Change node id to reflect node name

* record audio node

* sf

.ci/windows_base_files/README_VERY_IMPORTANT.txt

@@ -4,6 +4,9 @@ if you have a NVIDIA gpu:
 
 run_nvidia_gpu.bat
 
+if you want to enable the fast fp16 accumulation (faster for fp16 models with slightly less quality):
+
+run_nvidia_gpu_fast_fp16_accumulation.bat
 
 
 To run it in slow CPU mode:

.gitattributes

@@ -1,2 +1,3 @@
 /web/assets/** linguist-generated
 /web/** linguist-vendored
+comfy_api_nodes/apis/__init__.py linguist-generated

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -22,7 +22,7 @@ body:
       description: Please confirm you have tried to reproduce the issue with all custom nodes disabled.
       options:
         - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
-          required: true
+          required: false
   - type: textarea
     attributes:
       label: Expected Behavior

.github/ISSUE_TEMPLATE/user-support.yml

@@ -18,7 +18,7 @@ body:
         description: Please confirm you have tried to reproduce the issue with all custom nodes disabled.
         options:
           - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
-            required: true
+            required: false
     - type: textarea
       attributes:
             label: Your question

.github/workflows/check-line-endings.yml

@@ -0,0 +1,40 @@
+name: Check for Windows Line Endings
+
+on:
+  pull_request:
+    branches: ['*'] # Trigger on all pull requests to any branch
+
+jobs:
+  check-line-endings:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0 # Fetch all history to compare changes
+
+      - name: Check for Windows line endings (CRLF)
+        run: |
+          # Get the list of changed files in the PR
+          CHANGED_FILES=$(git diff --name-only ${{ github.event.pull_request.base.sha }}..${{ github.event.pull_request.head.sha }})
+
+          # Flag to track if CRLF is found
+          CRLF_FOUND=false
+
+          # Loop through each changed file
+          for FILE in $CHANGED_FILES; do
+            # Check if the file exists and is a text file
+            if [ -f "$FILE" ] && file "$FILE" | grep -q "text"; then
+              # Check for CRLF line endings
+              if grep -UP '\r$' "$FILE"; then
+                echo "Error: Windows line endings (CRLF) detected in $FILE"
+                CRLF_FOUND=true
+              fi
+            fi
+          done
+
+          # Exit with error if CRLF was found
+          if [ "$CRLF_FOUND" = true ]; then
+            exit 1
+          fi

.github/workflows/release-webhook.yml

@@ -0,0 +1,108 @@
+name: Release Webhook
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  send-webhook:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Send release webhook
+        env:
+          WEBHOOK_URL: ${{ secrets.RELEASE_GITHUB_WEBHOOK_URL }}
+          WEBHOOK_SECRET: ${{ secrets.RELEASE_GITHUB_WEBHOOK_SECRET }}
+        run: |
+          # Generate UUID for delivery ID
+          DELIVERY_ID=$(uuidgen)
+          HOOK_ID="release-webhook-$(date +%s)"
+          
+          # Create webhook payload matching GitHub release webhook format
+          PAYLOAD=$(cat <<EOF
+          {
+            "action": "published",
+            "release": {
+              "id": ${{ github.event.release.id }},
+              "node_id": "${{ github.event.release.node_id }}",
+              "url": "${{ github.event.release.url }}",
+              "html_url": "${{ github.event.release.html_url }}",
+              "assets_url": "${{ github.event.release.assets_url }}",
+              "upload_url": "${{ github.event.release.upload_url }}",
+              "tag_name": "${{ github.event.release.tag_name }}",
+              "target_commitish": "${{ github.event.release.target_commitish }}",
+              "name": ${{ toJSON(github.event.release.name) }},
+              "body": ${{ toJSON(github.event.release.body) }},
+              "draft": ${{ github.event.release.draft }},
+              "prerelease": ${{ github.event.release.prerelease }},
+              "created_at": "${{ github.event.release.created_at }}",
+              "published_at": "${{ github.event.release.published_at }}",
+              "author": {
+                "login": "${{ github.event.release.author.login }}",
+                "id": ${{ github.event.release.author.id }},
+                "node_id": "${{ github.event.release.author.node_id }}",
+                "avatar_url": "${{ github.event.release.author.avatar_url }}",
+                "url": "${{ github.event.release.author.url }}",
+                "html_url": "${{ github.event.release.author.html_url }}",
+                "type": "${{ github.event.release.author.type }}",
+                "site_admin": ${{ github.event.release.author.site_admin }}
+              },
+              "tarball_url": "${{ github.event.release.tarball_url }}",
+              "zipball_url": "${{ github.event.release.zipball_url }}",
+              "assets": ${{ toJSON(github.event.release.assets) }}
+            },
+            "repository": {
+              "id": ${{ github.event.repository.id }},
+              "node_id": "${{ github.event.repository.node_id }}",
+              "name": "${{ github.event.repository.name }}",
+              "full_name": "${{ github.event.repository.full_name }}",
+              "private": ${{ github.event.repository.private }},
+              "owner": {
+                "login": "${{ github.event.repository.owner.login }}",
+                "id": ${{ github.event.repository.owner.id }},
+                "node_id": "${{ github.event.repository.owner.node_id }}",
+                "avatar_url": "${{ github.event.repository.owner.avatar_url }}",
+                "url": "${{ github.event.repository.owner.url }}",
+                "html_url": "${{ github.event.repository.owner.html_url }}",
+                "type": "${{ github.event.repository.owner.type }}",
+                "site_admin": ${{ github.event.repository.owner.site_admin }}
+              },
+              "html_url": "${{ github.event.repository.html_url }}",
+              "clone_url": "${{ github.event.repository.clone_url }}",
+              "git_url": "${{ github.event.repository.git_url }}",
+              "ssh_url": "${{ github.event.repository.ssh_url }}",
+              "url": "${{ github.event.repository.url }}",
+              "created_at": "${{ github.event.repository.created_at }}",
+              "updated_at": "${{ github.event.repository.updated_at }}",
+              "pushed_at": "${{ github.event.repository.pushed_at }}",
+              "default_branch": "${{ github.event.repository.default_branch }}",
+              "fork": ${{ github.event.repository.fork }}
+            },
+            "sender": {
+              "login": "${{ github.event.sender.login }}",
+              "id": ${{ github.event.sender.id }},
+              "node_id": "${{ github.event.sender.node_id }}",
+              "avatar_url": "${{ github.event.sender.avatar_url }}",
+              "url": "${{ github.event.sender.url }}",
+              "html_url": "${{ github.event.sender.html_url }}",
+              "type": "${{ github.event.sender.type }}",
+              "site_admin": ${{ github.event.sender.site_admin }}
+            }
+          }
+          EOF
+          )
+          
+          # Generate HMAC-SHA256 signature
+          SIGNATURE=$(echo -n "$PAYLOAD" | openssl dgst -sha256 -hmac "$WEBHOOK_SECRET" -hex | cut -d' ' -f2)
+          
+          # Send webhook with required headers
+          curl -X POST "$WEBHOOK_URL" \
+            -H "Content-Type: application/json" \
+            -H "X-GitHub-Event: release" \
+            -H "X-GitHub-Delivery: $DELIVERY_ID" \
+            -H "X-GitHub-Hook-ID: $HOOK_ID" \
+            -H "X-Hub-Signature-256: sha256=$SIGNATURE" \
+            -H "User-Agent: GitHub-Actions-Webhook/1.0" \
+            -d "$PAYLOAD" \
+            --fail --silent --show-error
+          
+          echo "✅ Release webhook sent successfully"

.github/workflows/stable-release.yml

@@ -12,17 +12,17 @@ on:
         description: 'CUDA version'
         required: true
         type: string
-        default: "128"
+        default: "129"
       python_minor:
         description: 'Python minor version'
         required: true
         type: string
-        default: "12"
+        default: "13"
       python_patch:
         description: 'Python patch version'
         required: true
         type: string
-        default: "10"
+        default: "6"
 
 
 jobs:
@@ -66,8 +66,13 @@ jobs:
           curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
-            sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
-            cd ..
+          sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+          rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
+          rm ./Lib/site-packages/torch/lib/libprotoc.lib
+          rm ./Lib/site-packages/torch/lib/libprotobuf.lib
+
+          cd ..
 
           git clone --depth 1 https://github.com/comfyanonymous/taesd
           cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
@@ -85,7 +90,7 @@ jobs:
 
           cd ..
 
-          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
           mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
 
           cd ComfyUI_windows_portable
@@ -102,5 +107,4 @@ jobs:
           file: ComfyUI_windows_portable_nvidia.7z
           tag: ${{ inputs.git_tag }}
           overwrite: true
-          prerelease: true
-          make_latest: false
+          draft: true

.github/workflows/windows_release_dependencies.yml

@@ -17,19 +17,19 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
         required: true
         type: string
-        default: "12"
+        default: "13"
 
       python_patch:
         description: 'python patch version'
         required: true
         type: string
-        default: "10"
+        default: "6"
 #  push:
 #    branches:
 #      - master

.github/workflows/windows_release_nightly_pytorch.yml

@@ -7,7 +7,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
@@ -19,7 +19,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "2"
+        default: "5"
 #  push:
 #    branches:
 #      - master
@@ -53,6 +53,8 @@ jobs:
             ls ../temp_wheel_dir
             ./python.exe -s -m pip install --pre ../temp_wheel_dir/*
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd

.github/workflows/windows_release_package.yml

@@ -7,19 +7,19 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
         required: true
         type: string
-        default: "12"
+        default: "13"
 
       python_patch:
         description: 'python patch version'
         required: true
         type: string
-        default: "10"
+        default: "6"
 #  push:
 #    branches:
 #      - master
@@ -64,6 +64,10 @@ jobs:
             ./python.exe get-pip.py
             ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
+            rm ./Lib/site-packages/torch/lib/libprotoc.lib
+            rm ./Lib/site-packages/torch/lib/libprotobuf.lib
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
@@ -82,7 +86,7 @@ jobs:
 
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
             mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu${{ inputs.cu }}_or_cpu.7z
 
             cd ComfyUI_windows_portable

CODEOWNERS

@@ -5,20 +5,21 @@
 # Inlined the team members for now.
 
 # Maintainers
-*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
 
 # Python web server
-/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
 
 # Node developers
-/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
-/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
+/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy_api_nodes/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill

README.md

@@ -39,7 +39,7 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 ## Get Started
 
 #### [Desktop Application](https://www.comfy.org/download)
-- The easiest way to get started. 
+- The easiest way to get started.
 - Available on Windows & macOS.
 
 #### [Windows Portable Package](#installing)
@@ -55,7 +55,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
 - Image Models
-   - SD1.x, SD2.x,
+   - SD1.x, SD2.x ([unCLIP](https://comfyanonymous.github.io/ComfyUI_examples/unclip/))
    - [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), [SDXL Turbo](https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/)
    - [Stable Cascade](https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/)
    - [SD3 and SD3.5](https://comfyanonymous.github.io/ComfyUI_examples/sd3/)
@@ -66,6 +66,11 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Lumina Image 2.0](https://comfyanonymous.github.io/ComfyUI_examples/lumina2/)
    - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
    - [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
+   - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/)
+- Image Editing Models
+   - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
+   - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
+   - [HiDream E1.1](https://comfyanonymous.github.io/ComfyUI_examples/hidream/#hidream-e11)
 - Video Models
    - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/)
    - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/)
@@ -73,6 +78,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
    - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/) and [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
+   - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
    - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
@@ -80,9 +86,10 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2)
 - Asynchronous Queue system
 - Many optimizations: Only re-executes the parts of the workflow that changes between executions.
-- Smart memory management: can automatically run models on GPUs with as low as 1GB vram.
+- Smart memory management: can automatically run large models on GPUs with as low as 1GB vram with smart offloading.
 - Works even if you don't have a GPU with: ```--cpu``` (slow)
-- Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs and CLIP models.
+- Can load ckpt and safetensors: All in one checkpoints or standalone diffusion models, VAEs and CLIP models.
+- Safe loading of ckpt, pt, pth, etc.. files.
 - Embeddings/Textual inversion
 - [Loras (regular, locon and loha)](https://comfyanonymous.github.io/ComfyUI_examples/lora/)
 - [Hypernetworks](https://comfyanonymous.github.io/ComfyUI_examples/hypernetworks/)
@@ -93,12 +100,10 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - [Inpainting](https://comfyanonymous.github.io/ComfyUI_examples/inpaint/) with both regular and inpainting models.
 - [ControlNet and T2I-Adapter](https://comfyanonymous.github.io/ComfyUI_examples/controlnet/)
 - [Upscale Models (ESRGAN, ESRGAN variants, SwinIR, Swin2SR, etc...)](https://comfyanonymous.github.io/ComfyUI_examples/upscale_models/)
-- [unCLIP Models](https://comfyanonymous.github.io/ComfyUI_examples/unclip/)
 - [GLIGEN](https://comfyanonymous.github.io/ComfyUI_examples/gligen/)
 - [Model Merging](https://comfyanonymous.github.io/ComfyUI_examples/model_merging/)
 - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
 - Latent previews with [TAESD](#how-to-show-high-quality-previews)
-- Starts up very fast.
 - Works fully offline: core will never download anything unless you want to.
 - Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview).
 - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
@@ -107,7 +112,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 ## Release Process
 
-ComfyUI follows a weekly release cycle every Friday, with three interconnected repositories:
+ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
    - Releases a new stable version (e.g., v0.7.0)
@@ -175,10 +180,6 @@ If you have trouble extracting it, right click the file -> properties -> unblock
 
 See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor.
 
-## Jupyter Notebook
-
-To run it on services like paperspace, kaggle or colab you can use my [Jupyter Notebook](notebooks/comfyui_colab.ipynb)
-
 
 ## [comfy-cli](https://docs.comfy.org/comfy-cli/getting-started)
 
@@ -202,7 +203,7 @@ Put your VAE in: models/vae
 ### AMD GPUs (Linux only)
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.3```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4```
 
 This is the command to install the nightly with ROCm 6.4 which might have some performance improvements:
 
@@ -210,37 +211,29 @@ This is the command to install the nightly with ROCm 6.4 which might have some p
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip (currently available in PyTorch nightly builds). More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
-  
-1. To install PyTorch nightly, use the following command:
+(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
+
+1. To install PyTorch xpu, use the following command:
+
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu```
+
+This is the command to install the Pytorch xpu nightly which might have some performance improvements:
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-2. Launch ComfyUI by running `python main.py`
-
-
 (Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
 
-1. For Intel® Arc™ A-Series Graphics utilizing IPEX, create a conda environment and use the commands below:
-
-```
-conda install libuv
-pip install torch==2.3.1.post0+cxx11.abi torchvision==0.18.1.post0+cxx11.abi torchaudio==2.3.1.post0+cxx11.abi intel-extension-for-pytorch==2.3.110.post0+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/
-```
-
-For other supported Intel GPUs with IPEX, visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
-
-Additional discussion and help can be found [here](https://github.com/comfyanonymous/ComfyUI/discussions/476).
+1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
 
 ### NVIDIA
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129```
 
 This is the command to install pytorch nightly instead which might have performance improvements.
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu129```
 
 #### Troubleshooting
 
@@ -294,6 +287,13 @@ For models compatible with Cambricon Extension for PyTorch (torch_mlu). Here's a
 2. Next, install the PyTorch(torch_mlu) following the instructions on the [Installation](https://www.cambricon.com/docs/sdk_1.15.0/cambricon_pytorch_1.17.0/user_guide_1.9/index.html)
 3. Launch ComfyUI by running `python main.py`
 
