // [out_channels, out_channels, 3, 3]
-    struct ggml_tensor* out_layer_3_b;  // [out_channels, ]
-
-    // skip connection, only if out_channels != channels
-    struct ggml_tensor* skip_w;  // [out_channels, channels, 1, 1]
-    struct ggml_tensor* skip_b;  // [out_channels, ]
-
-    size_t calculate_mem_size(ggml_type wtype) {
-        size_t mem_size = 0;
-        mem_size += 2 * ggml_row_size(GGML_TYPE_F32, channels);                         // in_layer_0_w/b
-        mem_size += ggml_row_size(GGML_TYPE_F16, out_channels * channels * 3 * 3);      // in_layer_2_w
-        mem_size += 5 * ggml_row_size(GGML_TYPE_F32, out_channels);                     // in_layer_2_b/emb_layer_1_b/out_layer_0_w/out_layer_0_b/out_layer_3_b
-        mem_size += ggml_row_size(wtype, out_channels * emb_channels);                  // emb_layer_1_w
-        mem_size += ggml_row_size(GGML_TYPE_F16, out_channels * out_channels * 3 * 3);  // out_layer_3_w
-
-        if (out_channels != channels) {
-            mem_size += ggml_row_size(GGML_TYPE_F16, out_channels * channels * 1 * 1);  // skip_w
-            mem_size += ggml_row_size(GGML_TYPE_F32, out_channels);                     // skip_b
-        }
-        return mem_size;
-    }
-
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        in_layer_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
-        in_layer_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
-        in_layer_2_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, out_channels);
-        in_layer_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-
-        emb_layer_1_w = ggml_new_tensor_2d(ctx, wtype, emb_channels, out_channels);
-        emb_layer_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-
-        out_layer_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-        out_layer_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-        out_layer_3_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, out_channels, out_channels);
-        out_layer_3_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-
-        if (out_channels != channels) {
-            skip_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, channels, out_channels);
-            skip_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
+    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));
         }
     }
 
-    void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
-        tensors[prefix + "in_layers.0.weight"] = in_layer_0_w;
-        tensors[prefix + "in_layers.0.bias"]   = in_layer_0_b;
-        tensors[prefix + "in_layers.2.weight"] = in_layer_2_w;
-        tensors[prefix + "in_layers.2.bias"]   = in_layer_2_b;
+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);
 
-        tensors[prefix + "emb_layers.1.weight"] = emb_layer_1_w;
-        tensors[prefix + "emb_layers.1.bias"]   = emb_layer_1_b;
+        if (!skip_t_emb) {
+            // emb_layer_0 is nn.SILU()
+            blocks["emb_layers.1"] = std::shared_ptr<GGMLBlock>(new Linear(emb_channels, out_channels));
+        }
 
-        tensors[prefix + "out_layers.0.weight"] = out_layer_0_w;
-        tensors[prefix + "out_layers.0.bias"]   = out_layer_0_b;
-        tensors[prefix + "out_layers.3.weight"] = out_layer_3_w;
-        tensors[prefix + "out_layers.3.bias"]   = out_layer_3_b;
+        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) {
-            tensors[prefix + "skip_connection.weight"] = skip_w;
-            tensors[prefix + "skip_connection.bias"]   = skip_b;
+            blocks["skip_connection"] = conv_nd(dims, channels, out_channels, {1, 1}, {0, 0});
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb) {
-        // x: [N, channels, h, w]
-        // emb: [N, emb_channels]
+    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 = ggml_nn_group_norm(ctx, x, in_layer_0_w, in_layer_0_b);
+        auto h = in_layers_0->forward(ctx, x);
         h      = ggml_silu_inplace(ctx, h);
-        h      = ggml_nn_conv_2d(ctx, h, in_layer_2_w, in_layer_2_b, 1, 1, 1, 1);  // [N, out_channels, h, w]
+        h      = in_layers_2->forward(ctx, h);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
 
         // emb_layers
-        auto emb_out = ggml_silu(ctx, emb);
-        emb_out      = ggml_nn_linear(ctx, emb_out, emb_layer_1_w, emb_layer_1_b);           // [N, out_channels]
-        emb_out      = ggml_reshape_4d(ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]);  // [N, out_channels, 1, 1]
+        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 = ggml_add(ctx, h, emb_out);
-        h = ggml_nn_group_norm(ctx, h, out_layer_0_w, out_layer_0_b);
+        h = out_layers_0->forward(ctx, h);
         h = ggml_silu_inplace(ctx, h);
-
         // dropout, skip for inference
-
-        h = ggml_nn_conv_2d(ctx, h, out_layer_3_w, out_layer_3_b, 1, 1, 1, 1);  // [N, out_channels, h, w]
+        h = out_layers_3->forward(ctx, h);
 
         // skip connection
         if (out_channels != channels) {
-            x = ggml_nn_conv_2d(ctx, x, skip_w, skip_b);  // [N, out_channels, h, w]
+            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]
+        return h;  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
     }
 };
 
-struct SpatialTransformer {
-    int in_channels;        // mult * model_channels
-    int n_head;             // num_heads
-    int d_head;             // in_channels // n_heads
-    int depth       = 1;    // 1
-    int context_dim = 768;  // hidden_size, 1024 for VERSION_2_x
+class GEGLU : public GGMLBlock {
+protected:
+    int64_t dim_in;
+    int64_t dim_out;
 
-    // group norm
-    struct ggml_tensor* norm_w;  // [in_channels,]
-    struct ggml_tensor* norm_b;  // [in_channels,]
-
-    // proj_in
-    struct ggml_tensor* proj_in_w;  // [in_channels, in_channels, 1, 1]
-    struct ggml_tensor* proj_in_b;  // [in_channels,]
-
-    // transformer
-    struct Transformer {
-        // layer norm 1
-        struct ggml_tensor* norm1_w;  // [in_channels, ]
-        struct ggml_tensor* norm1_b;  // [in_channels, ]
-
-        // attn1
-        struct ggml_tensor* attn1_q_w;  // [in_channels, in_channels]
-        struct ggml_tensor* attn1_k_w;  // [in_channels, in_channels]
-        struct ggml_tensor* attn1_v_w;  // [in_channels, in_channels]
-
-        struct ggml_tensor* attn1_out_w;  // [in_channels, in_channels]
-        struct ggml_tensor* attn1_out_b;  // [in_channels, ]
-
-        // layer norm 2
-        struct ggml_tensor* norm2_w;  // [in_channels, ]
-        struct ggml_tensor* norm2_b;  // [in_channels, ]
-
-        // attn2
-        struct ggml_tensor* attn2_q_w;  // [in_channels, in_channels]
-        struct ggml_tensor* attn2_k_w;  // [in_channels, context_dim]
-        struct ggml_tensor* attn2_v_w;  // [in_channels, context_dim]
-
-        struct ggml_tensor* attn2_out_w;  // [in_channels, in_channels]
-        struct ggml_tensor* attn2_out_b;  // [in_channels, ]
-
-        // layer norm 3
-        struct ggml_tensor* norm3_w;  // [in_channels, ]
-        struct ggml_tensor* norm3_b;  // [in_channels, ]
-
-        // ff
-        struct ggml_tensor* ff_0_proj_w;  // [in_channels * 4 * 2, in_channels]
-        struct ggml_tensor* ff_0_proj_b;  // [in_channels * 4 * 2]
-
-        struct ggml_tensor* ff_2_w;  // [in_channels, in_channels * 4]
-        struct ggml_tensor* ff_2_b;  // [in_channels,]
-    };
-
-    std::vector<Transformer> transformers;
-
-    // proj_out
-    struct ggml_tensor* proj_out_w;  // [in_channels, in_channels, 1, 1]
-    struct ggml_tensor* proj_out_b;  // [in_channels,]
-
-    SpatialTransformer(int depth = 1)
-        : depth(depth) {
-        transformers.resize(depth);
+    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);
     }
 
-    int get_num_tensors() {
-        return depth * 20 + 7;
+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));
     }
 
-    size_t calculate_mem_size(ggml_type wtype) {
-        size_t mem_size = 0;
-        mem_size += 2 * ggml_row_size(GGML_TYPE_F32, in_channels);                        // norm_w/norm_b
-        mem_size += 2 * ggml_row_size(GGML_TYPE_F16, in_channels * in_channels * 1 * 1);  // proj_in_w/proj_out_w
-        mem_size += 2 * ggml_row_size(GGML_TYPE_F32, in_channels);                        // proj_in_b/proj_out_b
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [ne3, ne2, ne1, dim]
+        // return: [ne3, ne2, ne1, dim_out]
 
-        // transformer
-        for (auto& transformer : transformers) {
-            mem_size += 6 * ggml_row_size(GGML_TYPE_F32, in_channels);            // norm1-3_w/b
-            mem_size += 6 * ggml_row_size(wtype, in_channels * in_channels);      // attn1_q/k/v/out_w attn2_q/out_w
-            mem_size += 2 * ggml_row_size(wtype, in_channels * context_dim);      // attn2_k/v_w
-            mem_size += ggml_row_size(wtype, in_channels * 4 * 2 * in_channels);  // ff_0_proj_w
-            mem_size += ggml_row_size(GGML_TYPE_F32, in_channels * 4 * 2);        // ff_0_proj_b
-            mem_size += ggml_row_size(wtype, in_channels * 4 * in_channels);      // ff_2_w
-            mem_size += ggml_row_size(GGML_TYPE_F32, in_channels);                // ff_2_b
-        }
-        return mem_size;
+        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;
     }
+};
 
-    void init_params(struct ggml_context* ctx, ggml_allocr* alloc, ggml_type wtype) {
-        norm_w    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-        norm_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-        proj_in_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, in_channels, in_channels);
-        proj_in_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
+class CrossAttention : public GGMLBlock {
+protected:
+    int64_t query_dim;
+    int64_t context_dim;
+    int64_t n_head;
+    int64_t d_head;
 
-        proj_out_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, in_channels, in_channels);
-        proj_out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
+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;
 
-        // transformer
-        for (auto& transformer : transformers) {
-            transformer.norm1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-            transformer.norm1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
+        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));
 
