ComfyUI version 0.3.47

* [moonvalley] Update V2V node to match API specification

- Add exact resolution validation for supported resolutions (1920x1080, 1080x1920, 1152x1152, 1536x1152, 1152x1536)
- Change frame count validation from divisible by 32 to 16
- Add MP4 container format validation
- Remove internal parameters (steps, guidance_scale) from V2V inference params
- Update video duration handling to support only 5 seconds (auto-trim if longer)
- Add motion_intensity parameter (0-100) for Motion Transfer control type
- Add get_container_format() method to VideoInput classes

* update negative prompt

.ci/update_windows/update.py

@@ -63,7 +63,12 @@ except:
 print("checking out master branch")  # noqa: T201
 branch = repo.lookup_branch('master')
 if branch is None:
-    ref = repo.lookup_reference('refs/remotes/origin/master')
+    try:
+        ref = repo.lookup_reference('refs/remotes/origin/master')
+    except:
+        print("pulling.")  # noqa: T201
+        pull(repo)
+        ref = repo.lookup_reference('refs/remotes/origin/master')
     repo.checkout(ref)
     branch = repo.lookup_branch('master')
     if branch is None:

.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:

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -15,6 +15,14 @@ body:
         steps to replicate what went wrong and others will be able to repeat your steps and see the same issue happen.
 
         If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first.
+  - type: checkboxes
+    id: custom-nodes-test
+    attributes:
+      label: Custom Node Testing
+      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
   - type: textarea
     attributes:
       label: Expected Behavior

.github/ISSUE_TEMPLATE/user-support.yml

@@ -11,6 +11,14 @@ body:
             **2:** You have made an effort to find public answers to your question before asking here. In other words, you googled it first, and scrolled through recent help topics.
 
                 If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first.
+    - type: checkboxes
+      id: custom-nodes-test
+      attributes:
+        label: Custom Node Testing
+        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
     - 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

@@ -12,7 +12,7 @@ on:
         description: 'CUDA version'
         required: true
         type: string
-        default: "126"
+        default: "128"
       python_minor:
         description: 'Python minor version'
         required: true
@@ -22,7 +22,7 @@ on:
         description: 'Python patch version'
         required: true
         type: string
-        default: "9"
+        default: "10"
 
 
 jobs:
@@ -36,7 +36,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           ref: ${{ inputs.git_tag }}
-          fetch-depth: 0
+          fetch-depth: 150
           persist-credentials: false
       - uses: actions/cache/restore@v4
         id: cache
@@ -70,7 +70,7 @@ jobs:
             cd ..
 
           git clone --depth 1 https://github.com/comfyanonymous/taesd
-          cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+          cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
           mkdir ComfyUI_windows_portable
           mv python_embeded ComfyUI_windows_portable
@@ -85,12 +85,14 @@ jobs:
 
           cd ..
 
-          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -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=512m -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
           python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
+          python_embeded/python.exe -s ./update/update.py ComfyUI/
+
           ls
 
       - name: Upload binaries to release
@@ -100,5 +102,4 @@ jobs:
           file: ComfyUI_windows_portable_nvidia.7z
           tag: ${{ inputs.git_tag }}
           overwrite: true
-          prerelease: true
-          make_latest: false
+          draft: true

.github/workflows/test-launch.yml

@@ -17,7 +17,7 @@ jobs:
         path: "ComfyUI"
     - uses: actions/setup-python@v4
       with:
-        python-version: '3.9'
+        python-version: '3.10'
     - name: Install requirements
       run: |
         python -m pip install --upgrade pip

.github/workflows/update-api-stubs.yml

@@ -0,0 +1,56 @@
+name: Generate Pydantic Stubs from api.comfy.org
+
+on:
+  schedule:
+    - cron: '0 0 * * 1'
+  workflow_dispatch:
+
+jobs:
+  generate-models:
+    runs-on: ubuntu-latest
+    
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+      
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.10'
+      
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install 'datamodel-code-generator[http]'
+          npm install @redocly/cli
+      
+      - name: Download OpenAPI spec
+        run: |
+          curl -o openapi.yaml https://api.comfy.org/openapi
+      
+      - name: Filter OpenAPI spec with Redocly
+        run: |
+          npx @redocly/cli bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly.yaml --remove-unused-components
+      
+      - name: Generate API models
+        run: |
+          datamodel-codegen --use-subclass-enum --input filtered-openapi.yaml --output comfy_api_nodes/apis --output-model-type pydantic_v2.BaseModel
+      
+      - name: Check for changes
+        id: git-check
+        run: |
+          git diff --exit-code comfy_api_nodes/apis || echo "changes=true" >> $GITHUB_OUTPUT
+      
+      - name: Create Pull Request
+        if: steps.git-check.outputs.changes == 'true'
+        uses: peter-evans/create-pull-request@v5
+        with:
+          commit-message: 'chore: update API models from OpenAPI spec'
+          title: 'Update API models from api.comfy.org'
+          body: |
+            This PR updates the API models based on the latest api.comfy.org OpenAPI specification.
+            
+            Generated automatically by the a Github workflow.
+          branch: update-api-stubs
+          delete-branch: true
+          base: master

.github/workflows/windows_release_dependencies.yml

@@ -17,7 +17,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "126"
+        default: "128"
 
       python_minor:
         description: 'python minor version'
@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "9"
+        default: "10"
 #  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
@@ -53,10 +53,12 @@ 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
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+            cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable_nightly_pytorch
             mv python_embeded ComfyUI_windows_portable_nightly_pytorch

.github/workflows/windows_release_package.yml

@@ -7,7 +7,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "126"
+        default: "128"
 
       python_minor:
         description: 'python minor version'
@@ -19,7 +19,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "9"
+        default: "10"
 #  push:
 #    branches:
 #      - master
@@ -50,7 +50,7 @@ jobs:
 
         - uses: actions/checkout@v4
           with:
-            fetch-depth: 0
+            fetch-depth: 150
             persist-credentials: false
         - shell: bash
           run: |
@@ -67,7 +67,7 @@ jobs:
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+            cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable
             mv python_embeded ComfyUI_windows_portable
@@ -82,12 +82,14 @@ jobs:
 
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -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=512m -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
             python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
+            python_embeded/python.exe -s ./update/update.py ComfyUI/
+
             ls
 
         - name: Upload binaries to release

.gitignore

@@ -21,3 +21,6 @@ venv/
 *.log
 web_custom_versions/
 .DS_Store
+openapi.yaml
+filtered-openapi.yaml
+uv.lock

CODEOWNERS

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

README.md

@@ -6,6 +6,7 @@
 
 [![Website][website-shield]][website-url]
 [![Dynamic JSON Badge][discord-shield]][discord-url]
+[![Twitter][twitter-shield]][twitter-url]
 [![Matrix][matrix-shield]][matrix-url]
 <br>
 [![][github-release-shield]][github-release-link]
@@ -20,6 +21,8 @@
 <!-- Workaround to display total user from https://github.com/badges/shields/issues/4500#issuecomment-2060079995 -->
 [discord-shield]: https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fdiscord.com%2Fapi%2Finvites%2Fcomfyorg%3Fwith_counts%3Dtrue&query=%24.approximate_member_count&logo=discord&logoColor=white&label=Discord&color=green&suffix=%20total
 [discord-url]: https://www.comfy.org/discord
+[twitter-shield]: https://img.shields.io/twitter/follow/ComfyUI
+[twitter-url]: https://x.com/ComfyUI
 
 [github-release-shield]: https://img.shields.io/github/v/release/comfyanonymous/ComfyUI?style=flat&sort=semver
 [github-release-link]: https://github.com/comfyanonymous/ComfyUI/releases
@@ -49,11 +52,10 @@ Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon,
 ## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
 See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
-
 ## 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/)
@@ -63,21 +65,30 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
    - [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/)
+- 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/)
    - [LTX-Video](https://comfyanonymous.github.io/ComfyUI_examples/ltxv/)
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
-   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/)
+   - [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/)
 - 3D Models
    - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2)
-- [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
 - 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/)
@@ -88,17 +99,32 @@ 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: will never download anything.
+- 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.
 
 Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
 
+## Release Process
+
+ComfyUI follows a weekly release cycle every Friday, with three interconnected repositories:
+
+1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
+   - Releases a new stable version (e.g., v0.7.0)
+   - Serves as the foundation for the desktop release
+
+2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
+   - Builds a new release using the latest stable core version
+
+3. **[ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend)**
+   - Weekly frontend updates are merged into the core repository
+   - Features are frozen for the upcoming core release
+   - Development continues for the next release cycle
+
 ## Shortcuts
 
 | Keybind                            | Explanation                                                                                                        |
@@ -149,16 +175,10 @@ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
-If you have a 50 series Blackwell card like a 5090 or 5080 see [this discussion thread](https://github.com/comfyanonymous/ComfyUI/discussions/6643)
-
 #### How do I share models between another UI and ComfyUI?
 
 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)
 
@@ -182,11 +202,11 @@ 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/rocm6.2.4```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.3```
 
-This is the command to install the nightly with ROCm 6.3 which might have some performance improvements:
+This is the command to install the nightly with ROCm 6.4 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.4```
 
 ### Intel GPUs (Windows and Linux)
 
@@ -216,11 +236,11 @@ Additional discussion and help can be found [here](https://github.com/comfyanony
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu126```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128```
 
-This is the command to install pytorch nightly instead which supports the new blackwell 50xx series GPUs and might have performance improvements.
+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
 
@@ -253,6 +273,8 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve
 
 #### DirectML (AMD Cards on Windows)
 
+This is very badly supported and is not recommended. There are some unofficial builds of pytorch ROCm on windows that exist that will give you a much better experience than this. This readme will be updated once official pytorch ROCm builds for windows come out.
+
 ```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
 
 #### Ascend NPUs
@@ -272,6 +294,13 @@ 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```
@@ -286,7 +315,7 @@ For AMD 7600 and maybe other RDNA3 cards: ```HSA_OVERRIDE_GFX_VERSION=11.0.0 pyt
 
 ### AMD ROCm Tips
 
-You can enable experimental memory efficient attention on pytorch 2.5 in ComfyUI on RDNA3 and potentially other AMD GPUs using this command:
+You can enable experimental memory efficient attention on recent pytorch in ComfyUI on some AMD GPUs using this command, it should already be enabled by default on RDNA3. If this improves speed for you on latest pytorch on your GPU please report it so that I can enable it by default.
 
 ```TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention```
 

alembic.ini

@@ -0,0 +1,84 @@
+# A generic, single database configuration.
+
+[alembic]
+# path to migration scripts
+# Use forward slashes (/) also on windows to provide an os agnostic path
+script_location = alembic_db
+
+# template used to generate migration file names; The default value is %%(rev)s_%%(slug)s
+# Uncomment the line below if you want the files to be prepended with date and time
+# see https://alembic.sqlalchemy.org/en/latest/tutorial.html#editing-the-ini-file
+# for all available tokens
+# file_template = %%(year)d_%%(month).2d_%%(day).2d_%%(hour).2d%%(minute).2d-%%(rev)s_%%(slug)s
+
+# sys.path path, will be prepended to sys.path if present.
+# defaults to the current working directory.
+prepend_sys_path = .
+
+# timezone to use when rendering the date within the migration file
+# as well as the filename.
+# If specified, requires the python>=3.9 or backports.zoneinfo library and tzdata library.
+# Any required deps can installed by adding `alembic[tz]` to the pip requirements
+# string value is passed to ZoneInfo()
+# leave blank for localtime
+# timezone =
+
+# max length of characters to apply to the "slug" field
+# truncate_slug_length = 40
+
+# set to 'true' to run the environment during
+# the 'revision' command, regardless of autogenerate
+# revision_environment = false
+
+# set to 'true' to allow .pyc and .pyo files without
+# a source .py file to be detected as revisions in the
+# versions/ directory
+# sourceless = false
+
+# version location specification; This defaults
+# to alembic_db/versions.  When using multiple version
+# directories, initial revisions must be specified with --version-path.
+# The path separator used here should be the separator specified by "version_path_separator" below.
+# version_locations = %(here)s/bar:%(here)s/bat:alembic_db/versions
+
+# version path separator; As mentioned above, this is the character used to split
+# version_locations. The default within new alembic.ini files is "os", which uses os.pathsep.
+# If this key is omitted entirely, it falls back to the legacy behavior of splitting on spaces and/or commas.
+# Valid values for version_path_separator are:
+#
+# version_path_separator = :
+# version_path_separator = ;
+# version_path_separator = space
+# version_path_separator = newline
+#
+# Use os.pathsep. Default configuration used for new projects.
+version_path_separator = os
+
+# set to 'true' to search source files recursively
+# in each "version_locations" directory
+# new in Alembic version 1.10
+# recursive_version_locations = false
+
+# the output encoding used when revision files
+# are written from script.py.mako
+# output_encoding = utf-8
+
+sqlalchemy.url = sqlite:///user/comfyui.db
+
+
+[post_write_hooks]
+# post_write_hooks defines scripts or Python functions that are run
+# on newly generated revision scripts.  See the documentation for further
+# detail and examples
+
+# format using "black" - use the console_scripts runner, against the "black" entrypoint
+# hooks = black
+# black.type = console_scripts
+# black.entrypoint = black
+# black.options = -l 79 REVISION_SCRIPT_FILENAME
+
+# lint with attempts to fix using "ruff" - use the exec runner, execute a binary
+# hooks = ruff
+# ruff.type = exec
+# ruff.executable = %(here)s/.venv/bin/ruff
+# ruff.options = check --fix REVISION_SCRIPT_FILENAME

alembic_db/README.md

@@ -0,0 +1,4 @@
+## Generate new revision
+
+1. Update models in `/app/database/models.py`
+2. Run `alembic revision --autogenerate -m "{your message}"`

alembic_db/env.py

@@ -0,0 +1,64 @@
+from sqlalchemy import engine_from_config
+from sqlalchemy import pool
+
+from alembic import context
+
+# this is the Alembic Config object, which provides
+# access to the values within the .ini file in use.
+config = context.config
+
+
+from app.database.models import Base
+target_metadata = Base.metadata
+
+# other values from the config, defined by the needs of env.py,
+# can be acquired:
+# my_important_option = config.get_main_option("my_important_option")
+# ... etc.
+
+
+def run_migrations_offline() -> None:
+    """Run migrations in 'offline' mode.
+    This configures the context with just a URL
+    and not an Engine, though an Engine is acceptable
+    here as well.  By skipping the Engine creation
+    we don't even need a DBAPI to be available.
+    Calls to context.execute() here emit the given string to the
+    script output.
+    """
+    url = config.get_main_option("sqlalchemy.url")
+    context.configure(
+        url=url,
+        target_metadata=target_metadata,
+        literal_binds=True,
+        dialect_opts={"paramstyle": "named"},
+    )
+
+    with context.begin_transaction():
+        context.run_migrations()
+
+
+def run_migrations_online() -> None:
+    """Run migrations in 'online' mode.
+    In this scenario we need to create an Engine
+    and associate a connection with the context.
+    """
+    connectable = engine_from_config(
+        config.get_section(config.config_ini_section, {}),
+        prefix="sqlalchemy.",
+        poolclass=pool.NullPool,
+    )
+
+    with connectable.connect() as connection:
+        context.configure(
+            connection=connection, target_metadata=target_metadata
+        )
+
+        with context.begin_transaction():
+            context.run_migrations()
+
+
+if context.is_offline_mode():
+    run_migrations_offline()
+else:
+    run_migrations_online()

alembic_db/script.py.mako

@@ -0,0 +1,28 @@
+"""${message}
+
+Revision ID: ${up_revision}
+Revises: ${down_revision | comma,n}
+Create Date: ${create_date}
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+${imports if imports else ""}
+
+# revision identifiers, used by Alembic.
+revision: str = ${repr(up_revision)}
+down_revision: Union[str, None] = ${repr(down_revision)}
+branch_labels: Union[str, Sequence[str], None] = ${repr(branch_labels)}
+depends_on: Union[str, Sequence[str], None] = ${repr(depends_on)}
+
+
+def upgrade() -> None:
+    """Upgrade schema."""
+    ${upgrades if upgrades else "pass"}
+
+
+def downgrade() -> None:
+    """Downgrade schema."""
+    ${downgrades if downgrades else "pass"}

app/custom_node_manager.py

@@ -93,16 +93,20 @@ class CustomNodeManager:
 
     def add_routes(self, routes, webapp, loadedModules):
 
+        example_workflow_folder_names = ["example_workflows", "example", "examples", "workflow", "workflows"]
+
         @routes.get("/workflow_templates")
         async def get_workflow_templates(request):
             """Returns a web response that contains the map of custom_nodes names and their associated workflow templates. The ones without templates are omitted."""
-            files = [
-                file
-                for folder in folder_paths.get_folder_paths("custom_nodes")
-                for file in glob.glob(
-                    os.path.join(folder, "*/example_workflows/*.json")
-                )
-            ]
+
+            files = []
+
+            for folder in folder_paths.get_folder_paths("custom_nodes"):
+                for folder_name in example_workflow_folder_names:
+                    pattern = os.path.join(folder, f"*/{folder_name}/*.json")
+                    matched_files = glob.glob(pattern)
+                    files.extend(matched_files)
+
             workflow_templates_dict = (
                 {}
             )  # custom_nodes folder name -> example workflow names
@@ -118,15 +122,22 @@ class CustomNodeManager:
 
         # Serve workflow templates from custom nodes.
         for module_name, module_dir in loadedModules:
-            workflows_dir = os.path.join(module_dir, "example_workflows")
-            if os.path.exists(workflows_dir):
-                webapp.add_routes(
-                    [
-                        web.static(
-                            "/api/workflow_templates/" + module_name, workflows_dir
-                        )
-                    ]
-                )
+            for folder_name in example_workflow_folder_names:
+                workflows_dir = os.path.join(module_dir, folder_name)
+
+                if os.path.exists(workflows_dir):
+                    if folder_name != "example_workflows":
+                        logging.debug(
+                            "Found example workflow folder '%s' for custom node '%s', consider renaming it to 'example_workflows'",
+                            folder_name, module_name)
+
+                    webapp.add_routes(
+                        [
+                            web.static(
+                                "/api/workflow_templates/" + module_name, workflows_dir
+                            )
+                        ]
+                    )
 
         @routes.get("/i18n")
         async def get_i18n(request):

app/database/db.py

@@ -0,0 +1,112 @@
+import logging
+import os
+import shutil
+from app.logger import log_startup_warning
+from utils.install_util import get_missing_requirements_message
+from comfy.cli_args import args
+
+_DB_AVAILABLE = False
+Session = None
+
+
+try:
+    from alembic import command
+    from alembic.config import Config
+    from alembic.runtime.migration import MigrationContext
+    from alembic.script import ScriptDirectory
+    from sqlalchemy import create_engine
+    from sqlalchemy.orm import sessionmaker
+
+    _DB_AVAILABLE = True
+except ImportError as e:
+    log_startup_warning(
+        f"""
+------------------------------------------------------------------------
+Error importing dependencies: {e}
+{get_missing_requirements_message()}
+This error is happening because ComfyUI now uses a local sqlite database.
+------------------------------------------------------------------------
+""".strip()
+    )
+
+
+def dependencies_available():
+    """
+    Temporary function to check if the dependencies are available
+    """
+    return _DB_AVAILABLE
+
+
+def can_create_session():
+    """
+    Temporary function to check if the database is available to create a session
+    During initial release there may be environmental issues (or missing dependencies) that prevent the database from being created
+    """
+    return dependencies_available() and Session is not None
+
+
+def get_alembic_config():
+    root_path = os.path.join(os.path.dirname(__file__), "../..")
+    config_path = os.path.abspath(os.path.join(root_path, "alembic.ini"))
+    scripts_path = os.path.abspath(os.path.join(root_path, "alembic_db"))
+
+    config = Config(config_path)
+    config.set_main_option("script_location", scripts_path)
+    config.set_main_option("sqlalchemy.url", args.database_url)
+
+    return config
+
+
+def get_db_path():
+    url = args.database_url
+    if url.startswith("sqlite:///"):
+        return url.split("///")[1]
+    else:
+        raise ValueError(f"Unsupported database URL '{url}'.")
+
+
+def init_db():
+    db_url = args.database_url
+    logging.debug(f"Database URL: {db_url}")
+    db_path = get_db_path()
+    db_exists = os.path.exists(db_path)
+
+    config = get_alembic_config()
+
+    # Check if we need to upgrade
+    engine = create_engine(db_url)
+    conn = engine.connect()
+
+    context = MigrationContext.configure(conn)
+    current_rev = context.get_current_revision()
+
+    script = ScriptDirectory.from_config(config)
+    target_rev = script.get_current_head()
+
+    if target_rev is None:
+        logging.warning("No target revision found.")
+    elif current_rev != target_rev:
+        # Backup the database pre upgrade
+        backup_path = db_path + ".bkp"
+        if db_exists:
+            shutil.copy(db_path, backup_path)
+        else:
+            backup_path = None
+
+        try:
+            command.upgrade(config, target_rev)
+            logging.info(f"Database upgraded from {current_rev} to {target_rev}")
+        except Exception as e:
+            if backup_path:
+                # Restore the database from backup if upgrade fails
+                shutil.copy(backup_path, db_path)
+                os.remove(backup_path)
+            logging.exception("Error upgrading database: ")
+            raise e
+
+    global Session
+    Session = sessionmaker(bind=engine)
+
+
+def create_session():
+    return Session()

app/database/models.py

@@ -0,0 +1,14 @@
+from sqlalchemy.orm import declarative_base
+
+Base = declarative_base()
+
+
+def to_dict(obj):
+    fields = obj.__table__.columns.keys()
+    return {
+        field: (val.to_dict() if hasattr(val, "to_dict") else val)
+        for field in fields
+        if (val := getattr(obj, field))
+    }
+
+# TODO: Define models here

app/frontend_management.py

@@ -16,40 +16,61 @@ from importlib.metadata import version
 import requests
 from typing_extensions import NotRequired
 
+from utils.install_util import get_missing_requirements_message, requirements_path
+
 from comfy.cli_args import DEFAULT_VERSION_STRING
 import app.logger
 
-# The path to the requirements.txt file
-req_path = Path(__file__).parents[1] / "requirements.txt"
-
 
 def frontend_install_warning_message():
-    """The warning message to display when the frontend version is not up to date."""
-
-    extra = ""
-    if sys.flags.no_user_site:
-        extra = "-s "
     return f"""
-Please install the updated requirements.txt file by running:
-{sys.executable} {extra}-m pip install -r {req_path}
+{get_missing_requirements_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.
-
-If you are on the portable package you can run: update\\update_comfyui.bat to solve this problem
 """.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(req_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"""
@@ -121,9 +142,22 @@ class FrontEndProvider:
         response.raise_for_status()  # Raises an HTTPError if the response was an error
         return response.json()
 
+    @cached_property
+    def latest_prerelease(self) -> Release:
+        """Get the latest pre-release version - even if it's older than the latest release"""
+        release = [release for release in self.all_releases if release["prerelease"]]
+
+        if not release:
+            raise ValueError("No pre-releases found")
+
+        # GitHub returns releases in reverse chronological order, so first is latest
+        return release[0]
+
     def get_release(self, version: str) -> Release:
         if version == "latest":
             return self.latest_release
+        elif version == "prerelease":
+            return self.latest_prerelease
         else:
             for release in self.all_releases:
                 if release["tag_name"] in [version, f"v{version}"]:
@@ -164,6 +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:
@@ -205,6 +244,19 @@ comfyui-workflow-templates is not installed.
 """.strip()
             )
 