+#### Iluvatar Corex
+
+For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step guide tailored to your platform and installation method:
+
+1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536)
+2. Launch ComfyUI by running `python main.py`
+
 # Running
 
 ```python main.py```
@@ -344,7 +344,7 @@ Generate a self-signed certificate (not appropriate for shared/production use) a
 
 Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app will now be accessible with `https://...` instead of `http://...`.
 
-> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above. 
+> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above.
 <br/><br/>If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal.
 
 ## Support and dev channel

app/frontend_management.py

@@ -29,18 +29,48 @@ def frontend_install_warning_message():
 This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
 """.strip()
 
+def parse_version(version: str) -> tuple[int, int, int]:
+        return tuple(map(int, version.split(".")))
+
+def is_valid_version(version: str) -> bool:
+    """Validate if a string is a valid semantic version (X.Y.Z format)."""
+    pattern = r"^(\d+)\.(\d+)\.(\d+)$"
+    return bool(re.match(pattern, version))
+
+def get_installed_frontend_version():
+    """Get the currently installed frontend package version."""
+    frontend_version_str = version("comfyui-frontend-package")
+    return frontend_version_str
+
+def get_required_frontend_version():
+    """Get the required frontend version from requirements.txt."""
+    try:
+        with open(requirements_path, "r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith("comfyui-frontend-package=="):
+                    version_str = line.split("==")[-1]
+                    if not is_valid_version(version_str):
+                        logging.error(f"Invalid version format in requirements.txt: {version_str}")
+                        return None
+                    return version_str
+            logging.error("comfyui-frontend-package not found in requirements.txt")
+            return None
+    except FileNotFoundError:
+        logging.error("requirements.txt not found. Cannot determine required frontend version.")
+        return None
+    except Exception as e:
+        logging.error(f"Error reading requirements.txt: {e}")
+        return None
 
 def check_frontend_version():
     """Check if the frontend version is up to date."""
 
-    def parse_version(version: str) -> tuple[int, int, int]:
-        return tuple(map(int, version.split(".")))
-
     try:
-        frontend_version_str = version("comfyui-frontend-package")
+        frontend_version_str = get_installed_frontend_version()
         frontend_version = parse_version(frontend_version_str)
-        with open(requirements_path, "r", encoding="utf-8") as f:
-            required_frontend = parse_version(f.readline().split("=")[-1])
+        required_frontend_str = get_required_frontend_version()
+        required_frontend = parse_version(required_frontend_str)
         if frontend_version < required_frontend:
             app.logger.log_startup_warning(
                 f"""
@@ -168,6 +198,11 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None:
 class FrontendManager:
     CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
 
+    @classmethod
+    def get_required_frontend_version(cls) -> str:
+        """Get the required frontend package version."""
+        return get_required_frontend_version()
+
     @classmethod
     def default_frontend_path(cls) -> str:
         try:

app/model_manager.py

@@ -130,10 +130,21 @@ class ModelFileManager:
 
             for file_name in filenames:
                 try:
-                    relative_path = os.path.relpath(os.path.join(dirpath, file_name), directory)
-                    result.append(relative_path)
-                except:
-                    logging.warning(f"Warning: Unable to access {file_name}. Skipping this file.")
+                    full_path = os.path.join(dirpath, file_name)
+                    relative_path = os.path.relpath(full_path, directory)
+
+                    # Get file metadata
+                    file_info = {
+                        "name": relative_path,
+                        "pathIndex": pathIndex,
+                        "modified": os.path.getmtime(full_path),  # Add modification time
+                        "created": os.path.getctime(full_path),   # Add creation time
+                        "size": os.path.getsize(full_path)        # Add file size
+                    }
+                    result.append(file_info)
+
+                except Exception as e:
+                    logging.warning(f"Warning: Unable to access {file_name}. Error: {e}. Skipping this file.")
                     continue
 
             for d in subdirs:
@@ -144,7 +155,7 @@ class ModelFileManager:
                     logging.warning(f"Warning: Unable to access {path}. Skipping this path.")
                     continue
 
-        return [{"name": f, "pathIndex": pathIndex} for f in result], dirs, time.perf_counter()
+        return result, dirs, time.perf_counter()
 
     def get_model_previews(self, filepath: str) -> list[str | BytesIO]:
         dirname = os.path.dirname(filepath)

app/user_manager.py

@@ -20,13 +20,15 @@ class FileInfo(TypedDict):
     path: str
     size: int
     modified: int
+    created: int
 
 
 def get_file_info(path: str, relative_to: str) -> FileInfo:
     return {
         "path": os.path.relpath(path, relative_to).replace(os.sep, '/'),
         "size": os.path.getsize(path),
-        "modified": os.path.getmtime(path)
+        "modified": os.path.getmtime(path),
+        "created": os.path.getctime(path)
     }
 
 

comfy/cli_args.py

@@ -49,7 +49,8 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
 parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
-parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
+parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.")
+parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.")
 cm_group = parser.add_mutually_exclusive_group()
 cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
 cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")
@@ -131,6 +132,8 @@ parser.add_argument("--reserve-vram", type=float, default=None, help="Set the am
 
 parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
+parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
+
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
@@ -144,6 +147,7 @@ class PerformanceFeature(enum.Enum):
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: fp16_accumulation fp8_matrix_mult cublas_ops")
 
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
+parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")
 
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
@@ -151,6 +155,7 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win
 
 parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
 parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
+parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.")
 parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")

comfy/conds.py

@@ -1,6 +1,7 @@
 import torch
 import math
 import comfy.utils
+import logging
 
 
 class CONDRegular:
@@ -10,12 +11,15 @@ class CONDRegular:
     def _copy_with(self, cond):
         return self.__class__(cond)
 
-    def process_cond(self, batch_size, device, **kwargs):
-        return self._copy_with(comfy.utils.repeat_to_batch_size(self.cond, batch_size).to(device))
+    def process_cond(self, batch_size, **kwargs):
+        return self._copy_with(comfy.utils.repeat_to_batch_size(self.cond, batch_size))
 
     def can_concat(self, other):
         if self.cond.shape != other.cond.shape:
             return False
+        if self.cond.device != other.cond.device:
+            logging.warning("WARNING: conds not on same device, skipping concat.")
+            return False
         return True
 
     def concat(self, others):
@@ -29,14 +33,14 @@ class CONDRegular:
 
 
 class CONDNoiseShape(CONDRegular):
-    def process_cond(self, batch_size, device, area, **kwargs):
+    def process_cond(self, batch_size, area, **kwargs):
         data = self.cond
         if area is not None:
             dims = len(area) // 2
             for i in range(dims):
                 data = data.narrow(i + 2, area[i + dims], area[i])
 
-        return self._copy_with(comfy.utils.repeat_to_batch_size(data, batch_size).to(device))
+        return self._copy_with(comfy.utils.repeat_to_batch_size(data, batch_size))
 
 
 class CONDCrossAttn(CONDRegular):
@@ -51,6 +55,9 @@ class CONDCrossAttn(CONDRegular):
             diff = mult_min // min(s1[1], s2[1])
             if diff > 4: #arbitrary limit on the padding because it's probably going to impact performance negatively if it's too much
                 return False
+        if self.cond.device != other.cond.device:
+            logging.warning("WARNING: conds not on same device: skipping concat.")
+            return False
         return True
 
     def concat(self, others):
@@ -73,7 +80,7 @@ class CONDConstant(CONDRegular):
     def __init__(self, cond):
         self.cond = cond
 
-    def process_cond(self, batch_size, device, **kwargs):
+    def process_cond(self, batch_size, **kwargs):
         return self._copy_with(self.cond)
 
     def can_concat(self, other):
@@ -92,10 +99,10 @@ class CONDList(CONDRegular):
     def __init__(self, cond):
         self.cond = cond
 
-    def process_cond(self, batch_size, device, **kwargs):
+    def process_cond(self, batch_size, **kwargs):
         out = []
         for c in self.cond:
-            out.append(comfy.utils.repeat_to_batch_size(c, batch_size).to(device))
+            out.append(comfy.utils.repeat_to_batch_size(c, batch_size))
 
         return self._copy_with(out)
 

comfy/context_windows.py

@@ -0,0 +1,540 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Callable
+import torch
+import numpy as np
+import collections
+from dataclasses import dataclass
+from abc import ABC, abstractmethod
+import logging
+import comfy.model_management
+import comfy.patcher_extension
+if TYPE_CHECKING:
+    from comfy.model_base import BaseModel
+    from comfy.model_patcher import ModelPatcher
+    from comfy.controlnet import ControlBase
+
+
+class ContextWindowABC(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def get_tensor(self, full: torch.Tensor) -> torch.Tensor:
+        """
+        Get torch.Tensor applicable to current window.
+        """
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def add_window(self, full: torch.Tensor, to_add: torch.Tensor) -> torch.Tensor:
+        """
+        Apply torch.Tensor of window to the full tensor, in place. Returns reference to updated full tensor, not a copy.
+        """
+        raise NotImplementedError("Not implemented.")
+
+class ContextHandlerABC(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: ContextWindowABC, device=None) -> list:
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        raise NotImplementedError("Not implemented.")
+
+
+
+class IndexListContextWindow(ContextWindowABC):
+    def __init__(self, index_list: list[int], dim: int=0):
+        self.index_list = index_list
+        self.context_length = len(index_list)
+        self.dim = dim
+
+    def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor:
+        if dim is None:
+            dim = self.dim
+        if dim == 0 and full.shape[dim] == 1:
+            return full
+        idx = [slice(None)] * dim + [self.index_list]
+        return full[idx].to(device)
+
+    def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor:
+        if dim is None:
+            dim = self.dim
+        idx = [slice(None)] * dim + [self.index_list]
+        full[idx] += to_add
+        return full
+
+
+class IndexListCallbacks:
+    EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows"
+    COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
+    EXECUTE_START = "execute_start"
+    EXECUTE_CLEANUP = "execute_cleanup"
+
+    def init_callbacks(self):
+        return {}
+
+
+@dataclass
+class ContextSchedule:
+    name: str
+    func: Callable
+
+@dataclass
+class ContextFuseMethod:
+    name: str
+    func: Callable
+
+ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
+class IndexListContextHandler(ContextHandlerABC):
+    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, closed_loop=False, dim=0):
+        self.context_schedule = context_schedule
+        self.fuse_method = fuse_method
+        self.context_length = context_length
+        self.context_overlap = context_overlap
+        self.context_stride = context_stride
+        self.closed_loop = closed_loop
+        self.dim = dim
+        self._step = 0
+
+        self.callbacks = {}
+
+    def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
+        # for now, assume first dim is batch - should have stored on BaseModel in actual implementation
+        if x_in.size(self.dim) > self.context_length:
+            logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.")
+            return True
+        return False
+
+    def prepare_control_objects(self, control: ControlBase, device=None) -> ControlBase:
+        if control.previous_controlnet is not None:
+            self.prepare_control_objects(control.previous_controlnet, device)
+        return control
+
+    def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: IndexListContextWindow, device=None) -> list:
+        if cond_in is None:
+            return None
+        # reuse or resize cond items to match context requirements
+        resized_cond = []
+        # cond object is a list containing a dict - outer list is irrelevant, so just loop through it
+        for actual_cond in cond_in:
+            resized_actual_cond = actual_cond.copy()
+            # now we are in the inner dict - "pooled_output" is a tensor, "control" is a ControlBase object, "model_conds" is dictionary
+            for key in actual_cond:
+                try:
+                    cond_item = actual_cond[key]
+                    if isinstance(cond_item, torch.Tensor):
+                        # check that tensor is the expected length - x.size(0)
+                        if self.dim < cond_item.ndim and cond_item.size(self.dim) == x_in.size(self.dim):
+                            # if so, it's subsetting time - tell controls the expected indeces so they can handle them
+                            actual_cond_item = window.get_tensor(cond_item)
+                            resized_actual_cond[key] = actual_cond_item.to(device)
+                        else:
+                            resized_actual_cond[key] = cond_item.to(device)
+                    # look for control
+                    elif key == "control":
+                        resized_actual_cond[key] = self.prepare_control_objects(cond_item, device)
+                    elif isinstance(cond_item, dict):
+                        new_cond_item = cond_item.copy()
+                        # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
+                        for cond_key, cond_value in new_cond_item.items():
+                            if isinstance(cond_value, torch.Tensor):
+                                if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = window.get_tensor(cond_value, device)
+                            # if has cond that is a Tensor, check if needs to be subset
+                            elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device))
+                            elif cond_key == "num_video_frames": # for SVD
+                                new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond)
+                                new_cond_item[cond_key].cond = window.context_length
+                        resized_actual_cond[key] = new_cond_item
+                    else:
+                        resized_actual_cond[key] = cond_item
+                finally:
+                    del cond_item  # just in case to prevent VRAM issues
+            resized_cond.append(resized_actual_cond)
+        return resized_cond
+
+    def set_step(self, timestep: torch.Tensor, model_options: dict[str]):
+        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001)
+        matches = torch.nonzero(mask)
+        if torch.numel(matches) == 0:
+            raise Exception("No sample_sigmas matched current timestep; something went wrong.")
+        self._step = int(matches[0].item())
+
+    def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:
+        full_length = x_in.size(self.dim) # TODO: choose dim based on model
+        context_windows = self.context_schedule.func(full_length, self, model_options)
+        context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows]
+        return context_windows
+
+    def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        self.set_step(timestep, model_options)
+        context_windows = self.get_context_windows(model, x_in, model_options)
+        enumerated_context_windows = list(enumerate(context_windows))
+
+        conds_final = [torch.zeros_like(x_in) for _ in conds]
+        if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+            counts_final = [torch.ones(get_shape_for_dim(x_in, self.dim), device=x_in.device) for _ in conds]
+        else:
+            counts_final = [torch.zeros(get_shape_for_dim(x_in, self.dim), device=x_in.device) for _ in conds]
+        biases_final = [([0.0] * x_in.shape[self.dim]) for _ in conds]
+
+        for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EXECUTE_START, self.callbacks):
+            callback(self, model, x_in, conds, timestep, model_options)
+
+        for enum_window in enumerated_context_windows:
+            results = self.evaluate_context_windows(calc_cond_batch, model, x_in, conds, timestep, [enum_window], model_options)
+            for result in results:
+                self.combine_context_window_results(x_in, result.sub_conds_out, result.sub_conds, result.window, result.window_idx, len(enumerated_context_windows), timestep,
+                                            conds_final, counts_final, biases_final)
+        try:
+            # finalize conds
+            if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+                # relative is already normalized, so return as is
+                del counts_final
+                return conds_final
+            else:
+                # normalize conds via division by context usage counts
+                for i in range(len(conds_final)):
+                    conds_final[i] /= counts_final[i]
+                del counts_final
+                return conds_final
+        finally:
+            for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EXECUTE_CLEANUP, self.callbacks):
+                callback(self, model, x_in, conds, timestep, model_options)
+
+    def evaluate_context_windows(self, calc_cond_batch: Callable, model: BaseModel, x_in: torch.Tensor, conds, timestep: torch.Tensor, enumerated_context_windows: list[tuple[int, IndexListContextWindow]],
+                                model_options, device=None, first_device=None):
+        results: list[ContextResults] = []
+        for window_idx, window in enumerated_context_windows:
+            # allow processing to end between context window executions for faster Cancel
+            comfy.model_management.throw_exception_if_processing_interrupted()
+
+            for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EVALUATE_CONTEXT_WINDOWS, self.callbacks):
+                callback(self, model, x_in, conds, timestep, model_options, window_idx, window, model_options, device, first_device)
+
+            # update exposed params
+            model_options["transformer_options"]["context_window"] = window
+            # get subsections of x, timestep, conds
+            sub_x = window.get_tensor(x_in, device)
+            sub_timestep = window.get_tensor(timestep, device, dim=0)
+            sub_conds = [self.get_resized_cond(cond, x_in, window, device) for cond in conds]
+
+            sub_conds_out = calc_cond_batch(model, sub_conds, sub_x, sub_timestep, model_options)
+            if device is not None:
+                for i in range(len(sub_conds_out)):
+                    sub_conds_out[i] = sub_conds_out[i].to(x_in.device)
+            results.append(ContextResults(window_idx, sub_conds_out, sub_conds, window))
+        return results
+
+
+    def combine_context_window_results(self, x_in: torch.Tensor, sub_conds_out, sub_conds, window: IndexListContextWindow, window_idx: int, total_windows: int, timestep: torch.Tensor,
+                                    conds_final: list[torch.Tensor], counts_final: list[torch.Tensor], biases_final: list[torch.Tensor]):
+        if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+            for pos, idx in enumerate(window.index_list):
+                # bias is the influence of a specific index in relation to the whole context window
+                bias = 1 - abs(idx - (window.index_list[0] + window.index_list[-1]) / 2) / ((window.index_list[-1] - window.index_list[0] + 1e-2) / 2)
+                bias = max(1e-2, bias)
+                # take weighted average relative to total bias of current idx
+                for i in range(len(sub_conds_out)):
+                    bias_total = biases_final[i][idx]
+                    prev_weight = (bias_total / (bias_total + bias))
+                    new_weight = (bias / (bias_total + bias))
+                    # account for dims of tensors
+                    idx_window = [slice(None)] * self.dim + [idx]
+                    pos_window = [slice(None)] * self.dim + [pos]
+                    # apply new values
+                    conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight
+                    biases_final[i][idx] = bias_total + bias
+        else:
+            # add conds and counts based on weights of fuse method
+            weights = get_context_weights(window.context_length, x_in.shape[self.dim], window.index_list, self, sigma=timestep)
+            weights_tensor = match_weights_to_dim(weights, x_in, self.dim, device=x_in.device)
+            for i in range(len(sub_conds_out)):
+                window.add_window(conds_final[i], sub_conds_out[i] * weights_tensor)
+                window.add_window(counts_final[i], weights_tensor)
+
+        for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.COMBINE_CONTEXT_WINDOW_RESULTS, self.callbacks):
+            callback(self, x_in, sub_conds_out, sub_conds, window, window_idx, total_windows, timestep, conds_final, counts_final, biases_final)
+
+
+def _prepare_sampling_wrapper(executor, model, noise_shape: torch.Tensor, *args, **kwargs):
+    # limit noise_shape length to context_length for more accurate vram use estimation
+    model_options = kwargs.get("model_options", None)
+    if model_options is None:
+        raise Exception("model_options not found in prepare_sampling_wrapper; this should never happen, something went wrong.")
+    handler: IndexListContextHandler = model_options.get("context_handler", None)
+    if handler is not None:
+        noise_shape = list(noise_shape)
+        noise_shape[handler.dim] = min(noise_shape[handler.dim], handler.context_length)
+    return executor(model, noise_shape, *args, **kwargs)
+
+
+def create_prepare_sampling_wrapper(model: ModelPatcher):
+    model.add_wrapper_with_key(
+        comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING,
+        "ContextWindows_prepare_sampling",
+        _prepare_sampling_wrapper
+    )
+
+
+def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor:
+    total_dims = len(x_in.shape)
+    weights_tensor = torch.Tensor(weights).to(device=device)
+    for _ in range(dim):
+        weights_tensor = weights_tensor.unsqueeze(0)
+    for _ in range(total_dims - dim - 1):
+        weights_tensor = weights_tensor.unsqueeze(-1)
+    return weights_tensor
+
+def get_shape_for_dim(x_in: torch.Tensor, dim: int) -> list[int]:
+    total_dims = len(x_in.shape)
+    shape = []
+    for _ in range(dim):
+        shape.append(1)
+    shape.append(x_in.shape[dim])
+    for _ in range(total_dims - dim - 1):
+        shape.append(1)
+    return shape
+
+class ContextSchedules:
+    UNIFORM_LOOPED = "looped_uniform"
+    UNIFORM_STANDARD = "standard_uniform"
+    STATIC_STANDARD = "standard_static"
+    BATCHED = "batched"
+
+
+# from https://github.com/neggles/animatediff-cli/blob/main/src/animatediff/pipelines/context.py
+def create_windows_uniform_looped(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames < handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+
+    context_stride = min(handler.context_stride, int(np.ceil(np.log2(num_frames / handler.context_length))) + 1)
+    # obtain uniform windows as normal, looping and all
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(handler._step)))
+        for j in range(
+            int(ordered_halving(handler._step) * context_step) + pad,
+            num_frames + pad + (0 if handler.closed_loop else -handler.context_overlap),
+            (handler.context_length * context_step - handler.context_overlap),
+        ):
+            windows.append([e % num_frames for e in range(j, j + handler.context_length * context_step, context_step)])
+
+    return windows
+
+def create_windows_uniform_standard(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    # unlike looped, uniform_straight does NOT allow windows that loop back to the beginning;
+    # instead, they get shifted to the corresponding end of the frames.
+    # in the case that a window (shifted or not) is identical to the previous one, it gets skipped.
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+
+    context_stride = min(handler.context_stride, int(np.ceil(np.log2(num_frames / handler.context_length))) + 1)
+    # first, obtain uniform windows as normal, looping and all
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(handler._step)))
+        for j in range(
+            int(ordered_halving(handler._step) * context_step) + pad,
+            num_frames + pad + (-handler.context_overlap),
+            (handler.context_length * context_step - handler.context_overlap),
+        ):
+            windows.append([e % num_frames for e in range(j, j + handler.context_length * context_step, context_step)])
+
+    # now that windows are created, shift any windows that loop, and delete duplicate windows
+    delete_idxs = []
+    win_i = 0
+    while win_i < len(windows):
+        # if window is rolls over itself, need to shift it
+        is_roll, roll_idx = does_window_roll_over(windows[win_i], num_frames)
+        if is_roll:
+            roll_val = windows[win_i][roll_idx]  # roll_val might not be 0 for windows of higher strides
+            shift_window_to_end(windows[win_i], num_frames=num_frames)
+            # check if next window (cyclical) is missing roll_val
+            if roll_val not in windows[(win_i+1) % len(windows)]:
+                # need to insert new window here - just insert window starting at roll_val
+                windows.insert(win_i+1, list(range(roll_val, roll_val + handler.context_length)))
+        # delete window if it's not unique
+        for pre_i in range(0, win_i):
+            if windows[win_i] == windows[pre_i]:
+                delete_idxs.append(win_i)
+                break
+        win_i += 1
+
+    # reverse delete_idxs so that they will be deleted in an order that doesn't break idx correlation
+    delete_idxs.reverse()
+    for i in delete_idxs:
+        windows.pop(i)
+
+    return windows
+
+
+def create_windows_static_standard(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+    # always return the same set of windows
+    delta = handler.context_length - handler.context_overlap
+    for start_idx in range(0, num_frames, delta):
+        # if past the end of frames, move start_idx back to allow same context_length
+        ending = start_idx + handler.context_length
+        if ending >= num_frames:
+            final_delta = ending - num_frames
+            final_start_idx = start_idx - final_delta
+            windows.append(list(range(final_start_idx, final_start_idx + handler.context_length)))
+            break
+        windows.append(list(range(start_idx, start_idx + handler.context_length)))
+    return windows
+
+
+def create_windows_batched(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+    # always return the same set of windows;
+    # no overlap, just cut up based on context_length;
+    # last window size will be different if num_frames % opts.context_length != 0
+    for start_idx in range(0, num_frames, handler.context_length):
+        windows.append(list(range(start_idx, min(start_idx + handler.context_length, num_frames))))
+    return windows
+
+
+def create_windows_default(num_frames: int, handler: IndexListContextHandler):
+    return [list(range(num_frames))]
+
+
+CONTEXT_MAPPING = {
+    ContextSchedules.UNIFORM_LOOPED: create_windows_uniform_looped,
+    ContextSchedules.UNIFORM_STANDARD: create_windows_uniform_standard,
+    ContextSchedules.STATIC_STANDARD: create_windows_static_standard,
+    ContextSchedules.BATCHED: create_windows_batched,
+}
+
+
+def get_matching_context_schedule(context_schedule: str) -> ContextSchedule:
+    func = CONTEXT_MAPPING.get(context_schedule, None)
+    if func is None:
+        raise ValueError(f"Unknown context_schedule '{context_schedule}'.")
+    return ContextSchedule(context_schedule, func)
+
+
+def get_context_weights(length: int, full_length: int, idxs: list[int], handler: IndexListContextHandler, sigma: torch.Tensor=None):
+    return handler.fuse_method.func(length, sigma=sigma, handler=handler, full_length=full_length, idxs=idxs)
+
+
+def create_weights_flat(length: int, **kwargs) -> list[float]:
+    # weight is the same for all
+    return [1.0] * length
+
+def create_weights_pyramid(length: int, **kwargs) -> list[float]:
+    # weight is based on the distance away from the edge of the context window;
+    # based on weighted average concept in FreeNoise paper
+    if length % 2 == 0:
+        max_weight = length // 2
+        weight_sequence = list(range(1, max_weight + 1, 1)) + list(range(max_weight, 0, -1))
+    else:
+        max_weight = (length + 1) // 2
+        weight_sequence = list(range(1, max_weight, 1)) + [max_weight] + list(range(max_weight - 1, 0, -1))
+    return weight_sequence
+
+def create_weights_overlap_linear(length: int, full_length: int, idxs: list[int], handler: IndexListContextHandler, **kwargs):
+    # based on code in Kijai's WanVideoWrapper: https://github.com/kijai/ComfyUI-WanVideoWrapper/blob/dbb2523b37e4ccdf45127e5ae33e31362f755c8e/nodes.py#L1302
+    # only expected overlap is given different weights
+    weights_torch = torch.ones((length))
+    # blend left-side on all except first window
+    if min(idxs) > 0:
+        ramp_up = torch.linspace(1e-37, 1, handler.context_overlap)
+        weights_torch[:handler.context_overlap] = ramp_up
+    # blend right-side on all except last window
+    if max(idxs) < full_length-1:
+        ramp_down = torch.linspace(1, 1e-37, handler.context_overlap)
+        weights_torch[-handler.context_overlap:] = ramp_down
+    return weights_torch
+
+class ContextFuseMethods:
+    FLAT = "flat"
+    PYRAMID = "pyramid"
+    RELATIVE = "relative"
+    OVERLAP_LINEAR = "overlap-linear"
+
+    LIST = [PYRAMID, FLAT, OVERLAP_LINEAR]
+    LIST_STATIC = [PYRAMID, RELATIVE, FLAT, OVERLAP_LINEAR]
+
+
+FUSE_MAPPING = {
+    ContextFuseMethods.FLAT: create_weights_flat,
+    ContextFuseMethods.PYRAMID: create_weights_pyramid,
+    ContextFuseMethods.RELATIVE: create_weights_pyramid,
+    ContextFuseMethods.OVERLAP_LINEAR: create_weights_overlap_linear,
+}
+
+def get_matching_fuse_method(fuse_method: str) -> ContextFuseMethod:
+    func = FUSE_MAPPING.get(fuse_method, None)
+    if func is None:
+        raise ValueError(f"Unknown fuse_method '{fuse_method}'.")
+    return ContextFuseMethod(fuse_method, func)
+
+# Returns fraction that has denominator that is a power of 2
+def ordered_halving(val):
+    # get binary value, padded with 0s for 64 bits
+    bin_str = f"{val:064b}"
+    # flip binary value, padding included
+    bin_flip = bin_str[::-1]
+    # convert binary to int
+    as_int = int(bin_flip, 2)
+    # divide by 1 << 64, equivalent to 2**64, or 18446744073709551616,
+    # or b10000000000000000000000000000000000000000000000000000000000000000 (1 with 64 zero's)
+    return as_int / (1 << 64)
+
+
+def get_missing_indexes(windows: list[list[int]], num_frames: int) -> list[int]:
+    all_indexes = list(range(num_frames))
+    for w in windows:
+        for val in w:
+            try:
+                all_indexes.remove(val)
+            except ValueError:
+                pass
+    return all_indexes
+
+
+def does_window_roll_over(window: list[int], num_frames: int) -> tuple[bool, int]:
+    prev_val = -1
+    for i, val in enumerate(window):
+        val = val % num_frames
+        if val < prev_val:
+            return True, i
+        prev_val = val
+    return False, -1
+
+
+def shift_window_to_start(window: list[int], num_frames: int):
+    start_val = window[0]
+    for i in range(len(window)):
+        # 1) subtract each element by start_val to move vals relative to the start of all frames
+        # 2) add num_frames and take modulus to get adjusted vals
+        window[i] = ((window[i] - start_val) + num_frames) % num_frames
+
+
+def shift_window_to_end(window: list[int], num_frames: int):
+    # 1) shift window to start
+    shift_window_to_start(window, num_frames)
+    end_val = window[-1]
+    end_delta = num_frames - end_val - 1
+    for i in range(len(window)):
+        # 2) add end_delta to each val to slide windows to end
+        window[i] = window[i] + end_delta

comfy/controlnet.py

@@ -28,6 +28,7 @@ import comfy.model_detection
 import comfy.model_patcher
 import comfy.ops
 import comfy.latent_formats
+import comfy.model_base
 
 import comfy.cldm.cldm
 import comfy.t2i_adapter.adapter
@@ -43,7 +44,6 @@ if TYPE_CHECKING:
 
 def broadcast_image_to(tensor, target_batch_size, batched_number):
     current_batch_size = tensor.shape[0]
-    #print(current_batch_size, target_batch_size)
     if current_batch_size == 1:
         return tensor
 
@@ -265,12 +265,12 @@ class ControlNet(ControlBase):
         for c in self.extra_conds:
             temp = cond.get(c, None)
             if temp is not None:
-                extra[c] = temp.to(dtype)
+                extra[c] = comfy.model_base.convert_tensor(temp, dtype, x_noisy.device)
 
         timestep = self.model_sampling_current.timestep(t)
         x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
 
-        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=context.to(dtype), **extra)
+        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=comfy.model_management.cast_to_device(context, x_noisy.device, dtype), **extra)
         return self.control_merge(control, control_prev, output_dtype=None)
 
     def copy(self):

comfy/gligen.py

@@ -1,55 +1,10 @@
 import math
 import torch
 from torch import nn
-from .ldm.modules.attention import CrossAttention
-from inspect import isfunction
+from .ldm.modules.attention import CrossAttention, FeedForward
 import comfy.ops
 ops = comfy.ops.manual_cast
 
-def exists(val):
-    return val is not None
-
-
-def uniq(arr):
-    return{el: True for el in arr}.keys()
-
-
-def default(val, d):
-    if exists(val):
-        return val
-    return d() if isfunction(d) else d
-
-
-# feedforward
-class GEGLU(nn.Module):
-    def __init__(self, dim_in, dim_out):
-        super().__init__()
-        self.proj = ops.Linear(dim_in, dim_out * 2)
-
-    def forward(self, x):
-        x, gate = self.proj(x).chunk(2, dim=-1)
-        return x * torch.nn.functional.gelu(gate)
-
-
-class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
-        super().__init__()
-        inner_dim = int(dim * mult)
-        dim_out = default(dim_out, dim)
-        project_in = nn.Sequential(
-            ops.Linear(dim, inner_dim),
-            nn.GELU()
-        ) if not glu else GEGLU(dim, inner_dim)
-
-        self.net = nn.Sequential(
-            project_in,
-            nn.Dropout(dropout),
-            ops.Linear(inner_dim, dim_out)
-        )
-
-    def forward(self, x):
-        return self.net(x)
-
 
 class GatedCrossAttentionDense(nn.Module):
     def __init__(self, query_dim, context_dim, n_heads, d_head):

comfy/k_diffusion/sa_solver.py

@@ -0,0 +1,121 @@
+# SA-Solver: Stochastic Adams Solver (NeurIPS 2023, arXiv:2309.05019)
+# Conference: https://proceedings.neurips.cc/paper_files/paper/2023/file/f4a6806490d31216a3ba667eb240c897-Paper-Conference.pdf
+# Codebase ref: https://github.com/scxue/SA-Solver
+
+import math
+from typing import Union, Callable
+import torch
+
+
+def compute_exponential_coeffs(s: torch.Tensor, t: torch.Tensor, solver_order: int, tau_t: float) -> torch.Tensor:
+    """Compute (1 + tau^2) * integral of exp((1 + tau^2) * x) * x^p dx from s to t with exp((1 + tau^2) * t) factored out, using integration by parts.
+
+    Integral of exp((1 + tau^2) * x) * x^p dx
+        = product_terms[p] - (p / (1 + tau^2)) * integral of exp((1 + tau^2) * x) * x^(p-1) dx,
+    with base case p=0 where integral equals product_terms[0].
+
+    where
+        product_terms[p] = x^p * exp((1 + tau^2) * x) / (1 + tau^2).
+
+    Construct a recursive coefficient matrix following the above recursive relation to compute all integral terms up to p = (solver_order - 1).
+    Return coefficients used by the SA-Solver in data prediction mode.
+
+    Args:
+        s: Start time s.
+        t: End time t.
+        solver_order: Current order of the solver.
+        tau_t: Stochastic strength parameter in the SDE.
+
+    Returns:
+        Exponential coefficients used in data prediction, with exp((1 + tau^2) * t) factored out, ordered from p=0 to p=solver_order−1, shape (solver_order,).
+    """
+    tau_mul = 1 + tau_t ** 2
+    h = t - s
+    p = torch.arange(solver_order, dtype=s.dtype, device=s.device)
+
+    # product_terms after factoring out exp((1 + tau^2) * t)
+    # Includes (1 + tau^2) factor from outside the integral
+    product_terms_factored = (t ** p - s ** p * (-tau_mul * h).exp())
+
+    # Lower triangular recursive coefficient matrix
+    # Accumulates recursive coefficients based on p / (1 + tau^2)
+    recursive_depth_mat = p.unsqueeze(1) - p.unsqueeze(0)
+    log_factorial = (p + 1).lgamma()
+    recursive_coeff_mat = log_factorial.unsqueeze(1) - log_factorial.unsqueeze(0)
+    if tau_t > 0:
+        recursive_coeff_mat = recursive_coeff_mat - (recursive_depth_mat * math.log(tau_mul))
+    signs = torch.where(recursive_depth_mat % 2 == 0, 1.0, -1.0)
+    recursive_coeff_mat = (recursive_coeff_mat.exp() * signs).tril()
+
+    return recursive_coeff_mat @ product_terms_factored
+
+
+def compute_simple_stochastic_adams_b_coeffs(sigma_next: torch.Tensor, curr_lambdas: torch.Tensor, lambda_s: torch.Tensor, lambda_t: torch.Tensor, tau_t: float, is_corrector_step: bool = False) -> torch.Tensor:
+    """Compute simple order-2 b coefficients from SA-Solver paper (Appendix D. Implementation Details)."""
+    tau_mul = 1 + tau_t ** 2
+    h = lambda_t - lambda_s
+    alpha_t = sigma_next * lambda_t.exp()
+    if is_corrector_step:
+        # Simplified 1-step (order-2) corrector
+        b_1 = alpha_t * (0.5 * tau_mul * h)
+        b_2 = alpha_t * (-h * tau_mul).expm1().neg() - b_1
+    else:
+        # Simplified 2-step predictor
+        b_2 = alpha_t * (0.5 * tau_mul * h ** 2) / (curr_lambdas[-2] - lambda_s)
+        b_1 = alpha_t * (-h * tau_mul).expm1().neg() - b_2
+    return torch.stack([b_2, b_1])
+
+
+def compute_stochastic_adams_b_coeffs(sigma_next: torch.Tensor, curr_lambdas: torch.Tensor, lambda_s: torch.Tensor, lambda_t: torch.Tensor, tau_t: float, simple_order_2: bool = False, is_corrector_step: bool = False) -> torch.Tensor:
+    """Compute b_i coefficients for the SA-Solver (see eqs. 15 and 18).
+
+    The solver order corresponds to the number of input lambdas (half-logSNR points).
+
+    Args:
+        sigma_next: Sigma at end time t.
+        curr_lambdas: Lambda time points used to construct the Lagrange basis, shape (N,).
+        lambda_s: Lambda at start time s.
+        lambda_t: Lambda at end time t.
+        tau_t: Stochastic strength parameter in the SDE.
+        simple_order_2: Whether to enable the simple order-2 scheme.
+        is_corrector_step: Flag for corrector step in simple order-2 mode.
+
+    Returns:
+        b_i coefficients for the SA-Solver, shape (N,), where N is the solver order.
+    """
+    num_timesteps = curr_lambdas.shape[0]
+
+    if simple_order_2 and num_timesteps == 2:
+        return compute_simple_stochastic_adams_b_coeffs(sigma_next, curr_lambdas, lambda_s, lambda_t, tau_t, is_corrector_step)
+
+    # Compute coefficients by solving a linear system from Lagrange basis interpolation
+    exp_integral_coeffs = compute_exponential_coeffs(lambda_s, lambda_t, num_timesteps, tau_t)
+    vandermonde_matrix_T = torch.vander(curr_lambdas, num_timesteps, increasing=True).T
+    lagrange_integrals = torch.linalg.solve(vandermonde_matrix_T, exp_integral_coeffs)
+
+    # (sigma_t * exp(-tau^2 * lambda_t)) * exp((1 + tau^2) * lambda_t)
+    # = sigma_t * exp(lambda_t) = alpha_t
+    # exp((1 + tau^2) * lambda_t) is extracted from the integral
+    alpha_t = sigma_next * lambda_t.exp()
+    return alpha_t * lagrange_integrals
+
+
+def get_tau_interval_func(start_sigma: float, end_sigma: float, eta: float = 1.0) -> Callable[[Union[torch.Tensor, float]], float]:
+    """Return a function that controls the stochasticity of SA-Solver.
+
+    When eta = 0, SA-Solver runs as ODE. The official approach uses
+    time t to determine the SDE interval, while here we use sigma instead.
+
+    See:
+        https://github.com/scxue/SA-Solver/blob/main/README.md
+    """
+
+    def tau_func(sigma: Union[torch.Tensor, float]) -> float:
+        if eta <= 0:
+            return 0.0  # ODE
+
+        if isinstance(sigma, torch.Tensor):
+            sigma = sigma.item()
+        return eta if start_sigma >= sigma >= end_sigma else 0.0
+
+    return tau_func

comfy/k_diffusion/sampling.py

@@ -9,6 +9,7 @@ from tqdm.auto import trange, tqdm
 
 from . import utils
 from . import deis
+from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
 
@@ -412,9 +413,13 @@ def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, o
             ds.pop(0)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
-        cur_order = min(i + 1, order)
-        coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
-        x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            cur_order = min(i + 1, order)
+            coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
+            x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
     return x
 
 
@@ -1067,7 +1072,9 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
         d_cur = (x_cur - denoised) / t_cur
 
         order = min(max_order, i+1)
-        if order == 1:      # First Euler step.
+        if t_next == 0:     # Denoising step
+            x_next = denoised
+        elif order == 1:    # First Euler step.
             x_next = x_cur + (t_next - t_cur) * d_cur
         elif order == 2:    # Use one history point.
             x_next = x_cur + (t_next - t_cur) * (3 * d_cur - buffer_model[-1]) / 2
@@ -1085,6 +1092,7 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
 
     return x_next
 
+
 #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
 #under Apache 2 license
 def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
@@ -1108,7 +1116,9 @@ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=Non
         d_cur = (x_cur - denoised) / t_cur
 
         order = min(max_order, i+1)
-        if order == 1:      # First Euler step.
+        if t_next == 0:     # Denoising step
+            x_next = denoised
+        elif order == 1:    # First Euler step.
             x_next = x_cur + (t_next - t_cur) * d_cur
         elif order == 2:    # Use one history point.
             h_n = (t_next - t_cur)
@@ -1148,6 +1158,7 @@ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=Non
 
     return x_next
 
+
 #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
 #under Apache 2 license
 @torch.no_grad()
@@ -1198,39 +1209,22 @@ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None,
 
     return x_next
 
-@torch.no_grad()
-def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
-    extra_args = {} if extra_args is None else extra_args
-
-    temp = [0]
-    def post_cfg_function(args):
-        temp[0] = args["uncond_denoised"]
-        return args["denoised"]
-
-    model_options = extra_args.get("model_options", {}).copy()
-    extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
-
-    s_in = x.new_ones([x.shape[0]])
-    for i in trange(len(sigmas) - 1, disable=disable):
-        sigma_hat = sigmas[i]
-        denoised = model(x, sigma_hat * s_in, **extra_args)
-        d = to_d(x, sigma_hat, temp[0])
-        if callback is not None:
-            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
-        # Euler method
-        x = denoised + d * sigmas[i + 1]
-    return x
 
 @torch.no_grad()
 def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
-    """Ancestral sampling with Euler method steps."""
+    """Ancestral sampling with Euler method steps (CFG++)."""
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
 
-    temp = [0]
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+
+    uncond_denoised = None
+
     def post_cfg_function(args):
-        temp[0] = args["uncond_denoised"]
+        nonlocal uncond_denoised
+        uncond_denoised = args["uncond_denoised"]
         return args["denoised"]
 
     model_options = extra_args.get("model_options", {}).copy()
@@ -1239,15 +1233,33 @@ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=No
     s_in = x.new_ones([x.shape[0]])
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
-        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
-        d = to_d(x, sigmas[i], temp[0])
-        # Euler method
-        x = denoised + d * sigma_down
-        if sigmas[i + 1] > 0:
-            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            alpha_s = sigmas[i] * lambda_fn(sigmas[i]).exp()
+            alpha_t = sigmas[i + 1] * lambda_fn(sigmas[i + 1]).exp()
+            d = to_d(x, sigmas[i], alpha_s * uncond_denoised)   # to noise
+
+            # DDIM stochastic sampling
+            sigma_down, sigma_up = get_ancestral_step(sigmas[i] / alpha_s, sigmas[i + 1] / alpha_t, eta=eta)
+            sigma_down = alpha_t * sigma_down
+
+            # Euler method
+            x = alpha_t * denoised + sigma_down * d
+            if eta > 0 and s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
     return x
+
+
+@torch.no_grad()
+def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