-            transformer.attn1_q_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels);
-            transformer.attn1_k_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels);
-            transformer.attn1_v_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels);
-
-            transformer.attn1_out_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels);
-            transformer.attn1_out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-
-            transformer.norm2_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-            transformer.norm2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-
-            transformer.attn2_q_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels);
-            transformer.attn2_k_w = ggml_new_tensor_2d(ctx, wtype, context_dim, in_channels);
-            transformer.attn2_v_w = ggml_new_tensor_2d(ctx, wtype, context_dim, in_channels);
-
-            transformer.attn2_out_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels);
-            transformer.attn2_out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-
-            transformer.norm3_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-            transformer.norm3_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-
-            transformer.ff_0_proj_w = ggml_new_tensor_2d(ctx, wtype, in_channels, in_channels * 4 * 2);
-            transformer.ff_0_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels * 4 * 2);
-
-            transformer.ff_2_w = ggml_new_tensor_2d(ctx, wtype, in_channels * 4, in_channels);
-            transformer.ff_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-        }
-    }
-
-    void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
-        tensors[prefix + "norm.weight"]    = norm_w;
-        tensors[prefix + "norm.bias"]      = norm_b;
-        tensors[prefix + "proj_in.weight"] = proj_in_w;
-        tensors[prefix + "proj_in.bias"]   = proj_in_b;
-
-        // transformer
-        for (int i = 0; i < transformers.size(); i++) {
-            auto& transformer                                 = transformers[i];
-            std::string transformer_prefix                    = prefix + "transformer_blocks." + std::to_string(i) + ".";
-            tensors[transformer_prefix + "attn1.to_q.weight"] = transformer.attn1_q_w;
-            tensors[transformer_prefix + "attn1.to_k.weight"] = transformer.attn1_k_w;
-            tensors[transformer_prefix + "attn1.to_v.weight"] = transformer.attn1_v_w;
-
-            tensors[transformer_prefix + "attn1.to_out.0.weight"] = transformer.attn1_out_w;
-            tensors[transformer_prefix + "attn1.to_out.0.bias"]   = transformer.attn1_out_b;
-
-            tensors[transformer_prefix + "ff.net.0.proj.weight"] = transformer.ff_0_proj_w;
-            tensors[transformer_prefix + "ff.net.0.proj.bias"]   = transformer.ff_0_proj_b;
-            tensors[transformer_prefix + "ff.net.2.weight"]      = transformer.ff_2_w;
-            tensors[transformer_prefix + "ff.net.2.bias"]        = transformer.ff_2_b;
-
-            tensors[transformer_prefix + "attn2.to_q.weight"] = transformer.attn2_q_w;
-            tensors[transformer_prefix + "attn2.to_k.weight"] = transformer.attn2_k_w;
-            tensors[transformer_prefix + "attn2.to_v.weight"] = transformer.attn2_v_w;
-
-            tensors[transformer_prefix + "attn2.to_out.0.weight"] = transformer.attn2_out_w;
-            tensors[transformer_prefix + "attn2.to_out.0.bias"]   = transformer.attn2_out_b;
-
-            tensors[transformer_prefix + "norm1.weight"] = transformer.norm1_w;
-            tensors[transformer_prefix + "norm1.bias"]   = transformer.norm1_b;
-            tensors[transformer_prefix + "norm2.weight"] = transformer.norm2_w;
-            tensors[transformer_prefix + "norm2.bias"]   = transformer.norm2_b;
-            tensors[transformer_prefix + "norm3.weight"] = transformer.norm3_w;
-            tensors[transformer_prefix + "norm3.bias"]   = transformer.norm3_b;
-        }
-
-        tensors[prefix + "proj_out.weight"] = proj_out_w;
-        tensors[prefix + "proj_out.bias"]   = proj_out_b;
+        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, in_channels, h, w]
-        // context: [N, max_position, hidden_size(aka context_dim)]
-        auto x_in = x;
-        x         = ggml_nn_group_norm(ctx, x, norm_w, norm_b);
-        // proj_in
-        x = ggml_nn_conv_2d(ctx, x, proj_in_w, proj_in_b);  // [N, in_channels, h, w]
+        // 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"]);
 
-        // transformer
-        const int64_t n            = x->ne[3];
-        const int64_t c            = x->ne[2];
-        const int64_t h            = x->ne[1];
-        const int64_t w            = x->ne[0];
-        const int64_t max_position = context->ne[1];
-        x                          = ggml_cont(ctx, ggml_permute(ctx, x, 1, 2, 0, 3));  // [N, h, w, in_channels]
+        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;
 
-        for (auto& transformer : transformers) {
-            auto r = x;
-            // layer norm 1
-            x = ggml_reshape_2d(ctx, x, c, w * h * n);
-            x = ggml_nn_layer_norm(ctx, x, transformer.norm1_w, transformer.norm1_b);
+        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]
 
-            // self-attention
-            {
-                x                     = ggml_reshape_2d(ctx, x, c, h * w * n);        // [N * h * w, in_channels]
-                struct ggml_tensor* q = ggml_mul_mat(ctx, transformer.attn1_q_w, x);  // [N * h * w, in_channels]
-#if !defined(SD_USE_FLASH_ATTENTION) || defined(SD_USE_CUBLAS) || defined(SD_USE_METAL)
-                q = ggml_scale_inplace(ctx, q, 1.0f / sqrt((float)d_head));
-#endif
-                q = ggml_reshape_4d(ctx, q, d_head, n_head, h * w, n);   // [N, h * w, n_head, d_head]
-                q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));    // [N, n_head, h * w, d_head]
-                q = ggml_reshape_3d(ctx, q, d_head, h * w, n_head * n);  // [N * n_head, h * w, 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]
 
-                struct ggml_tensor* k = ggml_mul_mat(ctx, transformer.attn1_k_w, x);         // [N * h * w, in_channels]
-                k                     = ggml_reshape_4d(ctx, k, d_head, n_head, h * w, n);   // [N, h * w, n_head, d_head]
-                k                     = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3));    // [N, n_head, h * w, d_head]
-                k                     = ggml_reshape_3d(ctx, k, d_head, h * w, n_head * n);  // [N * n_head, h * w, 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]
 
-                struct ggml_tensor* v = ggml_mul_mat(ctx, transformer.attn1_v_w, x);         // [N * h * w, in_channels]
-                v                     = ggml_reshape_4d(ctx, v, d_head, n_head, h * w, n);   // [N, h * w, n_head, d_head]
-                v                     = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));    // [N, n_head, d_head, h * w]
-                v                     = ggml_reshape_3d(ctx, v, h * w, d_head, n_head * n);  // [N * n_head, d_head, h * w]
+        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]
 
-#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL)
-                struct ggml_tensor* kqv = ggml_flash_attn(ctx, q, k, v, false);  // [N * n_head, h * w, d_head]
-#else
-                struct ggml_tensor* kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, h * w, h * w]
-                // kq = ggml_diag_mask_inf_inplace(ctx, kq, 0);
-                kq = ggml_soft_max_inplace(ctx, kq);
+        x = ggml_reshape_3d(ctx, kqv, d_head * n_head, n_token, n);  // [N, n_token, inner_dim]
 
-                struct ggml_tensor* kqv = ggml_mul_mat(ctx, v, kq);  // [N * n_head, h * w, d_head]
-#endif
-                kqv = ggml_reshape_4d(ctx, kqv, d_head, h * w, n_head, n);
-                kqv = ggml_cont(ctx, ggml_permute(ctx, kqv, 0, 2, 1, 3));  // [N, h * w, n_head, d_head]
+        x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
+        return x;
+    }
+};
 
-                // x = ggml_cpy(ctx, kqv, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, d_head * n_head, h * w * n));
-                x = ggml_reshape_2d(ctx, kqv, d_head * n_head, h * w * n);
+class BasicTransformerBlock : public GGMLBlock {
+protected:
+    int64_t n_head;
+    int64_t d_head;
+    bool ff_in;
 
-                x = ggml_nn_linear(ctx, x, transformer.attn1_out_w, transformer.attn1_out_b);
+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));
 
-                x = ggml_reshape_4d(ctx, x, c, w, h, n);
-            }
+        if (ff_in) {
+            blocks["norm_in"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+            blocks["ff_in"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
+        }
+    }
 
-            x = ggml_add(ctx, x, r);
-            r = x;
+    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]
 
-            // layer norm 2
-            x = ggml_nn_layer_norm(ctx, x, transformer.norm2_w, transformer.norm2_b);
+        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"]);
 
-            // cross-attention
-            {
-                x                     = ggml_reshape_2d(ctx, x, c, h * w * n);                                           // [N * h * w, in_channels]
-                context               = ggml_reshape_2d(ctx, context, context->ne[0], context->ne[1] * context->ne[2]);  // [N * max_position, hidden_size]
-                struct ggml_tensor* q = ggml_mul_mat(ctx, transformer.attn2_q_w, x);                                     // [N * h * w, in_channels]
-#if !defined(SD_USE_FLASH_ATTENTION) || defined(SD_USE_CUBLAS) || defined(SD_USE_METAL)
-                q = ggml_scale_inplace(ctx, q, 1.0f / sqrt((float)d_head));
-#endif
-                q = ggml_reshape_4d(ctx, q, d_head, n_head, h * w, n);   // [N, h * w, n_head, d_head]
-                q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));    // [N, n_head, h * w, d_head]
-                q = ggml_reshape_3d(ctx, q, d_head, h * w, n_head * n);  // [N * n_head, h * w, d_head]
+        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"]);
 