+    @classmethod
+    def embedded_docs_path(cls) -> str:
+        """Get the path to embedded documentation"""
+        try:
+            import comfyui_embedded_docs
+
+            return str(
+                importlib.resources.files(comfyui_embedded_docs) / "docs"
+            )
+        except ImportError:
+            logging.info("comfyui-embedded-docs package not found")
+            return None
+
     @classmethod
     def parse_version_string(cls, value: str) -> tuple[str, str, str]:
         """
@@ -217,7 +269,7 @@ comfyui-workflow-templates is not installed.
         Raises:
             argparse.ArgumentTypeError: If the version string is invalid.
         """
-        VERSION_PATTERN = r"^([a-zA-Z0-9][a-zA-Z0-9-]{0,38})/([a-zA-Z0-9_.-]+)@(v?\d+\.\d+\.\d+|latest)$"
+        VERSION_PATTERN = r"^([a-zA-Z0-9][a-zA-Z0-9-]{0,38})/([a-zA-Z0-9_.-]+)@(v?\d+\.\d+\.\d+[-._a-zA-Z0-9]*|latest|prerelease)$"
         match_result = re.match(VERSION_PATTERN, value)
         if match_result is None:
             raise argparse.ArgumentTypeError(f"Invalid version string: {value}")

app/user_manager.py

@@ -197,6 +197,112 @@ class UserManager():
 
             return web.json_response(results)
 
+        @routes.get("/v2/userdata")
+        async def list_userdata_v2(request):
+            """
+            List files and directories in a user's data directory.
+
+            This endpoint provides a structured listing of contents within a specified
+            subdirectory of the user's data storage.
+
+            Query Parameters:
+            - path (optional): The relative path within the user's data directory
+                               to list. Defaults to the root ('').
+
+            Returns:
+            - 400: If the requested path is invalid, outside the user's data directory, or is not a directory.
+            - 404: If the requested path does not exist.
+            - 403: If the user is invalid.
+            - 500: If there is an error reading the directory contents.
+            - 200: JSON response containing a list of file and directory objects.
+                   Each object includes:
+                   - name: The name of the file or directory.
+                   - type: 'file' or 'directory'.
+                   - path: The relative path from the user's data root.
+                   - size (for files): The size in bytes.
+                   - modified (for files): The last modified timestamp (Unix epoch).
+            """
+            requested_rel_path = request.rel_url.query.get('path', '')
+
+            # URL-decode the path parameter
+            try:
+                requested_rel_path = parse.unquote(requested_rel_path)
+            except Exception as e:
+                logging.warning(f"Failed to decode path parameter: {requested_rel_path}, Error: {e}")
+                return web.Response(status=400, text="Invalid characters in path parameter")
+
+
+            # Check user validity and get the absolute path for the requested directory
+            try:
+                 base_user_path = self.get_request_user_filepath(request, None, create_dir=False)
+
+                 if requested_rel_path:
+                     target_abs_path = self.get_request_user_filepath(request, requested_rel_path, create_dir=False)
+                 else:
+                     target_abs_path = base_user_path
+
+            except KeyError as e:
+                 # Invalid user detected by get_request_user_id inside get_request_user_filepath
+                 logging.warning(f"Access denied for user: {e}")
+                 return web.Response(status=403, text="Invalid user specified in request")
+
+
+            if not target_abs_path:
+                 # Path traversal or other issue detected by get_request_user_filepath
+                 return web.Response(status=400, text="Invalid path requested")
+
+            # Handle cases where the user directory or target path doesn't exist
+            if not os.path.exists(target_abs_path):
+                # Check if it's the base user directory that's missing (new user case)
+                if target_abs_path == base_user_path:
+                    # It's okay if the base user directory doesn't exist yet, return empty list
+                     return web.json_response([])
+                else:
+                    # A specific subdirectory was requested but doesn't exist
+                     return web.Response(status=404, text="Requested path not found")
+
+            if not os.path.isdir(target_abs_path):
+                 return web.Response(status=400, text="Requested path is not a directory")
+
+            results = []
+            try:
+                for root, dirs, files in os.walk(target_abs_path, topdown=True):
+                    # Process directories
+                    for dir_name in dirs:
+                        dir_path = os.path.join(root, dir_name)
+                        rel_path = os.path.relpath(dir_path, base_user_path).replace(os.sep, '/')
+                        results.append({
+                            "name": dir_name,
+                            "path": rel_path,
+                            "type": "directory"
+                        })
+
+                    # Process files
+                    for file_name in files:
+                        file_path = os.path.join(root, file_name)
+                        rel_path = os.path.relpath(file_path, base_user_path).replace(os.sep, '/')
+                        entry_info = {
+                            "name": file_name,
+                            "path": rel_path,
+                            "type": "file"
+                        }
+                        try:
+                            stats = os.stat(file_path) # Use os.stat for potentially better performance with os.walk
+                            entry_info["size"] = stats.st_size
+                            entry_info["modified"] = stats.st_mtime
+                        except OSError as stat_error:
+                            logging.warning(f"Could not stat file {file_path}: {stat_error}")
+                            pass # Include file with available info
+                        results.append(entry_info)
+            except OSError as e:
+                logging.error(f"Error listing directory {target_abs_path}: {e}")
+                return web.Response(status=500, text="Error reading directory contents")
+
+            # Sort results alphabetically, directories first then files
+            results.sort(key=lambda x: (x['type'] != 'directory', x['name'].lower()))
+
+            return web.json_response(results)
+
         def get_user_data_path(request, check_exists = False, param = "file"):
             file = request.match_info.get(param, None)
             if not file:

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.")
@@ -66,6 +67,7 @@ fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the diff
 fpunet_group.add_argument("--fp16-unet", action="store_true", help="Run the diffusion model in fp16")
 fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
 fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")
+fpunet_group.add_argument("--fp8_e8m0fnu-unet", action="store_true", help="Store unet weights in fp8_e8m0fnu.")
 
 fpvae_group = parser.add_mutually_exclusive_group()
 fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")
@@ -87,6 +89,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
+parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"
@@ -127,6 +130,7 @@ vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for e
 
 parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.")
 
+parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
 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.")
 
@@ -140,12 +144,17 @@ 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.")
 parser.add_argument("--windows-standalone-build", action="store_true", help="Windows standalone build: Enable convenient things that most people using the standalone windows build will probably enjoy (like auto opening the page on startup).")
 
 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.")
 
@@ -190,6 +199,18 @@ parser.add_argument("--user-directory", type=is_valid_directory, default=None, h
 
 parser.add_argument("--enable-compress-response-body", action="store_true", help="Enable compressing response body.")
 
+parser.add_argument(
+    "--comfy-api-base",
+    type=str,
+    default="https://api.comfy.org",
+    help="Set the base URL for the ComfyUI API.  (default: https://api.comfy.org)",
+)
+
+database_default_path = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
+)
+parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
+
 if comfy.options.args_parsing:
     args = parser.parse_args()
 else:

comfy/clip_vision.py

@@ -18,6 +18,7 @@ class Output:
         setattr(self, key, item)
 
 def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
     mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
     std = torch.tensor(std, device=image.device, dtype=image.dtype)
     image = image.movedim(-1, 1)

comfy/comfy_types/node_typing.py

@@ -1,7 +1,7 @@
 """Comfy-specific type hinting"""
 
 from __future__ import annotations
-from typing import Literal, TypedDict
+from typing import Literal, TypedDict, Optional
 from typing_extensions import NotRequired
 from abc import ABC, abstractmethod
 from enum import Enum
@@ -37,6 +37,8 @@ class IO(StrEnum):
     CONTROL_NET = "CONTROL_NET"
     VAE = "VAE"
     MODEL = "MODEL"
+    LORA_MODEL = "LORA_MODEL"
+    LOSS_MAP = "LOSS_MAP"
     CLIP_VISION = "CLIP_VISION"
     CLIP_VISION_OUTPUT = "CLIP_VISION_OUTPUT"
     STYLE_MODEL = "STYLE_MODEL"
@@ -48,6 +50,7 @@ class IO(StrEnum):
     FACE_ANALYSIS = "FACE_ANALYSIS"
     BBOX = "BBOX"
     SEGS = "SEGS"
+    VIDEO = "VIDEO"
 
     ANY = "*"
     """Always matches any type, but at a price.
@@ -115,6 +118,15 @@ class InputTypeOptions(TypedDict):
     """When a link exists, rather than receiving the evaluated value, you will receive the link (i.e. `["nodeId", <outputIndex>]`). Designed for node expansion."""
     tooltip: NotRequired[str]
     """Tooltip for the input (or widget), shown on pointer hover"""
+    socketless: NotRequired[bool]
+    """All inputs (including widgets) have an input socket to connect links. When ``true``, if there is a widget for this input, no socket will be created.
+    Available from frontend v1.17.5
+    Ref: https://github.com/Comfy-Org/ComfyUI_frontend/pull/3548
+    """
+    widgetType: NotRequired[str]
+    """Specifies a type to be used for widget initialization if different from the input type.
+    Available from frontend v1.18.0
+    https://github.com/Comfy-Org/ComfyUI_frontend/pull/3550"""
     # class InputTypeNumber(InputTypeOptions):
     # default: float | int
     min: NotRequired[float]
@@ -224,6 +236,8 @@ class ComfyNodeABC(ABC):
     """Flags a node as experimental, informing users that it may change or not work as expected."""
     DEPRECATED: bool
     """Flags a node as deprecated, indicating to users that they should find alternatives to this node."""
+    API_NODE: Optional[bool]
+    """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview."""
 
     @classmethod
     @abstractmethod
@@ -262,7 +276,7 @@ class ComfyNodeABC(ABC):
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
-    OUTPUT_IS_LIST: tuple[bool]
+    OUTPUT_IS_LIST: tuple[bool, ...]
     """A tuple indicating which node outputs are lists, but will be connected to nodes that expect individual items.
 
     Connected nodes that do not implement `INPUT_IS_LIST` will be executed once for every item in the list.
@@ -281,7 +295,7 @@ class ComfyNodeABC(ABC):
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
 
-    RETURN_TYPES: tuple[IO]
+    RETURN_TYPES: tuple[IO, ...]
     """A tuple representing the outputs of this node.
 
     Usage::
@@ -290,12 +304,12 @@ class ComfyNodeABC(ABC):
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-types
     """
-    RETURN_NAMES: tuple[str]
+    RETURN_NAMES: tuple[str, ...]
     """The output slot names for each item in `RETURN_TYPES`, e.g. ``RETURN_NAMES = ("count", "filter_string")``
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-names
     """
-    OUTPUT_TOOLTIPS: tuple[str]
+    OUTPUT_TOOLTIPS: tuple[str, ...]
     """A tuple of strings to use as tooltips for node outputs, one for each item in `RETURN_TYPES`."""
     FUNCTION: str
     """The name of the function to execute as a literal string, e.g. `FUNCTION = "execute"`

comfy/conds.py

@@ -24,6 +24,10 @@ class CONDRegular:
             conds.append(x.cond)
         return torch.cat(conds)
 
+    def size(self):
+        return list(self.cond.size())
+
+
 class CONDNoiseShape(CONDRegular):
     def process_cond(self, batch_size, device, area, **kwargs):
         data = self.cond
@@ -64,6 +68,7 @@ class CONDCrossAttn(CONDRegular):
             out.append(c)
         return torch.cat(out)
 
+
 class CONDConstant(CONDRegular):
     def __init__(self, cond):
         self.cond = cond
@@ -78,3 +83,48 @@ class CONDConstant(CONDRegular):
 
     def concat(self, others):
         return self.cond
+
+    def size(self):
+        return [1]
+
+
+class CONDList(CONDRegular):
+    def __init__(self, cond):
+        self.cond = cond
+
+    def process_cond(self, batch_size, device, **kwargs):
+        out = []
+        for c in self.cond:
+            out.append(comfy.utils.repeat_to_batch_size(c, batch_size).to(device))
+
+        return self._copy_with(out)
+
+    def can_concat(self, other):
+        if len(self.cond) != len(other.cond):
+            return False
+        for i in range(len(self.cond)):
+            if self.cond[i].shape != other.cond[i].shape:
+                return False
+
+        return True
+
+    def concat(self, others):
+        out = []
+        for i in range(len(self.cond)):
+            o = [self.cond[i]]
+            for x in others:
+                o.append(x.cond[i])
+            out.append(torch.cat(o))
+
+        return out
+
+    def size(self):  # hackish implementation to make the mem estimation work
+        o = 0
+        c = 1
+        for c in self.cond:
+            size = c.size()
+            o += math.prod(size)
+            if len(size) > 1:
+                c = size[1]
+
+        return [1, c, o // c]

comfy/controlnet.py

@@ -390,8 +390,9 @@ class ControlLora(ControlNet):
                 pass
 
         for k in self.control_weights:
-            if k not in {"lora_controlnet"}:
-                comfy.utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device()))
+            if (k not in {"lora_controlnet"}):
+                if (k.endswith(".up") or k.endswith(".down") or k.endswith(".weight") or k.endswith(".bias")) and ("__" not in k):
+                    comfy.utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device()))
 
     def copy(self):
         c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling)
@@ -736,6 +737,7 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
     return control
 
 def load_controlnet(ckpt_path, model=None, model_options={}):
+    model_options = model_options.copy()
     if "global_average_pooling" not in model_options:
         filename = os.path.splitext(ckpt_path)[0]
         if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling

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/image_encoders/dino2.py

@@ -116,7 +116,7 @@ class Dino2Embeddings(torch.nn.Module):
     def forward(self, pixel_values):
         x = self.patch_embeddings(pixel_values)
         # TODO: mask_token?
-        x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
+        x = torch.cat((self.cls_token.to(device=x.device, dtype=x.dtype).expand(x.shape[0], -1, -1), x), dim=1)
         x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype)
         return x
 

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

@@ -1,4 +1,5 @@
 import math
+from functools import partial
 
 from scipy import integrate
 import torch
@@ -8,6 +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
 
@@ -142,6 +144,33 @@ class BrownianTreeNoiseSampler:
         return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
 
 
+def sigma_to_half_log_snr(sigma, model_sampling):
+    """Convert sigma to half-logSNR log(alpha_t / sigma_t)."""
+    if isinstance(model_sampling, comfy.model_sampling.CONST):
+        # log((1 - t) / t) = log((1 - sigma) / sigma)
+        return sigma.logit().neg()
+    return sigma.log().neg()
+
+
+def half_log_snr_to_sigma(half_log_snr, model_sampling):
+    """Convert half-logSNR log(alpha_t / sigma_t) to sigma."""
+    if isinstance(model_sampling, comfy.model_sampling.CONST):
+        # 1 / (1 + exp(half_log_snr))
+        return half_log_snr.neg().sigmoid()
+    return half_log_snr.neg().exp()
+
+
+def offset_first_sigma_for_snr(sigmas, model_sampling, percent_offset=1e-4):
+    """Adjust the first sigma to avoid invalid logSNR."""
+    if len(sigmas) <= 1:
+        return sigmas
+    if isinstance(model_sampling, comfy.model_sampling.CONST):
+        if sigmas[0] >= 1:
+            sigmas = sigmas.clone()
+            sigmas[0] = model_sampling.percent_to_sigma(percent_offset)
+    return sigmas
+
+
 @torch.no_grad()
 def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
     """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
@@ -384,9 +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
 
 
@@ -682,6 +715,7 @@ def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=Non
         # logged_x = torch.cat((logged_x, x.unsqueeze(0)), dim=0)
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     """DPM-Solver++ (stochastic)."""
@@ -693,38 +727,49 @@ def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=N
     seed = extra_args.get("seed", None)
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
     s_in = x.new_ones([x.shape[0]])
-    sigma_fn = lambda t: t.neg().exp()
-    t_fn = lambda sigma: sigma.log().neg()
+
+    model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
+    sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling)
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
 
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
         if sigmas[i + 1] == 0:
-            # Euler method
-            d = to_d(x, sigmas[i], denoised)
-            dt = sigmas[i + 1] - sigmas[i]
-            x = x + d * dt
+            # Denoising step
+            x = denoised
         else:
             # DPM-Solver++
-            t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
-            h = t_next - t
-            s = t + h * r
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
+            lambda_s_1 = lambda_s + r * h
             fac = 1 / (2 * r)
 
+            sigma_s_1 = sigma_fn(lambda_s_1)
+
+            alpha_s = sigmas[i] * lambda_s.exp()
+            alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
             # Step 1
-            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta)
-            s_ = t_fn(sd)
-            x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised
-            x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su
-            denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
+            sd, su = get_ancestral_step(lambda_s.neg().exp(), lambda_s_1.neg().exp(), eta)
+            lambda_s_1_ = sd.log().neg()
+            h_ = lambda_s_1_ - lambda_s
+            x_2 = (alpha_s_1 / alpha_s) * (-h_).exp() * x - alpha_s_1 * (-h_).expm1() * denoised
+            if eta > 0 and s_noise > 0:
+                x_2 = x_2 + alpha_s_1 * noise_sampler(sigmas[i], sigma_s_1) * s_noise * su
+            denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
             # Step 2
-            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta)
-            t_next_ = t_fn(sd)
+            sd, su = get_ancestral_step(lambda_s.neg().exp(), lambda_t.neg().exp(), eta)
+            lambda_t_ = sd.log().neg()
+            h_ = lambda_t_ - lambda_s
             denoised_d = (1 - fac) * denoised + fac * denoised_2
-            x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d
-            x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su
+            x = (alpha_t / alpha_s) * (-h_).exp() * x - alpha_t * (-h_).expm1() * denoised_d
+            if eta > 0 and s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * su
     return x
 
 
@@ -753,6 +798,7 @@ def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=No
         old_denoised = denoised
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     """DPM-Solver++(2M) SDE."""
@@ -768,9 +814,12 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) 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')
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+
     old_denoised = None
-    h_last = None
-    h = None
+    h, h_last = None, None
 
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
@@ -781,26 +830,29 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
             x = denoised
         else:
             # DPM-Solver++(2M) SDE
-            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = s - t
-            eta_h = eta * h
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
+            h_eta = h * (eta + 1)
 
-            x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_t * (-h_eta).expm1().neg() * denoised
 
             if old_denoised is not None:
                 r = h_last / h
                 if solver_type == 'heun':
-                    x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)
+                    x = x + alpha_t * ((-h_eta).expm1().neg() / (-h_eta) + 1) * (1 / r) * (denoised - old_denoised)
                 elif solver_type == 'midpoint':
-                    x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)
+                    x = x + 0.5 * alpha_t * (-h_eta).expm1().neg() * (1 / r) * (denoised - old_denoised)
 
-            if eta:
-                x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
+            if eta > 0 and s_noise > 0:
+                x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
 
         old_denoised = denoised
         h_last = h
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """DPM-Solver++(3M) SDE."""
@@ -814,6 +866,10 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) 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')
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+
     denoised_1, denoised_2 = None, None
     h, h_1, h_2 = None, None, None
 
@@ -825,13 +881,16 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
             # Denoising step
             x = denoised
         else:
-            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = s - t
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
             h_eta = h * (eta + 1)
 
-            x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_t * (-h_eta).expm1().neg() * denoised
 
             if h_2 is not None:
+                # DPM-Solver++(3M) SDE
                 r0 = h_1 / h
                 r1 = h_2 / h
                 d1_0 = (denoised - denoised_1) / r0
@@ -840,20 +899,22 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
                 d2 = (d1_0 - d1_1) / (r0 + r1)
                 phi_2 = h_eta.neg().expm1() / h_eta + 1
                 phi_3 = phi_2 / h_eta - 0.5
-                x = x + phi_2 * d1 - phi_3 * d2
+                x = x + (alpha_t * phi_2) * d1 - (alpha_t * phi_3) * d2
             elif h_1 is not None:
+                # DPM-Solver++(2M) SDE
                 r = h_1 / h
                 d = (denoised - denoised_1) / r
                 phi_2 = h_eta.neg().expm1() / h_eta + 1
-                x = x + phi_2 * d
+                x = x + (alpha_t * phi_2) * d
 
-            if eta:
+            if eta > 0 and s_noise > 0:
                 x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
 
         denoised_1, denoised_2 = denoised, denoised_1
         h_1, h_2 = h, h_1
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     if len(sigmas) <= 1:
@@ -863,6 +924,7 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
 
+
 @torch.no_grad()
 def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     if len(sigmas) <= 1:
@@ -872,6 +934,7 @@ def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
 
+
 @torch.no_grad()
 def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     if len(sigmas) <= 1:
@@ -1009,7 +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
@@ -1027,6 +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):
@@ -1050,7 +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)
@@ -1090,6 +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()
@@ -1140,39 +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()
@@ -1181,15 +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."""
@@ -1277,6 +1347,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
     phi1_fn = lambda t: torch.expm1(t) / t
     phi2_fn = lambda t: (phi1_fn(t) - 1.0) / t
 
+    old_sigma_down = None
     old_denoised = None
     uncond_denoised = None
     def post_cfg_function(args):
@@ -1304,9 +1375,9 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
                 x = x + d * dt
         else:
             # Second order multistep method in https://arxiv.org/pdf/2308.02157
-            t, t_next, t_prev = t_fn(sigmas[i]), t_fn(sigma_down), t_fn(sigmas[i - 1])
+            t, t_old, t_next, t_prev = t_fn(sigmas[i]), t_fn(old_sigma_down), t_fn(sigma_down), t_fn(sigmas[i - 1])
             h = t_next - t
-            c2 = (t_prev - t) / h
+            c2 = (t_prev - t_old) / h
 
             phi1_val, phi2_val = phi1_fn(-h), phi2_fn(-h)
             b1 = torch.nan_to_num(phi1_val - phi2_val / c2, nan=0.0)
@@ -1326,6 +1397,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
             old_denoised = uncond_denoised
         else:
             old_denoised = denoised
+        old_sigma_down = sigma_down
     return x
 
 @torch.no_grad()
@@ -1344,33 +1416,59 @@ 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.):
+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"""
     extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
     old_d = None
 
+    uncond_denoised = None
+    def post_cfg_function(args):
+        nonlocal uncond_denoised
+        uncond_denoised = args["uncond_denoised"]
+        return args["denoised"]
+
+    if cfg_pp:
+        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)
+
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
-        d = to_d(x, sigmas[i], denoised)
+        if cfg_pp:
+            d = to_d(x, sigmas[i], uncond_denoised)
+        else:
+            d = to_d(x, sigmas[i], denoised)
         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:
-            # Euler method
-            x = x + d * dt
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
         else:
-            # Gradient estimation
-            d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
-            x = x + d_bar * dt
+            # Euler method
+            if cfg_pp:
+                x = denoised + d * sigmas[i + 1]
+            else:
+                x = x + d * dt
+
+            if i >= 1:
+                # Gradient estimation
+                d_bar = (ge_gamma - 1) * (d - old_d)
+                x = x + d_bar * dt
         old_d = d
     return x
 
+
 @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_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=1.0, 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
@@ -1378,12 +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
 