+    """Euler method steps (CFG++)."""
+    return sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None)
+
+
 @torch.no_grad()
 def sample_dpmpp_2s_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
@@ -1404,6 +1416,7 @@ def sample_res_multistep_ancestral(model, x, sigmas, extra_args=None, callback=N
 def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
 
+
 @torch.no_grad()
 def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False):
     """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK"""
@@ -1430,31 +1443,32 @@ def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None,
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
         dt = sigmas[i + 1] - sigmas[i]
-        if i == 0:
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
             # Euler method
             if cfg_pp:
                 x = denoised + d * sigmas[i + 1]
             else:
                 x = x + d * dt
-        else:
-            # Gradient estimation
-            if cfg_pp:
+
+            if i >= 1:
+                # Gradient estimation
                 d_bar = (ge_gamma - 1) * (d - old_d)
-                x = denoised + d * sigmas[i + 1] + d_bar * dt
-            else:
-                d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
                 x = x + d_bar * dt
         old_d = d
     return x
 
+
 @torch.no_grad()
 def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
     return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True)
 
+
 @torch.no_grad()
-def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3):
-    """
-    Extended Reverse-Time SDE solver (VE ER-SDE-Solver-3). Arxiv: https://arxiv.org/abs/2309.06169.
+def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1.0, noise_sampler=None, noise_scaler=None, max_stage=3):
+    """Extended Reverse-Time SDE solver (VP ER-SDE-Solver-3). arXiv: https://arxiv.org/abs/2309.06169.
     Code reference: https://github.com/QinpengCui/ER-SDE-Solver/blob/main/er_sde_solver.py.
     """
     extra_args = {} if extra_args is None else extra_args
@@ -1462,12 +1476,18 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
     s_in = x.new_ones([x.shape[0]])
 
-    def default_noise_scaler(sigma):
-        return sigma * ((sigma ** 0.3).exp() + 10.0)
-    noise_scaler = default_noise_scaler if noise_scaler is None else noise_scaler
+    def default_er_sde_noise_scaler(x):
+        return x * ((x ** 0.3).exp() + 10.0)
+
+    noise_scaler = default_er_sde_noise_scaler if noise_scaler is None else noise_scaler
     num_integration_points = 200.0
     point_indice = torch.arange(0, num_integration_points, dtype=torch.float32, device=x.device)
 
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+    half_log_snrs = sigma_to_half_log_snr(sigmas, model_sampling)
+    er_lambdas = half_log_snrs.neg().exp()  # er_lambda_t = sigma_t / alpha_t
+
     old_denoised = None
     old_denoised_d = None
 
@@ -1478,32 +1498,36 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
         stage_used = min(max_stage, i + 1)
         if sigmas[i + 1] == 0:
             x = denoised
-        elif stage_used == 1:
-            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
-            x = r * x + (1 - r) * denoised
         else:
-            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
-            x = r * x + (1 - r) * denoised
+            er_lambda_s, er_lambda_t = er_lambdas[i], er_lambdas[i + 1]
+            alpha_s = sigmas[i] / er_lambda_s
+            alpha_t = sigmas[i + 1] / er_lambda_t
+            r_alpha = alpha_t / alpha_s
+            r = noise_scaler(er_lambda_t) / noise_scaler(er_lambda_s)
 
-            dt = sigmas[i + 1] - sigmas[i]
-            sigma_step_size = -dt / num_integration_points
-            sigma_pos = sigmas[i + 1] + point_indice * sigma_step_size
-            scaled_pos = noise_scaler(sigma_pos)
+            # Stage 1 Euler
+            x = r_alpha * r * x + alpha_t * (1 - r) * denoised
 
-            # Stage 2
-            s = torch.sum(1 / scaled_pos) * sigma_step_size
-            denoised_d = (denoised - old_denoised) / (sigmas[i] - sigmas[i - 1])
-            x = x + (dt + s * noise_scaler(sigmas[i + 1])) * denoised_d
+            if stage_used >= 2:
+                dt = er_lambda_t - er_lambda_s
+                lambda_step_size = -dt / num_integration_points
+                lambda_pos = er_lambda_t + point_indice * lambda_step_size
+                scaled_pos = noise_scaler(lambda_pos)
 
-            if stage_used >= 3:
-                # Stage 3
-                s_u = torch.sum((sigma_pos - sigmas[i]) / scaled_pos) * sigma_step_size
-                denoised_u = (denoised_d - old_denoised_d) / ((sigmas[i] - sigmas[i - 2]) / 2)
-                x = x + ((dt ** 2) / 2 + s_u * noise_scaler(sigmas[i + 1])) * denoised_u
-            old_denoised_d = denoised_d
+                # Stage 2
+                s = torch.sum(1 / scaled_pos) * lambda_step_size
+                denoised_d = (denoised - old_denoised) / (er_lambda_s - er_lambdas[i - 1])
+                x = x + alpha_t * (dt + s * noise_scaler(er_lambda_t)) * denoised_d
 
-        if s_noise != 0 and sigmas[i + 1] > 0:
-            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * (sigmas[i + 1] ** 2 - sigmas[i] ** 2 * r ** 2).sqrt().nan_to_num(nan=0.0)
+                if stage_used >= 3:
+                    # Stage 3
+                    s_u = torch.sum((lambda_pos - er_lambda_s) / scaled_pos) * lambda_step_size
+                    denoised_u = (denoised_d - old_denoised_d) / ((er_lambda_s - er_lambdas[i - 2]) / 2)
+                    x = x + alpha_t * ((dt ** 2) / 2 + s_u * noise_scaler(er_lambda_t)) * denoised_u
+                old_denoised_d = denoised_d
+
+            if s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * (er_lambda_t ** 2 - er_lambda_s ** 2 * r ** 2).sqrt().nan_to_num(nan=0.0)
         old_denoised = denoised
     return x
 
@@ -1625,3 +1649,113 @@ def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=Non
             if inject_noise:
                 x = x + sigmas[i + 1] * (noise_coeff_3 * noise_1 + noise_coeff_2 * noise_2 + noise_coeff_1 * noise_3) * s_noise
     return x
+
+
+@torch.no_grad()
+def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, use_pece=False, simple_order_2=False):
+    """Stochastic Adams Solver with predictor-corrector method (NeurIPS 2023)."""
+    if len(sigmas) <= 1:
+        return x
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+    lambdas = sigma_to_half_log_snr(sigmas, model_sampling=model_sampling)
+
+    if tau_func is None:
+        # Use default interval for stochastic sampling
+        start_sigma = model_sampling.percent_to_sigma(0.2)
+        end_sigma = model_sampling.percent_to_sigma(0.8)
+        tau_func = sa_solver.get_tau_interval_func(start_sigma, end_sigma, eta=1.0)
+
+    max_used_order = max(predictor_order, corrector_order)
+    x_pred = x  # x: current state, x_pred: predicted next state
+
+    h = 0.0
+    tau_t = 0.0
+    noise = 0.0
+    pred_list = []
+
+    # Lower order near the end to improve stability
+    lower_order_to_end = sigmas[-1].item() == 0
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        # Evaluation
+        denoised = model(x_pred, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({"x": x_pred, "i": i, "sigma": sigmas[i], "sigma_hat": sigmas[i], "denoised": denoised})
+        pred_list.append(denoised)
+        pred_list = pred_list[-max_used_order:]
+
+        predictor_order_used = min(predictor_order, len(pred_list))
+        if i == 0 or (sigmas[i + 1] == 0 and not use_pece):
+            corrector_order_used = 0
+        else:
+            corrector_order_used = min(corrector_order, len(pred_list))
+
+        if lower_order_to_end:
+            predictor_order_used = min(predictor_order_used, len(sigmas) - 2 - i)
+            corrector_order_used = min(corrector_order_used, len(sigmas) - 1 - i)
+
+        # Corrector
+        if corrector_order_used == 0:
+            # Update by the predicted state
+            x = x_pred
+        else:
+            curr_lambdas = lambdas[i - corrector_order_used + 1:i + 1]
+            b_coeffs = sa_solver.compute_stochastic_adams_b_coeffs(
+                sigmas[i],
+                curr_lambdas,
+                lambdas[i - 1],
+                lambdas[i],
+                tau_t,
+                simple_order_2,
+                is_corrector_step=True,
+            )
+            pred_mat = torch.stack(pred_list[-corrector_order_used:], dim=1)    # (B, K, ...)
+            corr_res = torch.tensordot(pred_mat, b_coeffs, dims=([1], [0]))  # (B, ...)
+            x = sigmas[i] / sigmas[i - 1] * (-(tau_t ** 2) * h).exp() * x + corr_res
+
+            if tau_t > 0 and s_noise > 0:
+                # The noise from the previous predictor step
+                x = x + noise
+
+            if use_pece:
+                # Evaluate the corrected state
+                denoised = model(x, sigmas[i] * s_in, **extra_args)
+                pred_list[-1] = denoised
+
+        # Predictor
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            tau_t = tau_func(sigmas[i + 1])
+            curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
+            b_coeffs = sa_solver.compute_stochastic_adams_b_coeffs(
+                sigmas[i + 1],
+                curr_lambdas,
+                lambdas[i],
+                lambdas[i + 1],
+                tau_t,
+                simple_order_2,
+                is_corrector_step=False,
+            )
+            pred_mat = torch.stack(pred_list[-predictor_order_used:], dim=1)    # (B, K, ...)
+            pred_res = torch.tensordot(pred_mat, b_coeffs, dims=([1], [0]))  # (B, ...)
+            h = lambdas[i + 1] - lambdas[i]
+            x_pred = sigmas[i + 1] / sigmas[i] * (-(tau_t ** 2) * h).exp() * x + pred_res
+
+            if tau_t > 0 and s_noise > 0:
+                noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise
+                x_pred = x_pred + noise
+    return x
+
+
+@torch.no_grad()
+def sample_sa_solver_pece(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, simple_order_2=False):
+    """Stochastic Adams Solver with PECE (Predict–Evaluate–Correct–Evaluate) mode (NeurIPS 2023)."""
+    return sample_sa_solver(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, tau_func=tau_func, s_noise=s_noise, noise_sampler=noise_sampler, predictor_order=predictor_order, corrector_order=corrector_order, use_pece=True, simple_order_2=simple_order_2)

comfy/latent_formats.py

@@ -457,6 +457,82 @@ class Wan21(LatentFormat):
         latents_std = self.latents_std.to(latent.device, latent.dtype)
         return latent * latents_std / self.scale_factor + latents_mean
 
+class Wan22(Wan21):
+    latent_channels = 48
+    latent_dimensions = 3
+
+    latent_rgb_factors = [
+            [ 0.0119,  0.0103,  0.0046],
+            [-0.1062, -0.0504,  0.0165],
+            [ 0.0140,  0.0409,  0.0491],
+            [-0.0813, -0.0677,  0.0607],
+            [ 0.0656,  0.0851,  0.0808],
+            [ 0.0264,  0.0463,  0.0912],
+            [ 0.0295,  0.0326,  0.0590],
+            [-0.0244, -0.0270,  0.0025],
+            [ 0.0443, -0.0102,  0.0288],
+            [-0.0465, -0.0090, -0.0205],
+            [ 0.0359,  0.0236,  0.0082],
+            [-0.0776,  0.0854,  0.1048],
+            [ 0.0564,  0.0264,  0.0561],
+            [ 0.0006,  0.0594,  0.0418],
+            [-0.0319, -0.0542, -0.0637],
+            [-0.0268,  0.0024,  0.0260],
+            [ 0.0539,  0.0265,  0.0358],
+            [-0.0359, -0.0312, -0.0287],
+            [-0.0285, -0.1032, -0.1237],
+            [ 0.1041,  0.0537,  0.0622],
+            [-0.0086, -0.0374, -0.0051],
+            [ 0.0390,  0.0670,  0.2863],
+            [ 0.0069,  0.0144,  0.0082],
+            [ 0.0006, -0.0167,  0.0079],
+            [ 0.0313, -0.0574, -0.0232],
+            [-0.1454, -0.0902, -0.0481],
+            [ 0.0714,  0.0827,  0.0447],
+            [-0.0304, -0.0574, -0.0196],
+            [ 0.0401,  0.0384,  0.0204],
+            [-0.0758, -0.0297, -0.0014],
+            [ 0.0568,  0.1307,  0.1372],
+            [-0.0055, -0.0310, -0.0380],
+            [ 0.0239, -0.0305,  0.0325],
+            [-0.0663, -0.0673, -0.0140],
+            [-0.0416, -0.0047, -0.0023],
+            [ 0.0166,  0.0112, -0.0093],
+            [-0.0211,  0.0011,  0.0331],
+            [ 0.1833,  0.1466,  0.2250],
+            [-0.0368,  0.0370,  0.0295],
+            [-0.3441, -0.3543, -0.2008],
+            [-0.0479, -0.0489, -0.0420],
+            [-0.0660, -0.0153,  0.0800],
+            [-0.0101,  0.0068,  0.0156],
+            [-0.0690, -0.0452, -0.0927],
+            [-0.0145,  0.0041,  0.0015],
+            [ 0.0421,  0.0451,  0.0373],
+            [ 0.0504, -0.0483, -0.0356],
+            [-0.0837,  0.0168,  0.0055]
+        ]
+
+    latent_rgb_factors_bias = [0.0317, -0.0878, -0.1388]
+
+    def __init__(self):
+        self.scale_factor = 1.0
+        self.latents_mean = torch.tensor([
+                -0.2289, -0.0052, -0.1323, -0.2339, -0.2799, 0.0174, 0.1838, 0.1557,
+                -0.1382, 0.0542, 0.2813, 0.0891, 0.1570, -0.0098, 0.0375, -0.1825,
+                -0.2246, -0.1207, -0.0698, 0.5109, 0.2665, -0.2108, -0.2158, 0.2502,
+                -0.2055, -0.0322, 0.1109, 0.1567, -0.0729, 0.0899, -0.2799, -0.1230,
+                -0.0313, -0.1649, 0.0117, 0.0723, -0.2839, -0.2083, -0.0520, 0.3748,
+                0.0152, 0.1957, 0.1433, -0.2944, 0.3573, -0.0548, -0.1681, -0.0667,
+            ]).view(1, self.latent_channels, 1, 1, 1)
+        self.latents_std = torch.tensor([
+                0.4765, 1.0364, 0.4514, 1.1677, 0.5313, 0.4990, 0.4818, 0.5013,
+                0.8158, 1.0344, 0.5894, 1.0901, 0.6885, 0.6165, 0.8454, 0.4978,
+                0.5759, 0.3523, 0.7135, 0.6804, 0.5833, 1.4146, 0.8986, 0.5659,
+                0.7069, 0.5338, 0.4889, 0.4917, 0.4069, 0.4999, 0.6866, 0.4093,
+                0.5709, 0.6065, 0.6415, 0.4944, 0.5726, 1.2042, 0.5458, 1.6887,
+                0.3971, 1.0600, 0.3943, 0.5537, 0.5444, 0.4089, 0.7468, 0.7744
+            ]).view(1, self.latent_channels, 1, 1, 1)
+
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1

comfy/ldm/chroma/model.py

@@ -254,13 +254,12 @@ class Chroma(nn.Module):
 
     def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
         bs, c, h, w = x.shape
-        patch_size = 2
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
 
-        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=self.patch_size, pw=self.patch_size)
 
-        h_len = ((h + (patch_size // 2)) // patch_size)
-        w_len = ((w + (patch_size // 2)) // patch_size)
+        h_len = ((h + (self.patch_size // 2)) // self.patch_size)
+        w_len = ((w + (self.patch_size // 2)) // self.patch_size)
         img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
         img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
         img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
@@ -268,4 +267,4 @@ class Chroma(nn.Module):
 
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
-        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h,:w]

comfy/ldm/cosmos/cosmos_tokenizer/utils.py

@@ -58,7 +58,8 @@ def is_odd(n: int) -> bool:
 
 
 def nonlinearity(x):
-    return x * torch.sigmoid(x)
+    # x * sigmoid(x)
+    return torch.nn.functional.silu(x)
 
 
 def Normalize(in_channels, num_groups=32):

comfy/ldm/flux/model.py

@@ -224,19 +224,27 @@ class Flux(nn.Module):
         if ref_latents is not None:
             h = 0
             w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "offset") == "index"
             for ref in ref_latents:
-                h_offset = 0
-                w_offset = 0
-                if ref.shape[-2] + h > ref.shape[-1] + w:
-                    w_offset = w
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
                 else:
-                    h_offset = h
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
 
-                kontext, kontext_ids = self.process_img(ref, index=1, h_offset=h_offset, w_offset=w_offset)
+                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
-                h = max(h, ref.shape[-2] + h_offset)
-                w = max(w, ref.shape[-1] + w_offset)
 
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))

comfy/ldm/hunyuan3d/vae.py

@@ -178,7 +178,7 @@ class FourierEmbedder(nn.Module):
 
 class CrossAttentionProcessor:
     def __call__(self, attn, q, k, v):
-        out = F.scaled_dot_product_attention(q, k, v)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v)
         return out
 
 

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -973,7 +973,7 @@ class VideoVAE(nn.Module):
             norm_layer=config.get("norm_layer", "group_norm"),
             causal=config.get("causal_decoder", False),
             timestep_conditioning=self.timestep_conditioning,
-            spatial_padding_mode=config.get("spatial_padding_mode", "zeros"),
+            spatial_padding_mode=config.get("spatial_padding_mode", "reflect"),
         )
 
         self.per_channel_statistics = processor()

comfy/ldm/models/autoencoder.py

@@ -11,7 +11,7 @@ from comfy.ldm.modules.ema import LitEma
 import comfy.ops
 
 class DiagonalGaussianRegularizer(torch.nn.Module):
-    def __init__(self, sample: bool = True):
+    def __init__(self, sample: bool = False):
         super().__init__()
         self.sample = sample
 
@@ -19,16 +19,12 @@ class DiagonalGaussianRegularizer(torch.nn.Module):
         yield from ()
 
     def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
-        log = dict()
         posterior = DiagonalGaussianDistribution(z)
         if self.sample:
             z = posterior.sample()
         else:
             z = posterior.mode()
-        kl_loss = posterior.kl()
-        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
-        log["kl_loss"] = kl_loss
-        return z, log
+        return z, None
 
 
 class AbstractAutoencoder(torch.nn.Module):

comfy/ldm/modules/attention.py

@@ -448,7 +448,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             mask = mask.unsqueeze(1)
 
     if SDP_BATCH_LIMIT >= b:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -461,7 +461,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 if mask.shape[0] > 1:
                     m = mask[i : i + SDP_BATCH_LIMIT]
 
-            out[i : i + SDP_BATCH_LIMIT] = torch.nn.functional.scaled_dot_product_attention(
+            out[i : i + SDP_BATCH_LIMIT] = comfy.ops.scaled_dot_product_attention(
                 q[i : i + SDP_BATCH_LIMIT],
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],

comfy/ldm/modules/diffusionmodules/model.py

@@ -36,7 +36,7 @@ def get_timestep_embedding(timesteps, embedding_dim):
 
 def nonlinearity(x):
     # swish
-    return x*torch.sigmoid(x)
+    return torch.nn.functional.silu(x)
 
 
 def Normalize(in_channels, num_groups=32):
@@ -285,7 +285,7 @@ def pytorch_attention(q, k, v):
     )
 
     try:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")

comfy/ldm/pixart/pixartms.py

@@ -1,256 +1,256 @@
-# Based on:
-# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
-# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
-import torch
-import torch.nn as nn
-
-from .blocks import (
-    t2i_modulate,
-    CaptionEmbedder,
-    AttentionKVCompress,
-    MultiHeadCrossAttention,
-    T2IFinalLayer,
-    SizeEmbedder,
-)
-from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
-
-
-def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
-    grid_h, grid_w = torch.meshgrid(
-        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
-        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
-        indexing='ij'
-    )
-    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
-    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
-    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
-    return emb
-
-class PixArtMSBlock(nn.Module):
-    """
-    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
-    """
-    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
-                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.attn = AttentionKVCompress(
-            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
-            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
-        )
-        self.cross_attn = MultiHeadCrossAttention(
-            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
-        )
-        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        # to be compatible with lower version pytorch
-        approx_gelu = lambda: nn.GELU(approximate="tanh")
-        self.mlp = Mlp(
-            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
-            dtype=dtype, device=device, operations=operations
-        )
-        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
-
-    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
-        B, N, C = x.shape
-
-        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
-        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
-        x = x + self.cross_attn(x, y, mask)
-        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
-
-        return x
-
-
-### Core PixArt Model ###
-class PixArtMS(nn.Module):
-    """
-    Diffusion model with a Transformer backbone.
-    """
-    def __init__(
-            self,
-            input_size=32,
-            patch_size=2,
-            in_channels=4,
-            hidden_size=1152,
-            depth=28,
-            num_heads=16,
-            mlp_ratio=4.0,
-            class_dropout_prob=0.1,
-            learn_sigma=True,
-            pred_sigma=True,
-            drop_path: float = 0.,
-            caption_channels=4096,
-            pe_interpolation=None,
-            pe_precision=None,
-            config=None,
-            model_max_length=120,
-            micro_condition=True,
-            qk_norm=False,
-            kv_compress_config=None,
-            dtype=None,
-            device=None,
-            operations=None,
-            **kwargs,
-    ):
-        nn.Module.__init__(self)
-        self.dtype = dtype
-        self.pred_sigma = pred_sigma
-        self.in_channels = in_channels
-        self.out_channels = in_channels * 2 if pred_sigma else in_channels
-        self.patch_size = patch_size
-        self.num_heads = num_heads
-        self.pe_interpolation = pe_interpolation
-        self.pe_precision = pe_precision
-        self.hidden_size = hidden_size
-        self.depth = depth
-
-        approx_gelu = lambda: nn.GELU(approximate="tanh")
-        self.t_block = nn.Sequential(
-            nn.SiLU(),
-            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
-        )
-        self.x_embedder = PatchEmbed(
-            patch_size=patch_size,
-            in_chans=in_channels,
-            embed_dim=hidden_size,
-            bias=True,
-            dtype=dtype,
-            device=device,
-            operations=operations
-        )
-        self.t_embedder = TimestepEmbedder(
-            hidden_size, dtype=dtype, device=device, operations=operations,
-        )
-        self.y_embedder = CaptionEmbedder(
-            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
-            act_layer=approx_gelu, token_num=model_max_length,
-            dtype=dtype, device=device, operations=operations,
-        )
-
-        self.micro_conditioning = micro_condition
-        if self.micro_conditioning:
-            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
-            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
-
-        # For fixed sin-cos embedding:
-        # num_patches = (input_size // patch_size) * (input_size // patch_size)
-        # self.base_size = input_size // self.patch_size
-        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
-
-        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
-        if kv_compress_config is None:
-            kv_compress_config = {
-                'sampling': None,
-                'scale_factor': 1,
-                'kv_compress_layer': [],
-            }
-        self.blocks = nn.ModuleList([
-            PixArtMSBlock(
-                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
-                sampling=kv_compress_config['sampling'],
-                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
-                qk_norm=qk_norm,
-                dtype=dtype,
-                device=device,
-                operations=operations,
-            )
-            for i in range(depth)
-        ])
-        self.final_layer = T2IFinalLayer(
-            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
-        )
-
-    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
-        """
-        Original forward pass of PixArt.
-        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
-        t: (N,) tensor of diffusion timesteps
-        y: (N, 1, 120, C) conditioning
-        ar: (N, 1): aspect ratio
-        cs: (N ,2) size conditioning for height/width
-        """
-        B, C, H, W = x.shape
-        c_res = (H + W) // 2
-        pe_interpolation = self.pe_interpolation
-        if pe_interpolation is None or self.pe_precision is not None:
-            # calculate pe_interpolation on-the-fly
-            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
-
-        pos_embed = get_2d_sincos_pos_embed_torch(
-            self.hidden_size,
-            h=(H // self.patch_size),
-            w=(W // self.patch_size),
-            pe_interpolation=pe_interpolation,
-            base_size=((round(c_res / 64) * 64) // self.patch_size),
-            device=x.device,
-            dtype=x.dtype,
-        ).unsqueeze(0)
-
-        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
-        t = self.t_embedder(timestep, x.dtype)  # (N, D)
-
-        if self.micro_conditioning and (c_size is not None and c_ar is not None):
-            bs = x.shape[0]
-            c_size = self.csize_embedder(c_size, bs)  # (N, D)
-            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
-            t = t + torch.cat([c_size, c_ar], dim=1)
-
-        t0 = self.t_block(t)
-        y = self.y_embedder(y, self.training)  # (N, D)
-
-        if mask is not None:
-            if mask.shape[0] != y.shape[0]:
-                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
-            mask = mask.squeeze(1).squeeze(1)
-            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
-            y_lens = mask.sum(dim=1).tolist()
-        else:
-            y_lens = None
-            y = y.squeeze(1).view(1, -1, x.shape[-1])
-        for block in self.blocks:
-            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
-
-        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
-        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
-
-        return x
-
-    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
-        B, C, H, W = x.shape
-
-        # Fallback for missing microconds
-        if self.micro_conditioning:
-            if c_size is None:
-                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
-
-            if c_ar is None:
-                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
-
-        ## Still accepts the input w/o that dim but returns garbage
-        if len(context.shape) == 3:
-            context = context.unsqueeze(1)
-
-        ## run original forward pass
-        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
-
-        ## only return EPS
-        if self.pred_sigma:
-            return out[:, :self.in_channels]
-        return out
-
-    def unpatchify(self, x, h, w):
-        """
-        x: (N, T, patch_size**2 * C)
-        imgs: (N, H, W, C)
-        """
-        c = self.out_channels
-        p = self.x_embedder.patch_size[0]
-        h = h // self.patch_size
-        w = w // self.patch_size
-        assert h * w == x.shape[1]
-
-        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
-        x = torch.einsum('nhwpqc->nchpwq', x)
-        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
-        return imgs
+# Based on:
+# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
+# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
+import torch
+import torch.nn as nn
+
+from .blocks import (
+    t2i_modulate,
+    CaptionEmbedder,
+    AttentionKVCompress,
+    MultiHeadCrossAttention,
+    T2IFinalLayer,
+    SizeEmbedder,
+)
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
+
+
+def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
+    grid_h, grid_w = torch.meshgrid(
+        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
+        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
+        indexing='ij'
+    )
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
+    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
+    return emb
+
+class PixArtMSBlock(nn.Module):
+    """
+    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
+    """
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
+                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.attn = AttentionKVCompress(
+            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
+            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.cross_attn = MultiHeadCrossAttention(
+            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        # to be compatible with lower version pytorch
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.mlp = Mlp(
+            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
+
+    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
+        B, N, C = x.shape
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
+        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
+        x = x + self.cross_attn(x, y, mask)
+        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
+
+        return x
+
+
+### Core PixArt Model ###
+class PixArtMS(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+    def __init__(
+            self,
+            input_size=32,
+            patch_size=2,
+            in_channels=4,
+            hidden_size=1152,
+            depth=28,
+            num_heads=16,
+            mlp_ratio=4.0,
+            class_dropout_prob=0.1,
+            learn_sigma=True,
+            pred_sigma=True,
+            drop_path: float = 0.,
+            caption_channels=4096,
+            pe_interpolation=None,
+            pe_precision=None,
+            config=None,
+            model_max_length=120,
+            micro_condition=True,
+            qk_norm=False,
+            kv_compress_config=None,
+            dtype=None,
+            device=None,
+            operations=None,
+            **kwargs,
+    ):
+        nn.Module.__init__(self)
+        self.dtype = dtype
+        self.pred_sigma = pred_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if pred_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.pe_interpolation = pe_interpolation
+        self.pe_precision = pe_precision
+        self.hidden_size = hidden_size
+        self.depth = depth
+
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.t_block = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+        self.x_embedder = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_channels,
+            embed_dim=hidden_size,
+            bias=True,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations,
+        )
+        self.y_embedder = CaptionEmbedder(
+            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
+            act_layer=approx_gelu, token_num=model_max_length,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.micro_conditioning = micro_condition
+        if self.micro_conditioning:
+            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+
+        # For fixed sin-cos embedding:
+        # num_patches = (input_size // patch_size) * (input_size // patch_size)
+        # self.base_size = input_size // self.patch_size
+        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
+
+        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
+        if kv_compress_config is None:
+            kv_compress_config = {
+                'sampling': None,
+                'scale_factor': 1,
+                'kv_compress_layer': [],
+            }
+        self.blocks = nn.ModuleList([
+            PixArtMSBlock(
+                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
+                sampling=kv_compress_config['sampling'],
+                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
+                qk_norm=qk_norm,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+            for i in range(depth)
+        ])
+        self.final_layer = T2IFinalLayer(
+            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
+        """
+        Original forward pass of PixArt.
+        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
+        t: (N,) tensor of diffusion timesteps
+        y: (N, 1, 120, C) conditioning
+        ar: (N, 1): aspect ratio
+        cs: (N ,2) size conditioning for height/width
+        """
+        B, C, H, W = x.shape
+        c_res = (H + W) // 2
+        pe_interpolation = self.pe_interpolation
+        if pe_interpolation is None or self.pe_precision is not None:
+            # calculate pe_interpolation on-the-fly
+            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
+
+        pos_embed = get_2d_sincos_pos_embed_torch(
+            self.hidden_size,
+            h=(H // self.patch_size),
+            w=(W // self.patch_size),
+            pe_interpolation=pe_interpolation,
+            base_size=((round(c_res / 64) * 64) // self.patch_size),
+            device=x.device,
+            dtype=x.dtype,
+        ).unsqueeze(0)
+
+        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
+        t = self.t_embedder(timestep, x.dtype)  # (N, D)
+
+        if self.micro_conditioning and (c_size is not None and c_ar is not None):
+            bs = x.shape[0]
+            c_size = self.csize_embedder(c_size, bs)  # (N, D)
+            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
+            t = t + torch.cat([c_size, c_ar], dim=1)
+
+        t0 = self.t_block(t)
+        y = self.y_embedder(y, self.training)  # (N, D)
+
+        if mask is not None:
+            if mask.shape[0] != y.shape[0]:
+                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
+            mask = mask.squeeze(1).squeeze(1)
+            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
+            y_lens = mask.sum(dim=1).tolist()
+        else:
+            y_lens = None
+            y = y.squeeze(1).view(1, -1, x.shape[-1])
+        for block in self.blocks:
+            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
+
+        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
+        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
+
+        return x
+
+    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
+        B, C, H, W = x.shape
+
+        # Fallback for missing microconds
+        if self.micro_conditioning:
+            if c_size is None:
+                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+            if c_ar is None:
+                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+        ## Still accepts the input w/o that dim but returns garbage
+        if len(context.shape) == 3:
+            context = context.unsqueeze(1)
+
+        ## run original forward pass
+        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
+
+        ## only return EPS
+        if self.pred_sigma:
+            return out[:, :self.in_channels]
+        return out
+
+    def unpatchify(self, x, h, w):
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        c = self.out_channels
+        p = self.x_embedder.patch_size[0]
+        h = h // self.patch_size
+        w = w // self.patch_size
+        assert h * w == x.shape[1]
+
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
+        x = torch.einsum('nhwpqc->nchpwq', x)
+        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
+        return imgs

comfy/ldm/qwen_image/model.py

@@ -0,0 +1,443 @@
+# https://github.com/QwenLM/Qwen-Image (Apache 2.0)
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple
+from einops import repeat
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+from comfy.ldm.modules.attention import optimized_attention_masked
+from comfy.ldm.flux.layers import EmbedND
+import comfy.ldm.common_dit
+
+class GELU(nn.Module):
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out, bias=bias, dtype=dtype, device=device)
+        self.approximate = approximate
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states = F.gelu(hidden_states, approximate=self.approximate)
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        inner_dim=None,
+        bias: bool = True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        self.net = nn.ModuleList([])
+        self.net.append(GELU(dim, inner_dim, approximate="tanh", bias=bias, dtype=dtype, device=device, operations=operations))
+        self.net.append(nn.Dropout(dropout))
+        self.net.append(operations.Linear(inner_dim, dim_out, bias=bias, dtype=dtype, device=device))
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+def apply_rotary_emb(x, freqs_cis):
+    if x.shape[1] == 0:
+        return x
+
+    t_ = x.reshape(*x.shape[:-1], -1, 1, 2)
+    t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+    return t_out.reshape(*x.shape)
+
+
+class QwenTimestepProjEmbeddings(nn.Module):
+    def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
+        self.timestep_embedder = TimestepEmbedding(
+            in_channels=256,
+            time_embed_dim=embedding_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+    def forward(self, timestep, hidden_states):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))
+        return timesteps_emb
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        dim_head: int = 64,
+        heads: int = 8,
+        dropout: float = 0.0,
+        bias: bool = False,
+        eps: float = 1e-5,
+        out_bias: bool = True,
+        out_dim: int = None,
+        out_context_dim: int = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.inner_kv_dim = self.inner_dim
+        self.heads = heads
+        self.dim_head = dim_head
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
+        self.dropout = dropout
+
+        # Q/K normalization
+        self.norm_q = operations.RMSNorm(dim_head, eps=eps, elementwise_affine=True, dtype=dtype, device=device)
+        self.norm_k = operations.RMSNorm(dim_head, eps=eps, elementwise_affine=True, dtype=dtype, device=device)
+        self.norm_added_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+        self.norm_added_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+
+        # Image stream projections
+        self.to_q = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device)
+        self.to_k = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+        self.to_v = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+
+        # Text stream projections
+        self.add_q_proj = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device)
+        self.add_k_proj = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+        self.add_v_proj = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+
+        # Output projections
+        self.to_out = nn.ModuleList([
+            operations.Linear(self.inner_dim, self.out_dim, bias=out_bias, dtype=dtype, device=device),
+            nn.Dropout(dropout)
+        ])
+        self.to_add_out = operations.Linear(self.inner_dim, self.out_context_dim, bias=out_bias, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,  # Image stream
+        encoder_hidden_states: torch.FloatTensor = None,  # Text stream
+        encoder_hidden_states_mask: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        seq_txt = encoder_hidden_states.shape[1]
+
+        img_query = self.to_q(hidden_states).unflatten(-1, (self.heads, -1))
+        img_key = self.to_k(hidden_states).unflatten(-1, (self.heads, -1))
+        img_value = self.to_v(hidden_states).unflatten(-1, (self.heads, -1))
+
+        txt_query = self.add_q_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_key = self.add_k_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_value = self.add_v_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+
+        img_query = self.norm_q(img_query)
+        img_key = self.norm_k(img_key)
+        txt_query = self.norm_added_q(txt_query)
+        txt_key = self.norm_added_k(txt_key)
+
+        joint_query = torch.cat([txt_query, img_query], dim=1)
+        joint_key = torch.cat([txt_key, img_key], dim=1)
+        joint_value = torch.cat([txt_value, img_value], dim=1)
+
+        joint_query = apply_rotary_emb(joint_query, image_rotary_emb)
+        joint_key = apply_rotary_emb(joint_key, image_rotary_emb)
+
+        joint_query = joint_query.flatten(start_dim=2)
+        joint_key = joint_key.flatten(start_dim=2)
+        joint_value = joint_value.flatten(start_dim=2)
+
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask)
+
+        txt_attn_output = joint_hidden_states[:, :seq_txt, :]
+        img_attn_output = joint_hidden_states[:, seq_txt:, :]
+
+        img_attn_output = self.to_out[0](img_attn_output)
+        img_attn_output = self.to_out[1](img_attn_output)
+        txt_attn_output = self.to_add_out(txt_attn_output)
+
+        return img_attn_output, txt_attn_output
+
+
+class QwenImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        eps: float = 1e-6,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+
+        self.img_mod = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device),
+        )
+        self.img_norm1 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.img_norm2 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.img_mlp = FeedForward(dim=dim, dim_out=dim, dtype=dtype, device=device, operations=operations)
+
+        self.txt_mod = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device),
+        )
+        self.txt_norm1 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.txt_norm2 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.txt_mlp = FeedForward(dim=dim, dim_out=dim, dtype=dtype, device=device, operations=operations)
+
+        self.attn = Attention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=True,
+            eps=eps,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def _modulate(self, x, mod_params):
+        shift, scale, gate = mod_params.chunk(3, dim=-1)
+        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_mask: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        img_mod_params = self.img_mod(temb)
+        txt_mod_params = self.txt_mod(temb)
+        img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
+        txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
+
+        img_normed = self.img_norm1(hidden_states)
+        img_modulated, img_gate1 = self._modulate(img_normed, img_mod1)
+        txt_normed = self.txt_norm1(encoder_hidden_states)
+        txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1)
+
+        img_attn_output, txt_attn_output = self.attn(
+            hidden_states=img_modulated,
+            encoder_hidden_states=txt_modulated,
+            encoder_hidden_states_mask=encoder_hidden_states_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+
+        hidden_states = hidden_states + img_gate1 * img_attn_output
+        encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
+
+        img_normed2 = self.img_norm2(hidden_states)
+        img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
+        hidden_states = hidden_states + img_gate2 * self.img_mlp(img_modulated2)
+
+        txt_normed2 = self.txt_norm2(encoder_hidden_states)
+        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
+        encoder_hidden_states = encoder_hidden_states + txt_gate2 * self.txt_mlp(txt_modulated2)
+
+        return encoder_hidden_states, hidden_states
+
+
+class LastLayer(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        conditioning_embedding_dim: int,
+        elementwise_affine=False,
+        eps=1e-6,
+        bias=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(embedding_dim, eps, elementwise_affine=False, bias=bias, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+        emb = self.linear(self.silu(conditioning_embedding))
+        scale, shift = torch.chunk(emb, 2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
+class QwenImageTransformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 64,
+        out_channels: Optional[int] = 16,
+        num_layers: int = 60,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 3584,
+        pooled_projection_dim: int = 768,
+        guidance_embeds: bool = False,
+        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.patch_size = patch_size
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope))
+
+        self.time_text_embed = QwenTimestepProjEmbeddings(
+            embedding_dim=self.inner_dim,
+            pooled_projection_dim=pooled_projection_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+        self.txt_norm = operations.RMSNorm(joint_attention_dim, eps=1e-6, dtype=dtype, device=device)
+        self.img_in = operations.Linear(in_channels, self.inner_dim, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(joint_attention_dim, self.inner_dim, dtype=dtype, device=device)
+
+        self.transformer_blocks = nn.ModuleList([
+            QwenImageTransformerBlock(
+                dim=self.inner_dim,
+                num_attention_heads=num_attention_heads,
+                attention_head_dim=attention_head_dim,
+                dtype=dtype,
+                device=device,
+                operations=operations
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
+        self.gradient_checkpointing = False
+
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
+        bs, c, t, h, w = x.shape
+        patch_size = self.patch_size
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+
+        h_offset = ((h_offset + (patch_size // 2)) // patch_size)
+        w_offset = ((w_offset + (patch_size // 2)) // patch_size)
+
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        ref_latents=None,
+        transformer_options={},
+        **kwargs
+    ):
+        timestep = timesteps
+        encoder_hidden_states = context
+        encoder_hidden_states_mask = attention_mask
+
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        num_embeds = hidden_states.shape[1]
+
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            for ref in ref_latents:
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
+                else:
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
+
+                kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