-                struct ggml_tensor* k = ggml_mul_mat(ctx, transformer.attn2_k_w, context);          // [N * max_position, in_channels]
-                k                     = ggml_reshape_4d(ctx, k, d_head, n_head, max_position, n);   // [N, max_position, n_head, d_head]
-                k                     = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3));           // [N, n_head, max_position, d_head]
-                k                     = ggml_reshape_3d(ctx, k, d_head, max_position, n_head * n);  // [N * n_head, max_position, d_head]
-
-                struct ggml_tensor* v = ggml_mul_mat(ctx, transformer.attn2_v_w, context);          // [N * max_position, in_channels]
-                v                     = ggml_reshape_4d(ctx, v, d_head, n_head, max_position, n);   // [N, max_position, n_head, d_head]
-                v                     = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));           // [N, n_head, d_head, max_position]
-                v                     = ggml_reshape_3d(ctx, v, max_position, d_head, n_head * n);  // [N * n_head, d_head, max_position]
-#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL)
-                struct ggml_tensor* kqv = ggml_flash_attn(ctx, q, k, v, false);  // [N * n_head, h * w, d_head]
-#else
-                struct ggml_tensor* kq  = ggml_mul_mat(ctx, k, q);   // [N * n_head, h * w, max_position]
-                // kq = ggml_diag_mask_inf_inplace(ctx, kq, 0);
-                kq = ggml_soft_max_inplace(ctx, kq);
-
-                struct ggml_tensor* kqv = ggml_mul_mat(ctx, v, kq);  // [N * n_head, h * w, d_head]
-#endif
-                kqv = ggml_reshape_4d(ctx, kqv, d_head, h * w, n_head, n);
-                kqv = ggml_cont(ctx, ggml_permute(ctx, kqv, 0, 2, 1, 3));
-
-                // x = ggml_cpy(ctx, kqv, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, d_head * n_head, h * w * n)); // [N * h * w, in_channels]
-                x = ggml_reshape_2d(ctx, kqv, d_head * n_head, h * w * n);  // [N * h * w, in_channels]
-
-                x = ggml_nn_linear(ctx, x, transformer.attn2_out_w, transformer.attn2_out_b);
-
-                x = ggml_reshape_4d(ctx, x, c, w, h, n);
-            }
-
-            x = ggml_add(ctx, x, r);
-            r = x;
-
-            // layer norm 3
-            x = ggml_reshape_2d(ctx, x, c, h * w * n);  // [N * h * w, in_channels]
-            x = ggml_nn_layer_norm(ctx, x, transformer.norm3_w, transformer.norm3_b);
-
-            // ff
-            {
-                // GEGLU
-                auto x_w    = ggml_view_2d(ctx,
-                                           transformer.ff_0_proj_w,
-                                           transformer.ff_0_proj_w->ne[0],
-                                           transformer.ff_0_proj_w->ne[1] / 2,
-                                           transformer.ff_0_proj_w->nb[1],
-                                           0);  // [in_channels * 4, in_channels]
-                auto x_b    = ggml_view_1d(ctx,
-                                           transformer.ff_0_proj_b,
-                                           transformer.ff_0_proj_b->ne[0] / 2,
-                                           0);  // [in_channels * 4, in_channels]
-                auto gate_w = ggml_view_2d(ctx,
-                                           transformer.ff_0_proj_w,
-                                           transformer.ff_0_proj_w->ne[0],
-                                           transformer.ff_0_proj_w->ne[1] / 2,
-                                           transformer.ff_0_proj_w->nb[1],
-                                           transformer.ff_0_proj_w->nb[1] * transformer.ff_0_proj_w->ne[1] / 2);  // [in_channels * 4, ]
-                auto gate_b = ggml_view_1d(ctx,
-                                           transformer.ff_0_proj_b,
-                                           transformer.ff_0_proj_b->ne[0] / 2,
-                                           transformer.ff_0_proj_b->nb[0] * transformer.ff_0_proj_b->ne[0] / 2);  // [in_channels * 4, ]
-                x           = ggml_reshape_2d(ctx, x, c, w * h * n);
-                auto x_in   = x;
-                x           = ggml_nn_linear(ctx, x_in, x_w, x_b);        // [N * h * w, in_channels * 4]
-                auto gate   = ggml_nn_linear(ctx, x_in, gate_w, gate_b);  // [N * h * w, in_channels * 4]
-
-                gate = ggml_gelu_inplace(ctx, gate);
-
-                x = ggml_mul(ctx, x, gate);  // [N * h * w, in_channels * 4]
-                // fc
-                x = ggml_nn_linear(ctx, x, transformer.ff_2_w, transformer.ff_2_b);  // [N * h * w, in_channels]
-            }
-
-            x = ggml_reshape_4d(ctx, x, c, w, h, n);  // [N, h, w, in_channels]
-
-            // residual
-            x = ggml_add(ctx, x, r);
+            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);
         }
 
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 2, 0, 1, 3));  // [N, in_channels, h, w]
+        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 = ggml_nn_conv_2d(ctx, x, proj_out_w, proj_out_b);  // [N, in_channels, h, w]
+        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__

esrgan.hpp

@@ -12,279 +12,160 @@
 
 */
 
-struct ResidualDenseBlock {
-    int num_features;
+class ResidualDenseBlock : public GGMLBlock {
+protected:
+    int num_feat;
     int num_grow_ch;
-    ggml_tensor* conv1_w;  // [num_grow_ch, num_features, 3, 3]
-    ggml_tensor* conv1_b;  // [num_grow_ch]
 
-    ggml_tensor* conv2_w;  // [num_grow_ch, num_features + num_grow_ch, 3, 3]
-    ggml_tensor* conv2_b;  // [num_grow_ch]
-
-    ggml_tensor* conv3_w;  // [num_grow_ch, num_features + 2 * num_grow_ch, 3, 3]
-    ggml_tensor* conv3_b;  // [num_grow_ch]
-
-    ggml_tensor* conv4_w;  // [num_grow_ch, num_features + 3 * num_grow_ch, 3, 3]
-    ggml_tensor* conv4_b;  // [num_grow_ch]
-
-    ggml_tensor* conv5_w;  // [num_features, num_features + 4 * num_grow_ch, 3, 3]
-    ggml_tensor* conv5_b;  // [num_features]
-
-    ResidualDenseBlock() {}
-
-    ResidualDenseBlock(int num_feat, int n_grow_ch) {
-        num_features = num_feat;
-        num_grow_ch  = n_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}));
     }
 
-    size_t calculate_mem_size() {
-        size_t mem_size = num_features * num_grow_ch * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv1_w
-        mem_size += num_grow_ch * ggml_type_size(GGML_TYPE_F32);                               // conv1_b
-
-        mem_size += (num_features + num_grow_ch) * num_grow_ch * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv2_w
-        mem_size += num_grow_ch * ggml_type_size(GGML_TYPE_F32);                                         // conv2_b
-
-        mem_size += (num_features + 2 * num_grow_ch) * num_grow_ch * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv3_w
-        mem_size += num_grow_ch * ggml_type_size(GGML_TYPE_F32);                                             // conv3_w
-
-        mem_size += (num_features + 3 * num_grow_ch) * num_grow_ch * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv4_w
-        mem_size += num_grow_ch * ggml_type_size(GGML_TYPE_F32);                                             // conv4_w
-
-        mem_size += (num_features + 4 * num_grow_ch) * num_features * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv5_w
-        mem_size += num_features * ggml_type_size(GGML_TYPE_F32);                                             // conv5_w
-
-        return mem_size;
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
     }
 
-    int get_num_tensors() {
-        int num_tensors = 10;
-        return num_tensors;
-    }
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
 
-    void init_params(ggml_context* ctx) {
-        conv1_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, num_features, num_grow_ch);
-        conv1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_grow_ch);
-        conv2_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, num_features + num_grow_ch, num_grow_ch);
-        conv2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_grow_ch);
-        conv3_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, num_features + 2 * num_grow_ch, num_grow_ch);
-        conv3_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_grow_ch);
-        conv4_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, num_features + 3 * num_grow_ch, num_grow_ch);
-        conv4_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_grow_ch);
-        conv5_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, num_features + 4 * num_grow_ch, num_features);
-        conv5_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_features);
-    }
+        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"]);
 
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors, std::string prefix) {
-        tensors[prefix + "conv1.weight"] = conv1_w;
-        tensors[prefix + "conv1.bias"]   = conv1_b;
+        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);
 
-        tensors[prefix + "conv2.weight"] = conv2_w;
-        tensors[prefix + "conv2.bias"]   = conv2_b;
-
-        tensors[prefix + "conv3.weight"] = conv3_w;
-        tensors[prefix + "conv3.bias"]   = conv3_b;
-
-        tensors[prefix + "conv4.weight"] = conv4_w;
-        tensors[prefix + "conv4.bias"]   = conv4_b;
-
-        tensors[prefix + "conv5.weight"] = conv5_w;
-        tensors[prefix + "conv5.bias"]   = conv5_b;
-    }
-
-    ggml_tensor* forward(ggml_context* ctx, float out_scale, ggml_tensor* x /* feat */) {
-        // x1 = self.lrelu(self.conv1(x))
-        ggml_tensor* x1 = ggml_nn_conv_2d(ctx, x, conv1_w, conv1_b, 1, 1, 1, 1);
-        x1              = ggml_leaky_relu(ctx, x1, 0.2f, true);
-
-        // x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
-        ggml_tensor* x_cat = ggml_concat(ctx, x, x1);
-        ggml_tensor* x2    = ggml_nn_conv_2d(ctx, x_cat, conv2_w, conv2_b, 1, 1, 1, 1);
-        x2                 = ggml_leaky_relu(ctx, x2, 0.2f, true);
-
-        // x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
-        x_cat           = ggml_concat(ctx, x_cat, x2);
-        ggml_tensor* x3 = ggml_nn_conv_2d(ctx, x_cat, conv3_w, conv3_b, 1, 1, 1, 1);
-        x3              = ggml_leaky_relu(ctx, x3, 0.2f, true);
-
-        // x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
-        x_cat           = ggml_concat(ctx, x_cat, x3);
-        ggml_tensor* x4 = ggml_nn_conv_2d(ctx, x_cat, conv4_w, conv4_b, 1, 1, 1, 1);
-        x4              = ggml_leaky_relu(ctx, x4, 0.2f, true);
-
-        // self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
-        x_cat           = ggml_concat(ctx, x_cat, x4);
-        ggml_tensor* x5 = ggml_nn_conv_2d(ctx, x_cat, conv5_w, conv5_b, 1, 1, 1, 1);
-
-        // return x5 * 0.2 + x
-        x5 = ggml_add(ctx, ggml_scale(ctx, x5, out_scale), x);
+        x5 = ggml_add(ctx, ggml_scale(ctx, x5, 0.2f), x);
         return x5;
     }
 };
 
-struct EsrganBlock {
-    ResidualDenseBlock rd_blocks[3];
-    int num_residual_blocks = 3;
-
-    EsrganBlock() {}
-
-    EsrganBlock(int num_feat, int num_grow_ch) {
-        for (int i = 0; i < num_residual_blocks; i++) {
-            rd_blocks[i] = ResidualDenseBlock(num_feat, num_grow_ch);
-        }
+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));
     }
 
-    int get_num_tensors() {
-        int num_tensors = 0;
-        for (int i = 0; i < num_residual_blocks; i++) {
-            num_tensors += rd_blocks[i].get_num_tensors();
+    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));
         }
-        return num_tensors;
+        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}));
     }
 
-    size_t calculate_mem_size() {
-        size_t mem_size = 0;
-        for (int i = 0; i < num_residual_blocks; i++) {
-            mem_size += rd_blocks[i].calculate_mem_size();
-        }
-        return mem_size;
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
     }
 
-    void init_params(ggml_context* ctx) {
-        for (int i = 0; i < num_residual_blocks; i++) {
-            rd_blocks[i].init_params(ctx);
-        }
-    }
+    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"]);
 
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors, std::string prefix) {
-        for (int i = 0; i < num_residual_blocks; i++) {
-            rd_blocks[i].map_by_name(tensors, prefix + "rdb" + std::to_string(i + 1) + ".");
-        }
-    }
+        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]);
 
-    ggml_tensor* forward(ggml_context* ctx, float out_scale, ggml_tensor* x) {
-        ggml_tensor* out = x;
-        for (int i = 0; i < num_residual_blocks; i++) {
-            // out = self.rdb...(x)
-            out = rd_blocks[i].forward(ctx, out_scale, out);
+            body_feat = block->forward(ctx, body_feat);
         }
-        // return out * 0.2 + x
-        out = ggml_add(ctx, ggml_scale(ctx, out, out_scale), x);
+        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 {
-    int scale        = 4;  // default RealESRGAN_x4plus_anime_6B
-    int num_blocks   = 6;  // default RealESRGAN_x4plus_anime_6B
-    int in_channels  = 3;
-    int out_channels = 3;
-    int num_features = 64;   // default RealESRGAN_x4plus_anime_6B
-    int num_grow_ch  = 32;   // default RealESRGAN_x4plus_anime_6B
-    int tile_size    = 128;  // avoid cuda OOM for 4gb VRAM
+    RRDBNet rrdb_net;
+    int scale     = 4;
+    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
 