@@ -1394,41 +1498,45 @@ 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
 
+
 @torch.no_grad()
 def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
-    '''
-    SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VE Data Prediction) stage 2
-    Arxiv: https://arxiv.org/abs/2305.14267
-    '''
+    """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
+    arXiv: https://arxiv.org/abs/2305.14267
+    """
     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
@@ -1436,6 +1544,11 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
 
     inject_noise = eta > 0 and s_noise > 0
 
+    model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
+    sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling)
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
         if callback is not None:
@@ -1443,80 +1556,206 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
         if sigmas[i + 1] == 0:
             x = denoised
         else:
-            t, t_next = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = t_next - t
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
             h_eta = h * (eta + 1)
-            s = t + r * h
+            lambda_s_1 = lambda_s + r * h
             fac = 1 / (2 * r)
-            sigma_s = s.neg().exp()
+            sigma_s_1 = sigma_fn(lambda_s_1)
+
+            # alpha_t = sigma_t * exp(log(alpha_t / sigma_t)) = sigma_t * exp(lambda_t)
+            alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
 
             coeff_1, coeff_2 = (-r * h_eta).expm1(), (-h_eta).expm1()
             if inject_noise:
+                # 0 < r < 1
                 noise_coeff_1 = (-2 * r * h * eta).expm1().neg().sqrt()
-                noise_coeff_2 = ((-2 * r * h * eta).expm1() - (-2 * h * eta).expm1()).sqrt()
-                noise_1, noise_2 = noise_sampler(sigmas[i], sigma_s), noise_sampler(sigma_s, sigmas[i + 1])
+                noise_coeff_2 = (-r * h * eta).exp() * (-2 * (1 - r) * h * eta).expm1().neg().sqrt()
+                noise_1, noise_2 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigmas[i + 1])
 
             # Step 1
-            x_2 = (coeff_1 + 1) * x - coeff_1 * denoised
-            if inject_noise:
-                x_2 = x_2 + sigma_s * (noise_coeff_1 * noise_1) * s_noise
-            denoised_2 = model(x_2, sigma_s * s_in, **extra_args)
-
-            # Step 2
-            denoised_d = (1 - fac) * denoised + fac * denoised_2
-            x = (coeff_2 + 1) * x - coeff_2 * denoised_d
-            if inject_noise:
-                x = x + sigmas[i + 1] * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
-    return x
-
-@torch.no_grad()
-def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
-    '''
-    SEEDS-3 - Stochastic Explicit Exponential Derivative-free Solvers (VE Data Prediction) stage 3
-    Arxiv: https://arxiv.org/abs/2305.14267
-    '''
-    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]])
-
-    inject_noise = eta > 0 and s_noise > 0
-
-    for i in trange(len(sigmas) - 1, disable=disable):
-        denoised = model(x, sigmas[i] * s_in, **extra_args)
-        if callback is not None:
-            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
-        if sigmas[i + 1] == 0:
-            x = denoised
-        else:
-            t, t_next = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = t_next - t
-            h_eta = h * (eta + 1)
-            s_1 = t + r_1 * h
-            s_2 = t + r_2 * h
-            sigma_s_1, sigma_s_2 = s_1.neg().exp(), s_2.neg().exp()
-
-            coeff_1, coeff_2, coeff_3 = (-r_1 * h_eta).expm1(), (-r_2 * h_eta).expm1(), (-h_eta).expm1()
-            if inject_noise:
-                noise_coeff_1 = (-2 * r_1 * h * eta).expm1().neg().sqrt()
-                noise_coeff_2 = ((-2 * r_1 * h * eta).expm1() - (-2 * r_2 * h * eta).expm1()).sqrt()
-                noise_coeff_3 = ((-2 * r_2 * h * eta).expm1() - (-2 * h * eta).expm1()).sqrt()
-                noise_1, noise_2, noise_3 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigma_s_2), noise_sampler(sigma_s_2, sigmas[i + 1])
-
-            # Step 1
-            x_2 = (coeff_1 + 1) * x - coeff_1 * denoised
+            x_2 = sigma_s_1 / sigmas[i] * (-r * h * eta).exp() * x - alpha_s_1 * coeff_1 * denoised
             if inject_noise:
                 x_2 = x_2 + sigma_s_1 * (noise_coeff_1 * noise_1) * s_noise
             denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
             # Step 2
-            x_3 = (coeff_2 + 1) * x - coeff_2 * denoised + (r_2 / r_1) * (coeff_2 / (r_2 * h_eta) + 1) * (denoised_2 - denoised)
+            denoised_d = (1 - fac) * denoised + fac * denoised_2
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * coeff_2 * denoised_d
+            if inject_noise:
+                x = x + sigmas[i + 1] * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
+    return x
+
+
+@torch.no_grad()
+def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
+    """SEEDS-3 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 3.
+    arXiv: https://arxiv.org/abs/2305.14267
+    """
+    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]])
+
+    inject_noise = eta > 0 and s_noise > 0
+
+    model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
+    sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling)
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            x = denoised
+        else:
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
+            h_eta = h * (eta + 1)
+            lambda_s_1 = lambda_s + r_1 * h
+            lambda_s_2 = lambda_s + r_2 * h
+            sigma_s_1, sigma_s_2 = sigma_fn(lambda_s_1), sigma_fn(lambda_s_2)
+
+            # alpha_t = sigma_t * exp(log(alpha_t / sigma_t)) = sigma_t * exp(lambda_t)
+            alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+            alpha_s_2 = sigma_s_2 * lambda_s_2.exp()
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
+            coeff_1, coeff_2, coeff_3 = (-r_1 * h_eta).expm1(), (-r_2 * h_eta).expm1(), (-h_eta).expm1()
+            if inject_noise:
+                # 0 < r_1 < r_2 < 1
+                noise_coeff_1 = (-2 * r_1 * h * eta).expm1().neg().sqrt()
+                noise_coeff_2 = (-r_1 * h * eta).exp() * (-2 * (r_2 - r_1) * h * eta).expm1().neg().sqrt()
+                noise_coeff_3 = (-r_2 * h * eta).exp() * (-2 * (1 - r_2) * h * eta).expm1().neg().sqrt()
+                noise_1, noise_2, noise_3 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigma_s_2), noise_sampler(sigma_s_2, sigmas[i + 1])
+
+            # Step 1
+            x_2 = sigma_s_1 / sigmas[i] * (-r_1 * h * eta).exp() * x - alpha_s_1 * coeff_1 * denoised
+            if inject_noise:
+                x_2 = x_2 + sigma_s_1 * (noise_coeff_1 * noise_1) * s_noise
+            denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
+
+            # Step 2
+            x_3 = sigma_s_2 / sigmas[i] * (-r_2 * h * eta).exp() * x - alpha_s_2 * coeff_2 * denoised + (r_2 / r_1) * alpha_s_2 * (coeff_2 / (r_2 * h_eta) + 1) * (denoised_2 - denoised)
             if inject_noise:
                 x_3 = x_3 + sigma_s_2 * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
             denoised_3 = model(x_3, sigma_s_2 * s_in, **extra_args)
 
             # Step 3
-            x = (coeff_3 + 1) * x - coeff_3 * denoised + (1. / r_2) * (coeff_3 / h_eta + 1) * (denoised_3 - denoised)
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * coeff_3 * denoised + (1. / r_2) * alpha_t * (coeff_3 / h_eta + 1) * (denoised_3 - denoised)
             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
@@ -466,3 +542,7 @@ class Hunyuan3Dv2mini(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1
     scale_factor = 1.0188137142395404
+
+class ACEAudio(LatentFormat):
+    latent_channels = 8
+    latent_dimensions = 2

comfy/ldm/ace/attention.py

@@ -0,0 +1,761 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/attention.py
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Tuple, Union, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+import comfy.model_management
+from comfy.ldm.modules.attention import optimized_attention
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        kv_heads: Optional[int] = None,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        qk_norm: Optional[str] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        out_bias: bool = True,
+        scale_qk: bool = True,
+        only_cross_attention: bool = False,
+        eps: float = 1e-5,
+        rescale_output_factor: float = 1.0,
+        residual_connection: bool = False,
+        processor=None,
+        out_dim: int = None,
+        out_context_dim: int = None,
+        context_pre_only=None,
+        pre_only=False,
+        elementwise_affine: bool = True,
+        is_causal: bool = False,
+        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 if kv_heads is None else dim_head * kv_heads
+        self.query_dim = query_dim
+        self.use_bias = bias
+        self.is_cross_attention = cross_attention_dim is not None
+        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+        self.rescale_output_factor = rescale_output_factor
+        self.residual_connection = residual_connection
+        self.dropout = dropout
+        self.fused_projections = False
+        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.context_pre_only = context_pre_only
+        self.pre_only = pre_only
+        self.is_causal = is_causal
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.only_cross_attention = only_cross_attention
+
+        if self.added_kv_proj_dim is None and self.only_cross_attention:
+            raise ValueError(
+                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
+            )
+
+        self.group_norm = None
+        self.spatial_norm = None
+
+        self.norm_q = None
+        self.norm_k = None
+
+        self.norm_cross = None
+        self.to_q = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device)
+
+        if not self.only_cross_attention:
+            # only relevant for the `AddedKVProcessor` classes
+            self.to_k = operations.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+            self.to_v = operations.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+        else:
+            self.to_k = None
+            self.to_v = None
+
+        self.added_proj_bias = added_proj_bias
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = operations.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias, dtype=dtype, device=device)
+            self.add_v_proj = operations.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias, dtype=dtype, device=device)
+            if self.context_pre_only is not None:
+                self.add_q_proj = operations.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias, dtype=dtype, device=device)
+        else:
+            self.add_q_proj = None
+            self.add_k_proj = None
+            self.add_v_proj = None
+
+        if not self.pre_only:
+            self.to_out = nn.ModuleList([])
+            self.to_out.append(operations.Linear(self.inner_dim, self.out_dim, bias=out_bias, dtype=dtype, device=device))
+            self.to_out.append(nn.Dropout(dropout))
+        else:
+            self.to_out = None
+
+        if self.context_pre_only is not None and not self.context_pre_only:
+            self.to_add_out = operations.Linear(self.inner_dim, self.out_context_dim, bias=out_bias, dtype=dtype, device=device)
+        else:
+            self.to_add_out = None
+
+        self.norm_added_q = None
+        self.norm_added_k = None
+        self.processor = processor
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **cross_attention_kwargs,
+    ) -> torch.Tensor:
+        return self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+
+class CustomLiteLAProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections. add rms norm for query and key and apply RoPE"""
+
+    def __init__(self):
+        self.kernel_func = nn.ReLU(inplace=False)
+        self.eps = 1e-15
+        self.pad_val = 1.0
+
+    def apply_rotary_emb(
+        self,
+        x: torch.Tensor,
+        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+        to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+        reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+        tensors contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            x (`torch.Tensor`):
+                Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+            freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+        """
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
+
+        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+        return out
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        hidden_states_len = hidden_states.shape[1]
+
+        input_ndim = hidden_states.ndim
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+        if encoder_hidden_states is not None:
+            context_input_ndim = encoder_hidden_states.ndim
+            if context_input_ndim == 4:
+                batch_size, channel, height, width = encoder_hidden_states.shape
+                encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size = hidden_states.shape[0]
+
+        # `sample` projections.
+        dtype = hidden_states.dtype
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        # `context` projections.
+        has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj")
+        if encoder_hidden_states is not None and has_encoder_hidden_state_proj:
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+            # attention
+            if not attn.is_cross_attention:
+                query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
+                key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
+                value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
+            else:
+                query = hidden_states
+                key = encoder_hidden_states
+                value = encoder_hidden_states
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1)
+        key = key.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1).transpose(-1, -2)
+        value = value.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1)
+
+        # RoPE需要 [B, H, S, D] 输入
+        # 此时 query是 [B, H, D, S], 需要转成 [B, H, S, D] 才能应用RoPE
+        query = query.permute(0, 1, 3, 2)  # [B, H, S, D]  (从 [B, H, D, S])
+
+        # Apply query and key normalization if needed
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if rotary_freqs_cis is not None:
+            query = self.apply_rotary_emb(query, rotary_freqs_cis)
+            if not attn.is_cross_attention:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis)
+            elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis_cross)
+
+        # 此时 query是 [B, H, S, D]，需要还原成 [B, H, D, S]
+        query = query.permute(0, 1, 3, 2)  # [B, H, D, S]
+
+        if attention_mask is not None:
+            # attention_mask: [B, S] -> [B, 1, S, 1]
+            attention_mask = attention_mask[:, None, :, None].to(key.dtype)  # [B, 1, S, 1]
+            query = query * attention_mask.permute(0, 1, 3, 2)  # [B, H, S, D] * [B, 1, S, 1]
+            if not attn.is_cross_attention:
+                key = key * attention_mask  # key: [B, h, S, D] 与 mask [B, 1, S, 1] 相乘
+                value = value * attention_mask.permute(0, 1, 3, 2)  # 如果 value 是 [B, h, D, S]，那么需调整mask以匹配S维度
+
+        if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj:
+            encoder_attention_mask = encoder_attention_mask[:, None, :, None].to(key.dtype)  # [B, 1, S_enc, 1]
+            # 此时 key: [B, h, S_enc, D], value: [B, h, D, S_enc]
+            key = key * encoder_attention_mask  # [B, h, S_enc, D] * [B, 1, S_enc, 1]
+            value = value * encoder_attention_mask.permute(0, 1, 3, 2)  # [B, h, D, S_enc] * [B, 1, 1, S_enc]
+
+        query = self.kernel_func(query)
+        key = self.kernel_func(key)
+
+        query, key, value = query.float(), key.float(), value.float()
+
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=self.pad_val)
+
+        vk = torch.matmul(value, key)
+
+        hidden_states = torch.matmul(vk, query)
+
+        if hidden_states.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.float()
+
+        hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps)
+
+        hidden_states = hidden_states.view(batch_size, attn.heads * head_dim, -1).permute(0, 2, 1)
+
+        hidden_states = hidden_states.to(dtype)
+        if encoder_hidden_states is not None:
+            encoder_hidden_states = encoder_hidden_states.to(dtype)
+
+        # Split the attention outputs.
+        if encoder_hidden_states is not None and not attn.is_cross_attention and has_encoder_hidden_state_proj:
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, : hidden_states_len],
+                hidden_states[:, hidden_states_len:],
+            )
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+        if encoder_hidden_states is not None and not attn.context_pre_only and not attn.is_cross_attention and hasattr(attn, "to_add_out"):
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if encoder_hidden_states is not None and context_input_ndim == 4:
+            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if torch.get_autocast_gpu_dtype() == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+            if encoder_hidden_states is not None:
+                encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+
+        return hidden_states, encoder_hidden_states
+
+
+class CustomerAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def apply_rotary_emb(
+        self,
+        x: torch.Tensor,
+        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+        to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+        reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+        tensors contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            x (`torch.Tensor`):
+                Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+            freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+        """
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
+
+        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+        return out
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+
+        residual = hidden_states
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj")
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if rotary_freqs_cis is not None:
+            query = self.apply_rotary_emb(query, rotary_freqs_cis)
+            if not attn.is_cross_attention:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis)
+            elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis_cross)
+
+        if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj:
+            # attention_mask: N x S1
+            # encoder_attention_mask: N x S2
+            # cross attention 整合attention_mask和encoder_attention_mask
+            combined_mask = attention_mask[:, :, None] * encoder_attention_mask[:, None, :]
+            attention_mask = torch.where(combined_mask == 1, 0.0, -torch.inf)
+            attention_mask = attention_mask[:, None, :, :].expand(-1, attn.heads, -1, -1).to(query.dtype)
+
+        elif not attn.is_cross_attention and attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        hidden_states = optimized_attention(
+            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True,
+        ).to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+def val2list(x: list or tuple or any, repeat_time=1) -> list:  # type: ignore
+    """Repeat `val` for `repeat_time` times and return the list or val if list/tuple."""
+    if isinstance(x, (list, tuple)):
+        return list(x)
+    return [x for _ in range(repeat_time)]
+
+
+def val2tuple(x: list or tuple or any, min_len: int = 1, idx_repeat: int = -1) -> tuple:  # type: ignore
+    """Return tuple with min_len by repeating element at idx_repeat."""
+    # convert to list first
+    x = val2list(x)
+
+    # repeat elements if necessary
+    if len(x) > 0:
+        x[idx_repeat:idx_repeat] = [x[idx_repeat] for _ in range(min_len - len(x))]
+
+    return tuple(x)
+
+
+def t2i_modulate(x, shift, scale):
+    return x * (1 + scale) + shift
+
+
+def get_same_padding(kernel_size: Union[int, Tuple[int, ...]]) -> Union[int, Tuple[int, ...]]:
+    if isinstance(kernel_size, tuple):
+        return tuple([get_same_padding(ks) for ks in kernel_size])
+    else:
+        assert kernel_size % 2 > 0, f"kernel size {kernel_size} should be odd number"
+        return kernel_size // 2
+
+class ConvLayer(nn.Module):
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        kernel_size=3,
+        stride=1,
+        dilation=1,
+        groups=1,
+        padding: Union[int, None] = None,
+        use_bias=False,
+        norm=None,
+        act=None,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        if padding is None:
+            padding = get_same_padding(kernel_size)
+            padding *= dilation
+
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.dilation = dilation
+        self.groups = groups
+        self.padding = padding
+        self.use_bias = use_bias
+
+        self.conv = operations.Conv1d(
+            in_dim,
+            out_dim,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=use_bias,
+            device=device,
+            dtype=dtype
+        )
+        if norm is not None:
+            self.norm = operations.RMSNorm(out_dim, elementwise_affine=False, dtype=dtype, device=device)
+        else:
+            self.norm = None
+        if act is not None:
+            self.act = nn.SiLU(inplace=True)
+        else:
+            self.act = None
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.conv(x)
+        if self.norm:
+            x = self.norm(x)
+        if self.act:
+            x = self.act(x)
+        return x
+
+
+class GLUMBConv(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: int,
+        out_feature=None,
+        kernel_size=3,
+        stride=1,
+        padding: Union[int, None] = None,
+        use_bias=False,
+        norm=(None, None, None),
+        act=("silu", "silu", None),
+        dilation=1,
+        dtype=None, device=None, operations=None
+    ):
+        out_feature = out_feature or in_features
+        super().__init__()
+        use_bias = val2tuple(use_bias, 3)
+        norm = val2tuple(norm, 3)
+        act = val2tuple(act, 3)
+
+        self.glu_act = nn.SiLU(inplace=False)
+        self.inverted_conv = ConvLayer(
+            in_features,
+            hidden_features * 2,
+            1,
+            use_bias=use_bias[0],
+            norm=norm[0],
+            act=act[0],
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.depth_conv = ConvLayer(
+            hidden_features * 2,
+            hidden_features * 2,
+            kernel_size,
+            stride=stride,
+            groups=hidden_features * 2,
+            padding=padding,
+            use_bias=use_bias[1],
+            norm=norm[1],
+            act=None,
+            dilation=dilation,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.point_conv = ConvLayer(
+            hidden_features,
+            out_feature,
+            1,
+            use_bias=use_bias[2],
+            norm=norm[2],
+            act=act[2],
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.transpose(1, 2)
+        x = self.inverted_conv(x)
+        x = self.depth_conv(x)
+
+        x, gate = torch.chunk(x, 2, dim=1)
+        gate = self.glu_act(gate)
+        x = x * gate
+
+        x = self.point_conv(x)
+        x = x.transpose(1, 2)
+
+        return x
+
+
+class LinearTransformerBlock(nn.Module):
+    """
+    A Sana block with global shared adaptive layer norm (adaLN-single) conditioning.
+    """
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        use_adaln_single=True,
+        cross_attention_dim=None,
+        added_kv_proj_dim=None,
+        context_pre_only=False,
+        mlp_ratio=4.0,
+        add_cross_attention=False,
+        add_cross_attention_dim=None,
+        qk_norm=None,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.norm1 = operations.RMSNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            added_kv_proj_dim=added_kv_proj_dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=True,
+            qk_norm=qk_norm,
+            processor=CustomLiteLAProcessor2_0(),
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.add_cross_attention = add_cross_attention
+        self.context_pre_only = context_pre_only
+
+        if add_cross_attention and add_cross_attention_dim is not None:
+            self.cross_attn = Attention(
+                query_dim=dim,
+                cross_attention_dim=add_cross_attention_dim,
+                added_kv_proj_dim=add_cross_attention_dim,
+                dim_head=attention_head_dim,
+                heads=num_attention_heads,
+                out_dim=dim,
+                context_pre_only=context_pre_only,
+                bias=True,
+                qk_norm=qk_norm,
+                processor=CustomerAttnProcessor2_0(),
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+
+        self.norm2 = operations.RMSNorm(dim, 1e-06, elementwise_affine=False)
+
+        self.ff = GLUMBConv(
+            in_features=dim,
+            hidden_features=int(dim * mlp_ratio),
+            use_bias=(True, True, False),
+            norm=(None, None, None),
+            act=("silu", "silu", None),
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.use_adaln_single = use_adaln_single
+        if use_adaln_single:
+            self.scale_shift_table = nn.Parameter(torch.empty(6, dim, dtype=dtype, device=device))
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: torch.FloatTensor = None,
+        encoder_attention_mask: torch.FloatTensor = None,
+        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        temb: torch.FloatTensor = None,
+    ):
+
+        N = hidden_states.shape[0]
+
+        # step 1: AdaLN single
+        if self.use_adaln_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                comfy.model_management.cast_to(self.scale_shift_table[None], dtype=temb.dtype, device=temb.device) + temb.reshape(N, 6, -1)
+            ).chunk(6, dim=1)
+
+        norm_hidden_states = self.norm1(hidden_states)
+        if self.use_adaln_single:
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+        # step 2: attention
+        if not self.add_cross_attention:
+            attn_output, encoder_hidden_states = self.attn(
+                hidden_states=norm_hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+            )
+        else:
+            attn_output, _ = self.attn(
+                hidden_states=norm_hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=None,
+                encoder_attention_mask=None,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=None,
+            )
+
+        if self.use_adaln_single:
+            attn_output = gate_msa * attn_output
+        hidden_states = attn_output + hidden_states
+
+        if self.add_cross_attention:
+            attn_output = self.cross_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # step 3: add norm
+        norm_hidden_states = self.norm2(hidden_states)
+        if self.use_adaln_single:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        # step 4: feed forward
+        ff_output = self.ff(norm_hidden_states)
+        if self.use_adaln_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = hidden_states + ff_output
+
+        return hidden_states