+                hidden_states = torch.cat([hidden_states, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size), ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size)))
+        txt_ids = torch.linspace(txt_start, txt_start + context.shape[1], steps=context.shape[1], device=x.device, dtype=x.dtype).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+
+        hidden_states = self.img_in(hidden_states)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb}, {"original_block": block_wrap})
+                hidden_states = out["img"]
+                encoder_hidden_states = out["txt"]
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                )
+
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
+        return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/wan/model.py

@@ -146,6 +146,15 @@ WAN_CROSSATTENTION_CLASSES = {
 }
 
 
+def repeat_e(e, x):
+    repeats = 1
+    if e.shape[1] > 1:
+        repeats = x.shape[1] // e.shape[1]
+    if repeats == 1:
+        return e
+    return torch.repeat_interleave(e, repeats, dim=1)
+
+
 class WanAttentionBlock(nn.Module):
 
     def __init__(self,
@@ -202,20 +211,23 @@ class WanAttentionBlock(nn.Module):
         """
         # assert e.dtype == torch.float32
 
-        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        if e.ndim < 4:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2)
         # assert e[0].dtype == torch.float32
 
         # self-attention
         y = self.self_attn(
-            self.norm1(x) * (1 + e[1]) + e[0],
+            self.norm1(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x),
             freqs)
 
-        x = x + y * e[2]
+        x = x + y * repeat_e(e[2], x)
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
-        y = self.ffn(self.norm2(x) * (1 + e[4]) + e[3])
-        x = x + y * e[5]
+        y = self.ffn(self.norm2(x) * (1 + repeat_e(e[4], x)) + repeat_e(e[3], x))
+        x = x + y * repeat_e(e[5], x)
         return x
 
 
@@ -325,8 +337,12 @@ class Head(nn.Module):
             e(Tensor): Shape [B, C]
         """
         # assert e.dtype == torch.float32
-        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
-        x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
+        if e.ndim < 3:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)
+
+        x = (self.head(self.norm(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x)))
         return x
 
 
@@ -375,6 +391,7 @@ class WanModel(torch.nn.Module):
                  cross_attn_norm=True,
                  eps=1e-6,
                  flf_pos_embed_token_number=None,
+                 in_dim_ref_conv=None,
                  image_model=None,
                  device=None,
                  dtype=None,
@@ -468,6 +485,11 @@ class WanModel(torch.nn.Module):
         else:
             self.img_emb = None
 
+        if in_dim_ref_conv is not None:
+            self.ref_conv = operations.Conv2d(in_dim_ref_conv, dim, kernel_size=patch_size[1:], stride=patch_size[1:], device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        else:
+            self.ref_conv = None
+
     def forward_orig(
         self,
         x,
@@ -506,8 +528,16 @@ class WanModel(torch.nn.Module):
 
         # time embeddings
         e = self.time_embedding(
-            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
-        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        full_ref = None
+        if self.ref_conv is not None:
+            full_ref = kwargs.get("reference_latent", None)
+            if full_ref is not None:
+                full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2)
+                x = torch.concat((full_ref, x), dim=1)
 
         # context
         context = self.text_embedding(context)
@@ -535,6 +565,9 @@ class WanModel(torch.nn.Module):
         # head
         x = self.head(x, e)
 
+        if full_ref is not None:
+            x = x[:, full_ref.shape[1]:]
+
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
@@ -553,6 +586,9 @@ class WanModel(torch.nn.Module):
             x = torch.cat([x, time_dim_concat], dim=2)
             t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0])
 
+        if self.ref_conv is not None and "reference_latent" in kwargs:
+            t_len += 1
+
         img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
         img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
         img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
@@ -732,7 +768,12 @@ class CameraWanModel(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        if model_type == 'camera':
+            model_type = 'i2v'
+        else:
+            model_type = 't2v'
+
+        super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
         self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings)
@@ -752,8 +793,7 @@ class CameraWanModel(WanModel):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         if self.control_adapter is not None and camera_conditions is not None:
-            x_camera = self.control_adapter(camera_conditions).to(x.dtype)
-            x = x + x_camera
+            x = x + self.control_adapter(camera_conditions).to(x.dtype)
         grid_sizes = x.shape[2:]
         x = x.flatten(2).transpose(1, 2)
 

comfy/ldm/wan/vae.py

@@ -24,12 +24,17 @@ class CausalConv3d(ops.Conv3d):
                          self.padding[1], 2 * self.padding[0], 0)
         self.padding = (0, 0, 0)
 
-    def forward(self, x, cache_x=None):
+    def forward(self, x, cache_x=None, cache_list=None, cache_idx=None):
+        if cache_list is not None:
+            cache_x = cache_list[cache_idx]
+            cache_list[cache_idx] = None
+
         padding = list(self._padding)
         if cache_x is not None and self._padding[4] > 0:
             cache_x = cache_x.to(x.device)
             x = torch.cat([cache_x, x], dim=2)
             padding[4] -= cache_x.shape[2]
+            del cache_x
         x = F.pad(x, padding)
 
         return super().forward(x)
@@ -52,15 +57,6 @@ class RMS_norm(nn.Module):
             x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
 
 
-class Upsample(nn.Upsample):
-
-    def forward(self, x):
-        """
-        Fix bfloat16 support for nearest neighbor interpolation.
-        """
-        return super().forward(x.float()).type_as(x)
-
-
 class Resample(nn.Module):
 
     def __init__(self, dim, mode):
@@ -73,11 +69,11 @@ class Resample(nn.Module):
         # layers
         if mode == 'upsample2d':
             self.resample = nn.Sequential(
-                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
                 ops.Conv2d(dim, dim // 2, 3, padding=1))
         elif mode == 'upsample3d':
             self.resample = nn.Sequential(
-                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
                 ops.Conv2d(dim, dim // 2, 3, padding=1))
             self.time_conv = CausalConv3d(
                 dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
@@ -157,29 +153,6 @@ class Resample(nn.Module):
                     feat_idx[0] += 1
         return x
 
-    def init_weight(self, conv):
-        conv_weight = conv.weight
-        nn.init.zeros_(conv_weight)
-        c1, c2, t, h, w = conv_weight.size()
-        one_matrix = torch.eye(c1, c2)
-        init_matrix = one_matrix
-        nn.init.zeros_(conv_weight)
-        #conv_weight.data[:,:,-1,1,1] = init_matrix * 0.5
-        conv_weight.data[:, :, 1, 0, 0] = init_matrix  #* 0.5
-        conv.weight.data.copy_(conv_weight)
-        nn.init.zeros_(conv.bias.data)
-
-    def init_weight2(self, conv):
-        conv_weight = conv.weight.data
-        nn.init.zeros_(conv_weight)
-        c1, c2, t, h, w = conv_weight.size()
-        init_matrix = torch.eye(c1 // 2, c2)
-        #init_matrix = repeat(init_matrix, 'o ... -> (o 2) ...').permute(1,0,2).contiguous().reshape(c1,c2)
-        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
-        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
-        conv.weight.data.copy_(conv_weight)
-        nn.init.zeros_(conv.bias.data)
-
 
 class ResidualBlock(nn.Module):
 
@@ -198,7 +171,7 @@ class ResidualBlock(nn.Module):
             if in_dim != out_dim else nn.Identity()
 
     def forward(self, x, feat_cache=None, feat_idx=[0]):
-        h = self.shortcut(x)
+        old_x = x
         for layer in self.residual:
             if isinstance(layer, CausalConv3d) and feat_cache is not None:
                 idx = feat_idx[0]
@@ -210,12 +183,12 @@ class ResidualBlock(nn.Module):
                             cache_x.device), cache_x
                     ],
                                         dim=2)
-                x = layer(x, feat_cache[idx])
+                x = layer(x, cache_list=feat_cache, cache_idx=idx)
                 feat_cache[idx] = cache_x
                 feat_idx[0] += 1
             else:
                 x = layer(x)
-        return x + h
+        return x + self.shortcut(old_x)
 
 
 class AttentionBlock(nn.Module):
@@ -494,12 +467,6 @@ class WanVAE(nn.Module):
         self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
-    def forward(self, x):
-        mu, log_var = self.encode(x)
-        z = self.reparameterize(mu, log_var)
-        x_recon = self.decode(z)
-        return x_recon, mu, log_var
-
     def encode(self, x):
         self.clear_cache()
         ## cache
@@ -545,18 +512,6 @@ class WanVAE(nn.Module):
         self.clear_cache()
         return out
 
-    def reparameterize(self, mu, log_var):
-        std = torch.exp(0.5 * log_var)
-        eps = torch.randn_like(std)
-        return eps * std + mu
-
-    def sample(self, imgs, deterministic=False):
-        mu, log_var = self.encode(imgs)
-        if deterministic:
-            return mu
-        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
-        return mu + std * torch.randn_like(std)
-
     def clear_cache(self):
         self._conv_num = count_conv3d(self.decoder)
         self._conv_idx = [0]

comfy/ldm/wan/vae2_2.py

@@ -0,0 +1,726 @@
+# original version: https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/vae2_2.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from .vae import AttentionBlock, CausalConv3d, RMS_norm
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+                # ops.Conv2d(dim, dim//2, 3, padding=1)
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == "upsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = "Rep"
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] != "Rep"):
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat(
+                            [
+                                feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                    cache_x.device),
+                                cache_x,
+                            ],
+                            dim=2,
+                        )
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] == "Rep"):
+                        cache_x = torch.cat(
+                            [
+                                torch.zeros_like(cache_x).to(cache_x.device),
+                                cache_x
+                            ],
+                            dim=2,
+                        )
+                    if feat_cache[idx] == "Rep":
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        x = self.resample(x)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=t)
+
+        if self.mode == "downsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1),
+        )
+        self.shortcut = (
+            CausalConv3d(in_dim, out_dim, 1)
+            if in_dim != out_dim else nn.Identity())
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        old_x = x
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, cache_list=feat_cache, cache_idx=idx)
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + self.shortcut(old_x)
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b c f (h q) (w r) -> b (c r q) f h w",
+            q=patch_size,
+            r=patch_size,
+        )
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b (c r q) f h w -> b c f (h q) (w r)",
+            q=patch_size,
+            r=patch_size,
+        )
+    return x
+
+
+class AvgDown3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1:, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_downsample=False,
+                 down_flag=False):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_upsample=False,
+                 up_flag=False):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i]
+                if i < len(temperal_downsample) else False)
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                ))
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_upsample=[False, True, True],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout),
+            AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout),
+        )
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(
+                temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks + 1,
+                    temperal_upsample=t_up_flag,
+                    up_flag=i != len(dim_mult) - 1,
+                ))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, 12, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(
+        self,
+        dim=160,
+        dec_dim=256,
+        z_dim=16,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(
+            dim,
+            z_dim * 2,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_downsample,
+            dropout,
+        )
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(
+            dec_dim,
+            z_dim,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_upsample,
+            dropout,
+        )
+
+    def encode(self, x):
+        self.clear_cache()
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                    first_chunk=True,
+                )
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        self.clear_cache()
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/lora.py

@@ -293,6 +293,16 @@ def model_lora_keys_unet(model, key_map={}):
                 key_lora = k[len("diffusion_model."):-len(".weight")]
                 key_map["{}".format(key_lora)] = k
 
+    if isinstance(model, comfy.model_base.QwenImage):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"): #QwenImage lora format
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                # Direct mapping for transformer_blocks format (QwenImage LoRA format)
+                key_map["{}".format(key_lora)] = k
+                # Support transformer prefix format
+                key_map["transformer.{}".format(key_lora)] = k
+                key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+
     return key_map
 
 

comfy/model_base.py

@@ -42,6 +42,7 @@ import comfy.ldm.hidream.model
 import comfy.ldm.chroma.model
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
+import comfy.ldm.qwen_image.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -106,10 +107,12 @@ def model_sampling(model_config, model_type):
     return ModelSampling(model_config)
 
 
-def convert_tensor(extra, dtype):
+def convert_tensor(extra, dtype, device):
     if hasattr(extra, "dtype"):
         if extra.dtype != torch.int and extra.dtype != torch.long:
-            extra = extra.to(dtype)
+            extra = comfy.model_management.cast_to_device(extra, device, dtype)
+        else:
+            extra = comfy.model_management.cast_to_device(extra, device, None)
     return extra
 
 
@@ -160,7 +163,7 @@ class BaseModel(torch.nn.Module):
         xc = self.model_sampling.calculate_input(sigma, x)
 
         if c_concat is not None:
-            xc = torch.cat([xc] + [c_concat], dim=1)
+            xc = torch.cat([xc] + [comfy.model_management.cast_to_device(c_concat, xc.device, xc.dtype)], dim=1)
 
         context = c_crossattn
         dtype = self.get_dtype()
@@ -169,20 +172,21 @@ class BaseModel(torch.nn.Module):
             dtype = self.manual_cast_dtype
 
         xc = xc.to(dtype)
+        device = xc.device
         t = self.model_sampling.timestep(t).float()
         if context is not None:
-            context = context.to(dtype)
+            context = comfy.model_management.cast_to_device(context, device, dtype)
 
         extra_conds = {}
         for o in kwargs:
             extra = kwargs[o]
 
             if hasattr(extra, "dtype"):
-                extra = convert_tensor(extra, dtype)
+                extra = convert_tensor(extra, dtype, device)
             elif isinstance(extra, list):
                 ex = []
                 for ext in extra:
-                    ex.append(convert_tensor(ext, dtype))
+                    ex.append(convert_tensor(ext, dtype, device))
                 extra = ex
             extra_conds[o] = extra
 
@@ -398,7 +402,7 @@ class SD21UNCLIP(BaseModel):
         unclip_conditioning = kwargs.get("unclip_conditioning", None)
         device = kwargs["device"]
         if unclip_conditioning is None:
-            return torch.zeros((1, self.adm_channels))
+            return torch.zeros((1, self.adm_channels), device=device)
         else:
             return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05), kwargs.get("seed", 0) - 10)
 
@@ -612,9 +616,11 @@ class IP2P:
 
         if image is None:
             image = torch.zeros_like(noise)
+        else:
+            image = image.to(device=device)
 
         if image.shape[1:] != noise.shape[1:]:
-            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            image = utils.common_upscale(image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
 
         image = utils.resize_to_batch_size(image, noise.shape[0])
         return self.process_ip2p_image_in(image)
@@ -693,7 +699,7 @@ class StableCascade_B(BaseModel):
         #size of prior doesn't really matter if zeros because it gets resized but I still want it to get batched
         prior = kwargs.get("stable_cascade_prior", torch.zeros((1, 16, (noise.shape[2] * 4) // 42, (noise.shape[3] * 4) // 42), dtype=noise.dtype, layout=noise.layout, device=noise.device))
 
-        out["effnet"] = comfy.conds.CONDRegular(prior)
+        out["effnet"] = comfy.conds.CONDRegular(prior.to(device=noise.device))
         out["sca"] = comfy.conds.CONDRegular(torch.zeros((1,)))
         return out
 
@@ -816,7 +822,7 @@ class PixArt(BaseModel):
 class Flux(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux):
         super().__init__(model_config, model_type, device=device, unet_model=unet_model)
-        self.memory_usage_factor_conds = ("kontext",)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def concat_cond(self, **kwargs):
         try:
@@ -884,6 +890,10 @@ class Flux(BaseModel):
             for lat in ref_latents:
                 latents.append(self.process_latent_in(lat))
             out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
         return out
 
     def extra_conds_shapes(self, **kwargs):
@@ -1097,8 +1107,9 @@ class WAN21(BaseModel):
                 image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
             image = utils.resize_to_batch_size(image, noise.shape[0])
 
-        if not self.image_to_video or extra_channels == image.shape[1]:
-            return image
+        if extra_channels != image.shape[1] + 4:
+            if not self.image_to_video or extra_channels == image.shape[1]:
+                return image
 
         if image.shape[1] > (extra_channels - 4):
             image = image[:, :(extra_channels - 4)]
@@ -1117,7 +1128,11 @@ class WAN21(BaseModel):
                 mask = mask.repeat(1, 4, 1, 1, 1)
             mask = utils.resize_to_batch_size(mask, noise.shape[0])
 
-        return torch.cat((mask, image), dim=1)
+        concat_mask_index = kwargs.get("concat_mask_index", 0)
+        if concat_mask_index != 0:
+            return torch.cat((image[:, :concat_mask_index], mask, image[:, concat_mask_index:]), dim=1)
+        else:
+            return torch.cat((mask, image), dim=1)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1133,6 +1148,10 @@ class WAN21(BaseModel):
         if time_dim_concat is not None:
             out['time_dim_concat'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_concat))
 
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1])[:, :, 0])
+
         return out
 
 
@@ -1157,10 +1176,10 @@ class WAN21_Vace(WAN21):
 
         vace_frames_out = []
         for j in range(len(vace_frames)):
-            vf = vace_frames[j].clone()
+            vf = vace_frames[j].to(device=noise.device, dtype=noise.dtype, copy=True)
             for i in range(0, vf.shape[1], 16):
                 vf[:, i:i + 16] = self.process_latent_in(vf[:, i:i + 16])
-            vf = torch.cat([vf, mask[j]], dim=1)
+            vf = torch.cat([vf, mask[j].to(device=noise.device, dtype=noise.dtype)], dim=1)
             vace_frames_out.append(vf)
 
         vace_frames = torch.stack(vace_frames_out, dim=1)
@@ -1182,6 +1201,31 @@ class WAN21_Camera(WAN21):
             out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
         return out
 
+class WAN22(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        temp_ts = (torch.mean(denoise_mask[:, :, :, :, :], dim=(1, 3, 4), keepdim=True) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1))).reshape(timestep.shape[0], -1)
+        return temp_ts
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
 class Hunyuan3Dv2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)
@@ -1277,3 +1321,24 @@ class Omnigen2(BaseModel):
         if ref_latents is not None:
             out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
         return out
+
+class QwenImage(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
+        return out

comfy/model_detection.py

@@ -346,7 +346,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config = {}
         dit_config["image_model"] = "wan2.1"
         dim = state_dict['{}head.modulation'.format(key_prefix)].shape[-1]
+        out_dim = state_dict['{}head.head.weight'.format(key_prefix)].shape[0] // 4
         dit_config["dim"] = dim
+        dit_config["out_dim"] = out_dim
         dit_config["num_heads"] = dim // 128
         dit_config["ffn_dim"] = state_dict['{}blocks.0.ffn.0.weight'.format(key_prefix)].shape[0]
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
@@ -362,7 +364,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1]
             dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.')
         elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys:
-            dit_config["model_type"] = "camera"
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "camera"
+            else:
+                dit_config["model_type"] = "camera_2.2"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"
@@ -371,6 +376,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         flf_weight = state_dict.get('{}img_emb.emb_pos'.format(key_prefix))
         if flf_weight is not None:
             dit_config["flf_pos_embed_token_number"] = flf_weight.shape[1]
+
+        ref_conv_weight = state_dict.get('{}ref_conv.weight'.format(key_prefix))
+        if ref_conv_weight is not None:
+            dit_config["in_dim_ref_conv"] = ref_conv_weight.shape[1]
+
         return dit_config
 
     if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
@@ -479,6 +489,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["timestep_scale"] = 1000.0
         return dit_config
 
+    if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image
+        dit_config = {}
+        dit_config["image_model"] = "qwen_image"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -865,7 +880,7 @@ def convert_diffusers_mmdit(state_dict, output_prefix=""):
         depth_single_blocks = count_blocks(state_dict, 'single_transformer_blocks.{}.')
         hidden_size = state_dict["x_embedder.bias"].shape[0]
         sd_map = comfy.utils.flux_to_diffusers({"depth": depth, "depth_single_blocks": depth_single_blocks, "hidden_size": hidden_size}, output_prefix=output_prefix)
-    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict: #SD3
+    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict and 'pos_embed.proj.weight' in state_dict: #SD3
         num_blocks = count_blocks(state_dict, 'transformer_blocks.{}.')
         depth = state_dict["pos_embed.proj.weight"].shape[0] // 64
         sd_map = comfy.utils.mmdit_to_diffusers({"depth": depth, "num_blocks": num_blocks}, output_prefix=output_prefix)

comfy/model_management.py

@@ -78,7 +78,6 @@ try:
     torch_version = torch.version.__version__
     temp = torch_version.split(".")
     torch_version_numeric = (int(temp[0]), int(temp[1]))
-    xpu_available = (torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] <= 4)) and torch.xpu.is_available()
 except:
     pass
 
@@ -101,11 +100,15 @@ if args.directml is not None:
     lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
 
 try:
-    import intel_extension_for_pytorch as ipex
-    _ = torch.xpu.device_count()
-    xpu_available = xpu_available or torch.xpu.is_available()
+    import intel_extension_for_pytorch as ipex  # noqa: F401
 except:
-    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
+    pass
+
+try:
+    _ = torch.xpu.device_count()
+    xpu_available = torch.xpu.is_available()
+except:
+    xpu_available = False
 
 try:
     if torch.backends.mps.is_available():
@@ -128,6 +131,11 @@ try:
 except:
     mlu_available = False
 
+try:
+    ixuca_available = hasattr(torch, "corex")
+except:
+    ixuca_available = False
+
 if args.cpu:
     cpu_state = CPUState.CPU
 
@@ -151,6 +159,12 @@ def is_mlu():
         return True
     return False
 
+def is_ixuca():
+    global ixuca_available
+    if ixuca_available:
+        return True
+    return False
+
 def get_torch_device():
     global directml_enabled
     global cpu_state
@@ -186,8 +200,9 @@ def get_total_memory(dev=None, torch_total_too=False):
         elif is_intel_xpu():
             stats = torch.xpu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
+            mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory
             mem_total_torch = mem_reserved
-            mem_total = torch.xpu.get_device_properties(dev).total_memory
+            mem_total = mem_total_xpu
         elif is_ascend_npu():
             stats = torch.npu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
@@ -288,7 +303,7 @@ try:
         if torch_version_numeric[0] >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
-    if is_intel_xpu() or is_ascend_npu() or is_mlu():
+    if is_intel_xpu() or is_ascend_npu() or is_mlu() or is_ixuca():
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             ENABLE_PYTORCH_ATTENTION = True
 except:
@@ -307,8 +322,11 @@ try:
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
-                if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx1201 and gfx950
+                if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                     ENABLE_PYTORCH_ATTENTION = True
+#            if torch_version_numeric >= (2, 8):
+#                if any((a in arch) for a in ["gfx1201"]):
+#                    ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
                 SUPPORT_FP8_OPS = True
@@ -325,7 +343,7 @@ if ENABLE_PYTORCH_ATTENTION:
 
 PRIORITIZE_FP16 = False  # TODO: remove and replace with something that shows exactly which dtype is faster than the other
 try:
-    if is_nvidia() and PerformanceFeature.Fp16Accumulation in args.fast:
+    if (is_nvidia() or is_amd()) and PerformanceFeature.Fp16Accumulation in args.fast:
         torch.backends.cuda.matmul.allow_fp16_accumulation = True
         PRIORITIZE_FP16 = True  # TODO: limit to cards where it actually boosts performance
         logging.info("Enabled fp16 accumulation.")
@@ -377,6 +395,8 @@ def get_torch_device_name(device):
             except:
                 allocator_backend = ""
             return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend)
+        elif device.type == "xpu":
+            return "{} {}".format(device, torch.xpu.get_device_name(device))
         else:
             return "{}".format(device.type)
     elif is_intel_xpu():
@@ -512,6 +532,8 @@ WINDOWS = any(platform.win32_ver())
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
 if WINDOWS:
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
+    if total_vram > (15 * 1024):  # more extra reserved vram on 16GB+ cards
+        EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -560,16 +582,23 @@ def free_memory(memory_required, device, keep_loaded=[]):
                 soft_empty_cache()
     return unloaded_models
 
+def get_models_memory_reserve(models):
+    total_reserve = 0
+    for model in models:
+        total_reserve += model.get_model_memory_reserve(convert_to_bytes=True)
+    return total_reserve
+
 def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
     inference_memory = minimum_inference_memory()
-    extra_mem = max(inference_memory, memory_required + extra_reserved_memory())
+    models_memory_reserve = get_models_memory_reserve(models)
+    extra_mem = max(inference_memory + models_memory_reserve, memory_required + extra_reserved_memory() + models_memory_reserve)
     if minimum_memory_required is None:
         minimum_memory_required = extra_mem
     else:
-        minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
+        minimum_memory_required = max(inference_memory + models_memory_reserve, minimum_memory_required + extra_reserved_memory() + models_memory_reserve)
 
     models = set(models)
 
@@ -876,6 +905,7 @@ def vae_dtype(device=None, allowed_dtypes=[]):
             return d
 
         # NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32
+        # slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3
         if d == torch.bfloat16 and (not is_amd()) and should_use_bf16(device):
             return d
 
@@ -926,9 +956,11 @@ def pick_weight_dtype(dtype, fallback_dtype, device=None):
     return dtype
 
 def device_supports_non_blocking(device):
+    if args.force_non_blocking:
+        return True
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
-    if is_intel_xpu():
+    if is_intel_xpu(): #xpu does support non blocking but it is slower on iGPUs for some reason so disable by default until situation changes
         return False
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
@@ -968,6 +1000,8 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         if is_device_cuda(device):
             ss[stream_counter].wait_stream(torch.cuda.current_stream())
+        elif is_device_xpu(device):
+            ss[stream_counter].wait_stream(torch.xpu.current_stream())
         stream_counters[device] = stream_counter
         return s
     elif is_device_cuda(device):
@@ -979,6 +1013,15 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
+    elif is_device_xpu(device):
+        ss = []
+        for k in range(NUM_STREAMS):
+            ss.append(torch.xpu.Stream(device=device, priority=0))
+        STREAMS[device] = ss
+        s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
+        stream_counters[device] = stream_counter
+        return s
     return None
 
 def sync_stream(device, stream):
@@ -986,6 +1029,8 @@ def sync_stream(device, stream):
         return
     if is_device_cuda(device):
         torch.cuda.current_stream().wait_stream(stream)
+    elif is_device_xpu(device):
+        torch.xpu.current_stream().wait_stream(stream)
 
 def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
     if device is None or weight.device == device:
@@ -1027,6 +1072,8 @@ def xformers_enabled():
         return False
     if is_mlu():
         return False
+    if is_ixuca():
+        return False
     if directml_enabled:
         return False
     return XFORMERS_IS_AVAILABLE
@@ -1062,6 +1109,8 @@ def pytorch_attention_flash_attention():
             return True
         if is_amd():
             return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
+        if is_ixuca():
+            return True
     return False
 
 def force_upcast_attention_dtype():
@@ -1092,8 +1141,8 @@ def get_free_memory(dev=None, torch_free_too=False):
             stats = torch.xpu.memory_stats(dev)
             mem_active = stats['active_bytes.all.current']
             mem_reserved = stats['reserved_bytes.all.current']
-            mem_free_torch = mem_reserved - mem_active
             mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved
+            mem_free_torch = mem_reserved - mem_active
             mem_free_total = mem_free_xpu + mem_free_torch
         elif is_ascend_npu():
             stats = torch.npu.memory_stats(dev)
@@ -1142,6 +1191,9 @@ def is_device_cpu(device):
 def is_device_mps(device):
     return is_device_type(device, 'mps')
 
+def is_device_xpu(device):
+    return is_device_type(device, 'xpu')
+
 def is_device_cuda(device):
     return is_device_type(device, 'cuda')
 
@@ -1173,7 +1225,10 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        return True
+        if torch_version_numeric < (2, 3):
+            return True
+        else:
+            return torch.xpu.get_device_properties(device).has_fp16
 
     if is_ascend_npu():
         return True
@@ -1181,6 +1236,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if is_mlu():
         return True
 
+    if is_ixuca():
+        return True
+
     if torch.version.hip:
         return True
 
@@ -1236,11 +1294,17 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        return True
+        if torch_version_numeric < (2, 3):
+            return True
+        else:
+            return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True
 
+    if is_ixuca():
+        return True
+
     if is_amd():
         arch = torch.cuda.get_device_properties(device).gcnArchName
         if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16

comfy/model_patcher.py

@@ -24,7 +24,7 @@ import inspect
 import logging
 import math
 import uuid
-from typing import Callable, Optional
+from typing import Callable, Optional, Union
 
 import torch
 
@@ -84,6 +84,12 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
         model_options["disable_cfg1_optimization"] = True
     return model_options
 
+def add_model_options_memory_reserve(model_options, memory_reserve_gb: float):
+    if "model_memory_reserve" not in model_options:
+        model_options["model_memory_reserve"] = []
+    model_options["model_memory_reserve"].append(memory_reserve_gb)
+    return model_options
+
 def create_model_options_clone(orig_model_options: dict):
     return comfy.patcher_extension.copy_nested_dicts(orig_model_options)
 
@@ -379,6 +385,9 @@ class ModelPatcher:
     def set_model_sampler_pre_cfg_function(self, pre_cfg_function, disable_cfg1_optimization=False):
         self.model_options = set_model_options_pre_cfg_function(self.model_options, pre_cfg_function, disable_cfg1_optimization)
 
+    def set_model_sampler_calc_cond_batch_function(self, sampler_calc_cond_batch_function):
+        self.model_options["sampler_calc_cond_batch_function"] = sampler_calc_cond_batch_function
+
     def set_model_unet_function_wrapper(self, unet_wrapper_function: UnetWrapperFunction):
         self.model_options["model_function_wrapper"] = unet_wrapper_function
 
@@ -436,6 +445,17 @@ class ModelPatcher:
             self.force_cast_weights = True
         self.patches_uuid = uuid.uuid4() #TODO: optimize by preventing a full model reload for this
 
+    def add_model_memory_reserve(self, memory_reserve_gb: float):
+        """Adds additional expected memory usage for the model, in gigabytes."""
+        self.model_options = add_model_options_memory_reserve(self.model_options, memory_reserve_gb)
+
+    def get_model_memory_reserve(self, convert_to_bytes: bool = False) -> Union[float, int]:
+        """Returns the total expected memory usage for the model in gigabytes, or bytes if convert_to_bytes is True."""
+        total_reserve = sum(self.model_options.get("model_memory_reserve", []))
+        if convert_to_bytes:
+            return total_reserve * 1024 * 1024 * 1024
+        return total_reserve
+
     def add_weight_wrapper(self, name, function):
         self.weight_wrapper_patches[name] = self.weight_wrapper_patches.get(name, []) + [function]
         self.patches_uuid = uuid.uuid4()

comfy/ops.py

@@ -24,6 +24,32 @@ import comfy.float
 import comfy.rmsnorm
 import contextlib
 
+
+def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+
+
+try:
+    if torch.cuda.is_available():
+        from torch.nn.attention import SDPBackend, sdpa_kernel
+        import inspect
+        if "set_priority" in inspect.signature(sdpa_kernel).parameters:
+            SDPA_BACKEND_PRIORITY = [
+                SDPBackend.FLASH_ATTENTION,
+                SDPBackend.EFFICIENT_ATTENTION,
+                SDPBackend.MATH,
+            ]
+
+            SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
+
+            def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+        else:
+            logging.warning("Torch version too old to set sdpa backend priority.")
+except (ModuleNotFoundError, TypeError):
+    logging.warning("Could not set sdpa backend priority.")
+
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
 def cast_to_input(weight, input, non_blocking=False, copy=True):
@@ -336,9 +362,12 @@ class fp8_ops(manual_cast):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            out = fp8_linear(self, input)
-            if out is not None:
-                return out
+            try:
+                out = fp8_linear(self, input)
+                if out is not None:
+                    return out
+            except Exception as e:
+                logging.info("Exception during fp8 op: {}".format(e))
 
             weight, bias = cast_bias_weight(self, input)
             return torch.nn.functional.linear(input, weight, bias)

comfy/rmsnorm.py

@@ -1,6 +1,7 @@
 import torch
 import comfy.model_management
 import numbers
+import logging
 
 RMSNorm = None
 
@@ -9,6 +10,7 @@ try:
     RMSNorm = torch.nn.RMSNorm
 except:
     rms_norm_torch = None
+    logging.warning("Please update pytorch to use native RMSNorm")
 
 
 def rms_norm(x, weight=None, eps=1e-6):

comfy/sampler_helpers.py

@@ -149,7 +149,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
+def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -158,8 +158,8 @@ def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
     for k in conds:
         get_hooks_from_cond(conds[k], hooks)
     # add wrappers and callbacks from ModelPatcher to transformer_options
-    model_options["transformer_options"]["wrappers"] = comfy.patcher_extension.copy_nested_dicts(model.wrappers)
-    model_options["transformer_options"]["callbacks"] = comfy.patcher_extension.copy_nested_dicts(model.callbacks)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("wrappers", {}), model.wrappers, copy_dict1=False)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("callbacks", {}), model.callbacks, copy_dict1=False)
     # begin registering hooks
     registered = comfy.hooks.HookGroup()
     target_dict = comfy.hooks.create_target_dict(comfy.hooks.EnumWeightTarget.Model)

comfy/samplers.py

@@ -16,6 +16,7 @@ import comfy.sampler_helpers
 import comfy.model_patcher
 import comfy.patcher_extension
 import comfy.hooks
+import comfy.context_windows
 import scipy.stats
 import numpy
 
@@ -89,7 +90,7 @@ def get_area_and_mult(conds, x_in, timestep_in):
     conditioning = {}
     model_conds = conds["model_conds"]
     for c in model_conds:
-        conditioning[c] = model_conds[c].process_cond(batch_size=x_in.shape[0], device=x_in.device, area=area)
+        conditioning[c] = model_conds[c].process_cond(batch_size=x_in.shape[0], area=area)
 
     hooks = conds.get('hooks', None)
     control = conds.get('control', None)
@@ -198,14 +199,20 @@ def finalize_default_conds(model: 'BaseModel', hooked_to_run: dict[comfy.hooks.H
             hooked_to_run.setdefault(p.hooks, list())
             hooked_to_run[p.hooks] += [(p, i)]
 
-def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options: dict[str]):
+    handler: comfy.context_windows.ContextHandlerABC = model_options.get("context_handler", None)
+    if handler is None or not handler.should_use_context(model, conds, x_in, timestep, model_options):
+        return _calc_cond_batch_outer(model, conds, x_in, timestep, model_options)
+    return handler.execute(_calc_cond_batch_outer, model, conds, x_in, timestep, model_options)
+
+def _calc_cond_batch_outer(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _calc_cond_batch,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, model_options, is_model_options=True)
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     out_conds = []
     out_counts = []
     # separate conds by matching hooks
@@ -373,7 +380,11 @@ def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_option
         uncond_ = uncond
 
     conds = [cond, uncond_]
-    out = calc_cond_batch(model, conds, x, timestep, model_options)
+    if "sampler_calc_cond_batch_function" in model_options:
+        args = {"conds": conds, "input": x, "sigma": timestep, "model": model, "model_options": model_options}
+        out = model_options["sampler_calc_cond_batch_function"](args)
+    else:
+        out = calc_cond_batch(model, conds, x, timestep, model_options)
 
     for fn in model_options.get("sampler_pre_cfg_function", []):
         args = {"conds":conds, "conds_out": out, "cond_scale": cond_scale, "timestep": timestep,
@@ -716,7 +727,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
-                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"]
+                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3", "sa_solver", "sa_solver_pece"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):
@@ -1039,13 +1050,13 @@ class SchedulerHandler(NamedTuple):
     use_ms: bool = True
 
 SCHEDULER_HANDLERS = {
-    "normal": SchedulerHandler(normal_scheduler),
+    "simple": SchedulerHandler(simple_scheduler),
+    "sgm_uniform": SchedulerHandler(partial(normal_scheduler, sgm=True)),
     "karras": SchedulerHandler(k_diffusion_sampling.get_sigmas_karras, use_ms=False),
     "exponential": SchedulerHandler(k_diffusion_sampling.get_sigmas_exponential, use_ms=False),
-    "sgm_uniform": SchedulerHandler(partial(normal_scheduler, sgm=True)),
-    "simple": SchedulerHandler(simple_scheduler),
     "ddim_uniform": SchedulerHandler(ddim_scheduler),
     "beta": SchedulerHandler(beta_scheduler),
+    "normal": SchedulerHandler(normal_scheduler),
     "linear_quadratic": SchedulerHandler(linear_quadratic_schedule),
     "kl_optimal": SchedulerHandler(kl_optimal_scheduler, use_ms=False),
 }

comfy/sd.py

@@ -14,10 +14,12 @@ import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
 import comfy.ldm.wan.vae
+import comfy.ldm.wan.vae2_2
 import comfy.ldm.hunyuan3d.vae
 import comfy.ldm.ace.vae.music_dcae_pipeline
 import yaml
 import math
+import os
 
 import comfy.utils
 
@@ -45,6 +47,7 @@ import comfy.text_encoders.wan
 import comfy.text_encoders.hidream
 import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
+import comfy.text_encoders.qwen_image
 
 import comfy.model_patcher
 import comfy.lora
@@ -419,17 +422,30 @@ class VAE:
                 self.memory_used_encode = lambda shape, dtype: (50 * (round((shape[2] + 7) / 8) * 8) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.working_dtypes = [torch.bfloat16, torch.float32]
             elif "decoder.middle.0.residual.0.gamma" in sd:
-                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
-                self.upscale_index_formula = (4, 8, 8)
-                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
-                self.downscale_index_formula = (4, 8, 8)
-                self.latent_dim = 3
-                self.latent_channels = 16
-                ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
-                self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
-                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
-                self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
-                self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
+                if "decoder.upsamples.0.upsamples.0.residual.2.weight" in sd:  # Wan 2.2 VAE
+                    self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16)
+                    self.upscale_index_formula = (4, 16, 16)
+                    self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16)
+                    self.downscale_index_formula = (4, 16, 16)
+                    self.latent_dim = 3
+                    self.latent_channels = 48
+                    ddconfig = {"dim": 160, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                    self.first_stage_model = comfy.ldm.wan.vae2_2.WanVAE(**ddconfig)
+                    self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                    self.memory_used_encode = lambda shape, dtype: 3300 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                    self.memory_used_decode = lambda shape, dtype: 8000 * shape[3] * shape[4] * (16 * 16) * model_management.dtype_size(dtype)
+                else:  # Wan 2.1 VAE
+                    self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                    self.upscale_index_formula = (4, 8, 8)
+                    self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                    self.downscale_index_formula = (4, 8, 8)
+                    self.latent_dim = 3
+                    self.latent_channels = 16
+                    ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                    self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
+                    self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                    self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                    self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
             elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd:
                 self.latent_dim = 1
                 ln_post = "geo_decoder.ln_post.weight" in sd
@@ -756,6 +772,7 @@ class CLIPType(Enum):
     CHROMA = 15
     ACE = 16
     OMNIGEN2 = 17
+    QWEN_IMAGE = 18
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -776,6 +793,7 @@ class TEModel(Enum):
     T5_XXL_OLD = 8
     GEMMA_2_2B = 9
     QWEN25_3B = 10
+    QWEN25_7B = 11
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -797,7 +815,11 @@ def detect_te_model(sd):
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
         return TEModel.GEMMA_2_2B
     if 'model.layers.0.self_attn.k_proj.bias' in sd:
-        return TEModel.QWEN25_3B
+        weight = sd['model.layers.0.self_attn.k_proj.bias']
+        if weight.shape[0] == 256:
+            return TEModel.QWEN25_3B
+        if weight.shape[0] == 512:
+            return TEModel.QWEN25_7B