-    ggml_tensor* conv_first_w;  // [num_features, in_channels, 3, 3]
-    ggml_tensor* conv_first_b;  // [num_features]
-
-    EsrganBlock body_blocks[6];
-    ggml_tensor* conv_body_w;  // [num_features, num_features, 3, 3]
-    ggml_tensor* conv_body_b;  // [num_features]
-
-    // upsample
-    ggml_tensor* conv_up1_w;  // [num_features, num_features, 3, 3]
-    ggml_tensor* conv_up1_b;  // [num_features]
-    ggml_tensor* conv_up2_w;  // [num_features, num_features, 3, 3]
-    ggml_tensor* conv_up2_b;  // [num_features]
-
-    ggml_tensor* conv_hr_w;    // [num_features, num_features, 3, 3]
-    ggml_tensor* conv_hr_b;    // [num_features]
-    ggml_tensor* conv_last_w;  // [out_channels, num_features, 3, 3]
-    ggml_tensor* conv_last_b;  // [out_channels]
-
-    bool decode_only = false;
-
-    ESRGAN() {
-        name = "esrgan";
-        for (int i = 0; i < num_blocks; i++) {
-            body_blocks[i] = EsrganBlock(num_features, num_grow_ch);
-        }
+    ESRGAN(ggml_backend_t backend,
+           ggml_type wtype)
+        : GGMLModule(backend, wtype) {
+        rrdb_net.init(params_ctx, wtype);
     }
 
-    size_t calculate_mem_size() {
-        size_t mem_size = num_features * in_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_first_w
-        mem_size += num_features * ggml_type_size(GGML_TYPE_F32);                              // conv_first_b
-
-        for (int i = 0; i < num_blocks; i++) {
-            mem_size += body_blocks[i].calculate_mem_size();
-        }
-
-        mem_size += num_features * num_features * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_body_w
-        mem_size += num_features * ggml_type_size(GGML_TYPE_F32);                         // conv_body_w
-
-        // upsample
-        mem_size += num_features * num_features * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_up1_w
-        mem_size += num_features * ggml_type_size(GGML_TYPE_F32);                         // conv_up1_b
-
-        mem_size += num_features * num_features * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_up2_w
-        mem_size += num_features * ggml_type_size(GGML_TYPE_F32);                         // conv_up2_b
-
-        mem_size += num_features * num_features * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_hr_w
-        mem_size += num_features * ggml_type_size(GGML_TYPE_F32);                         // conv_hr_b
-
-        mem_size += out_channels * num_features * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_last_w
-        mem_size += out_channels * ggml_type_size(GGML_TYPE_F32);                         // conv_last_b
-        return mem_size;
+    std::string get_desc() {
+        return "esrgan";
     }
 
-    size_t get_num_tensors() {
-        size_t num_tensors = 12;
-        for (int i = 0; i < num_blocks; i++) {
-            num_tensors += body_blocks[i].get_num_tensors();
-        }
-        return num_tensors;
+    size_t get_params_mem_size() {
+        return rrdb_net.get_params_mem_size();
     }
 
-    void init_params() {
-        ggml_allocr* alloc = ggml_allocr_new_from_buffer(params_buffer);
-        conv_first_w       = ggml_new_tensor_4d(params_ctx, GGML_TYPE_F16, 3, 3, in_channels, num_features);
-        conv_first_b       = ggml_new_tensor_1d(params_ctx, GGML_TYPE_F32, num_features);
-        conv_body_w        = ggml_new_tensor_4d(params_ctx, GGML_TYPE_F16, 3, 3, num_features, num_features);
-        conv_body_b        = ggml_new_tensor_1d(params_ctx, GGML_TYPE_F32, num_features);
-        conv_up1_w         = ggml_new_tensor_4d(params_ctx, GGML_TYPE_F16, 3, 3, num_features, num_features);
-        conv_up1_b         = ggml_new_tensor_1d(params_ctx, GGML_TYPE_F32, num_features);
-        conv_up2_w         = ggml_new_tensor_4d(params_ctx, GGML_TYPE_F16, 3, 3, num_features, num_features);
-        conv_up2_b         = ggml_new_tensor_1d(params_ctx, GGML_TYPE_F32, num_features);
-        conv_hr_w          = ggml_new_tensor_4d(params_ctx, GGML_TYPE_F16, 3, 3, num_features, num_features);
-        conv_hr_b          = ggml_new_tensor_1d(params_ctx, GGML_TYPE_F32, num_features);
-        conv_last_w        = ggml_new_tensor_4d(params_ctx, GGML_TYPE_F16, 3, 3, num_features, out_channels);
-        conv_last_b        = ggml_new_tensor_1d(params_ctx, GGML_TYPE_F32, out_channels);
-
-        for (int i = 0; i < num_blocks; i++) {
-            body_blocks[i].init_params(params_ctx);
-        }
-
-        // alloc all tensors linked to this context
-        for (struct ggml_tensor* t = ggml_get_first_tensor(params_ctx); t != NULL; t = ggml_get_next_tensor(params_ctx, t)) {
-            if (t->data == NULL) {
-                ggml_allocr_alloc(alloc, t);
-            }
-        }
-        ggml_allocr_free(alloc);
+    size_t get_params_num() {
+        return rrdb_net.get_params_num();
     }
 
-    bool load_from_file(const std::string& file_path, ggml_backend_t backend) {
+    bool load_from_file(const std::string& file_path) {
         LOG_INFO("loading esrgan from '%s'", file_path.c_str());
 
-        if (!alloc_params_buffer(backend)) {
-            return false;
-        }
-
+        alloc_params_buffer();
         std::map<std::string, ggml_tensor*> esrgan_tensors;
-
-        // prepare memory for the weights
-        {
-            init_params();
-            map_by_name(esrgan_tensors);
-        }
+        rrdb_net.get_param_tensors(esrgan_tensors);
 
         ModelLoader model_loader;
         if (!model_loader.init_from_file(file_path)) {
@@ -303,115 +184,22 @@ struct ESRGAN : public GGMLModule {
         return success;
     }
 
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors) {
-        tensors["conv_first.weight"] = conv_first_w;
-        tensors["conv_first.bias"]   = conv_first_b;
-
-        for (int i = 0; i < num_blocks; i++) {
-            body_blocks[i].map_by_name(tensors, "body." + std::to_string(i) + ".");
-        }
-
-        tensors["conv_body.weight"] = conv_body_w;
-        tensors["conv_body.bias"]   = conv_body_b;
-
-        tensors["conv_up1.weight"] = conv_up1_w;
-        tensors["conv_up1.bias"]   = conv_up1_b;
-        tensors["conv_up2.weight"] = conv_up2_w;
-        tensors["conv_up2.bias"]   = conv_up2_b;
-        tensors["conv_hr.weight"]  = conv_hr_w;
-        tensors["conv_hr.bias"]    = conv_hr_b;
-
-        tensors["conv_last.weight"] = conv_last_w;
-        tensors["conv_last.bias"]   = conv_last_b;
-    }
-
-    ggml_tensor* forward(ggml_context* ctx0, float out_scale, ggml_tensor* x /* feat */) {
-        // feat = self.conv_first(feat)
-        auto h = ggml_nn_conv_2d(ctx0, x, conv_first_w, conv_first_b, 1, 1, 1, 1);
-
-        auto body_h = h;
-        // self.body(feat)
-        for (int i = 0; i < num_blocks; i++) {
-            body_h = body_blocks[i].forward(ctx0, out_scale, body_h);
-        }
-
-        // body_feat = self.conv_body(self.body(feat))
-        body_h = ggml_nn_conv_2d(ctx0, body_h, conv_body_w, conv_body_b, 1, 1, 1, 1);
-
-        // feat = feat + body_feat
-        h = ggml_add(ctx0, h, body_h);
-
-        // upsample
-        // feat = self.lrelu(self.conv_up1(F.interpolate(feat, scale_factor=2, mode='nearest')))
-        h = ggml_upscale(ctx0, h, 2);
-        h = ggml_nn_conv_2d(ctx0, h, conv_up1_w, conv_up1_b, 1, 1, 1, 1);
-        h = ggml_leaky_relu(ctx0, h, 0.2f, true);
-
-        // feat = self.lrelu(self.conv_up2(F.interpolate(feat, scale_factor=2, mode='nearest')))
-        h = ggml_upscale(ctx0, h, 2);
-        h = ggml_nn_conv_2d(ctx0, h, conv_up2_w, conv_up2_b, 1, 1, 1, 1);
-        h = ggml_leaky_relu(ctx0, h, 0.2f, true);
-
-        // out = self.conv_last(self.lrelu(self.conv_hr(feat)))
-        h = ggml_nn_conv_2d(ctx0, h, conv_hr_w, conv_hr_b, 1, 1, 1, 1);
-        h = ggml_leaky_relu(ctx0, h, 0.2f, true);
-
-        h = ggml_nn_conv_2d(ctx0, h, conv_last_w, conv_last_b, 1, 1, 1, 1);
-        return h;
-    }
-
     struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
-        // since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
-        static size_t buf_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
-        static std::vector<uint8_t> buf(buf_size);
-
-        struct ggml_init_params params = {
-            /*.mem_size   =*/buf_size,
-            /*.mem_buffer =*/buf.data(),
-            /*.no_alloc   =*/true,  // the tensors will be allocated later by ggml_allocr_alloc_graph()
-        };
-
-        struct ggml_context* ctx0 = ggml_init(params);
-
-        struct ggml_cgraph* gf = ggml_new_graph(ctx0);
-
-        struct ggml_tensor* x_ = NULL;
-        float out_scale        = 0.2f;
-
-        // it's performing a compute, check if backend isn't cpu
-        if (!ggml_backend_is_cpu(backend)) {
-            // pass input tensors to gpu memory
-            x_ = ggml_dup_tensor(ctx0, x);
-            ggml_allocr_alloc(compute_allocr, x_);
-
-            // pass data to device backend
-            if (!ggml_allocr_is_measure(compute_allocr)) {
-                ggml_backend_tensor_set(x_, x->data, 0, ggml_nbytes(x));
-            }
-        } else {
-            x_ = x;
-        }
-
-        struct ggml_tensor* out = forward(ctx0, out_scale, 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);
-        ggml_free(ctx0);
-
         return gf;
     }
 
-    void alloc_compute_buffer(struct ggml_tensor* x) {
+    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::alloc_compute_buffer(get_graph);
-    }
-
-    void compute(struct ggml_tensor* work_result, const int n_threads, struct ggml_tensor* x) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(x);
-        };
-        GGMLModule::compute(get_graph, n_threads, work_result);
+        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
     }
 };
 

examples/cli/main.cpp

@@ -6,10 +6,11 @@
 #include <string>
 #include <vector>
 
-#include "preprocessing.hpp"
+// #include "preprocessing.hpp"
 #include "stable-diffusion.h"
 