comfy/ldm/ace/model.py

@@ -0,0 +1,385 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/ace_step_transformer.py
+
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, List, Union
+
+import torch
+from torch import nn
+
+import comfy.model_management
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+from .attention import LinearTransformerBlock, t2i_modulate
+from .lyric_encoder import ConformerEncoder as LyricEncoder
+
+
+def cross_norm(hidden_states, controlnet_input):
+    # input N x T x c
+    mean_hidden_states, std_hidden_states = hidden_states.mean(dim=(1,2), keepdim=True), hidden_states.std(dim=(1,2), keepdim=True)
+    mean_controlnet_input, std_controlnet_input = controlnet_input.mean(dim=(1,2), keepdim=True), controlnet_input.std(dim=(1,2), keepdim=True)
+    controlnet_input = (controlnet_input - mean_controlnet_input) * (std_hidden_states / (std_controlnet_input + 1e-12)) + mean_hidden_states
+    return controlnet_input
+
+
+# Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2
+class Qwen2RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, dtype=None, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=device).float() / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self._set_cos_sin_cache(
+            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.float32
+        )
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
+
+        freqs = torch.outer(t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype),
+            self.sin_cached[:seq_len].to(dtype=x.dtype),
+        )
+
+
+class T2IFinalLayer(nn.Module):
+    """
+    The final layer of Sana.
+    """
+
+    def __init__(self, hidden_size, patch_size=[16, 1], out_channels=256, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, patch_size[0] * patch_size[1] * out_channels, bias=True, dtype=dtype, device=device)
+        self.scale_shift_table = nn.Parameter(torch.empty(2, hidden_size, dtype=dtype, device=device))
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+
+    def unpatchfy(
+        self,
+        hidden_states: torch.Tensor,
+        width: int,
+    ):
+        # 4 unpatchify
+        new_height, new_width = 1, hidden_states.size(1)
+        hidden_states = hidden_states.reshape(
+            shape=(hidden_states.shape[0], new_height, new_width, self.patch_size[0], self.patch_size[1], self.out_channels)
+        ).contiguous()
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(hidden_states.shape[0], self.out_channels, new_height * self.patch_size[0], new_width * self.patch_size[1])
+        ).contiguous()
+        if width > new_width:
+            output = torch.nn.functional.pad(output, (0, width - new_width, 0, 0), 'constant', 0)
+        elif width < new_width:
+            output = output[:, :, :, :width]
+        return output
+
+    def forward(self, x, t, output_length):
+        shift, scale = (comfy.model_management.cast_to(self.scale_shift_table[None], device=t.device, dtype=t.dtype) + t[:, None]).chunk(2, dim=1)
+        x = t2i_modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        # unpatchify
+        output = self.unpatchfy(x, output_length)
+        return output
+
+
+class PatchEmbed(nn.Module):
+    """2D Image to Patch Embedding"""
+
+    def __init__(
+        self,
+        height=16,
+        width=4096,
+        patch_size=(16, 1),
+        in_channels=8,
+        embed_dim=1152,
+        bias=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        patch_size_h, patch_size_w = patch_size
+        self.early_conv_layers = nn.Sequential(
+            operations.Conv2d(in_channels, in_channels*256, kernel_size=patch_size, stride=patch_size, padding=0, bias=bias, dtype=dtype, device=device),
+            operations.GroupNorm(num_groups=32, num_channels=in_channels*256, eps=1e-6, affine=True, dtype=dtype, device=device),
+            operations.Conv2d(in_channels*256, embed_dim, kernel_size=1, stride=1, padding=0, bias=bias, dtype=dtype, device=device)
+        )
+        self.patch_size = patch_size
+        self.height, self.width = height // patch_size_h, width // patch_size_w
+        self.base_size = self.width
+
+    def forward(self, latent):
+        # early convolutions, N x C x H x W -> N x 256 * sqrt(patch_size) x H/patch_size x W/patch_size
+        latent = self.early_conv_layers(latent)
+        latent = latent.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        return latent
+
+
+class ACEStepTransformer2DModel(nn.Module):
+    # _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: Optional[int] = 8,
+        num_layers: int = 28,
+        inner_dim: int = 1536,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 24,
+        mlp_ratio: float = 4.0,
+        out_channels: int = 8,
+        max_position: int = 32768,
+        rope_theta: float = 1000000.0,
+        speaker_embedding_dim: int = 512,
+        text_embedding_dim: int = 768,
+        ssl_encoder_depths: List[int] = [9, 9],
+        ssl_names: List[str] = ["mert", "m-hubert"],
+        ssl_latent_dims: List[int] = [1024, 768],
+        lyric_encoder_vocab_size: int = 6681,
+        lyric_hidden_size: int = 1024,
+        patch_size: List[int] = [16, 1],
+        max_height: int = 16,
+        max_width: int = 4096,
+        audio_model=None,
+        dtype=None, device=None, operations=None
+
+    ):
+        super().__init__()
+
+        self.dtype = dtype
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+        self.inner_dim = inner_dim
+        self.out_channels = out_channels
+        self.max_position = max_position
+        self.patch_size = patch_size
+
+        self.rope_theta = rope_theta
+
+        self.rotary_emb = Qwen2RotaryEmbedding(
+            dim=self.attention_head_dim,
+            max_position_embeddings=self.max_position,
+            base=self.rope_theta,
+            dtype=dtype,
+            device=device,
+        )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.num_layers = num_layers
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                LinearTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=self.num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    mlp_ratio=mlp_ratio,
+                    add_cross_attention=True,
+                    add_cross_attention_dim=self.inner_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.t_block = nn.Sequential(nn.SiLU(), operations.Linear(self.inner_dim, 6 * self.inner_dim, bias=True, dtype=dtype, device=device))
+
+        # speaker
+        self.speaker_embedder = operations.Linear(speaker_embedding_dim, self.inner_dim, dtype=dtype, device=device)
+
+        # genre
+        self.genre_embedder = operations.Linear(text_embedding_dim, self.inner_dim, dtype=dtype, device=device)
+
+        # lyric
+        self.lyric_embs = operations.Embedding(lyric_encoder_vocab_size, lyric_hidden_size, dtype=dtype, device=device)
+        self.lyric_encoder = LyricEncoder(input_size=lyric_hidden_size, static_chunk_size=0, dtype=dtype, device=device, operations=operations)
+        self.lyric_proj = operations.Linear(lyric_hidden_size, self.inner_dim, dtype=dtype, device=device)
+
+        projector_dim = 2 * self.inner_dim
+
+        self.projectors = nn.ModuleList([
+            nn.Sequential(
+                operations.Linear(self.inner_dim, projector_dim, dtype=dtype, device=device),
+                nn.SiLU(),
+                operations.Linear(projector_dim, projector_dim, dtype=dtype, device=device),
+                nn.SiLU(),
+                operations.Linear(projector_dim, ssl_dim, dtype=dtype, device=device),
+            ) for ssl_dim in ssl_latent_dims
+        ])
+
+        self.proj_in = PatchEmbed(
+            height=max_height,
+            width=max_width,
+            patch_size=patch_size,
+            embed_dim=self.inner_dim,
+            bias=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.final_layer = T2IFinalLayer(self.inner_dim, patch_size=patch_size, out_channels=out_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward_lyric_encoder(
+        self,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        out_dtype=None,
+    ):
+        # N x T x D
+        lyric_embs = self.lyric_embs(lyric_token_idx, out_dtype=out_dtype)
+        prompt_prenet_out, _mask = self.lyric_encoder(lyric_embs, lyric_mask, decoding_chunk_size=1, num_decoding_left_chunks=-1)
+        prompt_prenet_out = self.lyric_proj(prompt_prenet_out)
+        return prompt_prenet_out
+
+    def encode(
+        self,
+        encoder_text_hidden_states: Optional[torch.Tensor] = None,
+        text_attention_mask: Optional[torch.LongTensor] = None,
+        speaker_embeds: Optional[torch.FloatTensor] = None,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        lyrics_strength=1.0,
+    ):
+
+        bs = encoder_text_hidden_states.shape[0]
+        device = encoder_text_hidden_states.device
+
+        # speaker embedding
+        encoder_spk_hidden_states = self.speaker_embedder(speaker_embeds).unsqueeze(1)
+
+        # genre embedding
+        encoder_text_hidden_states = self.genre_embedder(encoder_text_hidden_states)
+
+        # lyric
+        encoder_lyric_hidden_states = self.forward_lyric_encoder(
+            lyric_token_idx=lyric_token_idx,
+            lyric_mask=lyric_mask,
+            out_dtype=encoder_text_hidden_states.dtype,
+        )
+
+        encoder_lyric_hidden_states *= lyrics_strength
+
+        encoder_hidden_states = torch.cat([encoder_spk_hidden_states, encoder_text_hidden_states, encoder_lyric_hidden_states], dim=1)
+
+        encoder_hidden_mask = None
+        if text_attention_mask is not None:
+            speaker_mask = torch.ones(bs, 1, device=device)
+            encoder_hidden_mask = torch.cat([speaker_mask, text_attention_mask, lyric_mask], dim=1)
+
+        return encoder_hidden_states, encoder_hidden_mask
+
+    def decode(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_mask: torch.Tensor,
+        timestep: Optional[torch.Tensor],
+        output_length: int = 0,
+        block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
+        controlnet_scale: Union[float, torch.Tensor] = 1.0,
+    ):
+        embedded_timestep = self.timestep_embedder(self.time_proj(timestep).to(dtype=hidden_states.dtype))
+        temb = self.t_block(embedded_timestep)
+
+        hidden_states = self.proj_in(hidden_states)
+
+        # controlnet logic
+        if block_controlnet_hidden_states is not None:
+            control_condi = cross_norm(hidden_states, block_controlnet_hidden_states)
+            hidden_states = hidden_states + control_condi * controlnet_scale
+
+        # inner_hidden_states = []
+
+        rotary_freqs_cis = self.rotary_emb(hidden_states, seq_len=hidden_states.shape[1])
+        encoder_rotary_freqs_cis = self.rotary_emb(encoder_hidden_states, seq_len=encoder_hidden_states.shape[1])
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            hidden_states = block(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_hidden_mask,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=encoder_rotary_freqs_cis,
+                temb=temb,
+            )
+
+        output = self.final_layer(hidden_states, embedded_timestep, output_length)
+        return output
+
+    def forward(
+        self,
+        x,
+        timestep,
+        attention_mask=None,
+        context: Optional[torch.Tensor] = None,
+        text_attention_mask: Optional[torch.LongTensor] = None,
+        speaker_embeds: Optional[torch.FloatTensor] = None,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
+        controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        lyrics_strength=1.0,
+        **kwargs
+    ):
+        hidden_states = x
+        encoder_text_hidden_states = context
+        encoder_hidden_states, encoder_hidden_mask = self.encode(
+            encoder_text_hidden_states=encoder_text_hidden_states,
+            text_attention_mask=text_attention_mask,
+            speaker_embeds=speaker_embeds,
+            lyric_token_idx=lyric_token_idx,
+            lyric_mask=lyric_mask,
+            lyrics_strength=lyrics_strength,
+        )
+
+        output_length = hidden_states.shape[-1]
+
+        output = self.decode(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_hidden_mask=encoder_hidden_mask,
+            timestep=timestep,
+            output_length=output_length,
+            block_controlnet_hidden_states=block_controlnet_hidden_states,
+            controlnet_scale=controlnet_scale,
+        )
+
+        return output

comfy/ldm/ace/vae/autoencoder_dc.py

@@ -0,0 +1,644 @@
+# Rewritten from diffusers
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Tuple, Union
+
+import comfy.model_management
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class RMSNorm(ops.RMSNorm):
+    def __init__(self, dim, eps=1e-5, elementwise_affine=True, bias=False):
+        super().__init__(dim, eps=eps, elementwise_affine=elementwise_affine)
+        if elementwise_affine:
+            self.bias = nn.Parameter(torch.empty(dim)) if bias else None
+
+    def forward(self, x):
+        x = super().forward(x)
+        if self.elementwise_affine:
+            if self.bias is not None:
+                x = x + comfy.model_management.cast_to(self.bias, dtype=x.dtype, device=x.device)
+        return x
+
+
+def get_normalization(norm_type, num_features, num_groups=32, eps=1e-5):
+    if norm_type == "batch_norm":
+        return nn.BatchNorm2d(num_features)
+    elif norm_type == "group_norm":
+        return ops.GroupNorm(num_groups, num_features)
+    elif norm_type == "layer_norm":
+        return ops.LayerNorm(num_features)
+    elif norm_type == "rms_norm":
+        return RMSNorm(num_features, eps=eps, elementwise_affine=True, bias=True)
+    else:
+        raise ValueError(f"Unknown normalization type: {norm_type}")
+
+
+def get_activation(activation_type):
+    if activation_type == "relu":
+        return nn.ReLU()
+    elif activation_type == "relu6":
+        return nn.ReLU6()
+    elif activation_type == "silu":
+        return nn.SiLU()
+    elif activation_type == "leaky_relu":
+        return nn.LeakyReLU(0.2)
+    else:
+        raise ValueError(f"Unknown activation type: {activation_type}")
+
+
+class ResBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        norm_type: str = "batch_norm",
+        act_fn: str = "relu6",
+    ) -> None:
+        super().__init__()
+
+        self.norm_type = norm_type
+        self.nonlinearity = get_activation(act_fn) if act_fn is not None else nn.Identity()
+        self.conv1 = ops.Conv2d(in_channels, in_channels, 3, 1, 1)
+        self.conv2 = ops.Conv2d(in_channels, out_channels, 3, 1, 1, bias=False)
+        self.norm = get_normalization(norm_type, out_channels)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.conv1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.norm_type == "rms_norm":
+            # move channel to the last dimension so we apply RMSnorm across channel dimension
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+        else:
+            hidden_states = self.norm(hidden_states)
+
+        return hidden_states + residual
+
+class SanaMultiscaleAttentionProjection(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        num_attention_heads: int,
+        kernel_size: int,
+    ) -> None:
+        super().__init__()
+
+        channels = 3 * in_channels
+        self.proj_in = ops.Conv2d(
+            channels,
+            channels,
+            kernel_size,
+            padding=kernel_size // 2,
+            groups=channels,
+            bias=False,
+        )
+        self.proj_out = ops.Conv2d(channels, channels, 1, 1, 0, groups=3 * num_attention_heads, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.proj_in(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+        return hidden_states
+
+class SanaMultiscaleLinearAttention(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_attention_heads: int = None,
+        attention_head_dim: int = 8,
+        mult: float = 1.0,
+        norm_type: str = "batch_norm",
+        kernel_sizes: tuple = (5,),
+        eps: float = 1e-15,
+        residual_connection: bool = False,
+    ):
+        super().__init__()
+
+        self.eps = eps
+        self.attention_head_dim = attention_head_dim
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        num_attention_heads = (
+            int(in_channels // attention_head_dim * mult)
+            if num_attention_heads is None
+            else num_attention_heads
+        )
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.to_q = ops.Linear(in_channels, inner_dim, bias=False)
+        self.to_k = ops.Linear(in_channels, inner_dim, bias=False)
+        self.to_v = ops.Linear(in_channels, inner_dim, bias=False)
+
+        self.to_qkv_multiscale = nn.ModuleList()
+        for kernel_size in kernel_sizes:
+            self.to_qkv_multiscale.append(
+                SanaMultiscaleAttentionProjection(inner_dim, num_attention_heads, kernel_size)
+            )
+
+        self.nonlinearity = nn.ReLU()
+        self.to_out = ops.Linear(inner_dim * (1 + len(kernel_sizes)), out_channels, bias=False)
+        self.norm_out = get_normalization(norm_type, out_channels)
+
+    def apply_linear_attention(self, query, key, value):
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1)
+        scores = torch.matmul(value, key.transpose(-1, -2))
+        hidden_states = torch.matmul(scores, query)
+
+        hidden_states = hidden_states.to(dtype=torch.float32)
+        hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps)
+        return hidden_states
+
+    def apply_quadratic_attention(self, query, key, value):
+        scores = torch.matmul(key.transpose(-1, -2), query)
+        scores = scores.to(dtype=torch.float32)
+        scores = scores / (torch.sum(scores, dim=2, keepdim=True) + self.eps)
+        hidden_states = torch.matmul(value, scores.to(value.dtype))
+        return hidden_states
+
+    def forward(self, hidden_states):
+        height, width = hidden_states.shape[-2:]
+        if height * width > self.attention_head_dim:
+            use_linear_attention = True
+        else:
+            use_linear_attention = False
+
+        residual = hidden_states
+
+        batch_size, _, height, width = list(hidden_states.size())
+        original_dtype = hidden_states.dtype
+
+        hidden_states = hidden_states.movedim(1, -1)
+        query = self.to_q(hidden_states)
+        key = self.to_k(hidden_states)
+        value = self.to_v(hidden_states)
+        hidden_states = torch.cat([query, key, value], dim=3)
+        hidden_states = hidden_states.movedim(-1, 1)
+
+        multi_scale_qkv = [hidden_states]
+        for block in self.to_qkv_multiscale:
+            multi_scale_qkv.append(block(hidden_states))
+
+        hidden_states = torch.cat(multi_scale_qkv, dim=1)
+
+        if use_linear_attention:
+            # for linear attention upcast hidden_states to float32
+            hidden_states = hidden_states.to(dtype=torch.float32)
+
+        hidden_states = hidden_states.reshape(batch_size, -1, 3 * self.attention_head_dim, height * width)
+
+        query, key, value = hidden_states.chunk(3, dim=2)
+        query = self.nonlinearity(query)
+        key = self.nonlinearity(key)
+
+        if use_linear_attention:
+            hidden_states = self.apply_linear_attention(query, key, value)
+            hidden_states = hidden_states.to(dtype=original_dtype)
+        else:
+            hidden_states = self.apply_quadratic_attention(query, key, value)
+
+        hidden_states = torch.reshape(hidden_states, (batch_size, -1, height, width))
+        hidden_states = self.to_out(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.norm_type == "rms_norm":
+            hidden_states = self.norm_out(hidden_states.movedim(1, -1)).movedim(-1, 1)
+        else:
+            hidden_states = self.norm_out(hidden_states)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class EfficientViTBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        mult: float = 1.0,
+        attention_head_dim: int = 32,
+        qkv_multiscales: tuple = (5,),
+        norm_type: str = "batch_norm",
+    ) -> None:
+        super().__init__()
+
+        self.attn = SanaMultiscaleLinearAttention(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            mult=mult,
+            attention_head_dim=attention_head_dim,
+            norm_type=norm_type,
+            kernel_sizes=qkv_multiscales,
+            residual_connection=True,
+        )
+
+        self.conv_out = GLUMBConv(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            norm_type="rms_norm",
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.attn(x)
+        x = self.conv_out(x)
+        return x
+
+
+class GLUMBConv(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        expand_ratio: float = 4,
+        norm_type: str = None,
+        residual_connection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_channels = int(expand_ratio * in_channels)
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        self.nonlinearity = nn.SiLU()
+        self.conv_inverted = ops.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
+        self.conv_depth = ops.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
+        self.conv_point = ops.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
+
+        self.norm = None
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.residual_connection:
+            residual = hidden_states
+
+        hidden_states = self.conv_inverted(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv_depth(hidden_states)
+        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
+        hidden_states = hidden_states * self.nonlinearity(gate)
+
+        hidden_states = self.conv_point(hidden_states)
+
+        if self.norm_type == "rms_norm":
+            # move channel to the last dimension so we apply RMSnorm across channel dimension
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+def get_block(
+    block_type: str,
+    in_channels: int,
+    out_channels: int,
+    attention_head_dim: int,
+    norm_type: str,
+    act_fn: str,
+    qkv_mutliscales: tuple = (),
+):
+    if block_type == "ResBlock":
+        block = ResBlock(in_channels, out_channels, norm_type, act_fn)
+    elif block_type == "EfficientViTBlock":
+        block = EfficientViTBlock(
+            in_channels,
+            attention_head_dim=attention_head_dim,
+            norm_type=norm_type,
+            qkv_multiscales=qkv_mutliscales
+        )
+    else:
+        raise ValueError(f"Block with {block_type=} is not supported.")
+
+    return block
+
+
+class DCDownBlock2d(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int, downsample: bool = False, shortcut: bool = True) -> None:
+        super().__init__()
+
+        self.downsample = downsample
+        self.factor = 2
+        self.stride = 1 if downsample else 2
+        self.group_size = in_channels * self.factor**2 // out_channels
+        self.shortcut = shortcut
+
+        out_ratio = self.factor**2
+        if downsample:
+            assert out_channels % out_ratio == 0
+            out_channels = out_channels // out_ratio
+
+        self.conv = ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size=3,
+            stride=self.stride,
+            padding=1,
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        x = self.conv(hidden_states)
+        if self.downsample:
+            x = F.pixel_unshuffle(x, self.factor)
+
+        if self.shortcut:
+            y = F.pixel_unshuffle(hidden_states, self.factor)
+            y = y.unflatten(1, (-1, self.group_size))
+            y = y.mean(dim=2)
+            hidden_states = x + y
+        else:
+            hidden_states = x
+
+        return hidden_states
+
+
+class DCUpBlock2d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        interpolate: bool = False,
+        shortcut: bool = True,
+        interpolation_mode: str = "nearest",
+    ) -> None:
+        super().__init__()
+
+        self.interpolate = interpolate
+        self.interpolation_mode = interpolation_mode
+        self.shortcut = shortcut
+        self.factor = 2
+        self.repeats = out_channels * self.factor**2 // in_channels
+
+        out_ratio = self.factor**2
+        if not interpolate:
+            out_channels = out_channels * out_ratio
+
+        self.conv = ops.Conv2d(in_channels, out_channels, 3, 1, 1)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.interpolate:
+            x = F.interpolate(hidden_states, scale_factor=self.factor, mode=self.interpolation_mode)
+            x = self.conv(x)
+        else:
+            x = self.conv(hidden_states)
+            x = F.pixel_shuffle(x, self.factor)
+
+        if self.shortcut:
+            y = hidden_states.repeat_interleave(self.repeats, dim=1, output_size=hidden_states.shape[1] * self.repeats)
+            y = F.pixel_shuffle(y, self.factor)
+            hidden_states = x + y
+        else:
+            hidden_states = x
+
+        return hidden_states
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        attention_head_dim: int = 32,
+        block_type: str or tuple = "ResBlock",
+        block_out_channels: tuple = (128, 256, 512, 512, 1024, 1024),
+        layers_per_block: tuple = (2, 2, 2, 2, 2, 2),
+        qkv_multiscales: tuple = ((), (), (), (5,), (5,), (5,)),
+        downsample_block_type: str = "pixel_unshuffle",
+        out_shortcut: bool = True,
+    ):
+        super().__init__()
+
+        num_blocks = len(block_out_channels)
+
+        if isinstance(block_type, str):
+            block_type = (block_type,) * num_blocks
+
+        if layers_per_block[0] > 0:
+            self.conv_in = ops.Conv2d(
+                in_channels,
+                block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1],
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            )
+        else:
+            self.conv_in = DCDownBlock2d(
+                in_channels=in_channels,
+                out_channels=block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1],
+                downsample=downsample_block_type == "pixel_unshuffle",
+                shortcut=False,
+            )
+
+        down_blocks = []
+        for i, (out_channel, num_layers) in enumerate(zip(block_out_channels, layers_per_block)):
+            down_block_list = []
+
+            for _ in range(num_layers):
+                block = get_block(
+                    block_type[i],
+                    out_channel,
+                    out_channel,
+                    attention_head_dim=attention_head_dim,
+                    norm_type="rms_norm",
+                    act_fn="silu",
+                    qkv_mutliscales=qkv_multiscales[i],
+                )
+                down_block_list.append(block)
+
+            if i < num_blocks - 1 and num_layers > 0:
+                downsample_block = DCDownBlock2d(
+                    in_channels=out_channel,
+                    out_channels=block_out_channels[i + 1],
+                    downsample=downsample_block_type == "pixel_unshuffle",
+                    shortcut=True,
+                )
+                down_block_list.append(downsample_block)
+
+            down_blocks.append(nn.Sequential(*down_block_list))
+
+        self.down_blocks = nn.ModuleList(down_blocks)
+
+        self.conv_out = ops.Conv2d(block_out_channels[-1], latent_channels, 3, 1, 1)
+
+        self.out_shortcut = out_shortcut
+        if out_shortcut:
+            self.out_shortcut_average_group_size = block_out_channels[-1] // latent_channels
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.conv_in(hidden_states)
+        for down_block in self.down_blocks:
+            hidden_states = down_block(hidden_states)
+
+        if self.out_shortcut:
+            x = hidden_states.unflatten(1, (-1, self.out_shortcut_average_group_size))
+            x = x.mean(dim=2)
+            hidden_states = self.conv_out(hidden_states) + x
+        else:
+            hidden_states = self.conv_out(hidden_states)
+
+        return hidden_states
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        attention_head_dim: int = 32,
+        block_type: str or tuple = "ResBlock",
+        block_out_channels: tuple = (128, 256, 512, 512, 1024, 1024),
+        layers_per_block: tuple = (2, 2, 2, 2, 2, 2),
+        qkv_multiscales: tuple = ((), (), (), (5,), (5,), (5,)),
+        norm_type: str or tuple = "rms_norm",
+        act_fn: str or tuple = "silu",
+        upsample_block_type: str = "pixel_shuffle",
+        in_shortcut: bool = True,
+    ):
+        super().__init__()
+
+        num_blocks = len(block_out_channels)
+
+        if isinstance(block_type, str):
+            block_type = (block_type,) * num_blocks
+        if isinstance(norm_type, str):
+            norm_type = (norm_type,) * num_blocks
+        if isinstance(act_fn, str):
+            act_fn = (act_fn,) * num_blocks
+
+        self.conv_in = ops.Conv2d(latent_channels, block_out_channels[-1], 3, 1, 1)
+
+        self.in_shortcut = in_shortcut
+        if in_shortcut:
+            self.in_shortcut_repeats = block_out_channels[-1] // latent_channels
+
+        up_blocks = []
+        for i, (out_channel, num_layers) in reversed(list(enumerate(zip(block_out_channels, layers_per_block)))):
+            up_block_list = []
+
+            if i < num_blocks - 1 and num_layers > 0:
+                upsample_block = DCUpBlock2d(
+                    block_out_channels[i + 1],
+                    out_channel,
+                    interpolate=upsample_block_type == "interpolate",
+                    shortcut=True,
+                )
+                up_block_list.append(upsample_block)
+
+            for _ in range(num_layers):
+                block = get_block(
+                    block_type[i],
+                    out_channel,
+                    out_channel,
+                    attention_head_dim=attention_head_dim,
+                    norm_type=norm_type[i],
+                    act_fn=act_fn[i],
+                    qkv_mutliscales=qkv_multiscales[i],
+                )
+                up_block_list.append(block)
+
+            up_blocks.insert(0, nn.Sequential(*up_block_list))
+
+        self.up_blocks = nn.ModuleList(up_blocks)
+
+        channels = block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1]
+
+        self.norm_out = RMSNorm(channels, 1e-5, elementwise_affine=True, bias=True)
+        self.conv_act = nn.ReLU()
+        self.conv_out = None
+
+        if layers_per_block[0] > 0:
+            self.conv_out = ops.Conv2d(channels, in_channels, 3, 1, 1)
+        else:
+            self.conv_out = DCUpBlock2d(
+                channels, in_channels, interpolate=upsample_block_type == "interpolate", shortcut=False
+            )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.in_shortcut:
+            x = hidden_states.repeat_interleave(
+                self.in_shortcut_repeats, dim=1, output_size=hidden_states.shape[1] * self.in_shortcut_repeats
+            )
+            hidden_states = self.conv_in(hidden_states) + x
+        else:
+            hidden_states = self.conv_in(hidden_states)
+
+        for up_block in reversed(self.up_blocks):
+            hidden_states = up_block(hidden_states)
+
+        hidden_states = self.norm_out(hidden_states.movedim(1, -1)).movedim(-1, 1)
+        hidden_states = self.conv_act(hidden_states)
+        hidden_states = self.conv_out(hidden_states)
+        return hidden_states
+
+
+class AutoencoderDC(nn.Module):
+    def __init__(
+        self,
+        in_channels: int = 2,
+        latent_channels: int = 8,
+        attention_head_dim: int = 32,
+        encoder_block_types: Union[str, Tuple[str]] = ["ResBlock", "ResBlock", "ResBlock", "EfficientViTBlock"],
+        decoder_block_types: Union[str, Tuple[str]] = ["ResBlock", "ResBlock", "ResBlock", "EfficientViTBlock"],
+        encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024),
+        decoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024),
+        encoder_layers_per_block: Tuple[int] = (2, 2, 3, 3),
+        decoder_layers_per_block: Tuple[int] = (3, 3, 3, 3),
+        encoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (5,), (5,)),
+        decoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (5,), (5,)),
+        upsample_block_type: str = "interpolate",
+        downsample_block_type: str = "Conv",
+        decoder_norm_types: Union[str, Tuple[str]] = "rms_norm",
+        decoder_act_fns: Union[str, Tuple[str]] = "silu",
+        scaling_factor: float = 0.41407,
+    ) -> None:
+        super().__init__()
+
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            latent_channels=latent_channels,
+            attention_head_dim=attention_head_dim,
+            block_type=encoder_block_types,
+            block_out_channels=encoder_block_out_channels,
+            layers_per_block=encoder_layers_per_block,
+            qkv_multiscales=encoder_qkv_multiscales,
+            downsample_block_type=downsample_block_type,
+        )
+
+        self.decoder = Decoder(
+            in_channels=in_channels,
+            latent_channels=latent_channels,
+            attention_head_dim=attention_head_dim,
+            block_type=decoder_block_types,
+            block_out_channels=decoder_block_out_channels,
+            layers_per_block=decoder_layers_per_block,
+            qkv_multiscales=decoder_qkv_multiscales,
+            norm_type=decoder_norm_types,
+            act_fn=decoder_act_fns,
+            upsample_block_type=upsample_block_type,
+        )
+
+        self.scaling_factor = scaling_factor
+        self.spatial_compression_ratio = 2 ** (len(encoder_block_out_channels) - 1)
+
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
+        """Internal encoding function."""
+        encoded = self.encoder(x)
+        return encoded * self.scaling_factor
+
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        # Scale the latents back
+        z = z / self.scaling_factor
+        decoded = self.decoder(z)
+        return decoded
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        z = self.encode(x)
+        return self.decode(z)
+