     if "model.layers.0.post_attention_layernorm.weight" in sd:
         return TEModel.LLAMA3_8
     return None
@@ -902,6 +924,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif te_model == TEModel.QWEN25_3B:
             clip_target.clip = comfy.text_encoders.omnigen2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.omnigen2.Omnigen2Tokenizer
+        elif te_model == TEModel.QWEN25_7B:
+            clip_target.clip = comfy.text_encoders.qwen_image.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.qwen_image.QwenImageTokenizer
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
@@ -977,6 +1002,12 @@ def load_gligen(ckpt_path):
         model = model.half()
     return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
 
+def model_detection_error_hint(path, state_dict):
+    filename = os.path.basename(path)
+    if 'lora' in filename.lower():
+        return "\nHINT: This seems to be a Lora file and Lora files should be put in the lora folder and loaded with a lora loader node.."
+    return ""
+
 def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
     logging.warning("Warning: The load checkpoint with config function is deprecated and will eventually be removed, please use the other one.")
     model, clip, vae, _ = load_checkpoint_guess_config(ckpt_path, output_vae=output_vae, output_clip=output_clip, output_clipvision=False, embedding_directory=embedding_directory, output_model=True)
@@ -1005,7 +1036,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True)
     out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata)
     if out is None:
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+        raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd)))
     return out
 
 def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None):
@@ -1177,7 +1208,7 @@ def load_diffusion_model(unet_path, model_options={}):
     model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
         logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path))
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
+        raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(unet_path, model_detection_error_hint(unet_path, sd)))
     return model
 
 def load_unet(unet_path, dtype=None):

comfy/sd1_tokenizer/tokenizer_config.json

@@ -18,7 +18,7 @@
     "single_word": false
   },
   "errors": "replace",
-  "model_max_length": 77,
+  "model_max_length": 8192,
   "name_or_path": "openai/clip-vit-large-patch14",
   "pad_token": "<|endoftext|>",
   "special_tokens_map_file": "./special_tokens_map.json",

comfy/supported_models.py

@@ -19,6 +19,7 @@ import comfy.text_encoders.lumina2
 import comfy.text_encoders.wan
 import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
+import comfy.text_encoders.qwen_image
 
 from . import supported_models_base
 from . import latent_formats
@@ -1045,6 +1046,18 @@ class WAN21_Camera(WAN21_T2V):
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
         return out
+
+class WAN22_Camera(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "camera_2.2",
+        "in_dim": 36,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
+        return out
+
 class WAN21_Vace(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1059,6 +1072,19 @@ class WAN21_Vace(WAN21_T2V):
         out = model_base.WAN21_Vace(self, image_to_video=False, device=device)
         return out
 
+class WAN22_T2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "t2v",
+        "out_dim": 48,
+    }
+
+    latent_format = latent_formats.Wan22
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN22(self, image_to_video=True, device=device)
+        return out
+
 class Hunyuan3Dv2(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan3d2",
@@ -1214,9 +1240,38 @@ class Omnigen2(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_3b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.LuminaTokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect))
+        return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.Omnigen2Tokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect))
+
+class QwenImage(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "qwen_image",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 1.15,
+    }
+
+    memory_usage_factor = 1.8 #TODO
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.QwenImage(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.qwen_image.QwenImageTokenizer, comfy.text_encoders.qwen_image.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/text_encoders/llama.py

@@ -43,6 +43,23 @@ class Qwen25_3BConfig:
     mlp_activation = "silu"
     qkv_bias = True
 
+@dataclass
+class Qwen25_7BVLI_Config:
+    vocab_size: int = 152064
+    hidden_size: int = 3584
+    intermediate_size: int = 18944
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 28
+    num_key_value_heads: int = 4
+    max_position_embeddings: int = 128000
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = True
+
 @dataclass
 class Gemma2_2B_Config:
     vocab_size: int = 256000
@@ -348,6 +365,15 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen25_7BVLI_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Gemma2_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/pixart_t5.py

@@ -1,42 +1,42 @@
-import os
-
-from comfy import sd1_clip
-import comfy.text_encoders.t5
-import comfy.text_encoders.sd3_clip
-from comfy.sd1_clip import gen_empty_tokens
-
-from transformers import T5TokenizerFast
-
-class T5XXLModel(comfy.text_encoders.sd3_clip.T5XXLModel):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-
-    def gen_empty_tokens(self, special_tokens, *args, **kwargs):
-        # PixArt expects the negative to be all pad tokens
-        special_tokens = special_tokens.copy()
-        special_tokens.pop("end")
-        return gen_empty_tokens(special_tokens, *args, **kwargs)
-
-class PixArtT5XXL(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options)
-
-class T5XXLTokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) # no padding
-
-class PixArtTokenizer(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
-
-def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
-    class PixArtTEModel_(PixArtT5XXL):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
-                model_options = model_options.copy()
-                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
-            if dtype is None:
-                dtype = dtype_t5
-            super().__init__(device=device, dtype=dtype, model_options=model_options)
-    return PixArtTEModel_
+import os
+
+from comfy import sd1_clip
+import comfy.text_encoders.t5
+import comfy.text_encoders.sd3_clip
+from comfy.sd1_clip import gen_empty_tokens
+
+from transformers import T5TokenizerFast
+
+class T5XXLModel(comfy.text_encoders.sd3_clip.T5XXLModel):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def gen_empty_tokens(self, special_tokens, *args, **kwargs):
+        # PixArt expects the negative to be all pad tokens
+        special_tokens = special_tokens.copy()
+        special_tokens.pop("end")
+        return gen_empty_tokens(special_tokens, *args, **kwargs)
+
+class PixArtT5XXL(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options)
+
+class T5XXLTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) # no padding
+
+class PixArtTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
+
+def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
+    class PixArtTEModel_(PixArtT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype is None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return PixArtTEModel_

comfy/text_encoders/qwen_image.py

@@ -0,0 +1,71 @@
+from transformers import Qwen2Tokenizer
+from comfy import sd1_clip
+import comfy.text_encoders.llama
+import os
+import torch
+import numbers
+
+class Qwen25_7BVLITokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=3584, embedding_key='qwen25_7b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+
+
+class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
+        self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None,**kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+        return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+
+
+class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class QwenImageTEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        out, pooled, extra = super().encode_token_weights(token_weight_pairs)
+        tok_pairs = token_weight_pairs["qwen25_7b"][0]
+        count_im_start = 0
+        for i, v in enumerate(tok_pairs):
+            elem = v[0]
+            if not torch.is_tensor(elem):
+                if isinstance(elem, numbers.Integral):
+                    if elem == 151644 and count_im_start < 2:
+                        template_end = i
+                        count_im_start += 1
+
+        if out.shape[1] > (template_end + 3):
+            if tok_pairs[template_end + 1][0] == 872:
+                if tok_pairs[template_end + 2][0] == 198:
+                    template_end += 3
+
+        out = out[:, template_end:]
+
+        extra["attention_mask"] = extra["attention_mask"][:, template_end:]
+        if extra["attention_mask"].sum() == torch.numel(extra["attention_mask"]):
+            extra.pop("attention_mask")  # attention mask is useless if no masked elements
+
+        return out, pooled, extra
+
+
+def te(dtype_llama=None, llama_scaled_fp8=None):
+    class QwenImageTEModel_(QwenImageTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return QwenImageTEModel_

comfy/text_encoders/t5.py

@@ -146,7 +146,7 @@ class T5Attention(torch.nn.Module):
         )
         values = self.relative_attention_bias(relative_position_bucket, out_dtype=dtype)  # shape (query_length, key_length, num_heads)
         values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
-        return values
+        return values.contiguous()
 
     def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
         q = self.q(x)

comfy/utils.py

@@ -31,6 +31,7 @@ from einops import rearrange
 from comfy.cli_args import args
 
 MMAP_TORCH_FILES = args.mmap_torch_files
+DISABLE_MMAP = args.disable_mmap
 
 ALWAYS_SAFE_LOAD = False
 if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in pytorch 2.4, the unsafe path should be removed once earlier versions are deprecated
@@ -58,7 +59,10 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
             with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
                 sd = {}
                 for k in f.keys():
-                    sd[k] = f.get_tensor(k)
+                    tensor = f.get_tensor(k)
+                    if DISABLE_MMAP:  # TODO: Not sure if this is the best way to bypass the mmap issues
+                        tensor = tensor.to(device=device, copy=True)
+                    sd[k] = tensor
                 if return_metadata:
                     metadata = f.metadata()
         except Exception as e:
@@ -77,6 +81,7 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
         if safe_load or ALWAYS_SAFE_LOAD:
             pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args)
         else:
+            logging.warning("WARNING: loading {} unsafely, upgrade your pytorch to 2.4 or newer to load this file safely.".format(ckpt))
             pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
         if "state_dict" in pl_sd:
             sd = pl_sd["state_dict"]
@@ -693,6 +698,26 @@ def resize_to_batch_size(tensor, batch_size):
 
     return output
 
+def resize_list_to_batch_size(l, batch_size):
+    in_batch_size = len(l)
+    if in_batch_size == batch_size or in_batch_size == 0:
+        return l
+
+    if batch_size <= 1:
+        return l[:batch_size]
+
+    output = []
+    if batch_size < in_batch_size:
+        scale = (in_batch_size - 1) / (batch_size - 1)
+        for i in range(batch_size):
+            output.append(l[min(round(i * scale), in_batch_size - 1)])
+    else:
+        scale = in_batch_size / batch_size
+        for i in range(batch_size):
+           output.append(l[min(math.floor((i + 0.5) * scale), in_batch_size - 1)])
+
+    return output
+
 def convert_sd_to(state_dict, dtype):
     keys = list(state_dict.keys())
     for k in keys:
@@ -997,11 +1022,12 @@ def set_progress_bar_global_hook(function):
     PROGRESS_BAR_HOOK = function
 
 class ProgressBar:
-    def __init__(self, total):
+    def __init__(self, total, node_id=None):
         global PROGRESS_BAR_HOOK
         self.total = total
         self.current = 0
         self.hook = PROGRESS_BAR_HOOK
+        self.node_id = node_id
 
     def update_absolute(self, value, total=None, preview=None):
         if total is not None:
@@ -1010,7 +1036,7 @@ class ProgressBar:
             value = self.total
         self.current = value
         if self.hook is not None:
-            self.hook(self.current, self.total, preview)
+            self.hook(self.current, self.total, preview, node_id=self.node_id)
 
     def update(self, value):
         self.update_absolute(self.current + value)

comfy/weight_adapter/__init__.py

@@ -15,9 +15,20 @@ adapters: list[type[WeightAdapterBase]] = [
     OFTAdapter,
     BOFTAdapter,
 ]
+adapter_maps: dict[str, type[WeightAdapterBase]] = {
+    "LoRA": LoRAAdapter,
+    "LoHa": LoHaAdapter,
+    "LoKr": LoKrAdapter,
+    "OFT": OFTAdapter,
+    ## We disable not implemented algo for now
+    # "GLoRA": GLoRAAdapter,
+    # "BOFT": BOFTAdapter,
+}
+
 
 __all__ = [
     "WeightAdapterBase",
     "WeightAdapterTrainBase",
-    "adapters"
+    "adapters",
+    "adapter_maps",
 ] + [a.__name__ for a in adapters]

comfy/weight_adapter/base.py

@@ -133,3 +133,43 @@ def tucker_weight_from_conv(up, down, mid):
 def tucker_weight(wa, wb, t):
     temp = torch.einsum("i j ..., j r -> i r ...", t, wb)
     return torch.einsum("i j ..., i r -> r j ...", temp, wa)
+
+
+def factorization(dimension: int, factor: int = -1) -> tuple[int, int]:
+    """
+    return a tuple of two value of input dimension decomposed by the number closest to factor
+    second value is higher or equal than first value.
+
+    examples)
+    factor
+        -1               2                4               8               16               ...
+    127 -> 1, 127   127 -> 1, 127    127 -> 1, 127   127 -> 1, 127   127 -> 1, 127
+    128 -> 8, 16    128 -> 2, 64     128 -> 4, 32    128 -> 8, 16    128 -> 8, 16
+    250 -> 10, 25   250 -> 2, 125    250 -> 2, 125   250 -> 5, 50    250 -> 10, 25
+    360 -> 8, 45    360 -> 2, 180    360 -> 4, 90    360 -> 8, 45    360 -> 12, 30
+    512 -> 16, 32   512 -> 2, 256    512 -> 4, 128   512 -> 8, 64    512 -> 16, 32
+    1024 -> 32, 32  1024 -> 2, 512   1024 -> 4, 256  1024 -> 8, 128  1024 -> 16, 64
+    """
+
+    if factor > 0 and (dimension % factor) == 0 and dimension >= factor**2:
+        m = factor
+        n = dimension // factor
+        if m > n:
+            n, m = m, n
+        return m, n
+    if factor < 0:
+        factor = dimension
+    m, n = 1, dimension
+    length = m + n
+    while m < n:
+        new_m = m + 1
+        while dimension % new_m != 0:
+            new_m += 1
+        new_n = dimension // new_m
+        if new_m + new_n > length or new_m > factor:
+            break
+        else:
+            m, n = new_m, new_n
+    if m > n:
+        n, m = m, n
+    return m, n

comfy/weight_adapter/loha.py

@@ -3,7 +3,120 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose
+
+
+class HadaWeight(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, w1u, w1d, w2u, w2d, scale=torch.tensor(1)):
+        ctx.save_for_backward(w1d, w1u, w2d, w2u, scale)
+        diff_weight = ((w1u @ w1d) * (w2u @ w2d)) * scale
+        return diff_weight
+
+    @staticmethod
+    def backward(ctx, grad_out):
+        (w1d, w1u, w2d, w2u, scale) = ctx.saved_tensors
+        grad_out = grad_out * scale
+        temp = grad_out * (w2u @ w2d)
+        grad_w1u = temp @ w1d.T
+        grad_w1d = w1u.T @ temp
+
+        temp = grad_out * (w1u @ w1d)
+        grad_w2u = temp @ w2d.T
+        grad_w2d = w2u.T @ temp
+
+        del temp
+        return grad_w1u, grad_w1d, grad_w2u, grad_w2d, None
+
+
+class HadaWeightTucker(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, t1, w1u, w1d, t2, w2u, w2d, scale=torch.tensor(1)):
+        ctx.save_for_backward(t1, w1d, w1u, t2, w2d, w2u, scale)
+
+        rebuild1 = torch.einsum("i j ..., j r, i p -> p r ...", t1, w1d, w1u)
+        rebuild2 = torch.einsum("i j ..., j r, i p -> p r ...", t2, w2d, w2u)
+
+        return rebuild1 * rebuild2 * scale
+
+    @staticmethod
+    def backward(ctx, grad_out):
+        (t1, w1d, w1u, t2, w2d, w2u, scale) = ctx.saved_tensors
+        grad_out = grad_out * scale
+
+        temp = torch.einsum("i j ..., j r -> i r ...", t2, w2d)
+        rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w2u)
+
+        grad_w = rebuild * grad_out
+        del rebuild
+
+        grad_w1u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w)
+        grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w1u.T)
+        del grad_w, temp
+
+        grad_w1d = torch.einsum("i r ..., i j ... -> r j", t1, grad_temp)
+        grad_t1 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w1d.T)
+        del grad_temp
+
+        temp = torch.einsum("i j ..., j r -> i r ...", t1, w1d)
+        rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w1u)
+
+        grad_w = rebuild * grad_out
+        del rebuild
+
+        grad_w2u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w)
+        grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w2u.T)
+        del grad_w, temp
+
+        grad_w2d = torch.einsum("i r ..., i j ... -> r j", t2, grad_temp)
+        grad_t2 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w2d.T)
+        del grad_temp
+        return grad_t1, grad_w1u, grad_w1d, grad_t2, grad_w2u, grad_w2d, None
+
+
+class LohaDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        w1a, w1b, alpha, w2a, w2b, t1, t2, _ = weights
+
+        # Create trainable parameters
+        self.hada_w1_a = torch.nn.Parameter(w1a)
+        self.hada_w1_b = torch.nn.Parameter(w1b)
+        self.hada_w2_a = torch.nn.Parameter(w2a)
+        self.hada_w2_b = torch.nn.Parameter(w2b)
+
+        self.use_tucker = False
+        if t1 is not None and t2 is not None:
+            self.use_tucker = True
+            self.hada_t1 = torch.nn.Parameter(t1)
+            self.hada_t2 = torch.nn.Parameter(t2)
+        else:
+            # Keep the attributes for consistent access
+            self.hada_t1 = None
+            self.hada_t2 = None
+
+        # Store rank and non-trainable alpha
+        self.rank = w1b.shape[0]
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+
+        scale = self.alpha / self.rank
+        if self.use_tucker:
+            diff_weight = HadaWeightTucker.apply(self.hada_t1, self.hada_w1_a, self.hada_w1_b, self.hada_t2, self.hada_w2_a, self.hada_w2_b, scale)
+        else:
+            diff_weight = HadaWeight.apply(self.hada_w1_a, self.hada_w1_b, self.hada_w2_a, self.hada_w2_b, scale)
+
+        # Add the scaled difference to the original weight
+        weight = w.to(diff_weight) + diff_weight.reshape(w.shape)
+
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class LoHaAdapter(WeightAdapterBase):
@@ -13,6 +126,25 @@ class LoHaAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        in_dim = weight.shape[1:].numel()
+        mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.normal_(mat1, 0.1)
+        torch.nn.init.constant_(mat2, 0.0)
+        mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.normal_(mat3, 0.1)
+        torch.nn.init.normal_(mat4, 0.01)
+        return LohaDiff(
+            (mat1, mat2, alpha, mat3, mat4, None, None, None)
+        )
+
+    def to_train(self):
+        return LohaDiff(self.weights)
+
     @classmethod
     def load(
         cls,

comfy/weight_adapter/lokr.py

@@ -3,7 +3,77 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import (
+    WeightAdapterBase,
+    WeightAdapterTrainBase,
+    weight_decompose,
+    factorization,
+)
+
+
+class LokrDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) = weights
+        self.use_tucker = False
+        if lokr_w1_a is not None:
+            _, rank_a = lokr_w1_a.shape[0], lokr_w1_a.shape[1]
+            rank_a, _ = lokr_w1_b.shape[0], lokr_w1_b.shape[1]
+            self.lokr_w1_a = torch.nn.Parameter(lokr_w1_a)
+            self.lokr_w1_b = torch.nn.Parameter(lokr_w1_b)
+            self.w1_rebuild = True
+            self.ranka = rank_a
+
+        if lokr_w2_a is not None:
+            _, rank_b = lokr_w2_a.shape[0], lokr_w2_a.shape[1]
+            rank_b, _ = lokr_w2_b.shape[0], lokr_w2_b.shape[1]
+            self.lokr_w2_a = torch.nn.Parameter(lokr_w2_a)
+            self.lokr_w2_b = torch.nn.Parameter(lokr_w2_b)
+            if lokr_t2 is not None:
+                self.use_tucker = True
+                self.lokr_t2 = torch.nn.Parameter(lokr_t2)
+            self.w2_rebuild = True
+            self.rankb = rank_b
+
+        if lokr_w1 is not None:
+            self.lokr_w1 = torch.nn.Parameter(lokr_w1)
+            self.w1_rebuild = False
+
+        if lokr_w2 is not None:
+            self.lokr_w2 = torch.nn.Parameter(lokr_w2)
+            self.w2_rebuild = False
+
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    @property
+    def w1(self):
+        if self.w1_rebuild:
+            return (self.lokr_w1_a @ self.lokr_w1_b) * (self.alpha / self.ranka)
+        else:
+            return self.lokr_w1
+
+    @property
+    def w2(self):
+        if self.w2_rebuild:
+            if self.use_tucker:
+                w2 = torch.einsum(
+                    'i j k l, j r, i p -> p r k l',
+                    self.lokr_t2,
+                    self.lokr_w2_b,
+                    self.lokr_w2_a
+                )
+            else:
+                w2 = self.lokr_w2_a @ self.lokr_w2_b
+            return w2 * (self.alpha / self.rankb)
+        else:
+            return self.lokr_w2
+
+    def __call__(self, w):
+        diff = torch.kron(self.w1, self.w2)
+        return w + diff.reshape(w.shape).to(w)
+
+    def passive_memory_usage(self):
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class LoKrAdapter(WeightAdapterBase):
@@ -13,6 +83,20 @@ class LoKrAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        in_dim = weight.shape[1:].numel()
+        out1, out2 = factorization(out_dim, rank)
+        in1, in2 = factorization(in_dim, rank)
+        mat1 = torch.empty(out1, in1, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(out2, in2, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.kaiming_uniform_(mat2, a=5**0.5)
+        torch.nn.init.constant_(mat1, 0.0)
+        return LokrDiff(
+            (mat1, mat2, alpha, None, None, None, None, None, None)
+        )
+
     @classmethod
     def load(
         cls,

comfy/weight_adapter/lora.py

@@ -96,6 +96,7 @@ class LoRAAdapter(WeightAdapterBase):
         diffusers3_lora = "{}.lora.up.weight".format(x)
         mochi_lora = "{}.lora_B".format(x)
         transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
+        qwen_default_lora = "{}.lora_B.default.weight".format(x)
         A_name = None
 
         if regular_lora in lora.keys():
@@ -122,6 +123,10 @@ class LoRAAdapter(WeightAdapterBase):
             A_name = transformers_lora
             B_name = "{}.lora_linear_layer.down.weight".format(x)
             mid_name = None
+        elif qwen_default_lora in lora.keys():
+            A_name = qwen_default_lora
+            B_name = "{}.lora_A.default.weight".format(x)
+            mid_name = None
 
         if A_name is not None:
             mid = None

comfy/weight_adapter/oft.py

@@ -3,7 +3,58 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization
+
+
+class OFTDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        blocks, rescale, alpha, _ = weights
+
+        # Create trainable parameters
+        self.oft_blocks = torch.nn.Parameter(blocks)
+        if rescale is not None:
+            self.rescale = torch.nn.Parameter(rescale)
+            self.rescaled = True
+        else:
+            self.rescaled = False
+        self.block_num, self.block_size, _ = blocks.shape
+        self.constraint = float(alpha)
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+        I = torch.eye(self.block_size, device=self.oft_blocks.device)
+
+        ## generate r
+        # for Q = -Q^T
+        q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
+        normed_q = q
+        if self.constraint:
+            q_norm = torch.norm(q) + 1e-8
+            if q_norm > self.constraint:
+                normed_q = q * self.constraint / q_norm
+        # use float() to prevent unsupported type
+        r = (I + normed_q) @ (I - normed_q).float().inverse()
+
+        ## Apply chunked matmul on weight
+        _, *shape = w.shape
+        org_weight = w.to(dtype=r.dtype)
+        org_weight = org_weight.unflatten(0, (self.block_num, self.block_size))
+        # Init R=0, so add I on it to ensure the output of step0 is original model output
+        weight = torch.einsum(
+            "k n m, k n ... -> k m ...",
+            r,
+            org_weight,
+        ).flatten(0, 1)
+        if self.rescaled:
+            weight = self.rescale * weight
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class OFTAdapter(WeightAdapterBase):
@@ -13,6 +64,18 @@ class OFTAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        block_size, block_num = factorization(out_dim, rank)
+        block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=weight.dtype)
+        return OFTDiff(
+            (block, None, alpha, None)
+        )
+
+    def to_train(self):
+        return OFTDiff(self.weights)
+
     @classmethod
     def load(
         cls,
@@ -60,6 +123,8 @@ class OFTAdapter(WeightAdapterBase):
         blocks = v[0]
         rescale = v[1]
         alpha = v[2]
+        if alpha is None:
+            alpha = 0
         dora_scale = v[3]
 
         blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)

comfy_api/feature_flags.py

@@ -0,0 +1,69 @@
+"""
+Feature flags module for ComfyUI WebSocket protocol negotiation.
+
+This module handles capability negotiation between frontend and backend,
+allowing graceful protocol evolution while maintaining backward compatibility.
+"""
+
+from typing import Any, Dict
+
+from comfy.cli_args import args
+
+# Default server capabilities
+SERVER_FEATURE_FLAGS: Dict[str, Any] = {
+    "supports_preview_metadata": True,
+    "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
+}
+
+
+def get_connection_feature(
+    sockets_metadata: Dict[str, Dict[str, Any]],
+    sid: str,
+    feature_name: str,
+    default: Any = False
+) -> Any:
+    """
+    Get a feature flag value for a specific connection.
+
+    Args:
+        sockets_metadata: Dictionary of socket metadata
+        sid: Session ID of the connection
+        feature_name: Name of the feature to check
+        default: Default value if feature not found
+
+    Returns:
+        Feature value or default if not found
+    """
+    if sid not in sockets_metadata:
+        return default
+
+    return sockets_metadata[sid].get("feature_flags", {}).get(feature_name, default)
+
+
+def supports_feature(
+    sockets_metadata: Dict[str, Dict[str, Any]],
+    sid: str,
+    feature_name: str
+) -> bool:
+    """
+    Check if a connection supports a specific feature.
+
+    Args:
+        sockets_metadata: Dictionary of socket metadata
+        sid: Session ID of the connection
+        feature_name: Name of the feature to check
+
+    Returns:
+        Boolean indicating if feature is supported
+    """
+    return get_connection_feature(sockets_metadata, sid, feature_name, False) is True
+
+
+def get_server_features() -> Dict[str, Any]:
+    """
+    Get the server's feature flags.
+
+    Returns:
+        Dictionary of server feature flags
+    """
+    return SERVER_FEATURE_FLAGS.copy()

comfy_api/generate_api_stubs.py

@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+"""
+Script to generate .pyi stub files for the synchronous API wrappers.
+This allows generating stubs without running the full ComfyUI application.
+"""
+
+import os
+import sys
+import logging
+import importlib
+
+# Add ComfyUI to path so we can import modules
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from comfy_api.internal.async_to_sync import AsyncToSyncConverter
+from comfy_api.version_list import supported_versions
+
+
+def generate_stubs_for_module(module_name: str) -> None:
+    """Generate stub files for a specific module that exports ComfyAPI and ComfyAPISync."""
+    try:
+        # Import the module
+        module = importlib.import_module(module_name)
+
+        # Check if module has ComfyAPISync (the sync wrapper)
+        if hasattr(module, "ComfyAPISync"):
+            # Module already has a sync class
+            api_class = getattr(module, "ComfyAPI", None)
+            sync_class = getattr(module, "ComfyAPISync")
+
+            if api_class:
+                # Generate the stub file
+                AsyncToSyncConverter.generate_stub_file(api_class, sync_class)
+                logging.info(f"Generated stub file for {module_name}")
+            else:
+                logging.warning(
+                    f"Module {module_name} has ComfyAPISync but no ComfyAPI"
+                )
+
+        elif hasattr(module, "ComfyAPI"):
+            # Module only has async API, need to create sync wrapper first
+            from comfy_api.internal.async_to_sync import create_sync_class
+
+            api_class = getattr(module, "ComfyAPI")
+            sync_class = create_sync_class(api_class)
+
+            # Generate the stub file
+            AsyncToSyncConverter.generate_stub_file(api_class, sync_class)
+            logging.info(f"Generated stub file for {module_name}")
+        else:
+            logging.warning(
+                f"Module {module_name} does not export ComfyAPI or ComfyAPISync"
+            )
+
+    except Exception as e:
+        logging.error(f"Failed to generate stub for {module_name}: {e}")
+        import traceback
+
+        traceback.print_exc()
+
+
+def main():
+    """Main function to generate all API stub files."""
+    logging.basicConfig(level=logging.INFO)
+
+    logging.info("Starting stub generation...")
+
+    # Dynamically get module names from supported_versions
+    api_modules = []
+    for api_class in supported_versions:
+        # Extract module name from the class
+        module_name = api_class.__module__
+        if module_name not in api_modules:
+            api_modules.append(module_name)
+
+    logging.info(f"Found {len(api_modules)} API modules: {api_modules}")
+
+    # Generate stubs for each module
+    for module_name in api_modules:
+        generate_stubs_for_module(module_name)
+
+    logging.info("Stub generation complete!")
+
+
+if __name__ == "__main__":
+    main()

comfy_api/input/__init__.py

@@ -1,8 +1,16 @@
-from .basic_types import ImageInput, AudioInput
-from .video_types import VideoInput
+# This file only exists for backwards compatibility.
+from comfy_api.latest._input import (
+    ImageInput,
+    AudioInput,
+    MaskInput,
+    LatentInput,
+    VideoInput,
+)
 
 __all__ = [
     "ImageInput",
     "AudioInput",
+    "MaskInput",
+    "LatentInput",
     "VideoInput",
 ]

comfy_api/input/basic_types.py

@@ -1,20 +1,14 @@
-import torch
-from typing import TypedDict
-
-ImageInput = torch.Tensor
-"""
-An image in format [B, H, W, C] where B is the batch size, C is the number of channels,
-"""
-
-class AudioInput(TypedDict):
-    """
-    TypedDict representing audio input.
-    """
-
-    waveform: torch.Tensor
-    """
-    Tensor in the format [B, C, T] where B is the batch size, C is the number of channels,
-    """
-
-    sample_rate: int
+# This file only exists for backwards compatibility.
+from comfy_api.latest._input.basic_types import (
+    ImageInput,
+    AudioInput,
+    MaskInput,
+    LatentInput,
+)
 
+__all__ = [
+    "ImageInput",
+    "AudioInput",
+    "MaskInput",
+    "LatentInput",
+]

comfy_api/input/video_types.py

@@ -1,55 +1,6 @@
-from __future__ import annotations
-from abc import ABC, abstractmethod
-from typing import Optional
-from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
+# This file only exists for backwards compatibility.
+from comfy_api.latest._input.video_types import VideoInput
 
-class VideoInput(ABC):
-    """
-    Abstract base class for video input types.
-    """
-
-    @abstractmethod
-    def get_components(self) -> VideoComponents:
-        """
-        Abstract method to get the video components (images, audio, and frame rate).
-
-        Returns:
-            VideoComponents containing images, audio, and frame rate
-        """
-        pass
-
-    @abstractmethod
-    def save_to(
-        self,
-        path: str,
-        format: VideoContainer = VideoContainer.AUTO,
-        codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
-    ):
-        """
-        Abstract method to save the video input to a file.
-        """
-        pass
-
-    # Provide a default implementation, but subclasses can provide optimized versions
-    # if possible.
-    def get_dimensions(self) -> tuple[int, int]:
-        """
-        Returns the dimensions of the video input.
-
-        Returns:
-            Tuple of (width, height)
-        """
-        components = self.get_components()
-        return components.images.shape[2], components.images.shape[1]
-
-    def get_duration(self) -> float:
-        """
-        Returns the duration of the video in seconds.
-
-        Returns:
-            Duration in seconds
-        """
-        components = self.get_components()
-        frame_count = components.images.shape[0]
-        return float(frame_count / components.frame_rate)
+__all__ = [
+    "VideoInput",
+]

comfy_api/input_impl/__init__.py

@@ -1,7 +1,7 @@
-from .video_types import VideoFromFile, VideoFromComponents
+# This file only exists for backwards compatibility.
+from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents
 
 __all__ = [
-    # Implementations
     "VideoFromFile",
     "VideoFromComponents",
 ]

comfy_api/input_impl/video_types.py

@@ -1,303 +1,2 @@
-from __future__ import annotations
-from av.container import InputContainer
-from av.subtitles.stream import SubtitleStream
-from fractions import Fraction
-from typing import Optional
-from comfy_api.input import AudioInput
-import av
-import io
-import json
-import numpy as np
-import torch
-from comfy_api.input import VideoInput
-from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
-
-
-def container_to_output_format(container_format: str | None) -> str | None:
-    """
-    A container's `format` may be a comma-separated list of formats.
-    E.g., iso container's `format` may be `mov,mp4,m4a,3gp,3g2,mj2`.
-    However, writing to a file/stream with `av.open` requires a single format,
-    or `None` to auto-detect.
-    """
-    if not container_format:
-        return None  # Auto-detect
-
-    if "," not in container_format:
-        return container_format
-
-    formats = container_format.split(",")
-    return formats[0]
-
-
-def get_open_write_kwargs(
-    dest: str | io.BytesIO, container_format: str, to_format: str | None
-) -> dict:
-    """Get kwargs for writing a `VideoFromFile` to a file/stream with `av.open`"""
-    open_kwargs = {
-        "mode": "w",
-        # If isobmff, preserve custom metadata tags (workflow, prompt, extra_pnginfo)
-        "options": {"movflags": "use_metadata_tags"},
-    }
-
-    is_write_to_buffer = isinstance(dest, io.BytesIO)
-    if is_write_to_buffer:
-        # Set output format explicitly, since it cannot be inferred from file extension
-        if to_format == VideoContainer.AUTO:
-            to_format = container_format.lower()
-        elif isinstance(to_format, str):
-            to_format = to_format.lower()
-        open_kwargs["format"] = container_to_output_format(to_format)
-
-    return open_kwargs
-
-
-class VideoFromFile(VideoInput):
-    """
-    Class representing video input from a file.
-    """
-
-    def __init__(self, file: str | io.BytesIO):
-        """
-        Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object
-        containing the file contents.
-        """
-        self.__file = file
-
-    def get_dimensions(self) -> tuple[int, int]:
-        """
-        Returns the dimensions of the video input.
-
-        Returns:
-            Tuple of (width, height)
-        """
-        if isinstance(self.__file, io.BytesIO):
-            self.__file.seek(0)  # Reset the BytesIO object to the beginning
-        with av.open(self.__file, mode='r') as container:
-            for stream in container.streams:
-                if stream.type == 'video':
-                    assert isinstance(stream, av.VideoStream)
-                    return stream.width, stream.height
-        raise ValueError(f"No video stream found in file '{self.__file}'")
-
-    def get_duration(self) -> float:
-        """
-        Returns the duration of the video in seconds.
-
-        Returns:
-            Duration in seconds
-        """
-        if isinstance(self.__file, io.BytesIO):
-            self.__file.seek(0)
-        with av.open(self.__file, mode="r") as container:
-            if container.duration is not None:
-                return float(container.duration / av.time_base)
-
-            # Fallback: calculate from frame count and frame rate
-            video_stream = next(
-                (s for s in container.streams if s.type == "video"), None
-            )
-            if video_stream and video_stream.frames and video_stream.average_rate:
-                return float(video_stream.frames / video_stream.average_rate)
-
-            # Last resort: decode frames to count them
-            if video_stream and video_stream.average_rate:
-                frame_count = 0
-                container.seek(0)
-                for packet in container.demux(video_stream):
-                    for _ in packet.decode():
-                        frame_count += 1
-                if frame_count > 0:
-                    return float(frame_count / video_stream.average_rate)
-
-        raise ValueError(f"Could not determine duration for file '{self.__file}'")
-
-    def get_components_internal(self, container: InputContainer) -> VideoComponents:
-        # Get video frames
-        frames = []
-        for frame in container.decode(video=0):
-            img = frame.to_ndarray(format='rgb24')  # shape: (H, W, 3)
-            img = torch.from_numpy(img) / 255.0  # shape: (H, W, 3)
-            frames.append(img)
-
-        images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0)
-
-        # Get frame rate
-        video_stream = next(s for s in container.streams if s.type == 'video')
-        frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1)
-
-        # Get audio if available
-        audio = None
-        try:
-            container.seek(0)  # Reset the container to the beginning
-            for stream in container.streams:
-                if stream.type != 'audio':
-                    continue
-                assert isinstance(stream, av.AudioStream)
-                audio_frames = []
-                for packet in container.demux(stream):
-                    for frame in packet.decode():
-                        assert isinstance(frame, av.AudioFrame)
-                        audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
-                if len(audio_frames) > 0:
-                    audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
-                    audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
-                    audio = AudioInput({
-                        "waveform": audio_tensor,
-                        "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1,
-                    })
-        except StopIteration:
-            pass  # No audio stream
-
-        metadata = container.metadata
-        return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
-
-    def get_components(self) -> VideoComponents:
-        if isinstance(self.__file, io.BytesIO):
-            self.__file.seek(0)  # Reset the BytesIO object to the beginning
-        with av.open(self.__file, mode='r') as container:
-            return self.get_components_internal(container)
-        raise ValueError(f"No video stream found in file '{self.__file}'")
-
-    def save_to(
-        self,
-        path: str | io.BytesIO,
-        format: VideoContainer = VideoContainer.AUTO,
-        codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
-    ):
-        if isinstance(self.__file, io.BytesIO):
-            self.__file.seek(0)  # Reset the BytesIO object to the beginning
-        with av.open(self.__file, mode='r') as container:
-            container_format = container.format.name
-            video_encoding = container.streams.video[0].codec.name if len(container.streams.video) > 0 else None
-            reuse_streams = True
-            if format != VideoContainer.AUTO and format not in container_format.split(","):
-                reuse_streams = False
-            if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None:
-                reuse_streams = False
-
-            if not reuse_streams:
-                components = self.get_components_internal(container)
-                video = VideoFromComponents(components)
-                return video.save_to(
-                    path,
-                    format=format,
-                    codec=codec,
-                    metadata=metadata
-                )
-
-            streams = container.streams
-
-            open_kwargs = get_open_write_kwargs(path, container_format, format)
-            with av.open(path, **open_kwargs) as output_container:
-                # Copy over the original metadata
-                for key, value in container.metadata.items():
-                    if metadata is None or key not in metadata:
-                        output_container.metadata[key] = value
-
-                # Add our new metadata
-                if metadata is not None:
-                    for key, value in metadata.items():
-                        if isinstance(value, str):
-                            output_container.metadata[key] = value
-                        else:
-                            output_container.metadata[key] = json.dumps(value)
-
-                # Add streams to the new container
-                stream_map = {}
-                for stream in streams:
-                    if isinstance(stream, (av.VideoStream, av.AudioStream, SubtitleStream)):
-                        out_stream = output_container.add_stream_from_template(template=stream, opaque=True)
-                        stream_map[stream] = out_stream
-
-                # Write packets to the new container
-                for packet in container.demux():
-                    if packet.stream in stream_map and packet.dts is not None:
-                        packet.stream = stream_map[packet.stream]
-                        output_container.mux(packet)
-
-class VideoFromComponents(VideoInput):
-    """
-    Class representing video input from tensors.
-    """
-
-    def __init__(self, components: VideoComponents):
-        self.__components = components
-
-    def get_components(self) -> VideoComponents:
-        return VideoComponents(
-            images=self.__components.images,
-            audio=self.__components.audio,
-            frame_rate=self.__components.frame_rate
-        )
-
-    def save_to(
-        self,
-        path: str,
-        format: VideoContainer = VideoContainer.AUTO,
-        codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
-    ):
-        if format != VideoContainer.AUTO and format != VideoContainer.MP4:
-            raise ValueError("Only MP4 format is supported for now")
-        if codec != VideoCodec.AUTO and codec != VideoCodec.H264:
-            raise ValueError("Only H264 codec is supported for now")
-        with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}) as output:
-            # Add metadata before writing any streams
-            if metadata is not None:
-                for key, value in metadata.items():
-                    output.metadata[key] = json.dumps(value)
-
-            frame_rate = Fraction(round(self.__components.frame_rate * 1000), 1000)
-            # Create a video stream
-            video_stream = output.add_stream('h264', rate=frame_rate)
-            video_stream.width = self.__components.images.shape[2]
-            video_stream.height = self.__components.images.shape[1]
-            video_stream.pix_fmt = 'yuv420p'
-
-            # Create an audio stream
-            audio_sample_rate = 1
-            audio_stream: Optional[av.AudioStream] = None
-            if self.__components.audio:
-                audio_sample_rate = int(self.__components.audio['sample_rate'])
-                audio_stream = output.add_stream('aac', rate=audio_sample_rate)
-                audio_stream.sample_rate = audio_sample_rate
-                audio_stream.format = 'fltp'
-
-            # Encode video
-            for i, frame in enumerate(self.__components.images):
-                img = (frame * 255).clamp(0, 255).byte().cpu().numpy() # shape: (H, W, 3)
-                frame = av.VideoFrame.from_ndarray(img, format='rgb24')
-                frame = frame.reformat(format='yuv420p')  # Convert to YUV420P as required by h264
-                packet = video_stream.encode(frame)
-                output.mux(packet)
-
-            # Flush video
-            packet = video_stream.encode(None)
-            output.mux(packet)
-
-            if audio_stream and self.__components.audio:
-                # Encode audio
-                samples_per_frame = int(audio_sample_rate / frame_rate)
-                num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame
-                for i in range(num_frames):
-                    start = i * samples_per_frame
-                    end = start + samples_per_frame
-                    # TODO(Feature) - Add support for stereo audio
-                    chunk = (
-                        self.__components.audio["waveform"][0, 0, start:end]
-                        .unsqueeze(0)
-                        .contiguous()
-                        .numpy()
-                    )
-                    audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono')
-                    audio_frame.sample_rate = audio_sample_rate
-                    audio_frame.pts = i * samples_per_frame
-                    for packet in audio_stream.encode(audio_frame):
-                        output.mux(packet)
-
-                # Flush audio
-                for packet in audio_stream.encode(None):
-                    output.mux(packet)
-
+# This file only exists for backwards compatibility.
+from comfy_api.latest._input_impl.video_types import *  # noqa: F403

comfy_api/internal/__init__.py

@@ -0,0 +1,150 @@