 #define STB_IMAGE_IMPLEMENTATION
+#define STB_IMAGE_STATIC
 #include "stb_image.h"
 
 #define STB_IMAGE_WRITE_IMPLEMENTATION
@@ -43,12 +44,14 @@ const char* schedule_str[] = {
 const char* modes_str[] = {
     "txt2img",
     "img2img",
+    "img2vid",
     "convert",
 };
 
 enum SDMode {
     TXT2IMG,
     IMG2IMG,
+    IMG2VID,
     CONVERT,
     MODE_COUNT
 };
@@ -63,6 +66,8 @@ struct SDParams {
     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";
@@ -71,11 +76,18 @@ struct SDParams {
 
     std::string prompt;
     std::string negative_prompt;
-    float cfg_scale = 7.0f;
-    int clip_skip   = -1;  // <= 0 represents unspecified
-    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;
+
+    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;
@@ -87,7 +99,12 @@ struct SDParams {
     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) {
@@ -101,13 +118,20 @@ void print_params(SDParams params) {
     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);
@@ -120,6 +144,7 @@ void print_params(SDParams params) {
     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[]) {
@@ -135,7 +160,11 @@ void print_usage(int argc, const char* argv[]) {
     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");
@@ -146,6 +175,7 @@ void print_usage(int argc, const char* argv[]) {
     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");
@@ -190,7 +220,8 @@ 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);
             }
@@ -231,6 +262,18 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 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;
@@ -286,6 +329,16 @@ 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;
@@ -304,6 +357,12 @@ 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;
@@ -338,6 +397,12 @@ void parse_args(int argc, const char** argv, SDParams& params) {
             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") {
@@ -405,6 +470,8 @@ void parse_args(int argc, const char** argv, SDParams& params) {
             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);
@@ -420,7 +487,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         params.n_threads = get_num_physical_cores();
     }
 
-    if (params.mode != CONVERT && 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);
@@ -432,7 +499,7 @@ 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);
@@ -476,6 +543,18 @@ void parse_args(int argc, const char** argv, SDParams& params) {
     }
 }
 
+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) {
     std::string parameter_string = params.prompt + "\n";
     if (params.negative_prompt.size() != 0) {
@@ -496,18 +575,47 @@ 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;
-    if (!params->verbose && level <= SD_LOG_DEBUG) {
+    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;
     }
-    if (level <= SD_LOG_INFO) {
-        fputs(log, stdout);
-        fflush(stdout);
-    } else {
-        fputs(log, stderr);
-        fflush(stderr);
+
+    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[]) {
@@ -539,9 +647,14 @@ int main(int argc, const char* argv[]) {
         }
     }
 
+    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;
-    if (params.mode == IMG2IMG) {
+    if (params.mode == IMG2IMG || params.mode == IMG2VID) {
         vae_decode_only = false;
 
         int c              = 0;
@@ -573,6 +686,7 @@ int main(int argc, const char* argv[]) {
                                   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,
@@ -580,7 +694,9 @@ int main(int argc, const char* argv[]) {
                                   params.wtype,
                                   params.rng_type,
                                   params.schedule,
-                                  params.control_net_cpu);
+                                  params.clip_on_cpu,
+                                  params.control_net_cpu,
+                                  params.vae_on_cpu);
 
     if (sd_ctx == NULL) {
         printf("new_sd_ctx_t failed\n");
@@ -602,8 +718,14 @@ int main(int argc, const char* argv[]) {
                                            3,
                                            input_image_buffer};
             if (params.canny_preprocess) {  // apply preprocessor
-                LOG_INFO("Applying canny preprocessor");
-                control_image->data = preprocess_canny(control_image->data, control_image->width, control_image->height);
+                control_image->data = preprocess_canny(control_image->data,
+                                                       control_image->width,
+                                                       control_image->height,
+                                                       0.08f,
+                                                       0.08f,
+                                                       0.8f,
+                                                       1.0f,
+                                                       false);
             }
         }
         results = txt2img(sd_ctx,
@@ -618,26 +740,67 @@ int main(int argc, const char* argv[]) {
                           params.seed,
                           params.batch_count,
                           control_image,
-                          params.control_strength);
+                          params.control_strength,
+                          params.style_ratio,
+                          params.normalize_input,
+                          params.input_id_images_path.c_str());
     } else {
         sd_image_t input_image = {(uint32_t)params.width,
                                   (uint32_t)params.height,
                                   3,
                                   input_image_buffer};
 
-        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);
+        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);
+        }
     }
 
     if (results == NULL) {
@@ -647,7 +810,7 @@ int main(int argc, const char* argv[]) {
     }
 
     int upscale_factor = 4;  // unused for RealESRGAN_x4plus_anime_6B.pth
-    if (params.esrgan_path.size() > 0) {
+    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);
@@ -659,13 +822,17 @@ int main(int argc, const char* argv[]) {
                 if (results[i].data == NULL) {
                     continue;
                 }
-                sd_image_t upscaled_image = upscale(upscaler_ctx, results[i], upscale_factor);
-                if (upscaled_image.data == NULL) {
-                    printf("upscale failed\n");
-                    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;
                 }
-                free(results[i].data);
-                results[i] = upscaled_image;
+                results[i] = current_image;  // Set the final upscaled image as the result
             }
         }
     }

lora.hpp

@@ -12,26 +12,29 @@ struct LoraModel : public GGMLModule {
     ModelLoader model_loader;
     bool load_failed = false;
 
-    LoraModel(const std::string file_path = "")
-        : file_path(file_path) {
-        name = "lora";
-        if (!model_loader.init_from_file(file_path)) {
+    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;
         }
     }
 
-    size_t get_num_tensors() {
+    std::string get_desc() {
+        return "lora";
+    }
+
+    size_t get_params_num() {
         return LORA_GRAPH_SIZE;
     }
 
-    size_t calculate_mem_size() {
-        return model_loader.cal_mem_size(NULL);
+    size_t get_params_mem_size() {
+        return model_loader.get_params_mem_size(NULL);
     }
 
-    bool load_from_file(ggml_backend_t backend) {
-        if (!alloc_params_buffer(backend)) {
-            return false;
-        }
+    bool load_from_file(bool filter_tensor = false) {
         LOG_INFO("loading LoRA from '%s'", file_path.c_str());
 
         if (load_failed) {
@@ -39,42 +42,41 @@ struct LoraModel : public GGMLModule {
             return false;
         }
 
-        ggml_allocr* alloc = ggml_allocr_new_from_buffer(params_buffer);
-
+        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;
 
-            struct ggml_tensor* real = ggml_new_tensor(params_ctx, tensor_storage.type, tensor_storage.n_dims, tensor_storage.ne);
-            ggml_allocr_alloc(alloc, real);
+            if (filter_tensor && !contains(name, "lora")) {
+                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
+                return true;
+            }
 
-            *dst_tensor = real;
+            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;
+            }
 
-            lora_tensors[name] = 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");
-        ggml_allocr_free(alloc);
         return true;
     }
 
-    struct ggml_cgraph* build_graph(std::map<std::string, struct ggml_tensor*> model_tensors) {
-        // make a graph to compute all lora, expected lora and models tensors are in the same backend
-        // since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
-        static size_t buf_size = ggml_tensor_overhead() * LORA_GRAPH_SIZE + ggml_graph_overhead();
-        static std::vector<uint8_t> buf(buf_size);
-
-        struct ggml_init_params params = {
-            /*.mem_size   =*/buf_size,
-            /*.mem_buffer =*/buf.data(),
-            /*.no_alloc   =*/true,  // the tensors will be allocated later by ggml_allocr_alloc_graph()
-        };
-        // LOG_DEBUG("mem_size %u ", params.mem_size);
-
-        struct ggml_context* ctx0 = ggml_init(params);
-        struct ggml_cgraph* gf    = ggml_new_graph_custom(ctx0, LORA_GRAPH_SIZE, false);
+    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) {
@@ -87,6 +89,10 @@ struct LoraModel : public GGMLModule {
             }
             k_tensor = k_tensor.substr(0, k_pos);
             replace_all_chars(k_tensor, '.', '_');
+            // LOG_DEBUG("k_tensor %s", k_tensor.c_str());
+            if (k_tensor == "model_diffusion_model_output_blocks_2_2_conv") {  // fix for SDXL
+                k_tensor = "model_diffusion_model_output_blocks_2_1_conv";
+            }
             std::string 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";
@@ -125,54 +131,61 @@ struct LoraModel : public GGMLModule {
 
             // 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(ctx0, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
+            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(ctx0, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
+            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(ctx0, ggml_transpose(ctx0, lora_down));
-            struct ggml_tensor* updown = ggml_mul_mat(ctx0, lora_up, lora_down);
-            updown                     = ggml_cont(ctx0, ggml_transpose(ctx0, updown));
-            updown                     = ggml_reshape(ctx0, updown, weight);
+            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(ctx0, updown, scale_value);
+            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(ctx0, GGML_TYPE_F32, weight->n_dims, weight->ne);
-            //     final_weight = ggml_cpy_inplace(ctx0, weight, final_weight);
-            //     final_weight = ggml_add_inplace(ctx0, final_weight, updown);
-            //     final_weight = ggml_cpy_inplace(ctx0, final_weight, weight);
+            //     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(ctx0, weight, updown);
+            //     final_weight = ggml_add_inplace(compute_ctx, weight, updown);
             // }
-            final_weight = ggml_add_inplace(ctx0, weight, updown);  // apply directly
+            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 alloc_compute_buffer(std::map<std::string, struct ggml_tensor*> model_tensors) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(model_tensors);
-        };
-        GGMLModule::alloc_compute_buffer(get_graph);
-    }
-
     void apply(std::map<std::string, struct ggml_tensor*> model_tensors, int n_threads) {
-        alloc_compute_buffer(model_tensors);
-
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(model_tensors);
+            return build_lora_graph(model_tensors);
         };
-        GGMLModule::compute(get_graph, n_threads);
+        GGMLModule::compute(get_graph, n_threads, true);
     }
 };
 
-#endif  // __LORA_HPP__
+#endif  // __LORA_HPP__

model.cpp

@@ -108,6 +108,14 @@ std::unordered_map<std::string, std::string> open_clip_to_hf_clip_model = {
     {"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 = {
@@ -137,7 +145,10 @@ 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 prefix;
-    if (starts_with(new_name, "conditioner.embedders.0.")) {
+    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.")) {
@@ -146,28 +157,42 @@ std::string convert_open_clip_to_hf_clip(const std::string& name) {
     } 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;
     }
-    std::string open_clip_resblock_prefix = "model.transformer.resblocks.";
-    std::string hf_clip_resblock_prefix   = "transformer.text_model.encoder.layers.";
 
     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 (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);
+    std::string open_clip_resblock_prefix = "model.transformer.resblocks.";
+    std::string hf_clip_resblock_prefix   = "transformer.text_model.encoder.layers.";
 