comfy/ldm/ace/vae/music_dcae_pipeline.py

@@ -0,0 +1,109 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_dcae_pipeline.py
+import torch
+from .autoencoder_dc import AutoencoderDC
+import logging
+try:
+    import torchaudio
+except:
+    logging.warning("torchaudio missing, ACE model will be broken")
+
+import torchvision.transforms as transforms
+from .music_vocoder import ADaMoSHiFiGANV1
+
+
+class MusicDCAE(torch.nn.Module):
+    def __init__(self, source_sample_rate=None, dcae_config={}, vocoder_config={}):
+        super(MusicDCAE, self).__init__()
+
+        self.dcae = AutoencoderDC(**dcae_config)
+        self.vocoder = ADaMoSHiFiGANV1(**vocoder_config)
+
+        if source_sample_rate is None:
+            self.source_sample_rate = 48000
+        else:
+            self.source_sample_rate = source_sample_rate
+
+        # self.resampler = torchaudio.transforms.Resample(source_sample_rate, 44100)
+
+        self.transform = transforms.Compose([
+            transforms.Normalize(0.5, 0.5),
+        ])
+        self.min_mel_value = -11.0
+        self.max_mel_value = 3.0
+        self.audio_chunk_size = int(round((1024 * 512 / 44100 * 48000)))
+        self.mel_chunk_size = 1024
+        self.time_dimention_multiple = 8
+        self.latent_chunk_size = self.mel_chunk_size // self.time_dimention_multiple
+        self.scale_factor = 0.1786
+        self.shift_factor = -1.9091
+
+    def load_audio(self, audio_path):
+        audio, sr = torchaudio.load(audio_path)
+        return audio, sr
+
+    def forward_mel(self, audios):
+        mels = []
+        for i in range(len(audios)):
+            image = self.vocoder.mel_transform(audios[i])
+            mels.append(image)
+        mels = torch.stack(mels)
+        return mels
+
+    @torch.no_grad()
+    def encode(self, audios, audio_lengths=None, sr=None):
+        if audio_lengths is None:
+            audio_lengths = torch.tensor([audios.shape[2]] * audios.shape[0])
+            audio_lengths = audio_lengths.to(audios.device)
+
+        if sr is None:
+            sr = self.source_sample_rate
+
+        if sr != 44100:
+            audios = torchaudio.functional.resample(audios, sr, 44100)
+
+        max_audio_len = audios.shape[-1]
+        if max_audio_len % (8 * 512) != 0:
+            audios = torch.nn.functional.pad(audios, (0, 8 * 512 - max_audio_len % (8 * 512)))
+
+        mels = self.forward_mel(audios)
+        mels = (mels - self.min_mel_value) / (self.max_mel_value - self.min_mel_value)
+        mels = self.transform(mels)
+        latents = []
+        for mel in mels:
+            latent = self.dcae.encoder(mel.unsqueeze(0))
+            latents.append(latent)
+        latents = torch.cat(latents, dim=0)
+        # latent_lengths = (audio_lengths / sr * 44100 / 512 / self.time_dimention_multiple).long()
+        latents = (latents - self.shift_factor) * self.scale_factor
+        return latents
+        # return latents, latent_lengths
+
+    @torch.no_grad()
+    def decode(self, latents, audio_lengths=None, sr=None):
+        latents = latents / self.scale_factor + self.shift_factor
+
+        pred_wavs = []
+
+        for latent in latents:
+            mels = self.dcae.decoder(latent.unsqueeze(0))
+            mels = mels * 0.5 + 0.5
+            mels = mels * (self.max_mel_value - self.min_mel_value) + self.min_mel_value
+            wav = self.vocoder.decode(mels[0]).squeeze(1)
+
+            if sr is not None:
+                # resampler = torchaudio.transforms.Resample(44100, sr).to(latents.device).to(latents.dtype)
+                wav = torchaudio.functional.resample(wav, 44100, sr)
+                # wav = resampler(wav)
+            else:
+                sr = 44100
+            pred_wavs.append(wav)
+
+        if audio_lengths is not None:
+            pred_wavs = [wav[:, :length].cpu() for wav, length in zip(pred_wavs, audio_lengths)]
+        return torch.stack(pred_wavs)
+        # return sr, pred_wavs
+
+    def forward(self, audios, audio_lengths=None, sr=None):
+        latents, latent_lengths = self.encode(audios=audios, audio_lengths=audio_lengths, sr=sr)
+        sr, pred_wavs = self.decode(latents=latents, audio_lengths=audio_lengths, sr=sr)
+        return sr, pred_wavs, latents, latent_lengths

comfy/ldm/ace/vae/music_log_mel.py

@@ -0,0 +1,113 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_log_mel.py
+import torch
+import torch.nn as nn
+from torch import Tensor
+import logging
+try:
+    from torchaudio.transforms import MelScale
+except:
+    logging.warning("torchaudio missing, ACE model will be broken")
+
+import comfy.model_management
+
+class LinearSpectrogram(nn.Module):
+    def __init__(
+        self,
+        n_fft=2048,
+        win_length=2048,
+        hop_length=512,
+        center=False,
+        mode="pow2_sqrt",
+    ):
+        super().__init__()
+
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.hop_length = hop_length
+        self.center = center
+        self.mode = mode
+
+        self.register_buffer("window", torch.hann_window(win_length))
+
+    def forward(self, y: Tensor) -> Tensor:
+        if y.ndim == 3:
+            y = y.squeeze(1)
+
+        y = torch.nn.functional.pad(
+            y.unsqueeze(1),
+            (
+                (self.win_length - self.hop_length) // 2,
+                (self.win_length - self.hop_length + 1) // 2,
+            ),
+            mode="reflect",
+        ).squeeze(1)
+        dtype = y.dtype
+        spec = torch.stft(
+            y.float(),
+            self.n_fft,
+            hop_length=self.hop_length,
+            win_length=self.win_length,
+            window=comfy.model_management.cast_to(self.window, dtype=torch.float32, device=y.device),
+            center=self.center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+        spec = torch.view_as_real(spec)
+
+        if self.mode == "pow2_sqrt":
+            spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+        spec = spec.to(dtype)
+        return spec
+
+
+class LogMelSpectrogram(nn.Module):
+    def __init__(
+        self,
+        sample_rate=44100,
+        n_fft=2048,
+        win_length=2048,
+        hop_length=512,
+        n_mels=128,
+        center=False,
+        f_min=0.0,
+        f_max=None,
+    ):
+        super().__init__()
+
+        self.sample_rate = sample_rate
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.hop_length = hop_length
+        self.center = center
+        self.n_mels = n_mels
+        self.f_min = f_min
+        self.f_max = f_max or sample_rate // 2
+
+        self.spectrogram = LinearSpectrogram(n_fft, win_length, hop_length, center)
+        self.mel_scale = MelScale(
+            self.n_mels,
+            self.sample_rate,
+            self.f_min,
+            self.f_max,
+            self.n_fft // 2 + 1,
+            "slaney",
+            "slaney",
+        )
+
+    def compress(self, x: Tensor) -> Tensor:
+        return torch.log(torch.clamp(x, min=1e-5))
+
+    def decompress(self, x: Tensor) -> Tensor:
+        return torch.exp(x)
+
+    def forward(self, x: Tensor, return_linear: bool = False) -> Tensor:
+        linear = self.spectrogram(x)
+        x = self.mel_scale(linear)
+        x = self.compress(x)
+        # print(x.shape)
+        if return_linear:
+            return x, self.compress(linear)
+
+        return x

comfy/ldm/ace/vae/music_vocoder.py

@@ -0,0 +1,538 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_vocoder.py
+import torch
+from torch import nn
+
+from functools import partial
+from math import prod
+from typing import Callable, Tuple, List
+
+import numpy as np
+import torch.nn.functional as F
+from torch.nn.utils.parametrize import remove_parametrizations as remove_weight_norm
+
+from .music_log_mel import LogMelSpectrogram
+
+import comfy.model_management
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+def drop_path(
+    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
+):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+
+    """  # noqa: E501
+
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (
+        x.ndim - 1
+    )  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0 and scale_by_keep:
+        random_tensor.div_(keep_prob)
+    return x * random_tensor
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""  # noqa: E501
+
+    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
+
+    def extra_repr(self):
+        return f"drop_prob={round(self.drop_prob,3):0.3f}"
+
+
+class LayerNorm(nn.Module):
+    r"""LayerNorm that supports two data formats: channels_last (default) or channels_first.
+    The ordering of the dimensions in the inputs. channels_last corresponds to inputs with
+    shape (batch_size, height, width, channels) while channels_first corresponds to inputs
+    with shape (batch_size, channels, height, width).
+    """  # noqa: E501
+
+    def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(normalized_shape))
+        self.bias = nn.Parameter(torch.zeros(normalized_shape))
+        self.eps = eps
+        self.data_format = data_format
+        if self.data_format not in ["channels_last", "channels_first"]:
+            raise NotImplementedError
+        self.normalized_shape = (normalized_shape,)
+
+    def forward(self, x):
+        if self.data_format == "channels_last":
+            return F.layer_norm(
+                x, self.normalized_shape, comfy.model_management.cast_to(self.weight, dtype=x.dtype, device=x.device), comfy.model_management.cast_to(self.bias, dtype=x.dtype, device=x.device), self.eps
+            )
+        elif self.data_format == "channels_first":
+            u = x.mean(1, keepdim=True)
+            s = (x - u).pow(2).mean(1, keepdim=True)
+            x = (x - u) / torch.sqrt(s + self.eps)
+            x = comfy.model_management.cast_to(self.weight[:, None], dtype=x.dtype, device=x.device) * x + comfy.model_management.cast_to(self.bias[:, None], dtype=x.dtype, device=x.device)
+            return x
+
+
+class ConvNeXtBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.0.
+        kernel_size (int): Kernel size for depthwise conv. Default: 7.
+        dilation (int): Dilation for depthwise conv. Default: 1.
+    """  # noqa: E501
+
+    def __init__(
+        self,
+        dim: int,
+        drop_path: float = 0.0,
+        layer_scale_init_value: float = 1e-6,
+        mlp_ratio: float = 4.0,
+        kernel_size: int = 7,
+        dilation: int = 1,
+    ):
+        super().__init__()
+
+        self.dwconv = ops.Conv1d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=int(dilation * (kernel_size - 1) / 2),
+            groups=dim,
+        )  # depthwise conv
+        self.norm = LayerNorm(dim, eps=1e-6)
+        self.pwconv1 = ops.Linear(
+            dim, int(mlp_ratio * dim)
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = ops.Linear(int(mlp_ratio * dim), dim)
+        self.gamma = (
+            nn.Parameter(torch.empty((dim)), requires_grad=False)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(
+            drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x, apply_residual: bool = True):
+        input = x
+
+        x = self.dwconv(x)
+        x = x.permute(0, 2, 1)  # (N, C, L) -> (N, L, C)
+        x = self.norm(x)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+
+        if self.gamma is not None:
+            x = comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device) * x
+
+        x = x.permute(0, 2, 1)  # (N, L, C) -> (N, C, L)
+        x = self.drop_path(x)
+
+        if apply_residual:
+            x = input + x
+
+        return x
+
+
+class ParallelConvNeXtBlock(nn.Module):
+    def __init__(self, kernel_sizes: List[int], *args, **kwargs):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [
+                ConvNeXtBlock(kernel_size=kernel_size, *args, **kwargs)
+                for kernel_size in kernel_sizes
+            ]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.stack(
+            [block(x, apply_residual=False) for block in self.blocks] + [x],
+            dim=1,
+        ).sum(dim=1)
+
+
+class ConvNeXtEncoder(nn.Module):
+    def __init__(
+        self,
+        input_channels=3,
+        depths=[3, 3, 9, 3],
+        dims=[96, 192, 384, 768],
+        drop_path_rate=0.0,
+        layer_scale_init_value=1e-6,
+        kernel_sizes: Tuple[int] = (7,),
+    ):
+        super().__init__()
+        assert len(depths) == len(dims)
+
+        self.channel_layers = nn.ModuleList()
+        stem = nn.Sequential(
+            ops.Conv1d(
+                input_channels,
+                dims[0],
+                kernel_size=7,
+                padding=3,
+                padding_mode="replicate",
+            ),
+            LayerNorm(dims[0], eps=1e-6, data_format="channels_first"),
+        )
+        self.channel_layers.append(stem)
+
+        for i in range(len(depths) - 1):
+            mid_layer = nn.Sequential(
+                LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),
+                ops.Conv1d(dims[i], dims[i + 1], kernel_size=1),
+            )
+            self.channel_layers.append(mid_layer)
+
+        block_fn = (
+            partial(ConvNeXtBlock, kernel_size=kernel_sizes[0])
+            if len(kernel_sizes) == 1
+            else partial(ParallelConvNeXtBlock, kernel_sizes=kernel_sizes)
+        )
+
+        self.stages = nn.ModuleList()
+        drop_path_rates = [
+            x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
+        ]
+
+        cur = 0
+        for i in range(len(depths)):
+            stage = nn.Sequential(
+                *[
+                    block_fn(
+                        dim=dims[i],
+                        drop_path=drop_path_rates[cur + j],
+                        layer_scale_init_value=layer_scale_init_value,
+                    )
+                    for j in range(depths[i])
+                ]
+            )
+            self.stages.append(stage)
+            cur += depths[i]
+
+        self.norm = LayerNorm(dims[-1], eps=1e-6, data_format="channels_first")
+
+    def forward(
+        self,
+        x: torch.Tensor,
+    ) -> torch.Tensor:
+        for channel_layer, stage in zip(self.channel_layers, self.stages):
+            x = channel_layer(x)
+            x = stage(x)
+
+        return self.norm(x)
+
+
+def get_padding(kernel_size, dilation=1):
+    return (kernel_size * dilation - dilation) // 2
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super().__init__()
+
+        self.convs1 = nn.ModuleList(
+            [
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+
+        self.convs2 = nn.ModuleList(
+            [
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.silu(x)
+            xt = c1(xt)
+            xt = F.silu(xt)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for conv in self.convs1:
+            remove_weight_norm(conv)
+        for conv in self.convs2:
+            remove_weight_norm(conv)
+
+
+class HiFiGANGenerator(nn.Module):
+    def __init__(
+        self,
+        *,
+        hop_length: int = 512,
+        upsample_rates: Tuple[int] = (8, 8, 2, 2, 2),
+        upsample_kernel_sizes: Tuple[int] = (16, 16, 8, 2, 2),
+        resblock_kernel_sizes: Tuple[int] = (3, 7, 11),
+        resblock_dilation_sizes: Tuple[Tuple[int]] = (
+            (1, 3, 5), (1, 3, 5), (1, 3, 5)),
+        num_mels: int = 128,
+        upsample_initial_channel: int = 512,
+        use_template: bool = True,
+        pre_conv_kernel_size: int = 7,
+        post_conv_kernel_size: int = 7,
+        post_activation: Callable = partial(nn.SiLU, inplace=True),
+    ):
+        super().__init__()
+
+        assert (
+            prod(upsample_rates) == hop_length
+        ), f"hop_length must be {prod(upsample_rates)}"
+
+        self.conv_pre = torch.nn.utils.parametrizations.weight_norm(
+            ops.Conv1d(
+                num_mels,
+                upsample_initial_channel,
+                pre_conv_kernel_size,
+                1,
+                padding=get_padding(pre_conv_kernel_size),
+            )
+        )
+
+        self.num_upsamples = len(upsample_rates)
+        self.num_kernels = len(resblock_kernel_sizes)
+
+        self.noise_convs = nn.ModuleList()
+        self.use_template = use_template
+        self.ups = nn.ModuleList()
+
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+            if not use_template:
+                continue
+
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1:])
+                self.noise_convs.append(
+                    ops.Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(ops.Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes):
+                self.resblocks.append(ResBlock1(ch, k, d))
+
+        self.activation_post = post_activation()
+        self.conv_post = torch.nn.utils.parametrizations.weight_norm(
+            ops.Conv1d(
+                ch,
+                1,
+                post_conv_kernel_size,
+                1,
+                padding=get_padding(post_conv_kernel_size),
+            )
+        )
+
+    def forward(self, x, template=None):
+        x = self.conv_pre(x)
+
+        for i in range(self.num_upsamples):
+            x = F.silu(x, inplace=True)
+            x = self.ups[i](x)
+
+            if self.use_template:
+                x = x + self.noise_convs[i](template)
+
+            xs = None
+
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+
+            x = xs / self.num_kernels
+
+        x = self.activation_post(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for up in self.ups:
+            remove_weight_norm(up)
+        for block in self.resblocks:
+            block.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+
+class ADaMoSHiFiGANV1(nn.Module):
+    def __init__(
+        self,
+        input_channels: int = 128,
+        depths: List[int] = [3, 3, 9, 3],
+        dims: List[int] = [128, 256, 384, 512],
+        drop_path_rate: float = 0.0,
+        kernel_sizes: Tuple[int] = (7,),
+        upsample_rates: Tuple[int] = (4, 4, 2, 2, 2, 2, 2),
+        upsample_kernel_sizes: Tuple[int] = (8, 8, 4, 4, 4, 4, 4),
+        resblock_kernel_sizes: Tuple[int] = (3, 7, 11, 13),
+        resblock_dilation_sizes: Tuple[Tuple[int]] = (
+            (1, 3, 5), (1, 3, 5), (1, 3, 5), (1, 3, 5)),
+        num_mels: int = 512,
+        upsample_initial_channel: int = 1024,
+        use_template: bool = False,
+        pre_conv_kernel_size: int = 13,
+        post_conv_kernel_size: int = 13,
+        sampling_rate: int = 44100,
+        n_fft: int = 2048,
+        win_length: int = 2048,
+        hop_length: int = 512,
+        f_min: int = 40,
+        f_max: int = 16000,
+        n_mels: int = 128,
+    ):
+        super().__init__()
+
+        self.backbone = ConvNeXtEncoder(
+            input_channels=input_channels,
+            depths=depths,
+            dims=dims,
+            drop_path_rate=drop_path_rate,
+            kernel_sizes=kernel_sizes,
+        )
+
+        self.head = HiFiGANGenerator(
+            hop_length=hop_length,
+            upsample_rates=upsample_rates,
+            upsample_kernel_sizes=upsample_kernel_sizes,
+            resblock_kernel_sizes=resblock_kernel_sizes,
+            resblock_dilation_sizes=resblock_dilation_sizes,
+            num_mels=num_mels,
+            upsample_initial_channel=upsample_initial_channel,
+            use_template=use_template,
+            pre_conv_kernel_size=pre_conv_kernel_size,
+            post_conv_kernel_size=post_conv_kernel_size,
+        )
+        self.sampling_rate = sampling_rate
+        self.mel_transform = LogMelSpectrogram(
+            sample_rate=sampling_rate,
+            n_fft=n_fft,
+            win_length=win_length,
+            hop_length=hop_length,
+            f_min=f_min,
+            f_max=f_max,
+            n_mels=n_mels,
+        )
+        self.eval()
+
+    @torch.no_grad()
+    def decode(self, mel):
+        y = self.backbone(mel)
+        y = self.head(y)
+        return y
+
+    @torch.no_grad()
+    def encode(self, x):
+        return self.mel_transform(x)
+
+    def forward(self, mel):
+        y = self.backbone(mel)
+        y = self.head(y)
+        return y