+# Internal infrastructure for ComfyAPI
+from .api_registry import (
+    ComfyAPIBase as ComfyAPIBase,
+    ComfyAPIWithVersion as ComfyAPIWithVersion,
+    register_versions as register_versions,
+    get_all_versions as get_all_versions,
+)
+
+import asyncio
+from dataclasses import asdict
+from typing import Callable, Optional
+
+
+def first_real_override(cls: type, name: str, *, base: type=None) -> Optional[Callable]:
+    """Return the *callable* override of `name` visible on `cls`, or None if every
+    implementation up to (and including) `base` is the placeholder defined on `base`.
+
+    If base is not provided, it will assume cls has a GET_BASE_CLASS
+    """
+    if base is None:
+        if not hasattr(cls, "GET_BASE_CLASS"):
+            raise ValueError("base is required if cls does not have a GET_BASE_CLASS; is this a valid ComfyNode subclass?")
+        base = cls.GET_BASE_CLASS()
+    base_attr = getattr(base, name, None)
+    if base_attr is None:
+        return None
+    base_func = base_attr.__func__
+    for c in cls.mro():                       # NodeB, NodeA, ComfyNode, object …
+        if c is base:                         # reached the placeholder – we're done
+            break
+        if name in c.__dict__:                # first class that *defines* the attr
+            func = getattr(c, name).__func__
+            if func is not base_func:         # real override
+                return getattr(cls, name)     # bound to *cls*
+    return None
+
+
+class _ComfyNodeInternal:
+    """Class that all V3-based APIs inherit from for ComfyNode.
+
+    This is intended to only be referenced within execution.py, as it has to handle all V3 APIs going forward."""
+    @classmethod
+    def GET_NODE_INFO_V1(cls):
+        ...
+
+
+class _NodeOutputInternal:
+    """Class that all V3-based APIs inherit from for NodeOutput.
+
+    This is intended to only be referenced within execution.py, as it has to handle all V3 APIs going forward."""
+    ...
+
+
+def as_pruned_dict(dataclass_obj):
+    '''Return dict of dataclass object with pruned None values.'''
+    return prune_dict(asdict(dataclass_obj))
+
+def prune_dict(d: dict):
+    return {k: v for k,v in d.items() if v is not None}
+
+
+def is_class(obj):
+    '''
+    Returns True if is a class type.
+    Returns False if is a class instance.
+    '''
+    return isinstance(obj, type)
+
+
+def copy_class(cls: type) -> type:
+    '''
+    Copy a class and its attributes.
+    '''
+    if cls is None:
+        return None
+    cls_dict = {
+            k: v for k, v in cls.__dict__.items()
+            if k not in ('__dict__', '__weakref__', '__module__', '__doc__')
+        }
+    # new class
+    new_cls = type(
+        cls.__name__,
+        (cls,),
+        cls_dict
+    )
+    # metadata preservation
+    new_cls.__module__ = cls.__module__
+    new_cls.__doc__ = cls.__doc__
+    return new_cls
+
+
+class classproperty(object):
+    def __init__(self, f):
+        self.f = f
+    def __get__(self, obj, owner):
+        return self.f(owner)
+
+
+# NOTE: this was ai generated and validated by hand
+def shallow_clone_class(cls, new_name=None):
+    '''
+    Shallow clone a class while preserving super() functionality.
+    '''
+    new_name = new_name or f"{cls.__name__}Clone"
+    # Include the original class in the bases to maintain proper inheritance
+    new_bases = (cls,) + cls.__bases__
+    return type(new_name, new_bases, dict(cls.__dict__))
+
+# NOTE: this was ai generated and validated by hand
+def lock_class(cls):
+    '''
+    Lock a class so that its top-levelattributes cannot be modified.
+    '''
+    # Locked instance __setattr__
+    def locked_instance_setattr(self, name, value):
+        raise AttributeError(
+            f"Cannot set attribute '{name}' on immutable instance of {type(self).__name__}"
+        )
+    # Locked metaclass
+    class LockedMeta(type(cls)):
+        def __setattr__(cls_, name, value):
+            raise AttributeError(
+                f"Cannot modify class attribute '{name}' on locked class '{cls_.__name__}'"
+            )
+    # Rebuild class with locked behavior
+    locked_dict = dict(cls.__dict__)
+    locked_dict['__setattr__'] = locked_instance_setattr
+
+    return LockedMeta(cls.__name__, cls.__bases__, locked_dict)
+
+
+def make_locked_method_func(type_obj, func, class_clone):
+    """
+    Returns a function that, when called with **inputs, will execute:
+    getattr(type_obj, func).__func__(lock_class(class_clone), **inputs)
+
+    Supports both synchronous and asynchronous methods.
+    """
+    locked_class = lock_class(class_clone)
+    method = getattr(type_obj, func).__func__
+
+    # Check if the original method is async
+    if asyncio.iscoroutinefunction(method):
+        async def wrapped_async_func(**inputs):
+            return await method(locked_class, **inputs)
+        return wrapped_async_func
+    else:
+        def wrapped_func(**inputs):
+            return method(locked_class, **inputs)
+        return wrapped_func

comfy_api/internal/api_registry.py

@@ -0,0 +1,39 @@
+from typing import Type, List, NamedTuple
+from comfy_api.internal.singleton import ProxiedSingleton
+from packaging import version as packaging_version
+
+
+class ComfyAPIBase(ProxiedSingleton):
+    def __init__(self):
+        pass
+
+
+class ComfyAPIWithVersion(NamedTuple):
+    version: str
+    api_class: Type[ComfyAPIBase]
+
+
+def parse_version(version_str: str) -> packaging_version.Version:
+    """
+    Parses a version string into a packaging_version.Version object.
+    Raises ValueError if the version string is invalid.
+    """
+    if version_str == "latest":
+        return packaging_version.parse("9999999.9999999.9999999")
+    return packaging_version.parse(version_str)
+
+
+registered_versions: List[ComfyAPIWithVersion] = []
+
+
+def register_versions(versions: List[ComfyAPIWithVersion]):
+    versions.sort(key=lambda x: parse_version(x.version))
+    global registered_versions
+    registered_versions = versions
+
+
+def get_all_versions() -> List[ComfyAPIWithVersion]:
+    """
+    Returns a list of all registered ComfyAPI versions.
+    """
+    return registered_versions

comfy_api/internal/async_to_sync.py

@@ -0,0 +1,987 @@
+import asyncio
+import concurrent.futures
+import contextvars
+import functools
+import inspect
+import logging
+import os
+import textwrap
+import threading
+from enum import Enum
+from typing import Optional, Type, get_origin, get_args
+
+
+class TypeTracker:
+    """Tracks types discovered during stub generation for automatic import generation."""
+
+    def __init__(self):
+        self.discovered_types = {}  # type_name -> (module, qualname)
+        self.builtin_types = {
+            "Any",
+            "Dict",
+            "List",
+            "Optional",
+            "Tuple",
+            "Union",
+            "Set",
+            "Sequence",
+            "cast",
+            "NamedTuple",
+            "str",
+            "int",
+            "float",
+            "bool",
+            "None",
+            "bytes",
+            "object",
+            "type",
+            "dict",
+            "list",
+            "tuple",
+            "set",
+        }
+        self.already_imported = (
+            set()
+        )  # Track types already imported to avoid duplicates
+
+    def track_type(self, annotation):
+        """Track a type annotation and record its module/import info."""
+        if annotation is None or annotation is type(None):
+            return
+
+        # Skip builtins and typing module types we already import
+        type_name = getattr(annotation, "__name__", None)
+        if type_name and (
+            type_name in self.builtin_types or type_name in self.already_imported
+        ):
+            return
+
+        # Get module and qualname
+        module = getattr(annotation, "__module__", None)
+        qualname = getattr(annotation, "__qualname__", type_name or "")
+
+        # Skip types from typing module (they're already imported)
+        if module == "typing":
+            return
+
+        # Skip UnionType and GenericAlias from types module as they're handled specially
+        if module == "types" and type_name in ("UnionType", "GenericAlias"):
+            return
+
+        if module and module not in ["builtins", "__main__"]:
+            # Store the type info
+            if type_name:
+                self.discovered_types[type_name] = (module, qualname)
+
+    def get_imports(self, main_module_name: str) -> list[str]:
+        """Generate import statements for all discovered types."""
+        imports = []
+        imports_by_module = {}
+
+        for type_name, (module, qualname) in sorted(self.discovered_types.items()):
+            # Skip types from the main module (they're already imported)
+            if main_module_name and module == main_module_name:
+                continue
+
+            if module not in imports_by_module:
+                imports_by_module[module] = []
+            if type_name not in imports_by_module[module]:  # Avoid duplicates
+                imports_by_module[module].append(type_name)
+
+        # Generate import statements
+        for module, types in sorted(imports_by_module.items()):
+            if len(types) == 1:
+                imports.append(f"from {module} import {types[0]}")
+            else:
+                imports.append(f"from {module} import {', '.join(sorted(set(types)))}")
+
+        return imports
+
+
+class AsyncToSyncConverter:
+    """
+    Provides utilities to convert async classes to sync classes with proper type hints.
+    """
+
+    _thread_pool: Optional[concurrent.futures.ThreadPoolExecutor] = None
+    _thread_pool_lock = threading.Lock()
+    _thread_pool_initialized = False
+
+    @classmethod
+    def get_thread_pool(cls, max_workers=None) -> concurrent.futures.ThreadPoolExecutor:
+        """Get or create the shared thread pool with proper thread-safe initialization."""
+        # Fast path - check if already initialized without acquiring lock
+        if cls._thread_pool_initialized:
+            assert cls._thread_pool is not None, "Thread pool should be initialized"
+            return cls._thread_pool
+
+        # Slow path - acquire lock and create pool if needed
+        with cls._thread_pool_lock:
+            if not cls._thread_pool_initialized:
+                cls._thread_pool = concurrent.futures.ThreadPoolExecutor(
+                    max_workers=max_workers, thread_name_prefix="async_to_sync_"
+                )
+                cls._thread_pool_initialized = True
+
+        # This should never be None at this point, but add assertion for type checker
+        assert cls._thread_pool is not None
+        return cls._thread_pool
+
+    @classmethod
+    def run_async_in_thread(cls, coro_func, *args, **kwargs):
+        """
+        Run an async function in a separate thread from the thread pool.
+        Blocks until the async function completes.
+        Properly propagates contextvars between threads and manages event loops.
+        """
+        # Capture current context - this includes all context variables
+        context = contextvars.copy_context()
+
+        # Store the result and any exception that occurs
+        result_container: dict = {"result": None, "exception": None}
+
+        # Function that runs in the thread pool
+        def run_in_thread():
+            # Create new event loop for this thread
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+
+            try:
+                # Create the coroutine within the context
+                async def run_with_context():
+                    # The coroutine function might access context variables
+                    return await coro_func(*args, **kwargs)
+
+                # Run the coroutine with the captured context
+                # This ensures all context variables are available in the async function
+                result = context.run(loop.run_until_complete, run_with_context())
+                result_container["result"] = result
+            except Exception as e:
+                # Store the exception to re-raise in the calling thread
+                result_container["exception"] = e
+            finally:
+                # Ensure event loop is properly closed to prevent warnings
+                try:
+                    # Cancel any remaining tasks
+                    pending = asyncio.all_tasks(loop)
+                    for task in pending:
+                        task.cancel()
+
+                    # Run the loop briefly to handle cancellations
+                    if pending:
+                        loop.run_until_complete(
+                            asyncio.gather(*pending, return_exceptions=True)
+                        )
+                except Exception:
+                    pass  # Ignore errors during cleanup
+
+                # Close the event loop
+                loop.close()
+
+                # Clear the event loop from the thread
+                asyncio.set_event_loop(None)
+
+        # Submit to thread pool and wait for result
+        thread_pool = cls.get_thread_pool()
+        future = thread_pool.submit(run_in_thread)
+        future.result()  # Wait for completion
+
+        # Re-raise any exception that occurred in the thread
+        if result_container["exception"] is not None:
+            raise result_container["exception"]
+
+        return result_container["result"]
+
+    @classmethod
+    def create_sync_class(cls, async_class: Type, thread_pool_size=10) -> Type:
+        """
+        Creates a new class with synchronous versions of all async methods.
+
+        Args:
+            async_class: The async class to convert
+            thread_pool_size: Size of thread pool to use
+
+        Returns:
+            A new class with sync versions of all async methods
+        """
+        sync_class_name = "ComfyAPISyncStub"
+        cls.get_thread_pool(thread_pool_size)
+
+        # Create a proper class with docstrings and proper base classes
+        sync_class_dict = {
+            "__doc__": async_class.__doc__,
+            "__module__": async_class.__module__,
+            "__qualname__": sync_class_name,
+            "__orig_class__": async_class,  # Store original class for typing references
+        }
+
+        # Create __init__ method
+        def __init__(self, *args, **kwargs):
+            self._async_instance = async_class(*args, **kwargs)
+
+            # Handle annotated class attributes (like execution: Execution)
+            # Get all annotations from the class hierarchy
+            all_annotations = {}
+            for base_class in reversed(inspect.getmro(async_class)):
+                if hasattr(base_class, "__annotations__"):
+                    all_annotations.update(base_class.__annotations__)
+
+            # For each annotated attribute, check if it needs to be created or wrapped
+            for attr_name, attr_type in all_annotations.items():
+                if hasattr(self._async_instance, attr_name):
+                    # Attribute exists on the instance
+                    attr = getattr(self._async_instance, attr_name)
+                    # Check if this attribute needs a sync wrapper
+                    if hasattr(attr, "__class__"):
+                        from comfy_api.internal.singleton import ProxiedSingleton
+
+                        if isinstance(attr, ProxiedSingleton):
+                            # Create a sync version of this attribute
+                            try:
+                                sync_attr_class = cls.create_sync_class(attr.__class__)
+                                # Create instance of the sync wrapper with the async instance
+                                sync_attr = object.__new__(sync_attr_class)  # type: ignore
+                                sync_attr._async_instance = attr
+                                setattr(self, attr_name, sync_attr)
+                            except Exception:
+                                # If we can't create a sync version, keep the original
+                                setattr(self, attr_name, attr)
+                        else:
+                            # Not async, just copy the reference
+                            setattr(self, attr_name, attr)
+                else:
+                    # Attribute doesn't exist, but is annotated - create it
+                    # This handles cases like execution: Execution
+                    if isinstance(attr_type, type):
+                        # Check if the type is defined as an inner class
+                        if hasattr(async_class, attr_type.__name__):
+                            inner_class = getattr(async_class, attr_type.__name__)
+                            from comfy_api.internal.singleton import ProxiedSingleton
+
+                            # Create an instance of the inner class
+                            try:
+                                # For ProxiedSingleton classes, get or create the singleton instance
+                                if issubclass(inner_class, ProxiedSingleton):
+                                    async_instance = inner_class.get_instance()
+                                else:
+                                    async_instance = inner_class()
+
+                                # Create sync wrapper
+                                sync_attr_class = cls.create_sync_class(inner_class)
+                                sync_attr = object.__new__(sync_attr_class)  # type: ignore
+                                sync_attr._async_instance = async_instance
+                                setattr(self, attr_name, sync_attr)
+                                # Also set on the async instance for consistency
+                                setattr(self._async_instance, attr_name, async_instance)
+                            except Exception as e:
+                                logging.warning(
+                                    f"Failed to create instance for {attr_name}: {e}"
+                                )
+
+            # Handle other instance attributes that might not be annotated
+            for name, attr in inspect.getmembers(self._async_instance):
+                if name.startswith("_") or hasattr(self, name):
+                    continue
+
+                # If attribute is an instance of a class, and that class is defined in the original class
+                # we need to check if it needs a sync wrapper
+                if isinstance(attr, object) and not isinstance(
+                    attr, (str, int, float, bool, list, dict, tuple)
+                ):
+                    from comfy_api.internal.singleton import ProxiedSingleton
+
+                    if isinstance(attr, ProxiedSingleton):
+                        # Create a sync version of this nested class
+                        try:
+                            sync_attr_class = cls.create_sync_class(attr.__class__)
+                            # Create instance of the sync wrapper with the async instance
+                            sync_attr = object.__new__(sync_attr_class)  # type: ignore
+                            sync_attr._async_instance = attr
+                            setattr(self, name, sync_attr)
+                        except Exception:
+                            # If we can't create a sync version, keep the original
+                            setattr(self, name, attr)
+
+        sync_class_dict["__init__"] = __init__
+
+        # Process methods from the async class
+        for name, method in inspect.getmembers(
+            async_class, predicate=inspect.isfunction
+        ):
+            if name.startswith("_"):
+                continue
+
+            # Extract the actual return type from a coroutine
+            if inspect.iscoroutinefunction(method):
+                # Create sync version of async method with proper signature
+                @functools.wraps(method)
+                def sync_method(self, *args, _method_name=name, **kwargs):
+                    async_method = getattr(self._async_instance, _method_name)
+                    return AsyncToSyncConverter.run_async_in_thread(
+                        async_method, *args, **kwargs
+                    )
+
+                # Add to the class dict
+                sync_class_dict[name] = sync_method
+            else:
+                # For regular methods, create a proxy method
+                @functools.wraps(method)
+                def proxy_method(self, *args, _method_name=name, **kwargs):
+                    method = getattr(self._async_instance, _method_name)
+                    return method(*args, **kwargs)
+
+                # Add to the class dict
+                sync_class_dict[name] = proxy_method
+
+        # Handle property access
+        for name, prop in inspect.getmembers(
+            async_class, lambda x: isinstance(x, property)
+        ):
+
+            def make_property(name, prop_obj):
+                def getter(self):
+                    value = getattr(self._async_instance, name)
+                    if inspect.iscoroutinefunction(value):
+
+                        def sync_fn(*args, **kwargs):
+                            return AsyncToSyncConverter.run_async_in_thread(
+                                value, *args, **kwargs
+                            )
+
+                        return sync_fn
+                    return value
+
+                def setter(self, value):
+                    setattr(self._async_instance, name, value)
+
+                return property(getter, setter if prop_obj.fset else None)
+
+            sync_class_dict[name] = make_property(name, prop)
+
+        # Create the class
+        sync_class = type(sync_class_name, (object,), sync_class_dict)
+
+        return sync_class
+
+    @classmethod
+    def _format_type_annotation(
+        cls, annotation, type_tracker: Optional[TypeTracker] = None
+    ) -> str:
+        """Convert a type annotation to its string representation for stub files."""
+        if (
+            annotation is inspect.Parameter.empty
+            or annotation is inspect.Signature.empty
+        ):
+            return "Any"
+
+        # Handle None type
+        if annotation is type(None):
+            return "None"
+
+        # Track the type if we have a tracker
+        if type_tracker:
+            type_tracker.track_type(annotation)
+
+        # Try using typing.get_origin/get_args for Python 3.8+
+        try:
+            origin = get_origin(annotation)
+            args = get_args(annotation)
+
+            if origin is not None:
+                # Track the origin type
+                if type_tracker:
+                    type_tracker.track_type(origin)
+
+                # Get the origin name
+                origin_name = getattr(origin, "__name__", str(origin))
+                if "." in origin_name:
+                    origin_name = origin_name.split(".")[-1]
+
+                # Special handling for types.UnionType (Python 3.10+ pipe operator)
+                # Convert to old-style Union for compatibility
+                if str(origin) == "<class 'types.UnionType'>" or origin_name == "UnionType":
+                    origin_name = "Union"
+
+                # Format arguments recursively
+                if args:
+                    formatted_args = []
+                    for arg in args:
+                        # Track each type in the union
+                        if type_tracker:
+                            type_tracker.track_type(arg)
+                        formatted_args.append(cls._format_type_annotation(arg, type_tracker))
+                    return f"{origin_name}[{', '.join(formatted_args)}]"
+                else:
+                    return origin_name
+        except (AttributeError, TypeError):
+            # Fallback for older Python versions or non-generic types
+            pass
+
+        # Handle generic types the old way for compatibility
+        if hasattr(annotation, "__origin__") and hasattr(annotation, "__args__"):
+            origin = annotation.__origin__
+            origin_name = (
+                origin.__name__
+                if hasattr(origin, "__name__")
+                else str(origin).split("'")[1]
+            )
+
+            # Format each type argument
+            args = []
+            for arg in annotation.__args__:
+                args.append(cls._format_type_annotation(arg, type_tracker))
+
+            return f"{origin_name}[{', '.join(args)}]"
+
+        # Handle regular types with __name__
+        if hasattr(annotation, "__name__"):
+            return annotation.__name__
+
+        # Handle special module types (like types from typing module)
+        if hasattr(annotation, "__module__") and hasattr(annotation, "__qualname__"):
+            # For types like typing.Literal, typing.TypedDict, etc.
+            return annotation.__qualname__
+
+        # Last resort: string conversion with cleanup
+        type_str = str(annotation)
+
+        # Clean up common patterns more robustly
+        if type_str.startswith("<class '") and type_str.endswith("'>"):
+            type_str = type_str[8:-2]  # Remove "<class '" and "'>"
+
+        # Remove module prefixes for common modules
+        for prefix in ["typing.", "builtins.", "types."]:
+            if type_str.startswith(prefix):
+                type_str = type_str[len(prefix) :]
+
+        # Handle special cases
+        if type_str in ("_empty", "inspect._empty"):
+            return "None"
+
+        # Fix NoneType (this should rarely be needed now)
+        if type_str == "NoneType":
+            return "None"
+
+        return type_str
+
+    @classmethod
+    def _extract_coroutine_return_type(cls, annotation):
+        """Extract the actual return type from a Coroutine annotation."""
+        if hasattr(annotation, "__args__") and len(annotation.__args__) > 2:
+            # Coroutine[Any, Any, ReturnType] -> extract ReturnType
+            return annotation.__args__[2]
+        return annotation
+
+    @classmethod
+    def _format_parameter_default(cls, default_value) -> str:
+        """Format a parameter's default value for stub files."""
+        if default_value is inspect.Parameter.empty:
+            return ""
+        elif default_value is None:
+            return " = None"
+        elif isinstance(default_value, bool):
+            return f" = {default_value}"
+        elif default_value == {}:
+            return " = {}"
+        elif default_value == []:
+            return " = []"
+        else:
+            return f" = {default_value}"
+
+    @classmethod
+    def _format_method_parameters(
+        cls,
+        sig: inspect.Signature,
+        skip_self: bool = True,
+        type_hints: Optional[dict] = None,
+        type_tracker: Optional[TypeTracker] = None,
+    ) -> str:
+        """Format method parameters for stub files."""
+        params = []
+        if type_hints is None:
+            type_hints = {}
+
+        for i, (param_name, param) in enumerate(sig.parameters.items()):
+            if i == 0 and param_name == "self" and skip_self:
+                params.append("self")
+            else:
+                # Get type annotation from type hints if available, otherwise from signature
+                annotation = type_hints.get(param_name, param.annotation)
+                type_str = cls._format_type_annotation(annotation, type_tracker)
+
+                # Get default value
+                default_str = cls._format_parameter_default(param.default)
+
+                # Combine parameter parts
+                if annotation is inspect.Parameter.empty:
+                    params.append(f"{param_name}: Any{default_str}")
+                else:
+                    params.append(f"{param_name}: {type_str}{default_str}")
+
+        return ", ".join(params)
+
+    @classmethod
+    def _generate_method_signature(
+        cls,
+        method_name: str,
+        method,
+        is_async: bool = False,
+        type_tracker: Optional[TypeTracker] = None,
+    ) -> str:
+        """Generate a complete method signature for stub files."""
+        sig = inspect.signature(method)
+
+        # Try to get evaluated type hints to resolve string annotations
+        try:
+            from typing import get_type_hints
+            type_hints = get_type_hints(method)
+        except Exception:
+            # Fallback to empty dict if we can't get type hints
+            type_hints = {}
+
+        # For async methods, extract the actual return type
+        return_annotation = type_hints.get('return', sig.return_annotation)
+        if is_async and inspect.iscoroutinefunction(method):
+            return_annotation = cls._extract_coroutine_return_type(return_annotation)
+
+        # Format parameters with type hints
+        params_str = cls._format_method_parameters(sig, type_hints=type_hints, type_tracker=type_tracker)
+
+        # Format return type
+        return_type = cls._format_type_annotation(return_annotation, type_tracker)
+        if return_annotation is inspect.Signature.empty:
+            return_type = "None"
+
+        return f"def {method_name}({params_str}) -> {return_type}: ..."
+
+    @classmethod
+    def _generate_imports(
+        cls, async_class: Type, type_tracker: TypeTracker
+    ) -> list[str]:
+        """Generate import statements for the stub file."""
+        imports = []
+
+        # Add standard typing imports
+        imports.append(
+            "from typing import Any, Dict, List, Optional, Tuple, Union, Set, Sequence, cast, NamedTuple"
+        )
+
+        # Add imports from the original module
+        if async_class.__module__ != "builtins":
+            module = inspect.getmodule(async_class)
+            additional_types = []
+
+            if module:
+                # Check if module has __all__ defined
+                module_all = getattr(module, "__all__", None)
+
+                for name, obj in sorted(inspect.getmembers(module)):
+                    if isinstance(obj, type):
+                        # Skip if __all__ is defined and this name isn't in it
+                        # unless it's already been tracked as used in type annotations
+                        if module_all is not None and name not in module_all:
+                            # Check if this type was actually used in annotations
+                            if name not in type_tracker.discovered_types:
+                                continue
+
+                        # Check for NamedTuple
+                        if issubclass(obj, tuple) and hasattr(obj, "_fields"):
+                            additional_types.append(name)
+                            # Mark as already imported
+                            type_tracker.already_imported.add(name)
+                        # Check for Enum
+                        elif issubclass(obj, Enum) and name != "Enum":
+                            additional_types.append(name)
+                            # Mark as already imported
+                            type_tracker.already_imported.add(name)
+
+            if additional_types:
+                type_imports = ", ".join([async_class.__name__] + additional_types)
+                imports.append(f"from {async_class.__module__} import {type_imports}")
+            else:
+                imports.append(
+                    f"from {async_class.__module__} import {async_class.__name__}"
+                )
+
+        # Add imports for all discovered types
+        # Pass the main module name to avoid duplicate imports
+        imports.extend(
+            type_tracker.get_imports(main_module_name=async_class.__module__)
+        )
+
+        # Add base module import if needed
+        if hasattr(inspect.getmodule(async_class), "__name__"):
+            module_name = inspect.getmodule(async_class).__name__
+            if "." in module_name:
+                base_module = module_name.split(".")[0]
+                # Only add if not already importing from it
+                if not any(imp.startswith(f"from {base_module}") for imp in imports):
+                    imports.append(f"import {base_module}")
+
+        return imports
+
+    @classmethod
+    def _get_class_attributes(cls, async_class: Type) -> list[tuple[str, Type]]:
+        """Extract class attributes that are classes themselves."""
+        class_attributes = []
+
+        # Look for class attributes that are classes
+        for name, attr in sorted(inspect.getmembers(async_class)):
+            if isinstance(attr, type) and not name.startswith("_"):
+                class_attributes.append((name, attr))
+            elif (
+                hasattr(async_class, "__annotations__")
+                and name in async_class.__annotations__
+            ):
+                annotation = async_class.__annotations__[name]
+                if isinstance(annotation, type):
+                    class_attributes.append((name, annotation))
+
+        return class_attributes
+
+    @classmethod
+    def _generate_inner_class_stub(
+        cls,
+        name: str,
+        attr: Type,
+        indent: str = "    ",
+        type_tracker: Optional[TypeTracker] = None,
+    ) -> list[str]:
+        """Generate stub for an inner class."""
+        stub_lines = []
+        stub_lines.append(f"{indent}class {name}Sync:")
+
+        # Add docstring if available
+        if hasattr(attr, "__doc__") and attr.__doc__:
+            stub_lines.extend(
+                cls._format_docstring_for_stub(attr.__doc__, f"{indent}    ")
+            )
+
+        # Add __init__ if it exists
+        if hasattr(attr, "__init__"):
+            try:
+                init_method = getattr(attr, "__init__")
+                init_sig = inspect.signature(init_method)
+
+                # Try to get type hints
+                try:
+                    from typing import get_type_hints
+                    init_hints = get_type_hints(init_method)
+                except Exception:
+                    init_hints = {}
+
+                # Format parameters
+                params_str = cls._format_method_parameters(
+                    init_sig, type_hints=init_hints, type_tracker=type_tracker
+                )
+                # Add __init__ docstring if available (before the method)
+                if hasattr(init_method, "__doc__") and init_method.__doc__:
+                    stub_lines.extend(
+                        cls._format_docstring_for_stub(
+                            init_method.__doc__, f"{indent}    "
+                        )
+                    )
+                stub_lines.append(
+                    f"{indent}    def __init__({params_str}) -> None: ..."
+                )
+            except (ValueError, TypeError):
+                stub_lines.append(
+                    f"{indent}    def __init__(self, *args, **kwargs) -> None: ..."
+                )
+
+        # Add methods to the inner class
+        has_methods = False
+        for method_name, method in sorted(
+            inspect.getmembers(attr, predicate=inspect.isfunction)
+        ):
+            if method_name.startswith("_"):
+                continue
+
+            has_methods = True
+            try:
+                # Add method docstring if available (before the method signature)
+                if method.__doc__:
+                    stub_lines.extend(
+                        cls._format_docstring_for_stub(method.__doc__, f"{indent}    ")
+                    )
+
+                method_sig = cls._generate_method_signature(
+                    method_name, method, is_async=True, type_tracker=type_tracker
+                )
+                stub_lines.append(f"{indent}    {method_sig}")
+            except (ValueError, TypeError):
+                stub_lines.append(
+                    f"{indent}    def {method_name}(self, *args, **kwargs): ..."
+                )
+
+        if not has_methods:
+            stub_lines.append(f"{indent}    pass")
+
+        return stub_lines
+
+    @classmethod
+    def _format_docstring_for_stub(
+        cls, docstring: str, indent: str = "    "
+    ) -> list[str]:
+        """Format a docstring for inclusion in a stub file with proper indentation."""
+        if not docstring:
+            return []
+
+        # First, dedent the docstring to remove any existing indentation
+        dedented = textwrap.dedent(docstring).strip()
+
+        # Split into lines
+        lines = dedented.split("\n")
+
+        # Build the properly indented docstring
+        result = []
+        result.append(f'{indent}"""')
+
+        for line in lines:
+            if line.strip():  # Non-empty line
+                result.append(f"{indent}{line}")
+            else:  # Empty line
+                result.append("")
+
+        result.append(f'{indent}"""')
+        return result
+
+    @classmethod
+    def _post_process_stub_content(cls, stub_content: list[str]) -> list[str]:
+        """Post-process stub content to fix any remaining issues."""
+        processed = []
+
+        for line in stub_content:
+            # Skip processing imports
+            if line.startswith(("from ", "import ")):
+                processed.append(line)
+                continue
+
+            # Fix method signatures missing return types
+            if (
+                line.strip().startswith("def ")
+                and line.strip().endswith(": ...")
+                and ") -> " not in line
+            ):
+                # Add -> None for methods without return annotation
+                line = line.replace(": ...", " -> None: ...")
+
+            processed.append(line)
+
+        return processed
+
+    @classmethod
+    def generate_stub_file(cls, async_class: Type, sync_class: Type) -> None:
+        """
+        Generate a .pyi stub file for the sync class to help IDEs with type checking.
+        """
+        try:
+            # Only generate stub if we can determine module path
+            if async_class.__module__ == "__main__":
+                return
+
+            module = inspect.getmodule(async_class)
+            if not module:
+                return
+
+            module_path = module.__file__
+            if not module_path:
+                return
+
+            # Create stub file path in a 'generated' subdirectory
+            module_dir = os.path.dirname(module_path)
+            stub_dir = os.path.join(module_dir, "generated")
+
+            # Ensure the generated directory exists
+            os.makedirs(stub_dir, exist_ok=True)
+
+            module_name = os.path.basename(module_path)
+            if module_name.endswith(".py"):
+                module_name = module_name[:-3]
+
+            sync_stub_path = os.path.join(stub_dir, f"{sync_class.__name__}.pyi")
+
+            # Create a type tracker for this stub generation
+            type_tracker = TypeTracker()
+
+            stub_content = []
+
+            # We'll generate imports after processing all methods to capture all types
+            # Leave a placeholder for imports
+            imports_placeholder_index = len(stub_content)
+            stub_content.append("")  # Will be replaced with imports later
+
+            # Class definition
+            stub_content.append(f"class {sync_class.__name__}:")
+
+            # Docstring
+            if async_class.__doc__:
+                stub_content.extend(
+                    cls._format_docstring_for_stub(async_class.__doc__, "    ")
+                )
+
+            # Generate __init__
+            try:
+                init_method = async_class.__init__
+                init_signature = inspect.signature(init_method)
+
+                # Try to get type hints for __init__
+                try:
+                    from typing import get_type_hints
+                    init_hints = get_type_hints(init_method)
+                except Exception:
+                    init_hints = {}
+
+                # Format parameters
+                params_str = cls._format_method_parameters(
+                    init_signature, type_hints=init_hints, type_tracker=type_tracker
+                )
+                # Add __init__ docstring if available (before the method)
+                if hasattr(init_method, "__doc__") and init_method.__doc__:
+                    stub_content.extend(
+                        cls._format_docstring_for_stub(init_method.__doc__, "    ")
+                    )
+                stub_content.append(f"    def __init__({params_str}) -> None: ...")
+            except (ValueError, TypeError):
+                stub_content.append(
+                    "    def __init__(self, *args, **kwargs) -> None: ..."
+                )
+
+            stub_content.append("")  # Add newline after __init__
+
+            # Get class attributes
+            class_attributes = cls._get_class_attributes(async_class)
+
+            # Generate inner classes
+            for name, attr in class_attributes:
+                inner_class_stub = cls._generate_inner_class_stub(
+                    name, attr, type_tracker=type_tracker
+                )
+                stub_content.extend(inner_class_stub)
+                stub_content.append("")  # Add newline after the inner class
+
+            # Add methods to the main class
+            processed_methods = set()  # Keep track of methods we've processed
+            for name, method in sorted(
+                inspect.getmembers(async_class, predicate=inspect.isfunction)
+            ):
+                if name.startswith("_") or name in processed_methods:
+                    continue
+
+                processed_methods.add(name)
+
+                try:
+                    method_sig = cls._generate_method_signature(
+                        name, method, is_async=True, type_tracker=type_tracker
+                    )
+
+                    # Add docstring if available (before the method signature for proper formatting)
+                    if method.__doc__:
+                        stub_content.extend(
+                            cls._format_docstring_for_stub(method.__doc__, "    ")
+                        )
+
+                    stub_content.append(f"    {method_sig}")
+
+                    stub_content.append("")  # Add newline after each method
+
+                except (ValueError, TypeError):
+                    # If we can't get the signature, just add a simple stub
+                    stub_content.append(f"    def {name}(self, *args, **kwargs): ...")
+                    stub_content.append("")  # Add newline
+
+            # Add properties
+            for name, prop in sorted(
+                inspect.getmembers(async_class, lambda x: isinstance(x, property))
+            ):
+                stub_content.append("    @property")
+                stub_content.append(f"    def {name}(self) -> Any: ...")
+                if prop.fset:
+                    stub_content.append(f"    @{name}.setter")
+                    stub_content.append(
+                        f"    def {name}(self, value: Any) -> None: ..."
+                    )
+                stub_content.append("")  # Add newline after each property
+
+            # Add placeholders for the nested class instances
+            # Check the actual attribute names from class annotations and attributes
+            attribute_mappings = {}
+
+            # First check annotations for typed attributes (including from parent classes)
+            # Collect all annotations from the class hierarchy
+            all_annotations = {}
+            for base_class in reversed(inspect.getmro(async_class)):
+                if hasattr(base_class, "__annotations__"):
+                    all_annotations.update(base_class.__annotations__)
+
+            for attr_name, attr_type in sorted(all_annotations.items()):
+                for class_name, class_type in class_attributes:
+                    # If the class type matches the annotated type
+                    if (
+                        attr_type == class_type
+                        or (hasattr(attr_type, "__name__") and attr_type.__name__ == class_name)
+                        or (isinstance(attr_type, str) and attr_type == class_name)
+                    ):
+                        attribute_mappings[class_name] = attr_name
+
+            # Remove the extra checking - annotations should be sufficient
+
+            # Add the attribute declarations with proper names
+            for class_name, class_type in class_attributes:
+                # Check if there's a mapping from annotation
+                attr_name = attribute_mappings.get(class_name, class_name)
+                # Use the annotation name if it exists, even if the attribute doesn't exist yet
+                # This is because the attribute might be created at runtime
+                stub_content.append(f"    {attr_name}: {class_name}Sync")
+
+            stub_content.append("")  # Add a final newline
+
+            # Now generate imports with all discovered types
+            imports = cls._generate_imports(async_class, type_tracker)
+
+            # Deduplicate imports while preserving order
+            seen = set()
+            unique_imports = []
+            for imp in imports:
+                if imp not in seen:
+                    seen.add(imp)
+                    unique_imports.append(imp)
+                else:
+                    logging.warning(f"Duplicate import detected: {imp}")
+
+            # Replace the placeholder with actual imports
+            stub_content[imports_placeholder_index : imports_placeholder_index + 1] = (
+                unique_imports
+            )
+
+            # Post-process stub content
+            stub_content = cls._post_process_stub_content(stub_content)
+
+            # Write stub file
+            with open(sync_stub_path, "w") as f:
+                f.write("\n".join(stub_content))
+
+            logging.info(f"Generated stub file: {sync_stub_path}")
+
+        except Exception as e:
+            # If stub generation fails, log the error but don't break the main functionality
+            logging.error(
+                f"Error generating stub file for {sync_class.__name__}: {str(e)}"
+            )
+            import traceback
+
+            logging.error(traceback.format_exc())
+
+
+def create_sync_class(async_class: Type, thread_pool_size=10) -> Type:
+    """
+    Creates a sync version of an async class
+
+    Args:
+        async_class: The async class to convert
+        thread_pool_size: Size of thread pool to use
+
+    Returns:
+        A new class with sync versions of all async methods
+    """
+    return AsyncToSyncConverter.create_sync_class(async_class, thread_pool_size)

comfy_api/internal/singleton.py

@@ -0,0 +1,33 @@
+from typing import Type, TypeVar
+
+class SingletonMetaclass(type):
+    T = TypeVar("T", bound="SingletonMetaclass")
+    _instances = {}
+
+    def __call__(cls, *args, **kwargs):
+        if cls not in cls._instances:
+            cls._instances[cls] = super(SingletonMetaclass, cls).__call__(
+                *args, **kwargs
+            )
+        return cls._instances[cls]
+
+    def inject_instance(cls: Type[T], instance: T) -> None:
+        assert cls not in SingletonMetaclass._instances, (
+            "Cannot inject instance after first instantiation"
+        )
+        SingletonMetaclass._instances[cls] = instance
+
+    def get_instance(cls: Type[T], *args, **kwargs) -> T:
+        """
+        Gets the singleton instance of the class, creating it if it doesn't exist.
+        """
+        if cls not in SingletonMetaclass._instances:
+            SingletonMetaclass._instances[cls] = super(
+                SingletonMetaclass, cls
+            ).__call__(*args, **kwargs)
+        return cls._instances[cls]
+
+
+class ProxiedSingleton(object, metaclass=SingletonMetaclass):
+    def __init__(self):
+        super().__init__()

comfy_api/latest/__init__.py

@@ -0,0 +1,124 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Type, TYPE_CHECKING
+from comfy_api.internal import ComfyAPIBase
+from comfy_api.internal.singleton import ProxiedSingleton
+from comfy_api.internal.async_to_sync import create_sync_class
+from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
+from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents
+from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents
+from comfy_api.latest._io import _IO as io  #noqa: F401
+from comfy_api.latest._ui import _UI as ui  #noqa: F401
+# from comfy_api.latest._resources import _RESOURCES as resources  #noqa: F401
+from comfy_execution.utils import get_executing_context
+from comfy_execution.progress import get_progress_state, PreviewImageTuple
+from PIL import Image
+from comfy.cli_args import args
+import numpy as np
+
+
+class ComfyAPI_latest(ComfyAPIBase):
+    VERSION = "latest"
+    STABLE = False
+
+    class Execution(ProxiedSingleton):
+        async def set_progress(
+            self,
+            value: float,
+            max_value: float,
+            node_id: str | None = None,
+            preview_image: Image.Image | ImageInput | None = None,
+            ignore_size_limit: bool = False,
+        ) -> None:
+            """
+            Update the progress bar displayed in the ComfyUI interface.
+
+            This function allows custom nodes and API calls to report their progress
+            back to the user interface, providing visual feedback during long operations.
+
+            Migration from previous API: comfy.utils.PROGRESS_BAR_HOOK
+            """
+            executing_context = get_executing_context()
+            if node_id is None and executing_context is not None:
+                node_id = executing_context.node_id
+            if node_id is None:
+                raise ValueError("node_id must be provided if not in executing context")
+
+            # Convert preview_image to PreviewImageTuple if needed
+            to_display: PreviewImageTuple | Image.Image | ImageInput | None = preview_image
+            if to_display is not None:
+                # First convert to PIL Image if needed
+                if isinstance(to_display, ImageInput):
+                    # Convert ImageInput (torch.Tensor) to PIL Image
+                    # Handle tensor shape [B, H, W, C] -> get first image if batch
+                    tensor = to_display
+                    if len(tensor.shape) == 4:
+                        tensor = tensor[0]
+
+                    # Convert to numpy array and scale to 0-255
+                    image_np = (tensor.cpu().numpy() * 255).astype(np.uint8)
+                    to_display = Image.fromarray(image_np)
+
+                if isinstance(to_display, Image.Image):
+                    # Detect image format from PIL Image
+                    image_format = to_display.format if to_display.format else "JPEG"
+                    # Use None for preview_size if ignore_size_limit is True
+                    preview_size = None if ignore_size_limit else args.preview_size
+                    to_display = (image_format, to_display, preview_size)
+
+            get_progress_state().update_progress(
+                node_id=node_id,
+                value=value,
+                max_value=max_value,
+                image=to_display,
+            )
+
+    execution: Execution
+
+class ComfyExtension(ABC):
+    async def on_load(self) -> None:
+        """
+        Called when an extension is loaded.
+        This should be used to initialize any global resources neeeded by the extension.
+        """
+
+    @abstractmethod
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        """
+        Returns a list of nodes that this extension provides.
+        """
+
+class Input:
+    Image = ImageInput
+    Audio = AudioInput
+    Mask = MaskInput
+    Latent = LatentInput
+    Video = VideoInput
+
+class InputImpl:
+    VideoFromFile = VideoFromFile
+    VideoFromComponents = VideoFromComponents
+
+class Types:
+    VideoCodec = VideoCodec
+    VideoContainer = VideoContainer
+    VideoComponents = VideoComponents
+
+ComfyAPI = ComfyAPI_latest
+
+# Create a synchronous version of the API
+if TYPE_CHECKING:
+    import comfy_api.latest.generated.ComfyAPISyncStub  # type: ignore
+
+    ComfyAPISync: Type[comfy_api.latest.generated.ComfyAPISyncStub.ComfyAPISyncStub]
+ComfyAPISync = create_sync_class(ComfyAPI_latest)
+
+__all__ = [
+    "ComfyAPI",
+    "ComfyAPISync",
+    "Input",
+    "InputImpl",
+    "Types",
+    "ComfyExtension",
+]

comfy_api/latest/_input/__init__.py

@@ -0,0 +1,10 @@
+from .basic_types import ImageInput, AudioInput, MaskInput, LatentInput
+from .video_types import VideoInput
+
+__all__ = [
+    "ImageInput",
+    "AudioInput",
+    "VideoInput",
+    "MaskInput",
+    "LatentInput",
+]

comfy_api/latest/_input/basic_types.py

@@ -0,0 +1,42 @@
+import torch
+from typing import TypedDict, List, Optional
+
+ImageInput = torch.Tensor
+"""
+An image in format [B, H, W, C] where B is the batch size, C is the number of channels,
+"""
+
+MaskInput = torch.Tensor
+"""
+A mask in format [B, H, W] where B is the batch size
+"""
+
+class AudioInput(TypedDict):
+    """
+    TypedDict representing audio input.
+    """
+
+    waveform: torch.Tensor
+    """
+    Tensor in the format [B, C, T] where B is the batch size, C is the number of channels,
+    """
+
+    sample_rate: int
+
+class LatentInput(TypedDict):
+    """
+    TypedDict representing latent input.
+    """
+
+    samples: torch.Tensor
+    """
+    Tensor in the format [B, C, H, W] where B is the batch size, C is the number of channels,
+    H is the height, and W is the width.
+    """
+
+    noise_mask: Optional[MaskInput]
+    """
+    Optional noise mask tensor in the same format as samples.
+    """
+
+    batch_index: Optional[List[int]]

comfy_api/latest/_input/video_types.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+from abc import ABC, abstractmethod
+from typing import Optional, Union
+import io
+import av
+from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
+
+class VideoInput(ABC):
+    """
+    Abstract base class for video input types.
+    """
+
+    @abstractmethod
+    def get_components(self) -> VideoComponents:
+        """
+        Abstract method to get the video components (images, audio, and frame rate).
+
+        Returns:
+            VideoComponents containing images, audio, and frame rate
+        """
+        pass
+
+    @abstractmethod
+    def save_to(
+        self,
+        path: str,
+        format: VideoContainer = VideoContainer.AUTO,
+        codec: VideoCodec = VideoCodec.AUTO,
+        metadata: Optional[dict] = None
+    ):
+        """
+        Abstract method to save the video input to a file.
+        """
+        pass
+
+    def get_stream_source(self) -> Union[str, io.BytesIO]:
+        """
+        Get a streamable source for the video. This allows processing without
+        loading the entire video into memory.
+
+        Returns:
+            Either a file path (str) or a BytesIO object that can be opened with av.
+
+        Default implementation creates a BytesIO buffer, but subclasses should
+        override this for better performance when possible.
+        """
+        buffer = io.BytesIO()
+        self.save_to(buffer)
+        buffer.seek(0)
+        return buffer
+
+    # Provide a default implementation, but subclasses can provide optimized versions
+    # if possible.
+    def get_dimensions(self) -> tuple[int, int]:
+        """
+        Returns the dimensions of the video input.
+
+        Returns:
+            Tuple of (width, height)
+        """
+        components = self.get_components()
+        return components.images.shape[2], components.images.shape[1]
+
+    def get_duration(self) -> float:
+        """
+        Returns the duration of the video in seconds.
+
+        Returns:
+            Duration in seconds
+        """
+        components = self.get_components()
+        frame_count = components.images.shape[0]
+        return float(frame_count / components.frame_rate)
+
+    def get_container_format(self) -> str:
+        """
+        Returns the container format of the video (e.g., 'mp4', 'mov', 'avi').
+
+        Returns:
+            Container format as string
+        """
+        # Default implementation - subclasses should override for better performance
+        source = self.get_stream_source()
+        with av.open(source, mode="r") as container:
+            return container.format.name

comfy_api/latest/_input_impl/__init__.py

@@ -0,0 +1,7 @@
+from .video_types import VideoFromFile, VideoFromComponents
+
+__all__ = [
+    # Implementations
+    "VideoFromFile",
+    "VideoFromComponents",
+]

comfy_api/latest/_input_impl/video_types.py

@@ -0,0 +1,324 @@
+from __future__ import annotations
+from av.container import InputContainer
+from av.subtitles.stream import SubtitleStream
+from fractions import Fraction
+from typing import Optional
+from comfy_api.latest._input import AudioInput, VideoInput
+import av
+import io
+import json
+import numpy as np
+import torch
+from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents
+
+
+def container_to_output_format(container_format: str | None) -> str | None:
+    """
+    A container's `format` may be a comma-separated list of formats.
+    E.g., iso container's `format` may be `mov,mp4,m4a,3gp,3g2,mj2`.
+    However, writing to a file/stream with `av.open` requires a single format,
+    or `None` to auto-detect.
+    """
+    if not container_format:
+        return None  # Auto-detect
+
+    if "," not in container_format:
+        return container_format
+
+    formats = container_format.split(",")
+    return formats[0]
+
+
+def get_open_write_kwargs(
+    dest: str | io.BytesIO, container_format: str, to_format: str | None
+) -> dict:
+    """Get kwargs for writing a `VideoFromFile` to a file/stream with `av.open`"""
+    open_kwargs = {
+        "mode": "w",
+        # If isobmff, preserve custom metadata tags (workflow, prompt, extra_pnginfo)
+        "options": {"movflags": "use_metadata_tags"},
+    }
+
+    is_write_to_buffer = isinstance(dest, io.BytesIO)
+    if is_write_to_buffer:
+        # Set output format explicitly, since it cannot be inferred from file extension
+        if to_format == VideoContainer.AUTO:
+            to_format = container_format.lower()
+        elif isinstance(to_format, str):
+            to_format = to_format.lower()
+        open_kwargs["format"] = container_to_output_format(to_format)
+
+    return open_kwargs
+
+
+class VideoFromFile(VideoInput):
+    """
+    Class representing video input from a file.
+    """
+
+    def __init__(self, file: str | io.BytesIO):
+        """
+        Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object
+        containing the file contents.
+        """
+        self.__file = file
+
+    def get_stream_source(self) -> str | io.BytesIO:
+        """
+        Return the underlying file source for efficient streaming.
+        This avoids unnecessary memory copies when the source is already a file path.
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+        return self.__file
+
+    def get_dimensions(self) -> tuple[int, int]:
+        """
+        Returns the dimensions of the video input.
+
+        Returns:
+            Tuple of (width, height)
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)  # Reset the BytesIO object to the beginning
+        with av.open(self.__file, mode='r') as container:
+            for stream in container.streams:
+                if stream.type == 'video':
+                    assert isinstance(stream, av.VideoStream)
+                    return stream.width, stream.height
+        raise ValueError(f"No video stream found in file '{self.__file}'")
+
+    def get_duration(self) -> float:
+        """
+        Returns the duration of the video in seconds.
+
+        Returns:
+            Duration in seconds
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+        with av.open(self.__file, mode="r") as container:
+            if container.duration is not None:
+                return float(container.duration / av.time_base)
+
+            # Fallback: calculate from frame count and frame rate
+            video_stream = next(
+                (s for s in container.streams if s.type == "video"), None
+            )
+            if video_stream and video_stream.frames and video_stream.average_rate:
+                return float(video_stream.frames / video_stream.average_rate)
+
+            # Last resort: decode frames to count them
+            if video_stream and video_stream.average_rate:
+                frame_count = 0
+                container.seek(0)
+                for packet in container.demux(video_stream):
+                    for _ in packet.decode():
+                        frame_count += 1
+                if frame_count > 0:
+                    return float(frame_count / video_stream.average_rate)
+
+        raise ValueError(f"Could not determine duration for file '{self.__file}'")
+
+    def get_container_format(self) -> str:
+        """
+        Returns the container format of the video (e.g., 'mp4', 'mov', 'avi').
+
+        Returns:
+            Container format as string
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+        with av.open(self.__file, mode='r') as container:
+            return container.format.name
+
+    def get_components_internal(self, container: InputContainer) -> VideoComponents:
+        # Get video frames
+        frames = []
+        for frame in container.decode(video=0):
+            img = frame.to_ndarray(format='rgb24')  # shape: (H, W, 3)
+            img = torch.from_numpy(img) / 255.0  # shape: (H, W, 3)
+            frames.append(img)
+
+        images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0)
+
+        # Get frame rate
+        video_stream = next(s for s in container.streams if s.type == 'video')
+        frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1)
+
+        # Get audio if available
+        audio = None
+        try:
+            container.seek(0)  # Reset the container to the beginning
+            for stream in container.streams:
+                if stream.type != 'audio':
+                    continue
+                assert isinstance(stream, av.AudioStream)
+                audio_frames = []
+                for packet in container.demux(stream):
+                    for frame in packet.decode():
+                        assert isinstance(frame, av.AudioFrame)
+                        audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
+                if len(audio_frames) > 0:
+                    audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
+                    audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
+                    audio = AudioInput({
+                        "waveform": audio_tensor,
+                        "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1,
+                    })
+        except StopIteration:
+            pass  # No audio stream
+
+        metadata = container.metadata
+        return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
+
+    def get_components(self) -> VideoComponents:
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)  # Reset the BytesIO object to the beginning
+        with av.open(self.__file, mode='r') as container:
+            return self.get_components_internal(container)
+        raise ValueError(f"No video stream found in file '{self.__file}'")
+
+    def save_to(
+        self,
+        path: str | io.BytesIO,
+        format: VideoContainer = VideoContainer.AUTO,
+        codec: VideoCodec = VideoCodec.AUTO,
+        metadata: Optional[dict] = None
+    ):
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)  # Reset the BytesIO object to the beginning
+        with av.open(self.__file, mode='r') as container:
+            container_format = container.format.name
+            video_encoding = container.streams.video[0].codec.name if len(container.streams.video) > 0 else None
+            reuse_streams = True
+            if format != VideoContainer.AUTO and format not in container_format.split(","):
+                reuse_streams = False
+            if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None:
+                reuse_streams = False
+
+            if not reuse_streams:
+                components = self.get_components_internal(container)
+                video = VideoFromComponents(components)
+                return video.save_to(
+                    path,
+                    format=format,
+                    codec=codec,
+                    metadata=metadata
+                )
+
+            streams = container.streams
+
+            open_kwargs = get_open_write_kwargs(path, container_format, format)
+            with av.open(path, **open_kwargs) as output_container:
+                # Copy over the original metadata
+                for key, value in container.metadata.items():
+                    if metadata is None or key not in metadata:
+                        output_container.metadata[key] = value
+
+                # Add our new metadata
+                if metadata is not None:
+                    for key, value in metadata.items():
+                        if isinstance(value, str):
+                            output_container.metadata[key] = value
+                        else:
+                            output_container.metadata[key] = json.dumps(value)
+
+                # Add streams to the new container
+                stream_map = {}
+                for stream in streams:
+                    if isinstance(stream, (av.VideoStream, av.AudioStream, SubtitleStream)):
+                        out_stream = output_container.add_stream_from_template(template=stream, opaque=True)
+                        stream_map[stream] = out_stream
+
+                # Write packets to the new container
+                for packet in container.demux():
+                    if packet.stream in stream_map and packet.dts is not None:
+                        packet.stream = stream_map[packet.stream]
+                        output_container.mux(packet)
+
+class VideoFromComponents(VideoInput):
+    """
+    Class representing video input from tensors.
+    """
+
+    def __init__(self, components: VideoComponents):
+        self.__components = components
+
+    def get_components(self) -> VideoComponents:
+        return VideoComponents(
+            images=self.__components.images,
+            audio=self.__components.audio,
+            frame_rate=self.__components.frame_rate
+        )
+
+    def save_to(
+        self,
+        path: str,
+        format: VideoContainer = VideoContainer.AUTO,
+        codec: VideoCodec = VideoCodec.AUTO,
+        metadata: Optional[dict] = None
+    ):
+        if format != VideoContainer.AUTO and format != VideoContainer.MP4:
+            raise ValueError("Only MP4 format is supported for now")
+        if codec != VideoCodec.AUTO and codec != VideoCodec.H264:
+            raise ValueError("Only H264 codec is supported for now")
+        with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}) as output:
+            # Add metadata before writing any streams
+            if metadata is not None:
+                for key, value in metadata.items():
+                    output.metadata[key] = json.dumps(value)
+
+            frame_rate = Fraction(round(self.__components.frame_rate * 1000), 1000)
+            # Create a video stream
+            video_stream = output.add_stream('h264', rate=frame_rate)
+            video_stream.width = self.__components.images.shape[2]
+            video_stream.height = self.__components.images.shape[1]
+            video_stream.pix_fmt = 'yuv420p'
+
+            # Create an audio stream
+            audio_sample_rate = 1
+            audio_stream: Optional[av.AudioStream] = None
+            if self.__components.audio:
+                audio_sample_rate = int(self.__components.audio['sample_rate'])
+                audio_stream = output.add_stream('aac', rate=audio_sample_rate)
+                audio_stream.sample_rate = audio_sample_rate
+                audio_stream.format = 'fltp'
+
+            # Encode video
+            for i, frame in enumerate(self.__components.images):
+                img = (frame * 255).clamp(0, 255).byte().cpu().numpy() # shape: (H, W, 3)
+                frame = av.VideoFrame.from_ndarray(img, format='rgb24')
+                frame = frame.reformat(format='yuv420p')  # Convert to YUV420P as required by h264
+                packet = video_stream.encode(frame)
+                output.mux(packet)
+
+            # Flush video
+            packet = video_stream.encode(None)
+            output.mux(packet)
+
+            if audio_stream and self.__components.audio:
+                # Encode audio
+                samples_per_frame = int(audio_sample_rate / frame_rate)
+                num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame
+                for i in range(num_frames):
+                    start = i * samples_per_frame
+                    end = start + samples_per_frame
+                    # TODO(Feature) - Add support for stereo audio
+                    chunk = (
+                        self.__components.audio["waveform"][0, 0, start:end]
+                        .unsqueeze(0)
+                        .contiguous()
+                        .numpy()
+                    )
+                    audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono')
+                    audio_frame.sample_rate = audio_sample_rate
+                    audio_frame.pts = i * samples_per_frame
+                    for packet in audio_stream.encode(audio_frame):
+                        output.mux(packet)
+
+                # Flush audio
+                for packet in audio_stream.encode(None):
+                    output.mux(packet)
+
+

comfy_api/latest/_resources.py

@@ -0,0 +1,72 @@
+from __future__ import annotations
+import comfy.utils
+import folder_paths
+import logging
+from abc import ABC, abstractmethod
+from typing import Any
+import torch
+
+class ResourceKey(ABC):
+    Type = Any
+    def __init__(self):
+        ...
+
+class TorchDictFolderFilename(ResourceKey):
+    '''Key for requesting a torch file via file_name from a folder category.'''
+    Type = dict[str, torch.Tensor]
+    def __init__(self, folder_name: str, file_name: str):
+        self.folder_name = folder_name
+        self.file_name = file_name
+
+    def __hash__(self):
+        return hash((self.folder_name, self.file_name))
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, TorchDictFolderFilename):
+            return False
+        return self.folder_name == other.folder_name and self.file_name == other.file_name
+
+    def __str__(self):
+        return f"{self.folder_name} -> {self.file_name}"
+
+class Resources(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def get(self, key: ResourceKey, default: Any=...) -> Any:
+        pass
+
+class ResourcesLocal(Resources):
+    def __init__(self):
+        super().__init__()
+        self.local_resources: dict[ResourceKey, Any] = {}
+
+    def get(self, key: ResourceKey, default: Any=...) -> Any:
+        cached = self.local_resources.get(key, None)
+        if cached is not None:
+            logging.info(f"Using cached resource '{key}'")
+            return cached
+        logging.info(f"Loading resource '{key}'")
+        to_return = None
+        if isinstance(key, TorchDictFolderFilename):
+            if default is ...:
+                to_return = comfy.utils.load_torch_file(folder_paths.get_full_path_or_raise(key.folder_name, key.file_name), safe_load=True)
+            else:
+                full_path = folder_paths.get_full_path(key.folder_name, key.file_name)
+                if full_path is not None:
+                    to_return = comfy.utils.load_torch_file(full_path, safe_load=True)
+
+        if to_return is not None:
+            self.local_resources[key] = to_return
+            return to_return
+        if default is not ...:
+            return default
+        raise Exception(f"Unsupported resource key type: {type(key)}")
+
+
+class _RESOURCES:
+    ResourceKey = ResourceKey
+    TorchDictFolderFilename = TorchDictFolderFilename
+    Resources = Resources
+    ResourcesLocal = ResourcesLocal

comfy_api/latest/_ui.py

@@ -0,0 +1,463 @@
+from __future__ import annotations
+
+import json
+import os
+import random
+from io import BytesIO
+from typing import Type
+
+import av
+import numpy as np
+import torch
+try:
+    import torchaudio
+    TORCH_AUDIO_AVAILABLE = True
+except:
+    TORCH_AUDIO_AVAILABLE = False
+from PIL import Image as PILImage
+from PIL.PngImagePlugin import PngInfo
+
+import folder_paths
+
+# used for image preview
+from comfy.cli_args import args
+from comfy_api.latest._io import ComfyNode, FolderType, Image, _UIOutput
+
+
+class SavedResult(dict):
+    def __init__(self, filename: str, subfolder: str, type: FolderType):
+        super().__init__(filename=filename, subfolder=subfolder,type=type.value)
+
+    @property
+    def filename(self) -> str:
+        return self["filename"]
+
+    @property
+    def subfolder(self) -> str:
+        return self["subfolder"]
+
+    @property
+    def type(self) -> FolderType:
+        return FolderType(self["type"])
+
+
+class SavedImages(_UIOutput):
+    """A UI output class to represent one or more saved images, potentially animated."""
+    def __init__(self, results: list[SavedResult], is_animated: bool = False):
+        super().__init__()
+        self.results = results
+        self.is_animated = is_animated
+
+    def as_dict(self) -> dict:
+        data = {"images": self.results}
+        if self.is_animated:
+            data["animated"] = (True,)
+        return data
+
+
+class SavedAudios(_UIOutput):
+    """UI wrapper around one or more audio files on disk (FLAC / MP3 / Opus)."""
+    def __init__(self, results: list[SavedResult]):
+        super().__init__()
+        self.results = results
+
+    def as_dict(self) -> dict:
+        return {"audio": self.results}
+
+
+def _get_directory_by_folder_type(folder_type: FolderType) -> str:
+    if folder_type == FolderType.input:
+        return folder_paths.get_input_directory()
+    if folder_type == FolderType.output:
+        return folder_paths.get_output_directory()
+    return folder_paths.get_temp_directory()
+
+
+class ImageSaveHelper:
+    """A helper class with static methods to handle image saving and metadata."""
+
+    @staticmethod
+    def _convert_tensor_to_pil(image_tensor: torch.Tensor) -> PILImage.Image:
+        """Converts a single torch tensor to a PIL Image."""
+        return PILImage.fromarray(np.clip(255.0 * image_tensor.cpu().numpy(), 0, 255).astype(np.uint8))
+
+    @staticmethod
+    def _create_png_metadata(cls: Type[ComfyNode] | None) -> PngInfo | None:
+        """Creates a PngInfo object with prompt and extra_pnginfo."""
+        if args.disable_metadata or cls is None or not cls.hidden:
+            return None
+        metadata = PngInfo()
+        if cls.hidden.prompt:
+            metadata.add_text("prompt", json.dumps(cls.hidden.prompt))
+        if cls.hidden.extra_pnginfo:
+            for x in cls.hidden.extra_pnginfo:
+                metadata.add_text(x, json.dumps(cls.hidden.extra_pnginfo[x]))
+        return metadata
+
+    @staticmethod
+    def _create_animated_png_metadata(cls: Type[ComfyNode] | None) -> PngInfo | None:
+        """Creates a PngInfo object with prompt and extra_pnginfo for animated PNGs (APNG)."""
+        if args.disable_metadata or cls is None or not cls.hidden:
+            return None
+        metadata = PngInfo()
+        if cls.hidden.prompt:
+            metadata.add(
+                b"comf",
+                "prompt".encode("latin-1", "strict")
+                + b"\0"
+                + json.dumps(cls.hidden.prompt).encode("latin-1", "strict"),
+                after_idat=True,
+            )
+        if cls.hidden.extra_pnginfo:
+            for x in cls.hidden.extra_pnginfo:
+                metadata.add(
+                    b"comf",
+                    x.encode("latin-1", "strict")
+                    + b"\0"
+                    + json.dumps(cls.hidden.extra_pnginfo[x]).encode("latin-1", "strict"),
+                    after_idat=True,
+                )
+        return metadata
+
+    @staticmethod
+    def _create_webp_metadata(pil_image: PILImage.Image, cls: Type[ComfyNode] | None) -> PILImage.Exif:
+        """Creates EXIF metadata bytes for WebP images."""
+        exif_data = pil_image.getexif()
+        if args.disable_metadata or cls is None or cls.hidden is None:
+            return exif_data
+        if cls.hidden.prompt is not None:
+            exif_data[0x0110] = "prompt:{}".format(json.dumps(cls.hidden.prompt))  # EXIF 0x0110 = Model
+        if cls.hidden.extra_pnginfo is not None:
+            inital_exif_tag = 0x010F  # EXIF 0x010f = Make
+            for key, value in cls.hidden.extra_pnginfo.items():
+                exif_data[inital_exif_tag] = "{}:{}".format(key, json.dumps(value))
+                inital_exif_tag -= 1
+        return exif_data
+
+    @staticmethod
+    def save_images(
+        images, filename_prefix: str, folder_type: FolderType, cls: Type[ComfyNode] | None, compress_level = 4,
+    ) -> list[SavedResult]:
+        """Saves a batch of images as individual PNG files."""
+        full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path(
+            filename_prefix, _get_directory_by_folder_type(folder_type), images[0].shape[1], images[0].shape[0]
+        )
+        results = []
+        metadata = ImageSaveHelper._create_png_metadata(cls)
+        for batch_number, image_tensor in enumerate(images):
+            img = ImageSaveHelper._convert_tensor_to_pil(image_tensor)
+            filename_with_batch_num = filename.replace("%batch_num%", str(batch_number))
+            file = f"{filename_with_batch_num}_{counter:05}_.png"
+            img.save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=compress_level)
+            results.append(SavedResult(file, subfolder, folder_type))
+            counter += 1
+        return results
+
+    @staticmethod
+    def get_save_images_ui(images, filename_prefix: str, cls: Type[ComfyNode] | None, compress_level=4) -> SavedImages:
+        """Saves a batch of images and returns a UI object for the node output."""
+        return SavedImages(
+                ImageSaveHelper.save_images(
+                images,
+                filename_prefix=filename_prefix,
+                folder_type=FolderType.output,
+                cls=cls,
+                compress_level=compress_level,
+            )
+        )
+
+    @staticmethod
+    def save_animated_png(
+        images, filename_prefix: str, folder_type: FolderType, cls: Type[ComfyNode] | None, fps: float, compress_level: int
+    ) -> SavedResult:
+        """Saves a batch of images as a single animated PNG."""
+        full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path(
+            filename_prefix, _get_directory_by_folder_type(folder_type), images[0].shape[1], images[0].shape[0]
+        )
+        pil_images = [ImageSaveHelper._convert_tensor_to_pil(img) for img in images]
+        metadata = ImageSaveHelper._create_animated_png_metadata(cls)
+        file = f"{filename}_{counter:05}_.png"
+        save_path = os.path.join(full_output_folder, file)
+        pil_images[0].save(
+            save_path,
+            pnginfo=metadata,
+            compress_level=compress_level,
+            save_all=True,
+            duration=int(1000.0 / fps),
+            append_images=pil_images[1:],
+        )
+        return SavedResult(file, subfolder, folder_type)
+
+    @staticmethod
+    def get_save_animated_png_ui(
+        images, filename_prefix: str, cls: Type[ComfyNode] | None, fps: float, compress_level: int
+    ) -> SavedImages:
+        """Saves an animated PNG and returns a UI object for the node output."""
+        result = ImageSaveHelper.save_animated_png(
+            images,
+            filename_prefix=filename_prefix,
+            folder_type=FolderType.output,
+            cls=cls,
+            fps=fps,
+            compress_level=compress_level,
+        )
+        return SavedImages([result], is_animated=len(images) > 1)
+
+    @staticmethod
+    def save_animated_webp(
+        images,
+        filename_prefix: str,
+        folder_type: FolderType,
+        cls: Type[ComfyNode] | None,
+        fps: float,
+        lossless: bool,
+        quality: int,
+        method: int,
+    ) -> SavedResult:
+        """Saves a batch of images as a single animated WebP."""
+        full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path(
+            filename_prefix, _get_directory_by_folder_type(folder_type), images[0].shape[1], images[0].shape[0]
+        )
+        pil_images = [ImageSaveHelper._convert_tensor_to_pil(img) for img in images]
+        pil_exif = ImageSaveHelper._create_webp_metadata(pil_images[0], cls)
+        file = f"{filename}_{counter:05}_.webp"
+        pil_images[0].save(
+            os.path.join(full_output_folder, file),
+            save_all=True,
+            duration=int(1000.0 / fps),
+            append_images=pil_images[1:],
+            exif=pil_exif,
+            lossless=lossless,
+            quality=quality,
+            method=method,
+        )
+        return SavedResult(file, subfolder, folder_type)
+
+    @staticmethod
+    def get_save_animated_webp_ui(
+        images,
+        filename_prefix: str,
+        cls: Type[ComfyNode] | None,
+        fps: float,
+        lossless: bool,
+        quality: int,
+        method: int,
+    ) -> SavedImages:
+        """Saves an animated WebP and returns a UI object for the node output."""
+        result = ImageSaveHelper.save_animated_webp(
+            images,
+            filename_prefix=filename_prefix,
+            folder_type=FolderType.output,
+            cls=cls,
+            fps=fps,
+            lossless=lossless,
+            quality=quality,
+            method=method,
+        )
+        return SavedImages([result], is_animated=len(images) > 1)
+
+
+class AudioSaveHelper:
+    """A helper class with static methods to handle audio saving and metadata."""
+    _OPUS_RATES = [8000, 12000, 16000, 24000, 48000]
+
+    @staticmethod
+    def save_audio(
+        audio: dict,
+        filename_prefix: str,
+        folder_type: FolderType,
+        cls: Type[ComfyNode] | None,
+        format: str = "flac",
+        quality: str = "128k",
+    ) -> list[SavedResult]:
+        full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path(
+            filename_prefix, _get_directory_by_folder_type(folder_type)
+        )
+
+        metadata = {}
+        if not args.disable_metadata and cls is not None:
+            if cls.hidden.prompt is not None:
+                metadata["prompt"] = json.dumps(cls.hidden.prompt)
+            if cls.hidden.extra_pnginfo is not None:
+                for x in cls.hidden.extra_pnginfo:
+                    metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x])
+
+        results = []
+        for batch_number, waveform in enumerate(audio["waveform"].cpu()):
+            filename_with_batch_num = filename.replace("%batch_num%", str(batch_number))
+            file = f"{filename_with_batch_num}_{counter:05}_.{format}"
+            output_path = os.path.join(full_output_folder, file)
+
+            # Use original sample rate initially
+            sample_rate = audio["sample_rate"]
+
+            # Handle Opus sample rate requirements
+            if format == "opus":
+                if sample_rate > 48000:
+                    sample_rate = 48000
+                elif sample_rate not in AudioSaveHelper._OPUS_RATES:
+                    # Find the next highest supported rate
+                    for rate in sorted(AudioSaveHelper._OPUS_RATES):
+                        if rate > sample_rate:
+                            sample_rate = rate
+                            break
+                    if sample_rate not in AudioSaveHelper._OPUS_RATES:  # Fallback if still not supported
+                        sample_rate = 48000
+
+                # Resample if necessary
+                if sample_rate != audio["sample_rate"]:
+                    if not TORCH_AUDIO_AVAILABLE:
+                        raise Exception("torchaudio is not available; cannot resample audio.")
+                    waveform = torchaudio.functional.resample(waveform, audio["sample_rate"], sample_rate)
+
+            # Create output with specified format
+            output_buffer = BytesIO()
+            output_container = av.open(output_buffer, mode="w", format=format)
+
+            # Set metadata on the container
+            for key, value in metadata.items():
+                output_container.metadata[key] = value
+
+            # Set up the output stream with appropriate properties
+            if format == "opus":
+                out_stream = output_container.add_stream("libopus", rate=sample_rate)
+                if quality == "64k":
+                    out_stream.bit_rate = 64000
+                elif quality == "96k":
+                    out_stream.bit_rate = 96000
+                elif quality == "128k":
+                    out_stream.bit_rate = 128000
+                elif quality == "192k":
+                    out_stream.bit_rate = 192000
+                elif quality == "320k":
+                    out_stream.bit_rate = 320000
+            elif format == "mp3":
+                out_stream = output_container.add_stream("libmp3lame", rate=sample_rate)
+                if quality == "V0":
+                    # TODO i would really love to support V3 and V5 but there doesn't seem to be a way to set the qscale level, the property below is a bool
+                    out_stream.codec_context.qscale = 1
+                elif quality == "128k":
+                    out_stream.bit_rate = 128000
+                elif quality == "320k":
+                    out_stream.bit_rate = 320000
+            else:  # format == "flac":
+                out_stream = output_container.add_stream("flac", rate=sample_rate)
+
+            frame = av.AudioFrame.from_ndarray(
+                waveform.movedim(0, 1).reshape(1, -1).float().numpy(),
+                format="flt",
+                layout="mono" if waveform.shape[0] == 1 else "stereo",
+            )
+            frame.sample_rate = sample_rate
+            frame.pts = 0
+            output_container.mux(out_stream.encode(frame))
+
+            # Flush encoder
+            output_container.mux(out_stream.encode(None))
+
+            # Close containers
+            output_container.close()
+
+            # Write the output to file
+            output_buffer.seek(0)
+            with open(output_path, "wb") as f:
+                f.write(output_buffer.getbuffer())
+
+            results.append(SavedResult(file, subfolder, folder_type))
+            counter += 1
+
+        return results
+
+    @staticmethod
+    def get_save_audio_ui(
+        audio, filename_prefix: str, cls: Type[ComfyNode] | None, format: str = "flac", quality: str = "128k",
+    ) -> SavedAudios:
+        """Save and instantly wrap for UI."""
+        return SavedAudios(
+            AudioSaveHelper.save_audio(
+                audio,
+                filename_prefix=filename_prefix,
+                folder_type=FolderType.output,
+                cls=cls,
+                format=format,
+                quality=quality,
+            )
+        )
+
+
+class PreviewImage(_UIOutput):
+    def __init__(self, image: Image.Type, animated: bool = False, cls: Type[ComfyNode] = None, **kwargs):
+        self.values = ImageSaveHelper.save_images(
+            image,
+            filename_prefix="ComfyUI_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for _ in range(5)),
+            folder_type=FolderType.temp,
+            cls=cls,
+            compress_level=1,
+        )
+        self.animated = animated
+
+    def as_dict(self):
+        return {
+            "images": self.values,
+            "animated": (self.animated,)
+        }
+
+
+class PreviewMask(PreviewImage):
+    def __init__(self, mask: PreviewMask.Type, animated: bool=False, cls: ComfyNode=None, **kwargs):
+        preview = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3)
+        super().__init__(preview, animated, cls, **kwargs)
+
+
+class PreviewAudio(_UIOutput):
+    def __init__(self, audio: dict, cls: Type[ComfyNode] = None, **kwargs):
+        self.values = AudioSaveHelper.save_audio(
+            audio,
+            filename_prefix="ComfyUI_temp_" + "".join(random.choice("abcdefghijklmnopqrstuvwxyz") for _ in range(5)),
+            folder_type=FolderType.temp,
+            cls=cls,
+            format="flac",
+            quality="128k",
+        )
+
+    def as_dict(self) -> dict:
+        return {"audio": self.values}
+
+
+class PreviewVideo(_UIOutput):
+    def __init__(self, values: list[SavedResult | dict], **kwargs):
+        self.values = values
+
+    def as_dict(self):
+        return {"images": self.values, "animated": (True,)}
+
+
+class PreviewUI3D(_UIOutput):
+    def __init__(self, model_file, camera_info, **kwargs):
+        self.model_file = model_file
+        self.camera_info = camera_info
+
+    def as_dict(self):
+        return {"result": [self.model_file, self.camera_info]}
+
+
+class PreviewText(_UIOutput):
+    def __init__(self, value: str, **kwargs):
+        self.value = value
+
+    def as_dict(self):
+        return {"text": (self.value,)}
+
+
+class _UI:
+    SavedResult = SavedResult
+    SavedImages = SavedImages
+    SavedAudios = SavedAudios
+    ImageSaveHelper = ImageSaveHelper
+    AudioSaveHelper = AudioSaveHelper
+    PreviewImage = PreviewImage
+    PreviewMask = PreviewMask
+    PreviewAudio = PreviewAudio
+    PreviewVideo = PreviewVideo
+    PreviewUI3D = PreviewUI3D
+    PreviewText = PreviewText

comfy_api/latest/_util/__init__.py

@@ -0,0 +1,8 @@
+from .video_types import VideoContainer, VideoCodec, VideoComponents
+
+__all__ = [
+    # Utility Types
+    "VideoContainer",
+    "VideoCodec",
+    "VideoComponents",
+]

comfy_api/latest/_util/video_types.py

@@ -0,0 +1,52 @@
+from __future__ import annotations
+from dataclasses import dataclass
+from enum import Enum
+from fractions import Fraction
+from typing import Optional
+from comfy_api.latest._input import ImageInput, AudioInput
+
+class VideoCodec(str, Enum):
+    AUTO = "auto"
+    H264 = "h264"
+
+    @classmethod
+    def as_input(cls) -> list[str]:
+        """
+        Returns a list of codec names that can be used as node input.
+        """
+        return [member.value for member in cls]
+
+class VideoContainer(str, Enum):
+    AUTO = "auto"
+    MP4 = "mp4"
+
+    @classmethod
+    def as_input(cls) -> list[str]:
+        """
+        Returns a list of container names that can be used as node input.
+        """
+        return [member.value for member in cls]
+
+    @classmethod
+    def get_extension(cls, value) -> str:
+        """
+        Returns the file extension for the container.
+        """
+        if isinstance(value, str):
+            value = cls(value)
+        if value == VideoContainer.MP4 or value == VideoContainer.AUTO:
+            return "mp4"
+        return ""
+
+@dataclass
+class VideoComponents:
+    """
+    Dataclass representing the components of a video.
+    """
+
+    images: ImageInput
+    frame_rate: Fraction
+    audio: Optional[AudioInput] = None
+    metadata: Optional[dict] = None
+
+

comfy_api/latest/generated/ComfyAPISyncStub.pyi

@@ -0,0 +1,20 @@
+from typing import Any, Dict, List, Optional, Tuple, Union, Set, Sequence, cast, NamedTuple
+from comfy_api.latest import ComfyAPI_latest
+from PIL.Image import Image
+from torch import Tensor
+class ComfyAPISyncStub:
+    def __init__(self) -> None: ...
+
+    class ExecutionSync:
+        def __init__(self) -> None: ...
+        """
+        Update the progress bar displayed in the ComfyUI interface.
+
+        This function allows custom nodes and API calls to report their progress
+        back to the user interface, providing visual feedback during long operations.
+
+        Migration from previous API: comfy.utils.PROGRESS_BAR_HOOK
+        """
+        def set_progress(self, value: float, max_value: float, node_id: Union[str, None] = None, preview_image: Union[Image, Tensor, None] = None, ignore_size_limit: bool = False) -> None: ...
+
+    execution: ExecutionSync

comfy_api/util.py

@@ -0,0 +1,8 @@
+# This file only exists for backwards compatibility.
+from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents
+
+__all__ = [
+    "VideoCodec",
+    "VideoContainer",
+    "VideoComponents",
+]

comfy_api/util/__init__.py

@@ -1,7 +1,7 @@
-from .video_types import VideoContainer, VideoCodec, VideoComponents
+# This file only exists for backwards compatibility.
+from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents
 
 __all__ = [
-    # Utility Types
     "VideoContainer",
     "VideoCodec",
     "VideoComponents",

comfy_api/util/video_types.py

@@ -1,51 +1,12 @@
-from __future__ import annotations
-from dataclasses import dataclass
-from enum import Enum
-from fractions import Fraction
-from typing import Optional
-from comfy_api.input import ImageInput, AudioInput
-
-class VideoCodec(str, Enum):
-    AUTO = "auto"
-    H264 = "h264"
-
-    @classmethod
-    def as_input(cls) -> list[str]:
-        """
-        Returns a list of codec names that can be used as node input.
-        """
-        return [member.value for member in cls]
-
-class VideoContainer(str, Enum):
-    AUTO = "auto"
-    MP4 = "mp4"
-
-    @classmethod
-    def as_input(cls) -> list[str]:
-        """
-        Returns a list of container names that can be used as node input.
-        """
-        return [member.value for member in cls]
-