-        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;
+    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;
+            }
         }
-    }
+    };
+
+    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;
 }
@@ -179,6 +204,23 @@ 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"},
+    };
+    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",
@@ -227,7 +269,7 @@ 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) {
@@ -261,6 +303,11 @@ 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];
@@ -370,8 +417,8 @@ 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.") || starts_with(name, "conditioner.embedders.")) {
+    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);
@@ -385,8 +432,12 @@ std::string convert_tensor_name(const std::string& name) {
         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 {
@@ -395,6 +446,26 @@ 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")) {
+        size_t pos = new_name.find(".processor");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, strlen(".processor"), "");
+        }
+        pos = new_name.find_last_of('_');
+        if (pos != std::string::npos) {
+            std::string name_without_network_parts = new_name.substr(0, pos);
+            std::string network_part               = new_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, '.');
+            replace_all_chars(new_key, '.', '_');
+            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) {
@@ -437,7 +508,7 @@ void preprocess_tensor(TensorStorage tensor_storage,
 
     tensor_storage.name = new_name;
 
-    if (new_name.find("transformer.text_model.encoder.layers.") != std::string::npos &&
+    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);
@@ -449,7 +520,7 @@ void preprocess_tensor(TensorStorage tensor_storage,
 
         processed_tensor_storages.insert(processed_tensor_storages.end(), chunks.begin(), chunks.end());
 
-    } else if (new_name.find("transformer.text_model.encoder.layers.") != std::string::npos &&
+    } 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);
@@ -477,7 +548,13 @@ 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);
@@ -486,7 +563,9 @@ 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) {
@@ -515,7 +594,9 @@ 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);
         }
     }
 }
@@ -778,19 +859,27 @@ 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;
+            }
+        }
 
+        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;
@@ -803,6 +892,8 @@ 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;
@@ -946,7 +1037,7 @@ 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;
             }
@@ -1127,7 +1218,9 @@ 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
@@ -1181,7 +1274,6 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
 }
 
 SDVersion ModelLoader::get_sd_version() {
-    // return VERSION_1_x;
     TensorStorage token_embedding_weight;
     for (auto& tensor_storage : tensor_storages) {
         if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos) {
@@ -1190,6 +1282,10 @@ SDVersion ModelLoader::get_sd_version() {
         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" ||
@@ -1227,7 +1323,8 @@ std::string ModelLoader::load_merges() {
     return merges_utf8_str;
 }
 
-void remove_duplicates(std::vector<TensorStorage>& vec) {
+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) {
@@ -1235,13 +1332,16 @@ void remove_duplicates(std::vector<TensorStorage>& vec) {
         auto it                         = name_to_index_map.find(current_name);
 
         if (it != name_to_index_map.end()) {
-            vec[it->second] = vec[i];
+            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());
+    // 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) {
@@ -1255,7 +1355,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
 
         preprocess_tensor(tensor_storage, processed_tensor_storages);
     }
-    remove_duplicates(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];
@@ -1317,8 +1419,6 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
             if (tensor_storage.file_index != file_index) {
                 continue;
             }
-            // LOG_DEBUG("%s", tensor_storage.name.c_str());
-
             ggml_tensor* dst_tensor = NULL;
 
             success = on_new_tensor_cb(tensor_storage, &dst_tensor);
@@ -1353,7 +1453,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     }
 
                     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());
@@ -1372,7 +1472,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     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));
                 }
             }
@@ -1395,15 +1495,19 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
     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 {
-            if (ignore_tensors.find(name) == ignore_tensors.end()) {
-                LOG_WARN("unknown tensor '%s' in model file", name.c_str());
+            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;
         }
 
@@ -1462,7 +1566,7 @@ 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 += cal_mem_size(backend, type);
+    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});
 
@@ -1512,7 +1616,7 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
     return success;
 }
 
-int64_t ModelLoader::cal_mem_size(ggml_backend_t backend, ggml_type type) {
+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);
@@ -1558,4 +1662,4 @@ bool convert(const char* input_path, const char* vae_path, const char* output_pa
     }
     bool success = model_loader.save_to_gguf_file(output_path, (ggml_type)output_type);
     return success;
-}
+}

model.h

@@ -5,7 +5,9 @@
 #include <map>
 #include <memory>
 #include <set>
+#include <sstream>
 #include <string>
+#include <tuple>
 #include <vector>
 
 #include "ggml/ggml-backend.h"
@@ -13,19 +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
@@ -41,7 +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 {
@@ -69,6 +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;
@@ -82,7 +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];
         }
@@ -90,6 +100,24 @@ 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;
@@ -121,7 +149,7 @@ public:
                       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 cal_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
+    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 +1,305 @@
+#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";
+    }
+
+    size_t get_params_mem_size() {
+        size_t params_mem_size = id_encoder.get_params_mem_size();
+        return params_mem_size;
+    }
+
+    size_t get_params_num() {
+        size_t params_num = id_encoder.get_params_num();
+        return params_num;
+    }
+
+    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

@@ -117,15 +117,15 @@ void non_max_supression(struct ggml_tensor* result, struct ggml_tensor* G, struc
     }
 }
 
-void threshold_hystersis(struct ggml_tensor* img, float highThreshold, float lowThreshold, float weak, float strong) {
+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 * highThreshold;
-    float lt = ht * lowThreshold;
+    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
@@ -162,7 +162,7 @@ void threshold_hystersis(struct ggml_tensor* img, float highThreshold, float low
     }
 }
 
-uint8_t* preprocess_canny(uint8_t* img, int width, int height, float highThreshold = 0.08f, float lowThreshold = 0.08f, float weak = 0.8f, float strong = 1.0f, bool inverse = false) {
+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;
@@ -207,7 +207,7 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float highThresho
     normalize_tensor(G);
     prop_arctan2(iX, iY, tetha);
     non_max_supression(image_gray, G, tetha);
-    threshold_hystersis(image_gray, highThreshold, lowThreshold, weak, strong);
+    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++) {

stable-diffusion.h

@@ -72,6 +72,13 @@ enum sd_type_t {
     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,
@@ -88,8 +95,10 @@ enum sd_log_level_t {
 };
 
 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();
 
@@ -108,6 +117,7 @@ SD_API sd_ctx_t* new_sd_ctx(const char* model_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,
@@ -115,7 +125,9 @@ SD_API sd_ctx_t* new_sd_ctx(const char* model_path,
                             enum sd_type_t wtype,
                             enum rng_type_t rng_type,
                             enum schedule_t s,
-                            bool keep_control_net_cpu);
+                            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);
 
@@ -131,7 +143,10 @@ SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
                            int64_t seed,
                            int batch_count,
                            const sd_image_t* control_cond,
-                           float control_strength);
+                           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,
@@ -147,6 +162,21 @@ SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
                            int64_t seed,
                            int batch_count);
 
+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,
@@ -158,8 +188,17 @@ SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image,
 
 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__
+#endif  // __STABLE_DIFFUSION_H__

tae.hpp

@@ -8,88 +8,45 @@
 /*
     ===================================    TinyAutoEncoder  ===================================
     References:
-    https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/autoencoder_tiny.py
+    https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/autoencoders/vae.py
     https://github.com/madebyollin/taesd/blob/main/taesd.py
 
 */
-struct TAEBlock {
-    int in_channels;
-    int out_channels;
 
-    // conv
-    ggml_tensor* conv_0_w;  // [in_channels, out_channels, 3, 3]
-    ggml_tensor* conv_0_b;  // [in_channels]
-    ggml_tensor* conv_1_w;  // [out_channels, out_channels, 3, 3]
-    ggml_tensor* conv_1_b;  // [out_channels]
-    ggml_tensor* conv_2_w;  // [out_channels, out_channels, 3, 3]
-    ggml_tensor* conv_2_b;  // [out_channels]
+class TAEBlock : public UnaryBlock {
+protected:
+    int n_in;
+    int n_out;
 
-    // skip
-    ggml_tensor* conv_skip_w;  // [in_channels, out_channels, 1, 1]
-
-    size_t calculate_mem_size() {
-        size_t mem_size = in_channels * out_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_0_w
-        mem_size += in_channels * ggml_type_size(GGML_TYPE_F32);                               // conv_0_b
-        mem_size += out_channels * out_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);       // conv_1_w
-        mem_size += out_channels * ggml_type_size(GGML_TYPE_F32);                              // conv_1_b
-        mem_size += out_channels * out_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);       // conv_1_w
-        mem_size += out_channels * ggml_type_size(GGML_TYPE_F32);                              // conv_1_b
-        mem_size += out_channels * out_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);       // conv_2_w
-        mem_size += out_channels * ggml_type_size(GGML_TYPE_F32);                              // conv_2_b
-
-        if (in_channels != out_channels) {
-            mem_size += in_channels * out_channels * ggml_type_size(GGML_TYPE_F16);  // conv_skip_w
-        }
-        return mem_size;
-    }
-
-    int get_num_tensors() {
-        return 6 + (in_channels != out_channels ? 1 : 0);
-    }
-
-    void init_params(ggml_context* ctx) {
-        conv_0_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, out_channels, in_channels);
-        conv_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
-
-        conv_1_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, out_channels, out_channels);
-        conv_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-
-        conv_2_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, out_channels, out_channels);
-        conv_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
-
-        if (in_channels != out_channels) {
-            conv_skip_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, out_channels, in_channels);
+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));
         }
     }
 
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors, std::string prefix) {
-        tensors[prefix + "conv.0.weight"] = conv_0_w;
-        tensors[prefix + "conv.0.bias"]   = conv_0_b;
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, n_in, h, w]
+        // return: [n, n_out, h, w]
 
-        tensors[prefix + "conv.2.weight"] = conv_1_w;
-        tensors[prefix + "conv.2.bias"]   = conv_1_b;
+        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"]);
 
-        tensors[prefix + "conv.4.weight"] = conv_2_w;
-        tensors[prefix + "conv.4.bias"]   = conv_2_b;
+        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 (in_channels != out_channels) {
-            tensors[prefix + "skip.weight"] = conv_skip_w;
-        }
-    }
-
-    ggml_tensor* forward(ggml_context* ctx, ggml_tensor* x) {
-        // conv(n_in, n_out)
-        ggml_tensor* h;
-        h = ggml_nn_conv_2d(ctx, x, conv_0_w, conv_0_b, 1, 1, 1, 1);
-        h = ggml_relu_inplace(ctx, h);
-        h = ggml_nn_conv_2d(ctx, h, conv_1_w, conv_1_b, 1, 1, 1, 1);
-        h = ggml_relu_inplace(ctx, h);
-        h = ggml_nn_conv_2d(ctx, h, conv_2_w, conv_2_b, 1, 1, 1, 1);
-
-        // skip connection
-        if (in_channels != out_channels) {
-            // skip = nn.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
-            x = ggml_nn_conv_2d(ctx, x, conv_skip_w, NULL, 1, 1, 1, 1);
+        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);
@@ -98,412 +55,167 @@ struct TAEBlock {
     }
 };
 