comfy/ldm/audio/autoencoder.py

@@ -75,16 +75,10 @@ class SnakeBeta(nn.Module):
         return x
 
 def WNConv1d(*args, **kwargs):
-    try:
-        return torch.nn.utils.parametrizations.weight_norm(ops.Conv1d(*args, **kwargs))
-    except:
-        return torch.nn.utils.weight_norm(ops.Conv1d(*args, **kwargs)) #support pytorch 2.1 and older
+    return torch.nn.utils.parametrizations.weight_norm(ops.Conv1d(*args, **kwargs))
 
 def WNConvTranspose1d(*args, **kwargs):
-    try:
-        return torch.nn.utils.parametrizations.weight_norm(ops.ConvTranspose1d(*args, **kwargs))
-    except:
-        return torch.nn.utils.weight_norm(ops.ConvTranspose1d(*args, **kwargs)) #support pytorch 2.1 and older
+    return torch.nn.utils.parametrizations.weight_norm(ops.ConvTranspose1d(*args, **kwargs))
 
 def get_activation(activation: Literal["elu", "snake", "none"], antialias=False, channels=None) -> nn.Module:
     if activation == "elu":

comfy/ldm/chroma/layers.py

@@ -0,0 +1,181 @@
+import torch
+from torch import Tensor, nn
+
+from comfy.ldm.flux.math import attention
+from comfy.ldm.flux.layers import (
+    MLPEmbedder,
+    RMSNorm,
+    QKNorm,
+    SelfAttention,
+    ModulationOut,
+)
+
+
+
+class ChromaModulationOut(ModulationOut):
+    @classmethod
+    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
+        return cls(
+            shift=tensor[:, offset : offset + 1, :],
+            scale=tensor[:, offset + 1 : offset + 2, :],
+            gate=tensor[:, offset + 2 : offset + 3, :],
+        )
+
+
+
+
+class Approximator(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.in_proj(x)
+
+        for layer, norms in zip(self.layers, self.norms):
+            x = x + layer(norms(x))
+
+        x = self.out_proj(x)
+
+        return x
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.flipped_img_txt = flipped_img_txt
+
+    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None):
+        (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
+
+        # prepare image for attention
+        img_modulated = torch.addcmul(img_mod1.shift, 1 + img_mod1.scale, self.img_norm1(img))
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = torch.addcmul(txt_mod1.shift, 1 + txt_mod1.scale, self.txt_norm1(txt))
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        attn = attention(torch.cat((txt_q, img_q), dim=2),
+                         torch.cat((txt_k, img_k), dim=2),
+                         torch.cat((txt_v, img_v), dim=2),
+                         pe=pe, mask=attn_mask)
+
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img.addcmul_(img_mod1.gate, self.img_attn.proj(img_attn))
+        img.addcmul_(img_mod2.gate, self.img_mlp(torch.addcmul(img_mod2.shift, 1 + img_mod2.scale, self.img_norm2(img))))
+
+        # calculate the txt bloks
+        txt.addcmul_(txt_mod1.gate, self.txt_attn.proj(txt_attn))
+        txt.addcmul_(txt_mod2.gate, self.txt_mlp(torch.addcmul(txt_mod2.shift, 1 + txt_mod2.scale, self.txt_norm2(txt))))
+
+        if txt.dtype == torch.float16:
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
+
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+
+    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor:
+        mod = vec
+        x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x))
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        x.addcmul_(mod.gate, output)
+        if x.dtype == torch.float16:
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+        return x
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = vec
+        shift = shift.squeeze(1)
+        scale = scale.squeeze(1)
+        x = torch.addcmul(shift[:, None, :], 1 + scale[:, None, :], self.norm_final(x))
+        x = self.linear(x)
+        return x

comfy/ldm/chroma/model.py

@@ -0,0 +1,270 @@
+#Original code can be found on: https://github.com/black-forest-labs/flux
+
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import (
+    EmbedND,
+    timestep_embedding,
+)
+
+from .layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    SingleStreamBlock,
+    Approximator,
+    ChromaModulationOut,
+)
+
+
+@dataclass
+class ChromaParams:
+    in_channels: int
+    out_channels: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    patch_size: int
+    qkv_bias: bool
+    in_dim: int
+    out_dim: int
+    hidden_dim: int
+    n_layers: int
+
+
+
+
+class Chroma(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        params = ChromaParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.in_dim = params.in_dim
+        self.out_dim = params.out_dim
+        self.hidden_dim = params.hidden_dim
+        self.n_layers = params.n_layers
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        # set as nn identity for now, will overwrite it later.
+        self.distilled_guidance_layer = Approximator(
+                    in_dim=self.in_dim,
+                    hidden_dim=self.hidden_dim,
+                    out_dim=self.out_dim,
+                    n_layers=self.n_layers,
+                    dtype=dtype, device=device, operations=operations
+                )
+
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+    def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0):
+        # This function slices up the modulations tensor which has the following layout:
+        #   single     : num_single_blocks * 3 elements
+        #   double_img : num_double_blocks * 6 elements
+        #   double_txt : num_double_blocks * 6 elements
+        #   final      : 2 elements
+        if block_type == "final":
+            return (tensor[:, -2:-1, :], tensor[:, -1:, :])
+        single_block_count = self.params.depth_single_blocks
+        double_block_count = self.params.depth
+        offset = 3 * idx
+        if block_type == "single":
+            return ChromaModulationOut.from_offset(tensor, offset)
+        # Double block modulations are 6 elements so we double 3 * idx.
+        offset *= 2
+        if block_type in {"double_img", "double_txt"}:
+            # Advance past the single block modulations.
+            offset += 3 * single_block_count
+            if block_type == "double_txt":
+                # Advance past the double block img modulations.
+                offset += 6 * double_block_count
+            return (
+                ChromaModulationOut.from_offset(tensor, offset),
+                ChromaModulationOut.from_offset(tensor, offset + 3),
+            )
+        raise ValueError("Bad block_type")
+
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        guidance: Tensor = None,
+        control = None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+        patches_replace = transformer_options.get("patches_replace", {})
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        # distilled vector guidance
+        mod_index_length = 344
+        distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype)
+        # guidance = guidance *
+        distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype)
+
+        # get all modulation index
+        modulation_index = timestep_embedding(torch.arange(mod_index_length, device=img.device), 32).to(img.device, img.dtype)
+        # we need to broadcast the modulation index here so each batch has all of the index
+        modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype)
+        # and we need to broadcast timestep and guidance along too
+        timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype)
+        # then and only then we could concatenate it together
+        input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype)
+
+        mod_vectors = self.distilled_guidance_layer(input_vec)
+
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if i not in self.skip_mmdit:
+                double_mod = (
+                    self.get_modulations(mod_vectors, "double_img", idx=i),
+                    self.get_modulations(mod_vectors, "double_txt", idx=i),
+                )
+                if ("double_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"], out["txt"] = block(img=args["img"],
+                                                       txt=args["txt"],
+                                                       vec=args["vec"],
+                                                       pe=args["pe"],
+                                                       attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("double_block", i)]({"img": img,
+                                                               "txt": txt,
+                                                               "vec": double_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    txt = out["txt"]
+                    img = out["img"]
+                else:
+                    img, txt = block(img=img,
+                                     txt=txt,
+                                     vec=double_mod,
+                                     pe=pe,
+                                     attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_i = control.get("input")
+                    if i < len(control_i):
+                        add = control_i[i]
+                        if add is not None:
+                            img += add
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if i not in self.skip_dit:
+                single_mod = self.get_modulations(mod_vectors, "single", idx=i)
+                if ("single_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"] = block(args["img"],
+                                           vec=args["vec"],
+                                           pe=args["pe"],
+                                           attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("single_block", i)]({"img": img,
+                                                               "vec": single_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    img = out["img"]
+                else:
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_o = control.get("output")
+                    if i < len(control_o):
+                        add = control_o[i]
+                        if add is not None:
+                            img[:, txt.shape[1] :, ...] += add
+
+        img = img[:, txt.shape[1] :, ...]
+        final_mod = self.get_modulations(mod_vectors, "final")
+        img = self.final_layer(img, vec=final_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        bs, c, h, w = x.shape
+        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=self.patch_size, pw=self.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)
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        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=self.patch_size, pw=self.patch_size)[:,:,:h,:w]

comfy/ldm/cosmos/blocks.py

@@ -23,25 +23,14 @@ from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
 from torch import nn
 
-from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
 from comfy.ldm.modules.attention import optimized_attention
 
 
-def apply_rotary_pos_emb(
-    t: torch.Tensor,
-    freqs: torch.Tensor,
-) -> torch.Tensor:
-    t_ = t.reshape(*t.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).float()
-    t_out = freqs[..., 0] * t_[..., 0] + freqs[..., 1] * t_[..., 1]
-    t_out = t_out.movedim(-1, -2).reshape(*t.shape).type_as(t)
-    return t_out
-
-
-def get_normalization(name: str, channels: int, weight_args={}):
+def get_normalization(name: str, channels: int, weight_args={}, operations=None):
     if name == "I":
         return nn.Identity()
     elif name == "R":
-        return RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args)
+        return operations.RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args)
     else:
         raise ValueError(f"Normalization {name} not found")
 
@@ -120,15 +109,15 @@ class Attention(nn.Module):
 
         self.to_q = nn.Sequential(
             operations.Linear(query_dim, inner_dim, bias=qkv_bias, **weight_args),
-            get_normalization(qkv_norm[0], norm_dim),
+            get_normalization(qkv_norm[0], norm_dim, weight_args=weight_args, operations=operations),
         )
         self.to_k = nn.Sequential(
             operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args),
-            get_normalization(qkv_norm[1], norm_dim),
+            get_normalization(qkv_norm[1], norm_dim, weight_args=weight_args, operations=operations),
         )
         self.to_v = nn.Sequential(
             operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args),
-            get_normalization(qkv_norm[2], norm_dim),
+            get_normalization(qkv_norm[2], norm_dim, weight_args=weight_args, operations=operations),
         )
 
         self.to_out = nn.Sequential(

comfy/ldm/cosmos/model.py

@@ -27,8 +27,6 @@ from torchvision import transforms
 from enum import Enum
 import logging
 
-from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
-
 from .blocks import (
     FinalLayer,
     GeneralDITTransformerBlock,
@@ -195,7 +193,7 @@ class GeneralDIT(nn.Module):
 
         if self.affline_emb_norm:
             logging.debug("Building affine embedding normalization layer")
-            self.affline_norm = RMSNorm(model_channels, elementwise_affine=True, eps=1e-6)
+            self.affline_norm = operations.RMSNorm(model_channels, elementwise_affine=True, eps=1e-6, device=device, dtype=dtype)
         else:
             self.affline_norm = nn.Identity()
 

comfy/ldm/cosmos/position_embedding.py

@@ -66,15 +66,16 @@ class VideoRopePosition3DEmb(VideoPositionEmb):
         h_extrapolation_ratio: float = 1.0,
         w_extrapolation_ratio: float = 1.0,
         t_extrapolation_ratio: float = 1.0,
+        enable_fps_modulation: bool = True,
         device=None,
         **kwargs,  # used for compatibility with other positional embeddings; unused in this class
     ):
         del kwargs
         super().__init__()
-        self.register_buffer("seq", torch.arange(max(len_h, len_w, len_t), dtype=torch.float, device=device))
         self.base_fps = base_fps
         self.max_h = len_h
         self.max_w = len_w
+        self.enable_fps_modulation = enable_fps_modulation
 
         dim = head_dim
         dim_h = dim // 6 * 2
@@ -132,21 +133,19 @@ class VideoRopePosition3DEmb(VideoPositionEmb):
         temporal_freqs = 1.0 / (t_theta**self.dim_temporal_range.to(device=device))
 
         B, T, H, W, _ = B_T_H_W_C
+        seq = torch.arange(max(H, W, T), dtype=torch.float, device=device)
         uniform_fps = (fps is None) or isinstance(fps, (int, float)) or (fps.min() == fps.max())
         assert (
             uniform_fps or B == 1 or T == 1
         ), "For video batch, batch size should be 1 for non-uniform fps. For image batch, T should be 1"
-        assert (
-            H <= self.max_h and W <= self.max_w
-        ), f"Input dimensions (H={H}, W={W}) exceed the maximum dimensions (max_h={self.max_h}, max_w={self.max_w})"
-        half_emb_h = torch.outer(self.seq[:H].to(device=device), h_spatial_freqs)
-        half_emb_w = torch.outer(self.seq[:W].to(device=device), w_spatial_freqs)
+        half_emb_h = torch.outer(seq[:H].to(device=device), h_spatial_freqs)
+        half_emb_w = torch.outer(seq[:W].to(device=device), w_spatial_freqs)
 
         # apply sequence scaling in temporal dimension
-        if fps is None:  # image case
-            half_emb_t = torch.outer(self.seq[:T].to(device=device), temporal_freqs)
+        if fps is None or self.enable_fps_modulation is False:  # image case
+            half_emb_t = torch.outer(seq[:T].to(device=device), temporal_freqs)
         else:
-            half_emb_t = torch.outer(self.seq[:T].to(device=device) / fps * self.base_fps, temporal_freqs)
+            half_emb_t = torch.outer(seq[:T].to(device=device) / fps * self.base_fps, temporal_freqs)
 
         half_emb_h = torch.stack([torch.cos(half_emb_h), -torch.sin(half_emb_h), torch.sin(half_emb_h), torch.cos(half_emb_h)], dim=-1)
         half_emb_w = torch.stack([torch.cos(half_emb_w), -torch.sin(half_emb_w), torch.sin(half_emb_w), torch.cos(half_emb_w)], dim=-1)