-    @classmethod
-    def get_extension(cls, value) -> str:
-        """
-        Returns the file extension for the container.
-        """
-        if isinstance(value, str):
-            value = cls(value)
-        if value == VideoContainer.MP4 or value == VideoContainer.AUTO:
-            return "mp4"
-        return ""
-
-@dataclass
-class VideoComponents:
-    """
-    Dataclass representing the components of a video.
-    """
-
-    images: ImageInput
-    frame_rate: Fraction
-    audio: Optional[AudioInput] = None
-    metadata: Optional[dict] = None
+# This file only exists for backwards compatibility.
+from comfy_api.latest._util.video_types import (
+    VideoContainer,
+    VideoCodec,
+    VideoComponents,
+)
 
+__all__ = [
+    "VideoContainer",
+    "VideoCodec",
+    "VideoComponents",
+]

comfy_api/v0_0_1/__init__.py

@@ -0,0 +1,42 @@
+from comfy_api.v0_0_2 import (
+    ComfyAPIAdapter_v0_0_2,
+    Input as Input_v0_0_2,
+    InputImpl as InputImpl_v0_0_2,
+    Types as Types_v0_0_2,
+)
+from typing import Type, TYPE_CHECKING
+from comfy_api.internal.async_to_sync import create_sync_class
+
+
+# This version only exists to serve as a template for future version adapters.
+# There is no reason anyone should ever use it.
+class ComfyAPIAdapter_v0_0_1(ComfyAPIAdapter_v0_0_2):
+    VERSION = "0.0.1"
+    STABLE = True
+
+class Input(Input_v0_0_2):
+    pass
+
+class InputImpl(InputImpl_v0_0_2):
+    pass
+
+class Types(Types_v0_0_2):
+    pass
+
+ComfyAPI = ComfyAPIAdapter_v0_0_1
+
+# Create a synchronous version of the API
+if TYPE_CHECKING:
+    from comfy_api.v0_0_1.generated.ComfyAPISyncStub import ComfyAPISyncStub  # type: ignore
+
+    ComfyAPISync: Type[ComfyAPISyncStub]
+
+ComfyAPISync = create_sync_class(ComfyAPIAdapter_v0_0_1)
+
+__all__ = [
+    "ComfyAPI",
+    "ComfyAPISync",
+    "Input",
+    "InputImpl",
+    "Types",
+]

comfy_api/v0_0_1/generated/ComfyAPISyncStub.pyi

@@ -0,0 +1,20 @@
+from typing import Any, Dict, List, Optional, Tuple, Union, Set, Sequence, cast, NamedTuple
+from comfy_api.v0_0_1 import ComfyAPIAdapter_v0_0_1
+from PIL.Image import Image
+from torch import Tensor
+class ComfyAPISyncStub:
+    def __init__(self) -> None: ...
+
+    class ExecutionSync:
+        def __init__(self) -> None: ...
+        """
+        Update the progress bar displayed in the ComfyUI interface.
+
+        This function allows custom nodes and API calls to report their progress
+        back to the user interface, providing visual feedback during long operations.
+
+        Migration from previous API: comfy.utils.PROGRESS_BAR_HOOK
+        """
+        def set_progress(self, value: float, max_value: float, node_id: Union[str, None] = None, preview_image: Union[Image, Tensor, None] = None, ignore_size_limit: bool = False) -> None: ...
+
+    execution: ExecutionSync

comfy_api/v0_0_2/__init__.py

@@ -0,0 +1,45 @@
+from comfy_api.latest import (
+    ComfyAPI_latest,
+    Input as Input_latest,
+    InputImpl as InputImpl_latest,
+    Types as Types_latest,
+)
+from typing import Type, TYPE_CHECKING
+from comfy_api.internal.async_to_sync import create_sync_class
+from comfy_api.latest import io, ui, ComfyExtension  #noqa: F401
+
+
+class ComfyAPIAdapter_v0_0_2(ComfyAPI_latest):
+    VERSION = "0.0.2"
+    STABLE = False
+
+
+class Input(Input_latest):
+    pass
+
+
+class InputImpl(InputImpl_latest):
+    pass
+
+
+class Types(Types_latest):
+    pass
+
+
+ComfyAPI = ComfyAPIAdapter_v0_0_2
+
+# Create a synchronous version of the API
+if TYPE_CHECKING:
+    from comfy_api.v0_0_2.generated.ComfyAPISyncStub import ComfyAPISyncStub  # type: ignore
+
+    ComfyAPISync: Type[ComfyAPISyncStub]
+ComfyAPISync = create_sync_class(ComfyAPIAdapter_v0_0_2)
+
+__all__ = [
+    "ComfyAPI",
+    "ComfyAPISync",
+    "Input",
+    "InputImpl",
+    "Types",
+    "ComfyExtension",
+]

comfy_api/v0_0_2/generated/ComfyAPISyncStub.pyi

@@ -0,0 +1,20 @@
+from typing import Any, Dict, List, Optional, Tuple, Union, Set, Sequence, cast, NamedTuple
+from comfy_api.v0_0_2 import ComfyAPIAdapter_v0_0_2
+from PIL.Image import Image
+from torch import Tensor
+class ComfyAPISyncStub:
+    def __init__(self) -> None: ...
+
+    class ExecutionSync:
+        def __init__(self) -> None: ...
+        """
+        Update the progress bar displayed in the ComfyUI interface.
+
+        This function allows custom nodes and API calls to report their progress
+        back to the user interface, providing visual feedback during long operations.
+
+        Migration from previous API: comfy.utils.PROGRESS_BAR_HOOK
+        """
+        def set_progress(self, value: float, max_value: float, node_id: Union[str, None] = None, preview_image: Union[Image, Tensor, None] = None, ignore_size_limit: bool = False) -> None: ...
+
+    execution: ExecutionSync

comfy_api/version_list.py

@@ -0,0 +1,12 @@
+from comfy_api.latest import ComfyAPI_latest
+from comfy_api.v0_0_2 import ComfyAPIAdapter_v0_0_2
+from comfy_api.v0_0_1 import ComfyAPIAdapter_v0_0_1
+from comfy_api.internal import ComfyAPIBase
+from typing import List, Type
+
+supported_versions: List[Type[ComfyAPIBase]] = [
+    ComfyAPI_latest,
+    ComfyAPIAdapter_v0_0_2,
+    ComfyAPIAdapter_v0_0_1,
+]
+

comfy_api_nodes/README.md

@@ -2,7 +2,7 @@
 
 ## Introduction 
 
-Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview#api-nodes).
+Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview).
 
 ## Development
 

comfy_api_nodes/apinode_utils.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+import aiohttp
 import io
 import logging
 import mimetypes
@@ -21,7 +22,6 @@ from server import PromptServer
 
 import numpy as np
 from PIL import Image
-import requests
 import torch
 import math
 import base64
@@ -30,7 +30,7 @@ from io import BytesIO
 import av
 
 
-def download_url_to_video_output(video_url: str, timeout: int = None) -> VideoFromFile:
+async def download_url_to_video_output(video_url: str, timeout: int = None) -> VideoFromFile:
     """Downloads a video from a URL and returns a `VIDEO` output.
 
     Args:
@@ -39,7 +39,7 @@ def download_url_to_video_output(video_url: str, timeout: int = None) -> VideoFr
     Returns:
         A Comfy node `VIDEO` output.
     """
-    video_io = download_url_to_bytesio(video_url, timeout)
+    video_io = await download_url_to_bytesio(video_url, timeout)
     if video_io is None:
         error_msg = f"Failed to download video from {video_url}"
         logging.error(error_msg)
@@ -62,7 +62,7 @@ def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor:
     return s
 
 
-def validate_and_cast_response(
+async def validate_and_cast_response(
     response, timeout: int = None, node_id: Union[str, None] = None
 ) -> torch.Tensor:
     """Validates and casts a response to a torch.Tensor.
@@ -86,35 +86,24 @@ def validate_and_cast_response(
     image_tensors: list[torch.Tensor] = []
 
     # Process each image in the data array
-    for image_data in data:
-        image_url = image_data.url
-        b64_data = image_data.b64_json
+    async with aiohttp.ClientSession(timeout=aiohttp.ClientTimeout(total=timeout)) as session:
+        for img_data in data:
+            img_bytes: bytes
+            if img_data.b64_json:
+                img_bytes = base64.b64decode(img_data.b64_json)
+            elif img_data.url:
+                if node_id:
+                    PromptServer.instance.send_progress_text(f"Result URL: {img_data.url}", node_id)
+                async with session.get(img_data.url) as resp:
+                    if resp.status != 200:
+                        raise ValueError("Failed to download generated image")
+                    img_bytes = await resp.read()
+            else:
+                raise ValueError("Invalid image payload – neither URL nor base64 data present.")
 
-        if not image_url and not b64_data:
-            raise ValueError("No image was generated in the response")
-
-        if b64_data:
-            img_data = base64.b64decode(b64_data)
-            img = Image.open(io.BytesIO(img_data))
-
-        elif image_url:
-            if node_id:
-                PromptServer.instance.send_progress_text(
-                    f"Result URL: {image_url}", node_id
-                )
-            img_response = requests.get(image_url, timeout=timeout)
-            if img_response.status_code != 200:
-                raise ValueError("Failed to download the image")
-            img = Image.open(io.BytesIO(img_response.content))
-
-        img = img.convert("RGBA")
-
-        # Convert to numpy array, normalize to float32 between 0 and 1
-        img_array = np.array(img).astype(np.float32) / 255.0
-        img_tensor = torch.from_numpy(img_array)
-
-        # Add to list of tensors
-        image_tensors.append(img_tensor)
+            pil_img = Image.open(BytesIO(img_bytes)).convert("RGBA")
+            arr = np.asarray(pil_img).astype(np.float32) / 255.0
+            image_tensors.append(torch.from_numpy(arr))
 
     return torch.stack(image_tensors, dim=0)
 
@@ -175,7 +164,7 @@ def mimetype_to_extension(mime_type: str) -> str:
     return mime_type.split("/")[-1].lower()
 
 
-def download_url_to_bytesio(url: str, timeout: int = None) -> BytesIO:
+async def download_url_to_bytesio(url: str, timeout: int = None) -> BytesIO:
     """Downloads content from a URL using requests and returns it as BytesIO.
 
     Args:
@@ -185,9 +174,11 @@ def download_url_to_bytesio(url: str, timeout: int = None) -> BytesIO:
     Returns:
         BytesIO object containing the downloaded content.
     """
-    response = requests.get(url, stream=True, timeout=timeout)
-    response.raise_for_status()  # Raises HTTPError for bad responses (4XX or 5XX)
-    return BytesIO(response.content)
+    timeout_cfg = aiohttp.ClientTimeout(total=timeout) if timeout else None
+    async with aiohttp.ClientSession(timeout=timeout_cfg) as session:
+        async with session.get(url) as resp:
+            resp.raise_for_status()  # Raises HTTPError for bad responses (4XX or 5XX)
+            return BytesIO(await resp.read())
 
 
 def bytesio_to_image_tensor(image_bytesio: BytesIO, mode: str = "RGBA") -> torch.Tensor:
@@ -210,15 +201,15 @@ def bytesio_to_image_tensor(image_bytesio: BytesIO, mode: str = "RGBA") -> torch
     return torch.from_numpy(image_array).unsqueeze(0)
 
 
-def download_url_to_image_tensor(url: str, timeout: int = None) -> torch.Tensor:
+async def download_url_to_image_tensor(url: str, timeout: int = None) -> torch.Tensor:
     """Downloads an image from a URL and returns a [B, H, W, C] tensor."""
-    image_bytesio = download_url_to_bytesio(url, timeout)
+    image_bytesio = await download_url_to_bytesio(url, timeout)
     return bytesio_to_image_tensor(image_bytesio)
 
 
-def process_image_response(response: requests.Response) -> torch.Tensor:
+def process_image_response(response_content: bytes | str) -> torch.Tensor:
     """Uses content from a Response object and converts it to a torch.Tensor"""
-    return bytesio_to_image_tensor(BytesIO(response.content))
+    return bytesio_to_image_tensor(BytesIO(response_content))
 
 
 def _tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image.Image:
@@ -336,10 +327,10 @@ def text_filepath_to_data_uri(filepath: str) -> str:
     return f"data:{mime_type};base64,{base64_string}"
 
 
-def upload_file_to_comfyapi(
+async def upload_file_to_comfyapi(
     file_bytes_io: BytesIO,
     filename: str,
-    upload_mime_type: str,
+    upload_mime_type: Optional[str],
     auth_kwargs: Optional[dict[str, str]] = None,
 ) -> str:
     """
@@ -354,7 +345,10 @@ def upload_file_to_comfyapi(
     Returns:
         The download URL for the uploaded file.
     """
-    request_object = UploadRequest(file_name=filename, content_type=upload_mime_type)
+    if upload_mime_type is None:
+        request_object = UploadRequest(file_name=filename)
+    else:
+        request_object = UploadRequest(file_name=filename, content_type=upload_mime_type)
     operation = SynchronousOperation(
         endpoint=ApiEndpoint(
             path="/customers/storage",
@@ -366,12 +360,8 @@ def upload_file_to_comfyapi(
         auth_kwargs=auth_kwargs,
     )
 
-    response: UploadResponse = operation.execute()
-    upload_response = ApiClient.upload_file(
-        response.upload_url, file_bytes_io, content_type=upload_mime_type
-    )
-    upload_response.raise_for_status()
-
+    response: UploadResponse = await operation.execute()
+    await ApiClient.upload_file(response.upload_url, file_bytes_io, content_type=upload_mime_type)
     return response.download_url
 
 
@@ -399,7 +389,7 @@ def video_to_base64_string(
     return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8")
 
 
-def upload_video_to_comfyapi(
+async def upload_video_to_comfyapi(
     video: VideoInput,
     auth_kwargs: Optional[dict[str, str]] = None,
     container: VideoContainer = VideoContainer.MP4,
@@ -439,9 +429,7 @@ def upload_video_to_comfyapi(
     video.save_to(video_bytes_io, format=container, codec=codec)
     video_bytes_io.seek(0)
 
-    return upload_file_to_comfyapi(
-        video_bytes_io, filename, upload_mime_type, auth_kwargs
-    )
+    return await upload_file_to_comfyapi(video_bytes_io, filename, upload_mime_type, auth_kwargs)
 
 
 def audio_tensor_to_contiguous_ndarray(waveform: torch.Tensor) -> np.ndarray:
@@ -501,7 +489,7 @@ def audio_ndarray_to_bytesio(
     return audio_bytes_io
 
 
-def upload_audio_to_comfyapi(
+async def upload_audio_to_comfyapi(
     audio: AudioInput,
     auth_kwargs: Optional[dict[str, str]] = None,
     container_format: str = "mp4",
@@ -527,7 +515,7 @@ def upload_audio_to_comfyapi(
         audio_data_np, sample_rate, container_format, codec_name
     )
 
-    return upload_file_to_comfyapi(audio_bytes_io, filename, mime_type, auth_kwargs)
+    return await upload_file_to_comfyapi(audio_bytes_io, filename, mime_type, auth_kwargs)
 
 
 def audio_to_base64_string(
@@ -544,7 +532,7 @@ def audio_to_base64_string(
     return base64.b64encode(audio_bytes).decode("utf-8")
 
 
-def upload_images_to_comfyapi(
+async def upload_images_to_comfyapi(
     image: torch.Tensor,
     max_images=8,
     auth_kwargs: Optional[dict[str, str]] = None,
@@ -561,55 +549,15 @@ def upload_images_to_comfyapi(
         mime_type: Optional MIME type for the image.
     """
     # if batch, try to upload each file if max_images is greater than 0
-    idx_image = 0
     download_urls: list[str] = []
     is_batch = len(image.shape) > 3
-    batch_length = 1
-    if is_batch:
-        batch_length = image.shape[0]
-    while True:
-        curr_image = image
-        if len(image.shape) > 3:
-            curr_image = image[idx_image]
-        # get BytesIO version of image
-        img_binary = tensor_to_bytesio(curr_image, mime_type=mime_type)
-        # first, request upload/download urls from comfy API
-        if not mime_type:
-            request_object = UploadRequest(file_name=img_binary.name)
-        else:
-            request_object = UploadRequest(
-                file_name=img_binary.name, content_type=mime_type
-            )
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/customers/storage",
-                method=HttpMethod.POST,
-                request_model=UploadRequest,
-                response_model=UploadResponse,
-            ),
-            request=request_object,
-            auth_kwargs=auth_kwargs,
-        )
-        response = operation.execute()
+    batch_len = image.shape[0] if is_batch else 1
 
-        upload_response = ApiClient.upload_file(
-            response.upload_url, img_binary, content_type=mime_type
-        )
-        # verify success
-        try:
-            upload_response.raise_for_status()
-        except requests.exceptions.HTTPError as e:
-            raise ValueError(f"Could not upload one or more images: {e}") from e
-        # add download_url to list
-        download_urls.append(response.download_url)
-
-        idx_image += 1
-        # stop uploading additional files if done
-        if is_batch and max_images > 0:
-            if idx_image >= max_images:
-                break
-            if idx_image >= batch_length:
-                break
+    for idx in range(min(batch_len, max_images)):
+        tensor = image[idx] if is_batch else image
+        img_io = tensor_to_bytesio(tensor, mime_type=mime_type)
+        url = await upload_file_to_comfyapi(img_io, img_io.name, mime_type, auth_kwargs)
+        download_urls.append(url)
     return download_urls
 
 

comfy_api_nodes/apis/request_logger.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 import os
 import datetime
 import json

comfy_api_nodes/apis/tripo_api.py

@@ -127,7 +127,7 @@ class TripoTextToModelRequest(BaseModel):
     type: TripoTaskType = Field(TripoTaskType.TEXT_TO_MODEL, description='Type of task')
     prompt: str = Field(..., description='The text prompt describing the model to generate', max_length=1024)
     negative_prompt: Optional[str] = Field(None, description='The negative text prompt', max_length=1024)
-    model_version: Optional[TripoModelVersion] = TripoModelVersion.V2_5
+    model_version: Optional[TripoModelVersion] = TripoModelVersion.v2_5_20250123
     face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to')
     texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model')
     pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model')

comfy_api_nodes/nodes_bfl.py

@@ -1,3 +1,4 @@
+import asyncio
 import io
 from inspect import cleandoc
 from typing import Union, Optional
@@ -28,7 +29,7 @@ from comfy_api_nodes.apinode_utils import (
 
 import numpy as np
 from PIL import Image
-import requests
+import aiohttp
 import torch
 import base64
 import time
@@ -44,18 +45,18 @@ def convert_mask_to_image(mask: torch.Tensor):
     return mask
 
 
-def handle_bfl_synchronous_operation(
+async def handle_bfl_synchronous_operation(
     operation: SynchronousOperation,
     timeout_bfl_calls=360,
     node_id: Union[str, None] = None,
 ):
-    response_api: BFLFluxProGenerateResponse = operation.execute()
-    return _poll_until_generated(
+    response_api: BFLFluxProGenerateResponse = await operation.execute()
+    return await _poll_until_generated(
         response_api.polling_url, timeout=timeout_bfl_calls, node_id=node_id
     )
 
 
-def _poll_until_generated(
+async def _poll_until_generated(
     polling_url: str, timeout=360, node_id: Union[str, None] = None
 ):
     # used bfl-comfy-nodes to verify code implementation:
@@ -66,55 +67,56 @@ def _poll_until_generated(
     retry_404_seconds = 2
     retry_202_seconds = 2
     retry_pending_seconds = 1
-    request = requests.Request(method=HttpMethod.GET, url=polling_url)
-    # NOTE: should True loop be replaced with checking if workflow has been interrupted?
-    while True:
-        if node_id:
-            time_elapsed = time.time() - start_time
-            PromptServer.instance.send_progress_text(
-                f"Generating ({time_elapsed:.0f}s)", node_id
-            )
 
-        response = requests.Session().send(request.prepare())
-        if response.status_code == 200:
-            result = response.json()
-            if result["status"] == BFLStatus.ready:
-                img_url = result["result"]["sample"]
-                if node_id:
-                    PromptServer.instance.send_progress_text(
-                        f"Result URL: {img_url}", node_id
-                    )
-                img_response = requests.get(img_url)
-                return process_image_response(img_response)
-            elif result["status"] in [
-                BFLStatus.request_moderated,
-                BFLStatus.content_moderated,
-            ]:
-                status = result["status"]
-                raise Exception(
-                    f"BFL API did not return an image due to: {status}."
+    async with aiohttp.ClientSession() as session:
+        # NOTE: should True loop be replaced with checking if workflow has been interrupted?
+        while True:
+            if node_id:
+                time_elapsed = time.time() - start_time
+                PromptServer.instance.send_progress_text(
+                    f"Generating ({time_elapsed:.0f}s)", node_id
                 )
-            elif result["status"] == BFLStatus.error:
-                raise Exception(f"BFL API encountered an error: {result}.")
-            elif result["status"] == BFLStatus.pending:
-                time.sleep(retry_pending_seconds)
-                continue
-        elif response.status_code == 404:
-            if retries_404 < max_retries_404:
-                retries_404 += 1
-                time.sleep(retry_404_seconds)
-                continue
-            raise Exception(
-                f"BFL API could not find task after {max_retries_404} tries."
-            )
-        elif response.status_code == 202:
-            time.sleep(retry_202_seconds)
-        elif time.time() - start_time > timeout:
-            raise Exception(
-                f"BFL API experienced a timeout; could not return request under {timeout} seconds."
-            )
-        else:
-            raise Exception(f"BFL API encountered an error: {response.json()}")
+
+            async with session.get(polling_url) as response:
+                if response.status == 200:
+                    result = await response.json()
+                    if result["status"] == BFLStatus.ready:
+                        img_url = result["result"]["sample"]
+                        if node_id:
+                            PromptServer.instance.send_progress_text(
+                                f"Result URL: {img_url}", node_id
+                            )
+                        async with session.get(img_url) as img_resp:
+                            return process_image_response(await img_resp.content.read())
+                    elif result["status"] in [
+                        BFLStatus.request_moderated,
+                        BFLStatus.content_moderated,
+                    ]:
+                        status = result["status"]
+                        raise Exception(
+                            f"BFL API did not return an image due to: {status}."
+                        )
+                    elif result["status"] == BFLStatus.error:
+                        raise Exception(f"BFL API encountered an error: {result}.")
+                    elif result["status"] == BFLStatus.pending:
+                        await asyncio.sleep(retry_pending_seconds)
+                        continue
+                elif response.status == 404:
+                    if retries_404 < max_retries_404:
+                        retries_404 += 1
+                        await asyncio.sleep(retry_404_seconds)
+                        continue
+                    raise Exception(
+                        f"BFL API could not find task after {max_retries_404} tries."
+                    )
+                elif response.status == 202:
+                    await asyncio.sleep(retry_202_seconds)
+                elif time.time() - start_time > timeout:
+                    raise Exception(
+                        f"BFL API experienced a timeout; could not return request under {timeout} seconds."
+                    )
+                else:
+                    raise Exception(f"BFL API encountered an error: {response.json()}")
 
 def convert_image_to_base64(image: torch.Tensor):
     scaled_image = downscale_image_tensor(image, total_pixels=2048 * 2048)
@@ -222,7 +224,7 @@ class FluxProUltraImageNode(ComfyNodeABC):
     API_NODE = True
     CATEGORY = "api node/image/BFL"
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         aspect_ratio: str,
@@ -266,7 +268,7 @@ class FluxProUltraImageNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 
@@ -354,7 +356,7 @@ class FluxKontextProImageNode(ComfyNodeABC):
 
     BFL_PATH = "/proxy/bfl/flux-kontext-pro/generate"
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         aspect_ratio: str,
@@ -397,7 +399,7 @@ class FluxKontextProImageNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 
@@ -489,7 +491,7 @@ class FluxProImageNode(ComfyNodeABC):
     API_NODE = True
     CATEGORY = "api node/image/BFL"
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         prompt_upsampling,
@@ -524,7 +526,7 @@ class FluxProImageNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 
@@ -632,7 +634,7 @@ class FluxProExpandNode(ComfyNodeABC):
     API_NODE = True
     CATEGORY = "api node/image/BFL"
 
-    def api_call(
+    async def api_call(
         self,
         image: torch.Tensor,
         prompt: str,
@@ -670,7 +672,7 @@ class FluxProExpandNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 
@@ -744,7 +746,7 @@ class FluxProFillNode(ComfyNodeABC):
     API_NODE = True
     CATEGORY = "api node/image/BFL"
 
-    def api_call(
+    async def api_call(
         self,
         image: torch.Tensor,
         mask: torch.Tensor,
@@ -780,7 +782,7 @@ class FluxProFillNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 
@@ -879,7 +881,7 @@ class FluxProCannyNode(ComfyNodeABC):
     API_NODE = True
     CATEGORY = "api node/image/BFL"
 
-    def api_call(
+    async def api_call(
         self,
         control_image: torch.Tensor,
         prompt: str,
@@ -929,7 +931,7 @@ class FluxProCannyNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 
@@ -1008,7 +1010,7 @@ class FluxProDepthNode(ComfyNodeABC):
     API_NODE = True
     CATEGORY = "api node/image/BFL"
 
-    def api_call(
+    async def api_call(
         self,
         control_image: torch.Tensor,
         prompt: str,
@@ -1045,7 +1047,7 @@ class FluxProDepthNode(ComfyNodeABC):
             ),
             auth_kwargs=kwargs,
         )
-        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        output_image = await handle_bfl_synchronous_operation(operation, node_id=unique_id)
         return (output_image,)
 
 

comfy_api_nodes/nodes_gemini.py

@@ -2,6 +2,8 @@
 API Nodes for Gemini Multimodal LLM Usage via Remote API
 See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference
 """
+from __future__ import annotations
+
 
 import os
 from enum import Enum
@@ -301,7 +303,7 @@ class GeminiNode(ComfyNodeABC):
         """
         return GeminiPart(text=text)
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         model: GeminiModel,
@@ -330,7 +332,7 @@ class GeminiNode(ComfyNodeABC):
             parts.extend(files)
 
         # Create response
-        response = SynchronousOperation(
+        response = await SynchronousOperation(
             endpoint=get_gemini_endpoint(model),
             request=GeminiGenerateContentRequest(
                 contents=[
@@ -406,7 +408,7 @@ class GeminiInputFiles(ComfyNodeABC):
 
     def create_file_part(self, file_path: str) -> GeminiPart:
         mime_type = (
-            GeminiMimeType.pdf
+            GeminiMimeType.application_pdf
             if file_path.endswith(".pdf")
             else GeminiMimeType.text_plain
         )

comfy_api_nodes/nodes_ideogram.py

@@ -212,7 +212,7 @@ V3_RESOLUTIONS= [
     "1536x640"
 ]
 
-def download_and_process_images(image_urls):
+async def download_and_process_images(image_urls):
     """Helper function to download and process multiple images from URLs"""
 
     # Initialize list to store image tensors
@@ -220,7 +220,7 @@ def download_and_process_images(image_urls):
 
     for image_url in image_urls:
         # Using functions from apinode_utils.py to handle downloading and processing
-        image_bytesio = download_url_to_bytesio(image_url)  # Download image content to BytesIO
+        image_bytesio = await download_url_to_bytesio(image_url)  # Download image content to BytesIO
         img_tensor = bytesio_to_image_tensor(image_bytesio, mode="RGB")  # Convert to torch.Tensor with RGB mode
         image_tensors.append(img_tensor)
 
@@ -328,7 +328,7 @@ class IdeogramV1(ComfyNodeABC):
     DESCRIPTION = cleandoc(__doc__ or "")
     API_NODE = True
 
-    def api_call(
+    async def api_call(
         self,
         prompt,
         turbo=False,
@@ -367,7 +367,7 @@ class IdeogramV1(ComfyNodeABC):
             auth_kwargs=kwargs,
         )
 
-        response = operation.execute()
+        response = await operation.execute()
 
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
@@ -378,7 +378,7 @@ class IdeogramV1(ComfyNodeABC):
             raise Exception("No image URLs were generated in the response")
 
         display_image_urls_on_node(image_urls, unique_id)
-        return (download_and_process_images(image_urls),)
+        return (await download_and_process_images(image_urls),)
 
 
 class IdeogramV2(ComfyNodeABC):
@@ -487,7 +487,7 @@ class IdeogramV2(ComfyNodeABC):
     DESCRIPTION = cleandoc(__doc__ or "")
     API_NODE = True
 
-    def api_call(
+    async def api_call(
         self,
         prompt,
         turbo=False,
@@ -543,7 +543,7 @@ class IdeogramV2(ComfyNodeABC):
             auth_kwargs=kwargs,
         )
 
-        response = operation.execute()
+        response = await operation.execute()
 
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
@@ -554,7 +554,7 @@ class IdeogramV2(ComfyNodeABC):
             raise Exception("No image URLs were generated in the response")
 
         display_image_urls_on_node(image_urls, unique_id)
-        return (download_and_process_images(image_urls),)
+        return (await download_and_process_images(image_urls),)
 
 class IdeogramV3(ComfyNodeABC):
     """
@@ -653,7 +653,7 @@ class IdeogramV3(ComfyNodeABC):
     DESCRIPTION = cleandoc(__doc__ or "")
     API_NODE = True
 
-    def api_call(
+    async def api_call(
         self,
         prompt,
         image=None,
@@ -774,7 +774,7 @@ class IdeogramV3(ComfyNodeABC):
             )
 
         # Execute the operation and process response
-        response = operation.execute()
+        response = await operation.execute()
 
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
@@ -785,7 +785,7 @@ class IdeogramV3(ComfyNodeABC):
             raise Exception("No image URLs were generated in the response")
 
         display_image_urls_on_node(image_urls, unique_id)
-        return (download_and_process_images(image_urls),)
+        return (await download_and_process_images(image_urls),)
 
 
 NODE_CLASS_MAPPINGS = {

comfy_api_nodes/nodes_kling.py

@@ -109,7 +109,7 @@ class KlingApiError(Exception):
     pass
 
 
-def poll_until_finished(
+async def poll_until_finished(
     auth_kwargs: dict[str, str],
     api_endpoint: ApiEndpoint[Any, R],
     result_url_extractor: Optional[Callable[[R], str]] = None,
@@ -117,7 +117,7 @@ def poll_until_finished(
     node_id: Optional[str] = None,
 ) -> R:
     """Polls the Kling API endpoint until the task reaches a terminal state, then returns the response."""
-    return PollingOperation(
+    return await PollingOperation(
         poll_endpoint=api_endpoint,
         completed_statuses=[
             KlingTaskStatus.succeed.value,
@@ -132,6 +132,8 @@ def poll_until_finished(
         result_url_extractor=result_url_extractor,
         estimated_duration=estimated_duration,
         node_id=node_id,
+        poll_interval=16.0,
+        max_poll_attempts=256,
     ).execute()
 
 
@@ -276,18 +278,18 @@ def get_images_urls_from_response(response) -> Optional[str]:
         return None
 
 
-def video_result_to_node_output(
+async def video_result_to_node_output(
     video: KlingVideoResult,
 ) -> tuple[VideoFromFile, str, str]:
     """Converts a KlingVideoResult to a tuple of (VideoFromFile, str, str) to be used as a ComfyUI node output."""
     return (
-        download_url_to_video_output(video.url),
+        await download_url_to_video_output(str(video.url)),
         str(video.id),
         str(video.duration),
     )
 
 
-def image_result_to_node_output(
+async def image_result_to_node_output(
     images: list[KlingImageResult],
 ) -> torch.Tensor:
     """
@@ -295,9 +297,9 @@ def image_result_to_node_output(
     If multiple images are returned, they will be stacked along the batch dimension.
     """
     if len(images) == 1:
-        return download_url_to_image_tensor(images[0].url)
+        return await download_url_to_image_tensor(str(images[0].url))
     else:
-        return torch.cat([download_url_to_image_tensor(image.url) for image in images])
+        return torch.cat([await download_url_to_image_tensor(str(image.url)) for image in images])
 
 
 class KlingNodeBase(ComfyNodeABC):
@@ -465,10 +467,10 @@ class KlingTextToVideoNode(KlingNodeBase):
     RETURN_NAMES = ("VIDEO", "video_id", "duration")
     DESCRIPTION = "Kling Text to Video Node"
 
-    def get_response(
+    async def get_response(
         self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
     ) -> KlingText2VideoResponse:
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_TEXT_TO_VIDEO}/{task_id}",
@@ -481,7 +483,7 @@ class KlingTextToVideoNode(KlingNodeBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         negative_prompt: str,
@@ -517,17 +519,17 @@ class KlingTextToVideoNode(KlingNodeBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
 
         task_id = task_creation_response.data.task_id
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_video_result_response(final_response)
 
         video = get_video_from_response(final_response)
-        return video_result_to_node_output(video)
+        return await video_result_to_node_output(video)
 
 
 class KlingCameraControlT2VNode(KlingTextToVideoNode):
@@ -579,7 +581,7 @@ class KlingCameraControlT2VNode(KlingTextToVideoNode):
 
     DESCRIPTION = "Transform text into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original text."
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         negative_prompt: str,
@@ -589,7 +591,7 @@ class KlingCameraControlT2VNode(KlingTextToVideoNode):
         unique_id: Optional[str] = None,
         **kwargs,
     ):
-        return super().api_call(
+        return await super().api_call(
             model_name=KlingVideoGenModelName.kling_v1,
             cfg_scale=cfg_scale,
             mode=KlingVideoGenMode.std,
@@ -668,10 +670,10 @@ class KlingImage2VideoNode(KlingNodeBase):
     RETURN_NAMES = ("VIDEO", "video_id", "duration")
     DESCRIPTION = "Kling Image to Video Node"
 
-    def get_response(
+    async def get_response(
         self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
     ) -> KlingImage2VideoResponse:
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}",
@@ -684,7 +686,7 @@ class KlingImage2VideoNode(KlingNodeBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         start_frame: torch.Tensor,
         prompt: str,
@@ -731,17 +733,17 @@ class KlingImage2VideoNode(KlingNodeBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
         task_id = task_creation_response.data.task_id
 
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_video_result_response(final_response)
 
         video = get_video_from_response(final_response)
-        return video_result_to_node_output(video)
+        return await video_result_to_node_output(video)
 
 
 class KlingCameraControlI2VNode(KlingImage2VideoNode):
@@ -796,7 +798,7 @@ class KlingCameraControlI2VNode(KlingImage2VideoNode):
 
     DESCRIPTION = "Transform still images into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original image."
 
-    def api_call(
+    async def api_call(
         self,
         start_frame: torch.Tensor,
         prompt: str,
@@ -807,7 +809,7 @@ class KlingCameraControlI2VNode(KlingImage2VideoNode):
         unique_id: Optional[str] = None,
         **kwargs,
     ):
-        return super().api_call(
+        return await super().api_call(
             model_name=KlingVideoGenModelName.kling_v1_5,
             start_frame=start_frame,
             cfg_scale=cfg_scale,
@@ -895,7 +897,7 @@ class KlingStartEndFrameNode(KlingImage2VideoNode):
 
     DESCRIPTION = "Generate a video sequence that transitions between your provided start and end images. The node creates all frames in between, producing a smooth transformation from the first frame to the last."
 
-    def api_call(
+    async def api_call(
         self,
         start_frame: torch.Tensor,
         end_frame: torch.Tensor,
@@ -910,7 +912,7 @@ class KlingStartEndFrameNode(KlingImage2VideoNode):
         mode, duration, model_name = KlingStartEndFrameNode.get_mode_string_mapping()[
             mode
         ]
-        return super().api_call(
+        return await super().api_call(
             prompt=prompt,
             negative_prompt=negative_prompt,
             model_name=model_name,
@@ -962,10 +964,10 @@ class KlingVideoExtendNode(KlingNodeBase):
     RETURN_NAMES = ("VIDEO", "video_id", "duration")
     DESCRIPTION = "Kling Video Extend Node. Extend videos made by other Kling nodes. The video_id is created by using other Kling Nodes."
 
-    def get_response(
+    async def get_response(
         self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
     ) -> KlingVideoExtendResponse:
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_VIDEO_EXTEND}/{task_id}",
@@ -978,7 +980,7 @@ class KlingVideoExtendNode(KlingNodeBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         prompt: str,
         negative_prompt: str,
@@ -1004,17 +1006,17 @@ class KlingVideoExtendNode(KlingNodeBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
         task_id = task_creation_response.data.task_id
 
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_video_result_response(final_response)
 
         video = get_video_from_response(final_response)
-        return video_result_to_node_output(video)
+        return await video_result_to_node_output(video)
 
 
 class KlingVideoEffectsBase(KlingNodeBase):
@@ -1023,10 +1025,10 @@ class KlingVideoEffectsBase(KlingNodeBase):
     RETURN_TYPES = ("VIDEO", "STRING", "STRING")
     RETURN_NAMES = ("VIDEO", "video_id", "duration")
 
-    def get_response(
+    async def get_response(
         self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
     ) -> KlingVideoEffectsResponse:
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_VIDEO_EFFECTS}/{task_id}",
@@ -1039,7 +1041,7 @@ class KlingVideoEffectsBase(KlingNodeBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         dual_character: bool,
         effect_scene: KlingDualCharacterEffectsScene | KlingSingleImageEffectsScene,
@@ -1082,17 +1084,17 @@ class KlingVideoEffectsBase(KlingNodeBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
         task_id = task_creation_response.data.task_id
 
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_video_result_response(final_response)
 
         video = get_video_from_response(final_response)
-        return video_result_to_node_output(video)
+        return await video_result_to_node_output(video)
 
 
 class KlingDualCharacterVideoEffectNode(KlingVideoEffectsBase):
@@ -1140,7 +1142,7 @@ class KlingDualCharacterVideoEffectNode(KlingVideoEffectsBase):
     RETURN_TYPES = ("VIDEO", "STRING")
     RETURN_NAMES = ("VIDEO", "duration")
 
-    def api_call(
+    async def api_call(
         self,
         image_left: torch.Tensor,
         image_right: torch.Tensor,
@@ -1151,7 +1153,7 @@ class KlingDualCharacterVideoEffectNode(KlingVideoEffectsBase):
         unique_id: Optional[str] = None,
         **kwargs,
     ):
-        video, _, duration = super().api_call(
+        video, _, duration = await super().api_call(
             dual_character=True,
             effect_scene=effect_scene,
             model_name=model_name,
@@ -1206,7 +1208,7 @@ class KlingSingleImageVideoEffectNode(KlingVideoEffectsBase):
 
     DESCRIPTION = "Achieve different special effects when generating a video based on the effect_scene."
 
-    def api_call(
+    async def api_call(
         self,
         image: torch.Tensor,
         effect_scene: KlingSingleImageEffectsScene,
@@ -1215,7 +1217,7 @@ class KlingSingleImageVideoEffectNode(KlingVideoEffectsBase):
         unique_id: Optional[str] = None,
         **kwargs,
     ):
-        return super().api_call(
+        return await super().api_call(
             dual_character=False,
             effect_scene=effect_scene,
             model_name=model_name,
@@ -1251,11 +1253,11 @@ class KlingLipSyncBase(KlingNodeBase):
                 f"Text is too long. Maximum length is {MAX_PROMPT_LENGTH_LIP_SYNC} characters."
             )
 
-    def get_response(
+    async def get_response(
         self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
     ) -> KlingLipSyncResponse:
         """Polls the Kling API endpoint until the task reaches a terminal state."""
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_LIP_SYNC}/{task_id}",
@@ -1268,7 +1270,7 @@ class KlingLipSyncBase(KlingNodeBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         video: VideoInput,
         audio: Optional[AudioInput] = None,
@@ -1285,12 +1287,12 @@ class KlingLipSyncBase(KlingNodeBase):
         self.validate_lip_sync_video(video)
 
         # Upload video to Comfy API and get download URL
-        video_url = upload_video_to_comfyapi(video, auth_kwargs=kwargs)
+        video_url = await upload_video_to_comfyapi(video, auth_kwargs=kwargs)
         logging.info("Uploaded video to Comfy API. URL: %s", video_url)
 
         # Upload the audio file to Comfy API and get download URL
         if audio:
-            audio_url = upload_audio_to_comfyapi(audio, auth_kwargs=kwargs)
+            audio_url = await upload_audio_to_comfyapi(audio, auth_kwargs=kwargs)
             logging.info("Uploaded audio to Comfy API. URL: %s", audio_url)
         else:
             audio_url = None
@@ -1317,17 +1319,17 @@ class KlingLipSyncBase(KlingNodeBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
         task_id = task_creation_response.data.task_id
 
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_video_result_response(final_response)
 
         video = get_video_from_response(final_response)
-        return video_result_to_node_output(video)
+        return await video_result_to_node_output(video)
 
 
 class KlingLipSyncAudioToVideoNode(KlingLipSyncBase):
@@ -1355,7 +1357,7 @@ class KlingLipSyncAudioToVideoNode(KlingLipSyncBase):
 
     DESCRIPTION = "Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file. When using, ensure that the audio contains clearly distinguishable vocals and that the video contains a distinct face. The audio file should not be larger than 5MB. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length."
 
-    def api_call(
+    async def api_call(
         self,
         video: VideoInput,
         audio: AudioInput,
@@ -1363,7 +1365,7 @@ class KlingLipSyncAudioToVideoNode(KlingLipSyncBase):
         unique_id: Optional[str] = None,
         **kwargs,
     ):
-        return super().api_call(
+        return await super().api_call(
             video=video,
             audio=audio,
             voice_language=voice_language,
@@ -1467,7 +1469,7 @@ class KlingLipSyncTextToVideoNode(KlingLipSyncBase):
 
     DESCRIPTION = "Kling Lip Sync Text to Video Node. Syncs mouth movements in a video file to a text prompt. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length."
 
-    def api_call(
+    async def api_call(
         self,
         video: VideoInput,
         text: str,
@@ -1477,7 +1479,7 @@ class KlingLipSyncTextToVideoNode(KlingLipSyncBase):
         **kwargs,
     ):
         voice_id, voice_language = KlingLipSyncTextToVideoNode.get_voice_config()[voice]
-        return super().api_call(
+        return await super().api_call(
             video=video,
             text=text,
             voice_language=voice_language,
@@ -1531,10 +1533,10 @@ class KlingVirtualTryOnNode(KlingImageGenerationBase):
 
     DESCRIPTION = "Kling Virtual Try On Node. Input a human image and a cloth image to try on the cloth on the human. You can merge multiple clothing item pictures into one image with a white background."
 
-    def get_response(
+    async def get_response(
         self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
     ) -> KlingVirtualTryOnResponse:
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_VIRTUAL_TRY_ON}/{task_id}",
@@ -1547,7 +1549,7 @@ class KlingVirtualTryOnNode(KlingImageGenerationBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         human_image: torch.Tensor,
         cloth_image: torch.Tensor,
@@ -1570,17 +1572,17 @@ class KlingVirtualTryOnNode(KlingImageGenerationBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
         task_id = task_creation_response.data.task_id
 
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_image_result_response(final_response)
 
         images = get_images_from_response(final_response)
-        return (image_result_to_node_output(images),)
+        return (await image_result_to_node_output(images),)
 
 
 class KlingImageGenerationNode(KlingImageGenerationBase):
@@ -1653,13 +1655,13 @@ class KlingImageGenerationNode(KlingImageGenerationBase):
 
     DESCRIPTION = "Kling Image Generation Node. Generate an image from a text prompt with an optional reference image."
 
-    def get_response(
+    async def get_response(
         self,
         task_id: str,
         auth_kwargs: Optional[dict[str, str]],
         node_id: Optional[str] = None,
     ) -> KlingImageGenerationsResponse:
-        return poll_until_finished(
+        return await poll_until_finished(
             auth_kwargs,
             ApiEndpoint(
                 path=f"{PATH_IMAGE_GENERATIONS}/{task_id}",
@@ -1672,7 +1674,7 @@ class KlingImageGenerationNode(KlingImageGenerationBase):
             node_id=node_id,
         )
 
-    def api_call(
+    async def api_call(
         self,
         model_name: KlingImageGenModelName,
         prompt: str,
@@ -1688,7 +1690,11 @@ class KlingImageGenerationNode(KlingImageGenerationBase):
     ):
         self.validate_prompt(prompt, negative_prompt)
 
-        if image is not None:
+        if image is None:
+            image_type = None
+        elif model_name == KlingImageGenModelName.kling_v1:
+            raise ValueError(f"The model {KlingImageGenModelName.kling_v1.value} does not support reference images.")
+        else:
             image = tensor_to_base64_string(image)
 
         initial_operation = SynchronousOperation(
@@ -1712,17 +1718,17 @@ class KlingImageGenerationNode(KlingImageGenerationBase):
             auth_kwargs=kwargs,
         )
 
-        task_creation_response = initial_operation.execute()
+        task_creation_response = await initial_operation.execute()
         validate_task_creation_response(task_creation_response)
         task_id = task_creation_response.data.task_id
 
-        final_response = self.get_response(
+        final_response = await self.get_response(
             task_id, auth_kwargs=kwargs, node_id=unique_id
         )
         validate_image_result_response(final_response)
 
         images = get_images_from_response(final_response)
-        return (image_result_to_node_output(images),)
+        return (await image_result_to_node_output(images),)
 
 
 NODE_CLASS_MAPPINGS = {