-struct TinyEncoder {
+class TinyEncoder : public UnaryBlock {
     int in_channels = 3;
-    int z_channels  = 4;
     int channels    = 64;
+    int z_channels  = 4;
     int num_blocks  = 3;
 
-    // input
-    ggml_tensor* conv_input_w;  // [channels, in_channels, 3, 3]
-    ggml_tensor* conv_input_b;  // [channels]
-    TAEBlock initial_block;
-
-    ggml_tensor* conv_1_w;  // [channels, channels, 3, 3]
-    TAEBlock input_blocks[3];
-
-    // middle
-    ggml_tensor* conv_2_w;  // [channels, channels, 3, 3]
-    TAEBlock middle_blocks[3];
-
-    // output
-    ggml_tensor* conv_3_w;  // [channels, channels, 3, 3]
-    TAEBlock output_blocks[3];
-
-    // final
-    ggml_tensor* conv_final_w;  // [z_channels, channels, 3, 3]
-    ggml_tensor* conv_final_b;  // [z_channels]
-
+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++) {
-            input_blocks[i].in_channels  = channels;
-            input_blocks[i].out_channels = channels;
-
-            middle_blocks[i].in_channels  = channels;
-            middle_blocks[i].out_channels = channels;
-
-            output_blocks[i].in_channels  = channels;
-            output_blocks[i].out_channels = channels;
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
 
-        initial_block.in_channels  = channels;
-        initial_block.out_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}));
     }
 
-    size_t calculate_mem_size() {
-        size_t mem_size = channels * in_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_input_w
-        mem_size += channels * ggml_type_size(GGML_TYPE_F32);                              // conv_input_b
+    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]
 
-        mem_size += initial_block.calculate_mem_size();
+        for (int i = 0; i < num_blocks * 3 + 6; i++) {
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
 
-        mem_size += channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_1_w
-        mem_size += channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_2_w
-        mem_size += channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_3_w
-
-        for (int i = 0; i < num_blocks; i++) {
-            mem_size += input_blocks[i].calculate_mem_size();
-            mem_size += middle_blocks[i].calculate_mem_size();
-            mem_size += output_blocks[i].calculate_mem_size();
-        }
-        mem_size += z_channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_input_w
-        mem_size += z_channels * ggml_type_size(GGML_TYPE_F32);                     // conv_input_b
-        return mem_size;
-    }
-
-    int get_num_tensors() {
-        int num_tensors = 7;
-        for (int i = 0; i < num_blocks; i++) {
-            num_tensors += input_blocks[i].get_num_tensors();
-            num_tensors += middle_blocks[i].get_num_tensors();
-            num_tensors += output_blocks[i].get_num_tensors();
-        }
-        num_tensors += initial_block.get_num_tensors();
-        return num_tensors;
-    }
-
-    void init_params(ggml_context* ctx) {
-        conv_input_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, in_channels, channels);
-        conv_input_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
-
-        initial_block.init_params(ctx);
-
-        conv_1_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, channels);
-        conv_2_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, channels);
-        conv_3_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, channels);
-
-        conv_final_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, z_channels);
-        conv_final_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_channels);
-
-        for (int i = 0; i < num_blocks; i++) {
-            input_blocks[i].init_params(ctx);
-            middle_blocks[i].init_params(ctx);
-            output_blocks[i].init_params(ctx);
-        }
-    }
-
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors, std::string prefix) {
-        tensors[prefix + "0.weight"] = conv_input_w;
-        tensors[prefix + "0.bias"]   = conv_input_b;
-
-        initial_block.map_by_name(tensors, prefix + "1.");
-
-        tensors[prefix + "2.weight"] = conv_1_w;
-        for (int i = 0; i < num_blocks; i++) {
-            input_blocks[i].map_by_name(tensors, prefix + std::to_string(i + 3) + ".");
+            x = block->forward(ctx, x);
         }
 
-        tensors[prefix + "6.weight"] = conv_2_w;
-        for (int i = 0; i < num_blocks; i++) {
-            middle_blocks[i].map_by_name(tensors, prefix + std::to_string(i + 7) + ".");
-        }
-
-        tensors[prefix + "10.weight"] = conv_3_w;
-        for (int i = 0; i < num_blocks; i++) {
-            output_blocks[i].map_by_name(tensors, prefix + std::to_string(i + 11) + ".");
-        }
-
-        tensors[prefix + "14.weight"] = conv_final_w;
-        tensors[prefix + "14.bias"]   = conv_final_b;
-    }
-
-    ggml_tensor* forward(ggml_context* ctx, ggml_tensor* x) {
-        // conv(3, 64)
-        auto z = ggml_nn_conv_2d(ctx, x, conv_input_w, conv_input_b, 1, 1, 1, 1);
-
-        // Block(64, 64)
-        z = initial_block.forward(ctx, z);
-
-        // conv(64, 64, stride=2, bias=False)
-        z = ggml_nn_conv_2d(ctx, z, conv_1_w, NULL, 2, 2, 1, 1);
-
-        // Block(64, 64), Block(64, 64), Block(64, 64)
-        for (int i = 0; i < num_blocks; i++) {
-            z = input_blocks[i].forward(ctx, z);
-        }
-
-        // conv(64, 64, stride=2, bias=False)
-        z = ggml_nn_conv_2d(ctx, z, conv_2_w, NULL, 2, 2, 1, 1);
-
-        // Block(64, 64), Block(64, 64), Block(64, 64)
-        for (int i = 0; i < num_blocks; i++) {
-            z = middle_blocks[i].forward(ctx, z);
-        }
-
-        // conv(64, 64, stride=2, bias=False)
-        z = ggml_nn_conv_2d(ctx, z, conv_3_w, NULL, 2, 2, 1, 1);
-
-        // Block(64, 64), Block(64, 64), Block(64, 64)
-        for (int i = 0; i < num_blocks; i++) {
-            z = output_blocks[i].forward(ctx, z);
-        }
-
-        // conv(64, 4)
-        z = ggml_nn_conv_2d(ctx, z, conv_final_w, conv_final_b, 1, 1, 1, 1);
-        return z;
+        return x;
     }
 };
 
-struct TinyDecoder {
-    int z_channels      = 4;
-    int channels        = 64;
-    int output_channels = 3;
-    int num_blocks      = 3;
+class TinyDecoder : public UnaryBlock {
+    int z_channels   = 4;
+    int channels     = 64;
+    int out_channels = 3;
+    int num_blocks   = 3;
 
-    // input
-    ggml_tensor* conv_input_w;  // [channels, z_channels, 3, 3]
-    ggml_tensor* conv_input_b;  // [channels]
-    TAEBlock input_blocks[3];
-    ggml_tensor* conv_1_w;  // [channels, channels, 3, 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()
 
-    // middle
-    TAEBlock middle_blocks[3];
-    ggml_tensor* conv_2_w;  // [channels, channels, 3, 3]
-
-    // output
-    TAEBlock output_blocks[3];
-    ggml_tensor* conv_3_w;  // [channels, channels, 3, 3]
-
-    // final
-    TAEBlock final_block;
-    ggml_tensor* conv_final_w;  // [output_channels, channels, 3, 3]
-    ggml_tensor* conv_final_b;  // [output_channels]
-
-    TinyDecoder() {
         for (int i = 0; i < num_blocks; i++) {
-            input_blocks[i].in_channels  = channels;
-            input_blocks[i].out_channels = channels;
-
-            middle_blocks[i].in_channels  = channels;
-            middle_blocks[i].out_channels = channels;
-
-            output_blocks[i].in_channels  = channels;
-            output_blocks[i].out_channels = channels;
+            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));
 
-        final_block.in_channels  = channels;
-        final_block.out_channels = channels;
+        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}));
     }
 
-    size_t calculate_mem_size() {
-        size_t mem_size = channels * z_channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_input_w
-        mem_size += channels * ggml_type_size(GGML_TYPE_F32);                             // conv_input_b
+    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]
 
-        for (int i = 0; i < num_blocks; i++) {
-            mem_size += input_blocks[i].calculate_mem_size();
-        }
-        mem_size += channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_1_w
-
-        for (int i = 0; i < num_blocks; i++) {
-            mem_size += middle_blocks[i].calculate_mem_size();
-        }
-        mem_size += channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_2_w
-
-        for (int i = 0; i < num_blocks; i++) {
-            mem_size += output_blocks[i].calculate_mem_size();
-        }
-        mem_size += channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_3_w
-
-        mem_size += final_block.calculate_mem_size();
-        mem_size += output_channels * channels * 3 * 3 * ggml_type_size(GGML_TYPE_F16);  // conv_input_w
-        mem_size += output_channels * ggml_type_size(GGML_TYPE_F32);                     // conv_input_b
-        return mem_size;
-    }
-
-    int get_num_tensors() {
-        int num_tensors = 9;
-        for (int i = 0; i < num_blocks; i++) {
-            num_tensors += input_blocks[i].get_num_tensors();
-            num_tensors += middle_blocks[i].get_num_tensors();
-            num_tensors += output_blocks[i].get_num_tensors();
-        }
-        num_tensors += final_block.get_num_tensors();
-        return num_tensors;
-    }
-
-    void init_params(ggml_allocr* alloc, ggml_context* ctx) {
-        conv_input_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, z_channels, channels);
-        conv_input_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
-
-        conv_1_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, channels);
-        conv_2_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, channels);
-        conv_3_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, channels);
-
-        conv_final_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, output_channels);
-        conv_final_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, output_channels);
-
-        for (int i = 0; i < num_blocks; i++) {
-            input_blocks[i].init_params(ctx);
-            middle_blocks[i].init_params(ctx);
-            output_blocks[i].init_params(ctx);
-        }
-
-        final_block.init_params(ctx);
-    }
-
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors, std::string prefix) {
-        tensors[prefix + "0.weight"] = conv_input_w;
-        tensors[prefix + "0.bias"]   = conv_input_b;
-
-        for (int i = 0; i < num_blocks; i++) {
-            input_blocks[i].map_by_name(tensors, prefix + std::to_string(i + 2) + ".");
-        }
-
-        tensors[prefix + "6.weight"] = conv_1_w;
-        for (int i = 0; i < num_blocks; i++) {
-            middle_blocks[i].map_by_name(tensors, prefix + std::to_string(i + 7) + ".");
-        }
-
-        tensors[prefix + "11.weight"] = conv_2_w;
-        for (int i = 0; i < num_blocks; i++) {
-            output_blocks[i].map_by_name(tensors, prefix + std::to_string(i + 12) + ".");
-        }
-
-        tensors[prefix + "16.weight"] = conv_3_w;
-
-        final_block.map_by_name(tensors, prefix + "17.");
-
-        tensors[prefix + "18.weight"] = conv_final_w;
-        tensors[prefix + "18.bias"]   = conv_final_b;
-    }
-
-    ggml_tensor* forward(ggml_context* ctx, ggml_tensor* z) {
-        // torch.tanh(x / 3) * 3
         auto h = ggml_scale(ctx, z, 1.0f / 3.0f);
         h      = ggml_tanh_inplace(ctx, h);
         h      = ggml_scale(ctx, h, 3.0f);
 