comfy/ldm/cosmos/predict2.py

@@ -0,0 +1,864 @@
+# original code from: https://github.com/nvidia-cosmos/cosmos-predict2
+
+import torch
+from torch import nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+import logging
+from typing import Callable, Optional, Tuple
+import math
+
+from .position_embedding import VideoRopePosition3DEmb, LearnablePosEmbAxis
+from torchvision import transforms
+
+from comfy.ldm.modules.attention import optimized_attention
+
+def apply_rotary_pos_emb(
+    t: torch.Tensor,
+    freqs: torch.Tensor,
+) -> torch.Tensor:
+    t_ = t.reshape(*t.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).float()
+    t_out = freqs[..., 0] * t_[..., 0] + freqs[..., 1] * t_[..., 1]
+    t_out = t_out.movedim(-1, -2).reshape(*t.shape).type_as(t)
+    return t_out
+
+
+# ---------------------- Feed Forward Network -----------------------
+class GPT2FeedForward(nn.Module):
+    def __init__(self, d_model: int, d_ff: int, device=None, dtype=None, operations=None) -> None:
+        super().__init__()
+        self.activation = nn.GELU()
+        self.layer1 = operations.Linear(d_model, d_ff, bias=False, device=device, dtype=dtype)
+        self.layer2 = operations.Linear(d_ff, d_model, bias=False, device=device, dtype=dtype)
+
+        self._layer_id = None
+        self._dim = d_model
+        self._hidden_dim = d_ff
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.layer1(x)
+
+        x = self.activation(x)
+        x = self.layer2(x)
+        return x
+
+
+def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H_D: torch.Tensor) -> torch.Tensor:
+    """Computes multi-head attention using PyTorch's native implementation.
+
+    This function provides a PyTorch backend alternative to Transformer Engine's attention operation.
+    It rearranges the input tensors to match PyTorch's expected format, computes scaled dot-product
+    attention, and rearranges the output back to the original format.
+
+    The input tensor names use the following dimension conventions:
+
+    - B: batch size
+    - S: sequence length
+    - H: number of attention heads
+    - D: head dimension
+
+    Args:
+        q_B_S_H_D: Query tensor with shape (batch, seq_len, n_heads, head_dim)
+        k_B_S_H_D: Key tensor with shape (batch, seq_len, n_heads, head_dim)
+        v_B_S_H_D: Value tensor with shape (batch, seq_len, n_heads, head_dim)
+
+    Returns:
+        Attention output tensor with shape (batch, seq_len, n_heads * head_dim)
+    """
+    in_q_shape = q_B_S_H_D.shape
+    in_k_shape = k_B_S_H_D.shape
+    q_B_H_S_D = rearrange(q_B_S_H_D, "b ... h k -> b h ... k").view(in_q_shape[0], in_q_shape[-2], -1, in_q_shape[-1])
+    k_B_H_S_D = rearrange(k_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
+    v_B_H_S_D = rearrange(v_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
+    return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True)
+
+
+class Attention(nn.Module):
+    """
+    A flexible attention module supporting both self-attention and cross-attention mechanisms.
+
+    This module implements a multi-head attention layer that can operate in either self-attention
+    or cross-attention mode. The mode is determined by whether a context dimension is provided.
+    The implementation uses scaled dot-product attention and supports optional bias terms and
+    dropout regularization.
+
+    Args:
+        query_dim (int): The dimensionality of the query vectors.
+        context_dim (int, optional): The dimensionality of the context (key/value) vectors.
+            If None, the module operates in self-attention mode using query_dim. Default: None
+        n_heads (int, optional): Number of attention heads for multi-head attention. Default: 8
+        head_dim (int, optional): The dimension of each attention head. Default: 64
+        dropout (float, optional): Dropout probability applied to the output. Default: 0.0
+        qkv_format (str, optional): Format specification for QKV tensors. Default: "bshd"
+        backend (str, optional): Backend to use for the attention operation. Default: "transformer_engine"
+
+    Examples:
+        >>> # Self-attention with 512 dimensions and 8 heads
+        >>> self_attn = Attention(query_dim=512)
+        >>> x = torch.randn(32, 16, 512)  # (batch_size, seq_len, dim)
+        >>> out = self_attn(x)  # (32, 16, 512)
+
+        >>> # Cross-attention
+        >>> cross_attn = Attention(query_dim=512, context_dim=256)
+        >>> query = torch.randn(32, 16, 512)
+        >>> context = torch.randn(32, 8, 256)
+        >>> out = cross_attn(query, context)  # (32, 16, 512)
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        context_dim: Optional[int] = None,
+        n_heads: int = 8,
+        head_dim: int = 64,
+        dropout: float = 0.0,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        logging.debug(
+            f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
+            f"{n_heads} heads with a dimension of {head_dim}."
+        )
+        self.is_selfattn = context_dim is None  # self attention
+
+        context_dim = query_dim if context_dim is None else context_dim
+        inner_dim = head_dim * n_heads
+
+        self.n_heads = n_heads
+        self.head_dim = head_dim
+        self.query_dim = query_dim
+        self.context_dim = context_dim
+
+        self.q_proj = operations.Linear(query_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.k_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.v_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.v_norm = nn.Identity()
+
+        self.output_proj = operations.Linear(inner_dim, query_dim, bias=False, device=device, dtype=dtype)
+        self.output_dropout = nn.Dropout(dropout) if dropout > 1e-4 else nn.Identity()
+
+        self.attn_op = torch_attention_op
+
+        self._query_dim = query_dim
+        self._context_dim = context_dim
+        self._inner_dim = inner_dim
+
+    def compute_qkv(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        rope_emb: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        q = self.q_proj(x)
+        context = x if context is None else context
+        k = self.k_proj(context)
+        v = self.v_proj(context)
+        q, k, v = map(
+            lambda t: rearrange(t, "b ... (h d) -> b ... h d", h=self.n_heads, d=self.head_dim),
+            (q, k, v),
+        )
+
+        def apply_norm_and_rotary_pos_emb(
+            q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, rope_emb: Optional[torch.Tensor]
+        ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+            q = self.q_norm(q)
+            k = self.k_norm(k)
+            v = self.v_norm(v)
+            if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
+                q = apply_rotary_pos_emb(q, rope_emb)
+                k = apply_rotary_pos_emb(k, rope_emb)
+            return q, k, v
+
+        q, k, v = apply_norm_and_rotary_pos_emb(q, k, v, rope_emb)
+
+        return q, k, v
+
+    def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+        result = self.attn_op(q, k, v)  # [B, S, H, D]
+        return self.output_dropout(self.output_proj(result))
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        rope_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            x (Tensor): The query tensor of shape [B, Mq, K]
+            context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
+        """
+        q, k, v = self.compute_qkv(x, context, rope_emb=rope_emb)
+        return self.compute_attention(q, k, v)
+
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int):
+        super().__init__()
+        self.num_channels = num_channels
+
+    def forward(self, timesteps_B_T: torch.Tensor) -> torch.Tensor:
+        assert timesteps_B_T.ndim == 2, f"Expected 2D input, got {timesteps_B_T.ndim}"
+        timesteps = timesteps_B_T.flatten().float()
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device)
+        exponent = exponent / (half_dim - 0.0)
+
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+
+        sin_emb = torch.sin(emb)
+        cos_emb = torch.cos(emb)
+        emb = torch.cat([cos_emb, sin_emb], dim=-1)
+
+        return rearrange(emb, "(b t) d -> b t d", b=timesteps_B_T.shape[0], t=timesteps_B_T.shape[1])
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_features: int, out_features: int, use_adaln_lora: bool = False, device=None, dtype=None, operations=None):
+        super().__init__()
+        logging.debug(
+            f"Using AdaLN LoRA Flag:  {use_adaln_lora}. We enable bias if no AdaLN LoRA for backward compatibility."
+        )
+        self.in_dim = in_features
+        self.out_dim = out_features
+        self.linear_1 = operations.Linear(in_features, out_features, bias=not use_adaln_lora, device=device, dtype=dtype)
+        self.activation = nn.SiLU()
+        self.use_adaln_lora = use_adaln_lora
+        if use_adaln_lora:
+            self.linear_2 = operations.Linear(out_features, 3 * out_features, bias=False, device=device, dtype=dtype)
+        else:
+            self.linear_2 = operations.Linear(out_features, out_features, bias=False, device=device, dtype=dtype)
+
+    def forward(self, sample: torch.Tensor) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        emb = self.linear_1(sample)
+        emb = self.activation(emb)
+        emb = self.linear_2(emb)
+
+        if self.use_adaln_lora:
+            adaln_lora_B_T_3D = emb
+            emb_B_T_D = sample
+        else:
+            adaln_lora_B_T_3D = None
+            emb_B_T_D = emb
+
+        return emb_B_T_D, adaln_lora_B_T_3D
+
+
+class PatchEmbed(nn.Module):
+    """
+    PatchEmbed is a module for embedding patches from an input tensor by applying either 3D or 2D convolutional layers,
+    depending on the . This module can process inputs with temporal (video) and spatial (image) dimensions,
+    making it suitable for video and image processing tasks. It supports dividing the input into patches
+    and embedding each patch into a vector of size `out_channels`.
+
+    Parameters:
+    - spatial_patch_size (int): The size of each spatial patch.
+    - temporal_patch_size (int): The size of each temporal patch.
+    - in_channels (int): Number of input channels. Default: 3.
+    - out_channels (int): The dimension of the embedding vector for each patch. Default: 768.
+    - bias (bool): If True, adds a learnable bias to the output of the convolutional layers. Default: True.
+    """
+
+    def __init__(
+        self,
+        spatial_patch_size: int,
+        temporal_patch_size: int,
+        in_channels: int = 3,
+        out_channels: int = 768,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.spatial_patch_size = spatial_patch_size
+        self.temporal_patch_size = temporal_patch_size
+
+        self.proj = nn.Sequential(
+            Rearrange(
+                "b c (t r) (h m) (w n) -> b t h w (c r m n)",
+                r=temporal_patch_size,
+                m=spatial_patch_size,
+                n=spatial_patch_size,
+            ),
+            operations.Linear(
+                in_channels * spatial_patch_size * spatial_patch_size * temporal_patch_size, out_channels, bias=False, device=device, dtype=dtype
+            ),
+        )
+        self.dim = in_channels * spatial_patch_size * spatial_patch_size * temporal_patch_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass of the PatchEmbed module.
+
+        Parameters:
+        - x (torch.Tensor): The input tensor of shape (B, C, T, H, W) where
+            B is the batch size,
+            C is the number of channels,
+            T is the temporal dimension,
+            H is the height, and
+            W is the width of the input.
+
+        Returns:
+        - torch.Tensor: The embedded patches as a tensor, with shape b t h w c.
+        """
+        assert x.dim() == 5
+        _, _, T, H, W = x.shape
+        assert (
+            H % self.spatial_patch_size == 0 and W % self.spatial_patch_size == 0
+        ), f"H,W {(H, W)} should be divisible by spatial_patch_size {self.spatial_patch_size}"
+        assert T % self.temporal_patch_size == 0
+        x = self.proj(x)
+        return x
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of video DiT.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        spatial_patch_size: int,
+        temporal_patch_size: int,
+        out_channels: int,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = operations.Linear(
+            hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
+        )
+        self.hidden_size = hidden_size
+        self.n_adaln_chunks = 2
+        self.use_adaln_lora = use_adaln_lora
+        self.adaln_lora_dim = adaln_lora_dim
+        if use_adaln_lora:
+            self.adaln_modulation = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(hidden_size, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, self.n_adaln_chunks * hidden_size, bias=False, device=device, dtype=dtype),
+            )
+        else:
+            self.adaln_modulation = nn.Sequential(
+                nn.SiLU(), operations.Linear(hidden_size, self.n_adaln_chunks * hidden_size, bias=False, device=device, dtype=dtype)
+            )
+
+    def forward(
+        self,
+        x_B_T_H_W_D: torch.Tensor,
+        emb_B_T_D: torch.Tensor,
+        adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
+    ):
+        if self.use_adaln_lora:
+            assert adaln_lora_B_T_3D is not None
+            shift_B_T_D, scale_B_T_D = (
+                self.adaln_modulation(emb_B_T_D) + adaln_lora_B_T_3D[:, :, : 2 * self.hidden_size]
+            ).chunk(2, dim=-1)
+        else:
+            shift_B_T_D, scale_B_T_D = self.adaln_modulation(emb_B_T_D).chunk(2, dim=-1)
+
+        shift_B_T_1_1_D, scale_B_T_1_1_D = rearrange(shift_B_T_D, "b t d -> b t 1 1 d"), rearrange(
+            scale_B_T_D, "b t d -> b t 1 1 d"
+        )
+
+        def _fn(
+            _x_B_T_H_W_D: torch.Tensor,
+            _norm_layer: nn.Module,
+            _scale_B_T_1_1_D: torch.Tensor,
+            _shift_B_T_1_1_D: torch.Tensor,
+        ) -> torch.Tensor:
+            return _norm_layer(_x_B_T_H_W_D) * (1 + _scale_B_T_1_1_D) + _shift_B_T_1_1_D
+
+        x_B_T_H_W_D = _fn(x_B_T_H_W_D, self.layer_norm, scale_B_T_1_1_D, shift_B_T_1_1_D)
+        x_B_T_H_W_O = self.linear(x_B_T_H_W_D)
+        return x_B_T_H_W_O
+
+
+class Block(nn.Module):
+    """
+    A transformer block that combines self-attention, cross-attention and MLP layers with AdaLN modulation.
+    Each component (self-attention, cross-attention, MLP) has its own layer normalization and AdaLN modulation.
+
+    Parameters:
+        x_dim (int): Dimension of input features
+        context_dim (int): Dimension of context features for cross-attention
+        num_heads (int): Number of attention heads
+        mlp_ratio (float): Multiplier for MLP hidden dimension. Default: 4.0
+        use_adaln_lora (bool): Whether to use AdaLN-LoRA modulation. Default: False
+        adaln_lora_dim (int): Hidden dimension for AdaLN-LoRA layers. Default: 256
+
+    The block applies the following sequence:
+    1. Self-attention with AdaLN modulation
+    2. Cross-attention with AdaLN modulation
+    3. MLP with AdaLN modulation
+
+    Each component uses skip connections and layer normalization.
+    """
+
+    def __init__(
+        self,
+        x_dim: int,
+        context_dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        device=None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.x_dim = x_dim
+        self.layer_norm_self_attn = operations.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype)
+        self.self_attn = Attention(x_dim, None, num_heads, x_dim // num_heads, device=device, dtype=dtype, operations=operations)
+
+        self.layer_norm_cross_attn = operations.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype)
+        self.cross_attn = Attention(
+            x_dim, context_dim, num_heads, x_dim // num_heads, device=device, dtype=dtype, operations=operations
+        )
+
+        self.layer_norm_mlp = operations.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype)
+        self.mlp = GPT2FeedForward(x_dim, int(x_dim * mlp_ratio), device=device, dtype=dtype, operations=operations)
+
+        self.use_adaln_lora = use_adaln_lora
+        if self.use_adaln_lora:
+            self.adaln_modulation_self_attn = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, 3 * x_dim, bias=False, device=device, dtype=dtype),
+            )
+            self.adaln_modulation_cross_attn = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, 3 * x_dim, bias=False, device=device, dtype=dtype),
+            )
+            self.adaln_modulation_mlp = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, 3 * x_dim, bias=False, device=device, dtype=dtype),
+            )
+        else:
+            self.adaln_modulation_self_attn = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, 3 * x_dim, bias=False, device=device, dtype=dtype))
+            self.adaln_modulation_cross_attn = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, 3 * x_dim, bias=False, device=device, dtype=dtype))
+            self.adaln_modulation_mlp = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, 3 * x_dim, bias=False, device=device, dtype=dtype))
+
+    def forward(
+        self,
+        x_B_T_H_W_D: torch.Tensor,
+        emb_B_T_D: torch.Tensor,
+        crossattn_emb: torch.Tensor,
+        rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
+        extra_per_block_pos_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if extra_per_block_pos_emb is not None:
+            x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
+
+        if self.use_adaln_lora:
+            shift_self_attn_B_T_D, scale_self_attn_B_T_D, gate_self_attn_B_T_D = (
+                self.adaln_modulation_self_attn(emb_B_T_D) + adaln_lora_B_T_3D
+            ).chunk(3, dim=-1)
+            shift_cross_attn_B_T_D, scale_cross_attn_B_T_D, gate_cross_attn_B_T_D = (
+                self.adaln_modulation_cross_attn(emb_B_T_D) + adaln_lora_B_T_3D
+            ).chunk(3, dim=-1)
+            shift_mlp_B_T_D, scale_mlp_B_T_D, gate_mlp_B_T_D = (
+                self.adaln_modulation_mlp(emb_B_T_D) + adaln_lora_B_T_3D
+            ).chunk(3, dim=-1)
+        else:
+            shift_self_attn_B_T_D, scale_self_attn_B_T_D, gate_self_attn_B_T_D = self.adaln_modulation_self_attn(
+                emb_B_T_D
+            ).chunk(3, dim=-1)
+            shift_cross_attn_B_T_D, scale_cross_attn_B_T_D, gate_cross_attn_B_T_D = self.adaln_modulation_cross_attn(
+                emb_B_T_D
+            ).chunk(3, dim=-1)
+            shift_mlp_B_T_D, scale_mlp_B_T_D, gate_mlp_B_T_D = self.adaln_modulation_mlp(emb_B_T_D).chunk(3, dim=-1)
+
+        # Reshape tensors from (B, T, D) to (B, T, 1, 1, D) for broadcasting
+        shift_self_attn_B_T_1_1_D = rearrange(shift_self_attn_B_T_D, "b t d -> b t 1 1 d")
+        scale_self_attn_B_T_1_1_D = rearrange(scale_self_attn_B_T_D, "b t d -> b t 1 1 d")
+        gate_self_attn_B_T_1_1_D = rearrange(gate_self_attn_B_T_D, "b t d -> b t 1 1 d")
+
+        shift_cross_attn_B_T_1_1_D = rearrange(shift_cross_attn_B_T_D, "b t d -> b t 1 1 d")
+        scale_cross_attn_B_T_1_1_D = rearrange(scale_cross_attn_B_T_D, "b t d -> b t 1 1 d")
+        gate_cross_attn_B_T_1_1_D = rearrange(gate_cross_attn_B_T_D, "b t d -> b t 1 1 d")
+
+        shift_mlp_B_T_1_1_D = rearrange(shift_mlp_B_T_D, "b t d -> b t 1 1 d")
+        scale_mlp_B_T_1_1_D = rearrange(scale_mlp_B_T_D, "b t d -> b t 1 1 d")
+        gate_mlp_B_T_1_1_D = rearrange(gate_mlp_B_T_D, "b t d -> b t 1 1 d")
+
+        B, T, H, W, D = x_B_T_H_W_D.shape
+
+        def _fn(_x_B_T_H_W_D, _norm_layer, _scale_B_T_1_1_D, _shift_B_T_1_1_D):
+            return _norm_layer(_x_B_T_H_W_D) * (1 + _scale_B_T_1_1_D) + _shift_B_T_1_1_D
+
+        normalized_x_B_T_H_W_D = _fn(
+            x_B_T_H_W_D,
+            self.layer_norm_self_attn,
+            scale_self_attn_B_T_1_1_D,
+            shift_self_attn_B_T_1_1_D,
+        )
+        result_B_T_H_W_D = rearrange(
+            self.self_attn(
+                # normalized_x_B_T_HW_D,
+                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                None,
+                rope_emb=rope_emb_L_1_1_D,
+            ),
+            "b (t h w) d -> b t h w d",
+            t=T,
+            h=H,
+            w=W,
+        )
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
+
+        def _x_fn(
+            _x_B_T_H_W_D: torch.Tensor,
+            layer_norm_cross_attn: Callable,
+            _scale_cross_attn_B_T_1_1_D: torch.Tensor,
+            _shift_cross_attn_B_T_1_1_D: torch.Tensor,
+        ) -> torch.Tensor:
+            _normalized_x_B_T_H_W_D = _fn(
+                _x_B_T_H_W_D, layer_norm_cross_attn, _scale_cross_attn_B_T_1_1_D, _shift_cross_attn_B_T_1_1_D
+            )
+            _result_B_T_H_W_D = rearrange(
+                self.cross_attn(
+                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                    crossattn_emb,
+                    rope_emb=rope_emb_L_1_1_D,
+                ),
+                "b (t h w) d -> b t h w d",
+                t=T,
+                h=H,
+                w=W,
+            )
+            return _result_B_T_H_W_D
+
+        result_B_T_H_W_D = _x_fn(
+            x_B_T_H_W_D,
+            self.layer_norm_cross_attn,
+            scale_cross_attn_B_T_1_1_D,
+            shift_cross_attn_B_T_1_1_D,
+        )
+        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
+
+        normalized_x_B_T_H_W_D = _fn(
+            x_B_T_H_W_D,
+            self.layer_norm_mlp,
+            scale_mlp_B_T_1_1_D,
+            shift_mlp_B_T_1_1_D,
+        )
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
+        return x_B_T_H_W_D
+
+
+class MiniTrainDIT(nn.Module):
+    """
+    A clean impl of DIT that can load and  reproduce the training results of the original DIT model in~(cosmos 1)
+    A general implementation of adaln-modulated VIT-like~(DiT) transformer for video processing.
+
+    Args:
+        max_img_h (int): Maximum height of the input images.
+        max_img_w (int): Maximum width of the input images.
+        max_frames (int): Maximum number of frames in the video sequence.
+        in_channels (int): Number of input channels (e.g., RGB channels for color images).
+        out_channels (int): Number of output channels.
+        patch_spatial (tuple): Spatial resolution of patches for input processing.
+        patch_temporal (int): Temporal resolution of patches for input processing.
+        concat_padding_mask (bool): If True, includes a mask channel in the input to handle padding.
+        model_channels (int): Base number of channels used throughout the model.
+        num_blocks (int): Number of transformer blocks.
+        num_heads (int): Number of heads in the multi-head attention layers.
+        mlp_ratio (float): Expansion ratio for MLP blocks.
+        crossattn_emb_channels (int): Number of embedding channels for cross-attention.
+        pos_emb_cls (str): Type of positional embeddings.
+        pos_emb_learnable (bool): Whether positional embeddings are learnable.
+        pos_emb_interpolation (str): Method for interpolating positional embeddings.
+        min_fps (int): Minimum frames per second.
+        max_fps (int): Maximum frames per second.
+        use_adaln_lora (bool): Whether to use AdaLN-LoRA.
+        adaln_lora_dim (int): Dimension for AdaLN-LoRA.
+        rope_h_extrapolation_ratio (float): Height extrapolation ratio for RoPE.
+        rope_w_extrapolation_ratio (float): Width extrapolation ratio for RoPE.
+        rope_t_extrapolation_ratio (float): Temporal extrapolation ratio for RoPE.
+        extra_per_block_abs_pos_emb (bool): Whether to use extra per-block absolute positional embeddings.
+        extra_h_extrapolation_ratio (float): Height extrapolation ratio for extra embeddings.
+        extra_w_extrapolation_ratio (float): Width extrapolation ratio for extra embeddings.
+        extra_t_extrapolation_ratio (float): Temporal extrapolation ratio for extra embeddings.
+    """
+
+    def __init__(
+        self,
+        max_img_h: int,
+        max_img_w: int,
+        max_frames: int,
+        in_channels: int,
+        out_channels: int,
+        patch_spatial: int,  # tuple,
+        patch_temporal: int,
+        concat_padding_mask: bool = True,
+        # attention settings
+        model_channels: int = 768,
+        num_blocks: int = 10,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.0,
+        # cross attention settings
+        crossattn_emb_channels: int = 1024,
+        # positional embedding settings
+        pos_emb_cls: str = "sincos",
+        pos_emb_learnable: bool = False,
+        pos_emb_interpolation: str = "crop",
+        min_fps: int = 1,
+        max_fps: int = 30,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        rope_h_extrapolation_ratio: float = 1.0,
+        rope_w_extrapolation_ratio: float = 1.0,
+        rope_t_extrapolation_ratio: float = 1.0,
+        extra_per_block_abs_pos_emb: bool = False,
+        extra_h_extrapolation_ratio: float = 1.0,
+        extra_w_extrapolation_ratio: float = 1.0,
+        extra_t_extrapolation_ratio: float = 1.0,
+        rope_enable_fps_modulation: bool = True,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        self.dtype = dtype
+        self.max_img_h = max_img_h
+        self.max_img_w = max_img_w
+        self.max_frames = max_frames
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_spatial = patch_spatial
+        self.patch_temporal = patch_temporal
+        self.num_heads = num_heads
+        self.num_blocks = num_blocks
+        self.model_channels = model_channels
+        self.concat_padding_mask = concat_padding_mask
+        # positional embedding settings
+        self.pos_emb_cls = pos_emb_cls
+        self.pos_emb_learnable = pos_emb_learnable
+        self.pos_emb_interpolation = pos_emb_interpolation
+        self.min_fps = min_fps
+        self.max_fps = max_fps
+        self.rope_h_extrapolation_ratio = rope_h_extrapolation_ratio
+        self.rope_w_extrapolation_ratio = rope_w_extrapolation_ratio
+        self.rope_t_extrapolation_ratio = rope_t_extrapolation_ratio
+        self.extra_per_block_abs_pos_emb = extra_per_block_abs_pos_emb
+        self.extra_h_extrapolation_ratio = extra_h_extrapolation_ratio
+        self.extra_w_extrapolation_ratio = extra_w_extrapolation_ratio
+        self.extra_t_extrapolation_ratio = extra_t_extrapolation_ratio
+        self.rope_enable_fps_modulation = rope_enable_fps_modulation
+
+        self.build_pos_embed(device=device, dtype=dtype)
+        self.use_adaln_lora = use_adaln_lora
+        self.adaln_lora_dim = adaln_lora_dim
+        self.t_embedder = nn.Sequential(
+            Timesteps(model_channels),
+            TimestepEmbedding(model_channels, model_channels, use_adaln_lora=use_adaln_lora, device=device, dtype=dtype, operations=operations,),
+        )
+
+        in_channels = in_channels + 1 if concat_padding_mask else in_channels
+        self.x_embedder = PatchEmbed(
+            spatial_patch_size=patch_spatial,
+            temporal_patch_size=patch_temporal,
+            in_channels=in_channels,
+            out_channels=model_channels,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    x_dim=model_channels,
+                    context_dim=crossattn_emb_channels,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    use_adaln_lora=use_adaln_lora,
+                    adaln_lora_dim=adaln_lora_dim,
+                    device=device, dtype=dtype, operations=operations,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+
+        self.final_layer = FinalLayer(
+            hidden_size=self.model_channels,
+            spatial_patch_size=self.patch_spatial,
+            temporal_patch_size=self.patch_temporal,
+            out_channels=self.out_channels,
+            use_adaln_lora=self.use_adaln_lora,
+            adaln_lora_dim=self.adaln_lora_dim,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        self.t_embedding_norm = operations.RMSNorm(model_channels, eps=1e-6, device=device, dtype=dtype)
+
+    def build_pos_embed(self, device=None, dtype=None) -> None:
+        if self.pos_emb_cls == "rope3d":
+            cls_type = VideoRopePosition3DEmb
+        else:
+            raise ValueError(f"Unknown pos_emb_cls {self.pos_emb_cls}")
+
+        logging.debug(f"Building positional embedding with {self.pos_emb_cls} class, impl {cls_type}")
+        kwargs = dict(
+            model_channels=self.model_channels,
+            len_h=self.max_img_h // self.patch_spatial,
+            len_w=self.max_img_w // self.patch_spatial,
+            len_t=self.max_frames // self.patch_temporal,
+            max_fps=self.max_fps,
+            min_fps=self.min_fps,
+            is_learnable=self.pos_emb_learnable,
+            interpolation=self.pos_emb_interpolation,
+            head_dim=self.model_channels // self.num_heads,
+            h_extrapolation_ratio=self.rope_h_extrapolation_ratio,
+            w_extrapolation_ratio=self.rope_w_extrapolation_ratio,
+            t_extrapolation_ratio=self.rope_t_extrapolation_ratio,
+            enable_fps_modulation=self.rope_enable_fps_modulation,
+            device=device,
+        )
+        self.pos_embedder = cls_type(
+            **kwargs,  # type: ignore
+        )
+
+        if self.extra_per_block_abs_pos_emb:
+            kwargs["h_extrapolation_ratio"] = self.extra_h_extrapolation_ratio
+            kwargs["w_extrapolation_ratio"] = self.extra_w_extrapolation_ratio
+            kwargs["t_extrapolation_ratio"] = self.extra_t_extrapolation_ratio
+            kwargs["device"] = device
+            kwargs["dtype"] = dtype
+            self.extra_pos_embedder = LearnablePosEmbAxis(
+                **kwargs,  # type: ignore
+            )
+
+    def prepare_embedded_sequence(
+        self,
+        x_B_C_T_H_W: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """
+        Prepares an embedded sequence tensor by applying positional embeddings and handling padding masks.
+
+        Args:
+            x_B_C_T_H_W (torch.Tensor): video
+            fps (Optional[torch.Tensor]): Frames per second tensor to be used for positional embedding when required.
+                                    If None, a default value (`self.base_fps`) will be used.
+            padding_mask (Optional[torch.Tensor]): current it is not used
+
+        Returns:
+            Tuple[torch.Tensor, Optional[torch.Tensor]]:
+                - A tensor of shape (B, T, H, W, D) with the embedded sequence.
+                - An optional positional embedding tensor, returned only if the positional embedding class
+                (`self.pos_emb_cls`) includes 'rope'. Otherwise, None.
+
+        Notes:
+            - If `self.concat_padding_mask` is True, a padding mask channel is concatenated to the input tensor.
+            - The method of applying positional embeddings depends on the value of `self.pos_emb_cls`.
+            - If 'rope' is in `self.pos_emb_cls` (case insensitive), the positional embeddings are generated using
+                the `self.pos_embedder` with the shape [T, H, W].
+            - If "fps_aware" is in `self.pos_emb_cls`, the positional embeddings are generated using the
+            `self.pos_embedder` with the fps tensor.
+            - Otherwise, the positional embeddings are generated without considering fps.
+        """
+        if self.concat_padding_mask:
+            if padding_mask is None:
+                padding_mask = torch.zeros(x_B_C_T_H_W.shape[0], 1, x_B_C_T_H_W.shape[3], x_B_C_T_H_W.shape[4], dtype=x_B_C_T_H_W.dtype, device=x_B_C_T_H_W.device)
+            else:
+                padding_mask = transforms.functional.resize(
+                    padding_mask, list(x_B_C_T_H_W.shape[-2:]), interpolation=transforms.InterpolationMode.NEAREST
+                )
+            x_B_C_T_H_W = torch.cat(
+                [x_B_C_T_H_W, padding_mask.unsqueeze(1).repeat(1, 1, x_B_C_T_H_W.shape[2], 1, 1)], dim=1
+            )
+        x_B_T_H_W_D = self.x_embedder(x_B_C_T_H_W)
+
+        if self.extra_per_block_abs_pos_emb:
+            extra_pos_emb = self.extra_pos_embedder(x_B_T_H_W_D, fps=fps, device=x_B_C_T_H_W.device, dtype=x_B_C_T_H_W.dtype)
+        else:
+            extra_pos_emb = None
+
+        if "rope" in self.pos_emb_cls.lower():
+            return x_B_T_H_W_D, self.pos_embedder(x_B_T_H_W_D, fps=fps, device=x_B_C_T_H_W.device), extra_pos_emb
+        x_B_T_H_W_D = x_B_T_H_W_D + self.pos_embedder(x_B_T_H_W_D, device=x_B_C_T_H_W.device)  # [B, T, H, W, D]
+
+        return x_B_T_H_W_D, None, extra_pos_emb
+
+    def unpatchify(self, x_B_T_H_W_M: torch.Tensor) -> torch.Tensor:
+        x_B_C_Tt_Hp_Wp = rearrange(
+            x_B_T_H_W_M,
+            "B T H W (p1 p2 t C) -> B C (T t) (H p1) (W p2)",
+            p1=self.patch_spatial,
+            p2=self.patch_spatial,
+            t=self.patch_temporal,
+        )
+        return x_B_C_Tt_Hp_Wp
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ):
+        x_B_C_T_H_W = x
+        timesteps_B_T = timesteps
+        crossattn_emb = context
+        """
+        Args:
+            x: (B, C, T, H, W) tensor of spatial-temp inputs
+            timesteps: (B, ) tensor of timesteps
+            crossattn_emb: (B, N, D) tensor of cross-attention embeddings
+        """
+        x_B_T_H_W_D, rope_emb_L_1_1_D, extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D = self.prepare_embedded_sequence(
+            x_B_C_T_H_W,
+            fps=fps,
+            padding_mask=padding_mask,
+        )
+
+        if timesteps_B_T.ndim == 1:
+            timesteps_B_T = timesteps_B_T.unsqueeze(1)
+        t_embedding_B_T_D, adaln_lora_B_T_3D = self.t_embedder[1](self.t_embedder[0](timesteps_B_T).to(x_B_T_H_W_D.dtype))
+        t_embedding_B_T_D = self.t_embedding_norm(t_embedding_B_T_D)
+
+        # for logging purpose
+        affline_scale_log_info = {}
+        affline_scale_log_info["t_embedding_B_T_D"] = t_embedding_B_T_D.detach()
+        self.affline_scale_log_info = affline_scale_log_info
+        self.affline_emb = t_embedding_B_T_D
+        self.crossattn_emb = crossattn_emb
+
+        if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
+            assert (
+                x_B_T_H_W_D.shape == extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape
+            ), f"{x_B_T_H_W_D.shape} != {extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape}"
+
+        block_kwargs = {
+            "rope_emb_L_1_1_D": rope_emb_L_1_1_D.unsqueeze(1).unsqueeze(0),
+            "adaln_lora_B_T_3D": adaln_lora_B_T_3D,
+            "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
+        }
+        for block in self.blocks:
+            x_B_T_H_W_D = block(
+                x_B_T_H_W_D,
+                t_embedding_B_T_D,
+                crossattn_emb,
+                **block_kwargs,
+            )
+
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)
+        return x_B_C_Tt_Hp_Wp

comfy/ldm/flux/controlnet.py

@@ -121,6 +121,11 @@ class ControlNetFlux(Flux):
         if img.ndim != 3 or txt.ndim != 3:
             raise ValueError("Input img and txt tensors must have 3 dimensions.")
 