-        // conv(4, 64)
-        h = ggml_nn_conv_2d(ctx, h, conv_input_w, conv_input_b, 1, 1, 1, 1);
+        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)]);
 
-        // nn.ReLU()
-        h = ggml_relu_inplace(ctx, h);
-
-        // Block(64, 64), Block(64, 64), Block(64, 64)
-        for (int i = 0; i < num_blocks; i++) {
-            h = input_blocks[i].forward(ctx, h);
+            h = block->forward(ctx, h);
         }
 
-        // nn.Upsample(scale_factor=2)
-        h = ggml_upscale(ctx, h, 2);
-
-        // conv(64, 64, bias=False)
-        h = ggml_nn_conv_2d(ctx, h, conv_1_w, NULL, 1, 1, 1, 1);
-
-        // Block(64, 64), Block(64, 64), Block(64, 64)
-        for (int i = 0; i < num_blocks; i++) {
-            h = middle_blocks[i].forward(ctx, h);
-        }
-
-        // nn.Upsample(scale_factor=2)
-        h = ggml_upscale(ctx, h, 2);
-
-        // conv(64, 64, bias=False)
-        h = ggml_nn_conv_2d(ctx, h, conv_2_w, NULL, 1, 1, 1, 1);
-
-        // Block(64, 64), Block(64, 64), Block(64, 64)
-        for (int i = 0; i < num_blocks; i++) {
-            h = output_blocks[i].forward(ctx, h);
-        }
-
-        // nn.Upsample(scale_factor=2)
-        h = ggml_upscale(ctx, h, 2);
-
-        // conv(64, 64, bias=False)
-        h = ggml_nn_conv_2d(ctx, h, conv_3_w, NULL, 1, 1, 1, 1);
-
-        // Block(64, 64)
-        h = final_block.forward(ctx, h);
-
-        // conv(64, 3)
-        h = ggml_nn_conv_2d(ctx, h, conv_final_w, conv_final_b, 1, 1, 1, 1);
         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 {
-    TinyEncoder encoder;
-    TinyDecoder decoder;
+    TAESD taesd;
     bool decode_only = false;
 
-    TinyAutoEncoder(bool decoder_only_ = true)
-        : decode_only(decoder_only_) {
-        name = "tae";
+    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);
     }
 
-    size_t calculate_mem_size() {
-        size_t mem_size = decoder.calculate_mem_size();
-        if (!decode_only) {
-            mem_size += encoder.calculate_mem_size();
-        }
-        mem_size += 1024;  // padding
-        return mem_size;
+    std::string get_desc() {
+        return "taesd";
     }
 
-    size_t get_num_tensors() {
-        size_t num_tensors = decoder.get_num_tensors();
-        if (!decode_only) {
-            num_tensors += encoder.get_num_tensors();
-        }
-        return num_tensors;
+    size_t get_params_mem_size() {
+        return taesd.get_params_mem_size();
     }
 
-    void init_params() {
-        ggml_allocr* alloc = ggml_allocr_new_from_buffer(params_buffer);
-        decoder.init_params(alloc, params_ctx);
-        if (!decode_only) {
-            encoder.init_params(params_ctx);
-        }
-
-        // alloc all tensors linked to this context
-        for (struct ggml_tensor* t = ggml_get_first_tensor(params_ctx); t != NULL; t = ggml_get_next_tensor(params_ctx, t)) {
-            if (t->data == NULL) {
-                ggml_allocr_alloc(alloc, t);
-            }
-        }
-        ggml_allocr_free(alloc);
+    size_t get_params_num() {
+        return taesd.get_params_num();
     }
 
-    void map_by_name(std::map<std::string, ggml_tensor*>& tensors) {
-        decoder.map_by_name(tensors, "decoder.layers.");
-        encoder.map_by_name(tensors, "encoder.layers.");
-    }
-
-    bool load_from_file(const std::string& file_path, ggml_backend_t backend) {
+    bool load_from_file(const std::string& file_path) {
         LOG_INFO("loading taesd from '%s'", file_path.c_str());
-
-        if (!alloc_params_buffer(backend)) {
-            return false;
-        }
-
+        alloc_params_buffer();
         std::map<std::string, ggml_tensor*> taesd_tensors;
-
-        // prepare memory for the weights
-        {
-            init_params();
-            map_by_name(taesd_tensors);
-        }
-
-        std::map<std::string, struct ggml_tensor*> tensors_need_to_load;
+        taesd.get_param_tensors(taesd_tensors);
         std::set<std::string> ignore_tensors;
-        for (auto& pair : taesd_tensors) {
-            const std::string& name = pair.first;
-
-            if (decode_only && starts_with(name, "encoder")) {
-                ignore_tensors.insert(name);
-                continue;
-            }
-
-            tensors_need_to_load.insert(pair);
+        if (decode_only) {
+            ignore_tensors.insert("encoder.");
         }
 
         ModelLoader model_loader;
@@ -512,7 +224,7 @@ struct TinyAutoEncoder : public GGMLModule {
             return false;
         }
 
-        bool success = model_loader.load_tensors(tensors_need_to_load, backend, ignore_tensors);
+        bool success = model_loader.load_tensors(taesd_tensors, backend, ignore_tensors);
 
         if (!success) {
             LOG_ERROR("load tae tensors from model loader failed");
@@ -524,57 +236,23 @@ struct TinyAutoEncoder : public GGMLModule {
     }
 
     struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
-        // since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
-        static size_t buf_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
-        static std::vector<uint8_t> buf(buf_size);
-
-        struct ggml_init_params params = {
-            /*.mem_size   =*/buf_size,
-            /*.mem_buffer =*/buf.data(),
-            /*.no_alloc   =*/true,  // the tensors will be allocated later by ggml_allocr_alloc_graph()
-        };
-        // LOG_DEBUG("mem_size %u ", params.mem_size);
-
-        struct ggml_context* ctx0 = ggml_init(params);
-
-        struct ggml_cgraph* gf = ggml_new_graph(ctx0);
-
-        struct ggml_tensor* z_ = NULL;
-
-        // it's performing a compute, check if backend isn't cpu
-        if (!ggml_backend_is_cpu(backend)) {
-            // pass input tensors to gpu memory
-            z_ = ggml_dup_tensor(ctx0, z);
-            ggml_allocr_alloc(compute_allocr, z_);
-
-            // pass data to device backend
-            if (!ggml_allocr_is_measure(compute_allocr)) {
-                ggml_backend_tensor_set(z_, z->data, 0, ggml_nbytes(z));
-            }
-        } else {
-            z_ = z;
-        }
-
-        struct ggml_tensor* out = decode_graph ? decoder.forward(ctx0, z_) : encoder.forward(ctx0, z_);
-
+        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);
-        ggml_free(ctx0);
-
         return gf;
     }
 
-    void alloc_compute_buffer(struct ggml_tensor* x, bool decode) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(x, decode);
-        };
-        GGMLModule::alloc_compute_buffer(get_graph);
-    }
-
-    void compute(struct ggml_tensor* work_result, int n_threads, struct ggml_tensor* z, bool decode_graph) {
+    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, work_result);
+
+        GGMLModule::compute(get_graph, n_threads, false, output, output_ctx);
     }
 };
 

upscaler.cpp

@@ -6,7 +6,7 @@
 struct UpscalerGGML {
     ggml_backend_t backend    = NULL;  // general backend
     ggml_type model_data_type = GGML_TYPE_F16;
-    ESRGAN esrgan_upscaler;
+    std::shared_ptr<ESRGAN> esrgan_upscaler;
     std::string esrgan_path;
     int n_threads;
 
@@ -21,7 +21,7 @@ struct UpscalerGGML {
 #endif
 #ifdef SD_USE_METAL
         LOG_DEBUG("Using Metal backend");
-        ggml_metal_log_set_callback(ggml_log_callback_default, nullptr);
+        ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
 
@@ -30,7 +30,8 @@ struct UpscalerGGML {
             backend = ggml_backend_cpu_init();
         }
         LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
-        if (!esrgan_upscaler.load_from_file(esrgan_path, backend)) {
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_data_type);
+        if (!esrgan_upscaler->load_from_file(esrgan_path)) {
             return false;
         }
         return true;
@@ -39,8 +40,8 @@ struct UpscalerGGML {
     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;
+        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);
 
@@ -62,15 +63,11 @@ struct UpscalerGGML {
 
         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) {
-            if (init) {
-                esrgan_upscaler.alloc_compute_buffer(in);
-            } else {
-                esrgan_upscaler.compute(out, n_threads, in);
-            }
+            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();
+        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);

util.cpp

@@ -1,6 +1,7 @@
 #include "util.h"
 #include <stdarg.h>
 #include <algorithm>
+#include <cmath>
 #include <codecvt>
 #include <fstream>
 #include <locale>
@@ -9,6 +10,7 @@
 #include <thread>
 #include <unordered_set>
 #include <vector>
+#include "preprocessing.hpp"
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/sysctl.h>
@@ -23,6 +25,9 @@
 #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);
@@ -38,6 +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) {
@@ -86,6 +98,43 @@ std::string get_full_path(const std::string& dir, const std::string& filename) {
     }
 }
 
+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>
@@ -100,6 +149,7 @@ 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());
 
@@ -119,6 +169,27 @@ std::string get_full_path(const std::string& dir, const std::string& filename) {
     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
@@ -159,6 +230,9 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
+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);
@@ -174,7 +248,7 @@ std::u32string unicode_value_to_utf32(int unicode_value) {
     return utf32_string;
 }
 
-std::string sd_basename(const std::string& path) {
+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);
@@ -202,7 +276,47 @@ std::string path_join(const std::string& p1, const std::string& p2) {
     return p1 + "/" + p2;
 }
 
+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;
+
+    // 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 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);
@@ -252,23 +366,13 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
     va_list args;
     va_start(args, format);
 
-    const char* level_str = "DEBUG";
-    if (level == SD_LOG_INFO) {
-        level_str = "INFO ";
-    } else if (level == SD_LOG_WARN) {
-        level_str = "WARN ";
-    } else if (level == SD_LOG_ERROR) {
-        level_str = "ERROR";
-    }
-
-    static char log_buffer[LOG_BUFFER_SIZE];
-
-    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "[%s] %s:%-4d - ", level_str, sd_basename(file).c_str(), line);
+    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) - 1);
     }
+    strncat(log_buffer, "\n", LOG_BUFFER_SIZE - strlen(log_buffer));
 
     if (sd_log_cb) {
         sd_log_cb(level, log_buffer, sd_log_cb_data);
@@ -281,7 +385,10 @@ 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;
@@ -307,3 +414,152 @@ const char* sd_get_system_info() {
 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

@@ -3,11 +3,13 @@
 
 #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, ...);
 
@@ -17,11 +19,30 @@ 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::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);
 
-std::string sd_basename(const std::string& path);
+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);