+        if y is None:
+            y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+        else:
+            y = y[:, :self.params.vec_in_dim]
+
         # running on sequences img
         img = self.img_in(img)
 
@@ -174,7 +179,7 @@ class ControlNetFlux(Flux):
             out["output"] = out_output[:self.main_model_single]
         return out
 
-    def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
+    def forward(self, x, timesteps, context, y=None, guidance=None, hint=None, **kwargs):
         patch_size = 2
         if self.latent_input:
             hint = comfy.ldm.common_dit.pad_to_patch_size(hint, (patch_size, patch_size))

comfy/ldm/flux/layers.py

@@ -118,7 +118,7 @@ class Modulation(nn.Module):
 def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
     if modulation_dims is None:
         if m_add is not None:
-            return tensor * m_mult + m_add
+            return torch.addcmul(m_add, tensor, m_mult)
         else:
             return tensor * m_mult
     else:

comfy/ldm/flux/model.py

@@ -101,6 +101,10 @@ class Flux(nn.Module):
         transformer_options={},
         attn_mask: Tensor = None,
     ) -> Tensor:
+
+        if y is None:
+            y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+
         patches_replace = transformer_options.get("patches_replace", {})
         if img.ndim != 3 or txt.ndim != 3:
             raise ValueError("Input img and txt tensors must have 3 dimensions.")
@@ -155,6 +159,9 @@ class Flux(nn.Module):
                     if add is not None:
                         img += add
 
+        if img.dtype == torch.float16:
+            img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
+
         img = torch.cat((txt, img), 1)
 
         for i, block in enumerate(self.single_blocks):
@@ -188,20 +195,50 @@ class Flux(nn.Module):
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def forward(self, x, timestep, context, y, guidance=None, control=None, transformer_options={}, **kwargs):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
         bs, c, h, w = x.shape
         patch_size = self.patch_size
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
 
         img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
-
         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[:, :, 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)
-        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+        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 img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
+        bs, c, h_orig, w_orig = x.shape
+        patch_size = self.patch_size
+
+        h_len = ((h_orig + (patch_size // 2)) // patch_size)
+        w_len = ((w_orig + (patch_size // 2)) // patch_size)
+        img, img_ids = self.process_img(x)
+        img_tokens = img.shape[1]
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            for ref in ref_latents:
+                h_offset = 0
+                w_offset = 0
+                if ref.shape[-2] + h > ref.shape[-1] + w:
+                    w_offset = w
+                else:
+                    h_offset = h
+
+                kontext, kontext_ids = self.process_img(ref, index=1, 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))
-        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]
+        out = out[:, :img_tokens]
+        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_orig,:w_orig]

comfy/ldm/genmo/joint_model/asymm_models_joint.py

@@ -13,7 +13,6 @@ from comfy.ldm.modules.attention import optimized_attention
 from .layers import (
     FeedForward,
     PatchEmbed,
-    RMSNorm,
     TimestepEmbedder,
 )
 
@@ -90,10 +89,10 @@ class AsymmetricAttention(nn.Module):
 
         # Query and key normalization for stability.
         assert qk_norm
-        self.q_norm_x = RMSNorm(self.head_dim, device=device, dtype=dtype)
-        self.k_norm_x = RMSNorm(self.head_dim, device=device, dtype=dtype)
-        self.q_norm_y = RMSNorm(self.head_dim, device=device, dtype=dtype)
-        self.k_norm_y = RMSNorm(self.head_dim, device=device, dtype=dtype)
+        self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
 
         # Output layers. y features go back down from dim_x -> dim_y.
         self.proj_x = operations.Linear(dim_x, dim_x, bias=out_bias, device=device, dtype=dtype)

comfy/ldm/genmo/joint_model/layers.py

@@ -151,14 +151,3 @@ class PatchEmbed(nn.Module):
 
         x = self.norm(x)
         return x
-
-
-class RMSNorm(torch.nn.Module):
-    def __init__(self, hidden_size, eps=1e-5, device=None, dtype=None):
-        super().__init__()
-        self.eps = eps
-        self.weight = torch.nn.Parameter(torch.empty(hidden_size, device=device, dtype=dtype))
-        self.register_parameter("bias", None)
-
-    def forward(self, x):
-        return comfy.ldm.common_dit.rms_norm(x, self.weight, self.eps)

comfy/ldm/hidream/model.py

@@ -699,10 +699,13 @@ class HiDreamImageTransformer2DModel(nn.Module):
         y: Optional[torch.Tensor] = None,
         context: Optional[torch.Tensor] = None,
         encoder_hidden_states_llama3=None,
+        image_cond=None,
         control = None,
         transformer_options = {},
     ) -> torch.Tensor:
         bs, c, h, w = x.shape
+        if image_cond is not None:
+            x = torch.cat([x, image_cond], dim=-1)
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
         timesteps = t
         pooled_embeds = y

comfy/ldm/hunyuan_video/model.py

@@ -228,6 +228,7 @@ class HunyuanVideo(nn.Module):
         y: Tensor,
         guidance: Tensor = None,
         guiding_frame_index=None,
+        ref_latent=None,
         control=None,
         transformer_options={},
     ) -> Tensor:
@@ -238,6 +239,14 @@ class HunyuanVideo(nn.Module):
         img = self.img_in(img)
         vec = self.time_in(timestep_embedding(timesteps, 256, time_factor=1.0).to(img.dtype))
 
+        if ref_latent is not None:
+            ref_latent_ids = self.img_ids(ref_latent)
+            ref_latent = self.img_in(ref_latent)
+            img = torch.cat([ref_latent, img], dim=-2)
+            ref_latent_ids[..., 0] = -1
+            ref_latent_ids[..., 2] += (initial_shape[-1] // self.patch_size[-1])
+            img_ids = torch.cat([ref_latent_ids, img_ids], dim=-2)
+
         if guiding_frame_index is not None:
             token_replace_vec = self.time_in(timestep_embedding(guiding_frame_index, 256, time_factor=1.0))
             vec_ = self.vector_in(y[:, :self.params.vec_in_dim])
@@ -313,6 +322,8 @@ class HunyuanVideo(nn.Module):
                         img[:, : img_len] += add
 
         img = img[:, : img_len]
+        if ref_latent is not None:
+            img = img[:, ref_latent.shape[1]:]
 
         img = self.final_layer(img, vec, modulation_dims=modulation_dims)  # (N, T, patch_size ** 2 * out_channels)
 
@@ -324,7 +335,7 @@ class HunyuanVideo(nn.Module):
         img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
         return img
 
-    def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, control=None, transformer_options={}, **kwargs):
+    def img_ids(self, x):
         bs, c, t, h, w = x.shape
         patch_size = self.patch_size
         t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
@@ -334,7 +345,11 @@ class HunyuanVideo(nn.Module):
         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)
         img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
-        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+        return repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        img_ids = self.img_ids(x)
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
-        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, guiding_frame_index, control, transformer_options)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, guiding_frame_index, ref_latent, control=control, transformer_options=transformer_options)
         return out

comfy/ldm/hydit/models.py

@@ -3,7 +3,7 @@ import torch
 import torch.nn as nn
 
 import comfy.ops
-from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed, RMSNorm
+from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed
 from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
 from torch.utils import checkpoint
 
@@ -51,7 +51,7 @@ class HunYuanDiTBlock(nn.Module):
         if norm_type == "layer":
             norm_layer = operations.LayerNorm
         elif norm_type == "rms":
-            norm_layer = RMSNorm
+            norm_layer = operations.RMSNorm
         else:
             raise ValueError(f"Unknown norm_type: {norm_type}")
 

comfy/ldm/lightricks/model.py

@@ -1,7 +1,6 @@
 import torch
 from torch import nn
 import comfy.ldm.modules.attention
-from comfy.ldm.genmo.joint_model.layers import RMSNorm
 import comfy.ldm.common_dit
 from einops import rearrange
 import math
@@ -262,8 +261,8 @@ class CrossAttention(nn.Module):
         self.heads = heads
         self.dim_head = dim_head
 
-        self.q_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
-        self.k_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(inner_dim, eps=1e-5, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(inner_dim, eps=1e-5, dtype=dtype, device=device)
 
         self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)

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/lumina/model.py

@@ -8,7 +8,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 import comfy.ldm.common_dit
 
-from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, RMSNorm
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 
@@ -64,8 +64,8 @@ class JointAttention(nn.Module):
         )
 
         if qk_norm:
-            self.q_norm = RMSNorm(self.head_dim, elementwise_affine=True, **operation_settings)
-            self.k_norm = RMSNorm(self.head_dim, elementwise_affine=True, **operation_settings)
+            self.q_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+            self.k_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         else:
             self.q_norm = self.k_norm = nn.Identity()
 
@@ -242,11 +242,11 @@ class JointTransformerBlock(nn.Module):
             operation_settings=operation_settings,
         )
         self.layer_id = layer_id
-        self.attention_norm1 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings)
-        self.ffn_norm1 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings)
+        self.attention_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
 
-        self.attention_norm2 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings)
-        self.ffn_norm2 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings)
+        self.attention_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
 
         self.modulation = modulation
         if modulation:
@@ -431,7 +431,7 @@ class NextDiT(nn.Module):
 
         self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings)
         self.cap_embedder = nn.Sequential(
-            RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, **operation_settings),
+            operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
             operation_settings.get("operations").Linear(
                 cap_feat_dim,
                 dim,
@@ -457,7 +457,7 @@ class NextDiT(nn.Module):
                 for layer_id in range(n_layers)
             ]
         )
-        self.norm_final = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings)
+        self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings)
 
         assert (dim // n_heads) == sum(axes_dims)

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

@@ -20,8 +20,11 @@ if model_management.xformers_enabled():
 if model_management.sage_attention_enabled():
     try:
         from sageattention import sageattn
-    except ModuleNotFoundError:
-        logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
+    except ModuleNotFoundError as e:
+        if e.name == "sageattention":
+            logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
+        else:
+            raise e
         exit(-1)
 
 if model_management.flash_attention_enabled():
@@ -750,7 +753,7 @@ class BasicTransformerBlock(nn.Module):
             for p in patch:
                 n = p(n, extra_options)
 
-        x += n
+        x = n + x
         if "middle_patch" in transformer_patches:
             patch = transformer_patches["middle_patch"]
             for p in patch:
@@ -790,12 +793,12 @@ class BasicTransformerBlock(nn.Module):
             for p in patch:
                 n = p(n, extra_options)
 
-        x += n
+        x = n + x
         if self.is_res:
             x_skip = x
         x = self.ff(self.norm3(x))
         if self.is_res:
-            x += x_skip
+            x = x_skip + x
 
         return x
 

comfy/ldm/modules/sub_quadratic_attention.py

@@ -31,7 +31,7 @@ def dynamic_slice(
     starts: List[int],
     sizes: List[int],
 ) -> Tensor:
-    slicing = [slice(start, start + size) for start, size in zip(starts, sizes)]
+    slicing = tuple(slice(start, start + size) for start, size in zip(starts, sizes))
     return x[slicing]
 
 class AttnChunk(NamedTuple):

comfy/ldm/omnigen/omnigen2.py

@@ -0,0 +1,469 @@
+# Original code: https://github.com/VectorSpaceLab/OmniGen2
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from comfy.ldm.lightricks.model import Timesteps
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.modules.attention import optimized_attention_masked
+import comfy.model_management
+import comfy.ldm.common_dit
+
+
+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).to(dtype=x.dtype)
+
+
+def swiglu(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+    return F.silu(x) * y
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear_1 = operations.Linear(in_channels, time_embed_dim, dtype=dtype, device=device)
+        self.act = nn.SiLU()
+        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+
+class LuminaRMSNormZero(nn.Module):
+    def __init__(self, embedding_dim: int, norm_eps: float = 1e-5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(min(embedding_dim, 1024), 4 * embedding_dim, dtype=dtype, device=device)
+        self.norm = operations.RMSNorm(embedding_dim, eps=norm_eps, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor, emb: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
+        x = self.norm(x) * (1 + scale_msa[:, None])
+        return x, gate_msa, scale_mlp, gate_mlp
+
+
+class LuminaLayerNormContinuous(nn.Module):
+    def __init__(self, embedding_dim: int, conditioning_embedding_dim: int, elementwise_affine: bool = False, eps: float = 1e-6, out_dim: Optional[int] = None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear_1 = operations.Linear(conditioning_embedding_dim, embedding_dim, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(embedding_dim, eps, elementwise_affine, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(embedding_dim, out_dim, bias=True, dtype=dtype, device=device) if out_dim is not None else None
+
+    def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+        emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))
+        x = self.norm(x) * (1 + emb)[:, None, :]
+        if self.linear_2 is not None:
+            x = self.linear_2(x)
+        return x
+
+
+class LuminaFeedForward(nn.Module):
+    def __init__(self, dim: int, inner_dim: int, multiple_of: int = 256, dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)
+        self.linear_1 = operations.Linear(dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(inner_dim, dim, bias=False, dtype=dtype, device=device)
+        self.linear_3 = operations.Linear(dim, inner_dim, bias=False, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h1, h2 = self.linear_1(x), self.linear_3(x)
+        return self.linear_2(swiglu(h1, h2))
+
+
+class Lumina2CombinedTimestepCaptionEmbedding(nn.Module):
+    def __init__(self, hidden_size: int = 4096, text_feat_dim: int = 2048, frequency_embedding_size: int = 256, norm_eps: float = 1e-5, timestep_scale: float = 1.0, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0.0, scale=timestep_scale)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=min(hidden_size, 1024), dtype=dtype, device=device, operations=operations)
+        self.caption_embedder = nn.Sequential(
+            operations.RMSNorm(text_feat_dim, eps=norm_eps, dtype=dtype, device=device),
+            operations.Linear(text_feat_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, timestep: torch.Tensor, text_hidden_states: torch.Tensor, dtype: torch.dtype) -> Tuple[torch.Tensor, torch.Tensor]:
+        timestep_proj = self.time_proj(timestep).to(dtype=dtype)
+        time_embed = self.timestep_embedder(timestep_proj)
+        caption_embed = self.caption_embedder(text_hidden_states)
+        return time_embed, caption_embed
+
+
+class Attention(nn.Module):
+    def __init__(self, query_dim: int, dim_head: int, heads: int, kv_heads: int, eps: float = 1e-5, bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.heads = heads
+        self.kv_heads = kv_heads
+        self.dim_head = dim_head
+        self.scale = dim_head ** -0.5
+
+        self.to_q = operations.Linear(query_dim, heads * dim_head, bias=bias, dtype=dtype, device=device)
+        self.to_k = operations.Linear(query_dim, kv_heads * dim_head, bias=bias, dtype=dtype, device=device)
+        self.to_v = operations.Linear(query_dim, kv_heads * dim_head, bias=bias, dtype=dtype, device=device)
+
+        self.norm_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+        self.norm_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+
+        self.to_out = nn.Sequential(
+            operations.Linear(heads * dim_head, query_dim, bias=bias, dtype=dtype, device=device),
+            nn.Dropout(0.0)
+        )
+
+    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, image_rotary_emb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        query = self.to_q(hidden_states)
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        query = query.view(batch_size, -1, self.heads, self.dim_head)
+        key = key.view(batch_size, -1, self.kv_heads, self.dim_head)
+        value = value.view(batch_size, -1, self.kv_heads, self.dim_head)
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        if self.kv_heads < self.heads:
+            key = key.repeat_interleave(self.heads // self.kv_heads, dim=1)
+            value = value.repeat_interleave(self.heads // self.kv_heads, dim=1)
+
+        hidden_states = optimized_attention_masked(query, key, value, self.heads, attention_mask, skip_reshape=True)
+        hidden_states = self.to_out[0](hidden_states)
+        return hidden_states
+
+
+class OmniGen2TransformerBlock(nn.Module):
+    def __init__(self, dim: int, num_attention_heads: int, num_kv_heads: int, multiple_of: int, ffn_dim_multiplier: float, norm_eps: float, modulation: bool = True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.modulation = modulation
+
+        self.attn = Attention(
+            query_dim=dim,
+            dim_head=dim // num_attention_heads,
+            heads=num_attention_heads,
+            kv_heads=num_kv_heads,
+            eps=1e-5,
+            bias=False,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.feed_forward = LuminaFeedForward(
+            dim=dim,
+            inner_dim=4 * dim,
+            multiple_of=multiple_of,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        if modulation:
+            self.norm1 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations)
+        else:
+            self.norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+
+        self.ffn_norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+        self.norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+        self.ffn_norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, image_rotary_emb: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        if self.modulation:
+            norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
+            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
+            hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            hidden_states = hidden_states + self.norm2(attn_output)
+            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
+            hidden_states = hidden_states + self.ffn_norm2(mlp_output)
+        return hidden_states
+
+
+class OmniGen2RotaryPosEmbed(nn.Module):
+    def __init__(self, theta: int, axes_dim: Tuple[int, int, int], axes_lens: Tuple[int, int, int] = (300, 512, 512), patch_size: int = 2):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        self.axes_lens = axes_lens
+        self.patch_size = patch_size
+        self.rope_embedder = EmbedND(dim=sum(axes_dim), theta=self.theta, axes_dim=axes_dim)
+
+    def forward(self, batch_size, encoder_seq_len, l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, ref_img_sizes, img_sizes, device):
+        p = self.patch_size
+
+        seq_lengths = [cap_len + sum(ref_img_len) + img_len for cap_len, ref_img_len, img_len in zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len)]
+
+        max_seq_len = max(seq_lengths)
+        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
+        max_img_len = max(l_effective_img_len)
+
+        position_ids = torch.zeros(batch_size, max_seq_len, 3, dtype=torch.int32, device=device)
+
+        for i, (cap_seq_len, seq_len) in enumerate(zip(l_effective_cap_len, seq_lengths)):
+            position_ids[i, :cap_seq_len] = repeat(torch.arange(cap_seq_len, dtype=torch.int32, device=device), "l -> l 3")
+
+            pe_shift = cap_seq_len
+            pe_shift_len = cap_seq_len
+
+            if ref_img_sizes[i] is not None:
+                for ref_img_size, ref_img_len in zip(ref_img_sizes[i], l_effective_ref_img_len[i]):
+                    H, W = ref_img_size
+                    ref_H_tokens, ref_W_tokens = H // p, W // p
+
+                    row_ids = repeat(torch.arange(ref_H_tokens, dtype=torch.int32, device=device), "h -> h w", w=ref_W_tokens).flatten()
+                    col_ids = repeat(torch.arange(ref_W_tokens, dtype=torch.int32, device=device), "w -> h w", h=ref_H_tokens).flatten()
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 0] = pe_shift
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 1] = row_ids
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 2] = col_ids
+
+                    pe_shift += max(ref_H_tokens, ref_W_tokens)
+                    pe_shift_len += ref_img_len
+
+            H, W = img_sizes[i]
+            H_tokens, W_tokens = H // p, W // p
+
+            row_ids = repeat(torch.arange(H_tokens, dtype=torch.int32, device=device), "h -> h w", w=W_tokens).flatten()
+            col_ids = repeat(torch.arange(W_tokens, dtype=torch.int32, device=device), "w -> h w", h=H_tokens).flatten()
+
+            position_ids[i, pe_shift_len: seq_len, 0] = pe_shift
+            position_ids[i, pe_shift_len: seq_len, 1] = row_ids
+            position_ids[i, pe_shift_len: seq_len, 2] = col_ids
+
+        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2)
+
+        cap_freqs_cis_shape = list(freqs_cis.shape)
+        cap_freqs_cis_shape[1] = encoder_seq_len
+        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        ref_img_freqs_cis_shape = list(freqs_cis.shape)
+        ref_img_freqs_cis_shape[1] = max_ref_img_len
+        ref_img_freqs_cis = torch.zeros(*ref_img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        img_freqs_cis_shape = list(freqs_cis.shape)
+        img_freqs_cis_shape[1] = max_img_len
+        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        for i, (cap_seq_len, ref_img_len, img_len, seq_len) in enumerate(zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, seq_lengths)):
+            cap_freqs_cis[i, :cap_seq_len] = freqs_cis[i, :cap_seq_len]
+            ref_img_freqs_cis[i, :sum(ref_img_len)] = freqs_cis[i, cap_seq_len:cap_seq_len + sum(ref_img_len)]
+            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_seq_len + sum(ref_img_len):cap_seq_len + sum(ref_img_len) + img_len]
+
+        return cap_freqs_cis, ref_img_freqs_cis, img_freqs_cis, freqs_cis, l_effective_cap_len, seq_lengths
+
+
+class OmniGen2Transformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        out_channels: Optional[int] = None,
+        hidden_size: int = 2304,
+        num_layers: int = 26,
+        num_refiner_layers: int = 2,
+        num_attention_heads: int = 24,
+        num_kv_heads: int = 8,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        norm_eps: float = 1e-5,
+        axes_dim_rope: Tuple[int, int, int] = (32, 32, 32),
+        axes_lens: Tuple[int, int, int] = (300, 512, 512),
+        text_feat_dim: int = 1024,
+        timestep_scale: float = 1.0,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        self.patch_size = patch_size
+        self.out_channels = out_channels or in_channels
+        self.hidden_size = hidden_size
+        self.dtype = dtype
+
+        self.rope_embedder = OmniGen2RotaryPosEmbed(
+            theta=10000,
+            axes_dim=axes_dim_rope,
+            axes_lens=axes_lens,
+            patch_size=patch_size,
+        )
+
+        self.x_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device)
+        self.ref_image_patch_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device)
+
+        self.time_caption_embed = Lumina2CombinedTimestepCaptionEmbedding(
+            hidden_size=hidden_size,
+            text_feat_dim=text_feat_dim,
+            norm_eps=norm_eps,
+            timestep_scale=timestep_scale, dtype=dtype, device=device, operations=operations
+        )
+
+        self.noise_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.ref_image_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.context_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=False, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.layers = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_layers)
+        ])
+
+        self.norm_out = LuminaLayerNormContinuous(
+            embedding_dim=hidden_size,
+            conditioning_embedding_dim=min(hidden_size, 1024),
+            elementwise_affine=False,
+            eps=1e-6,
+            out_dim=patch_size * patch_size * self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+        self.image_index_embedding = nn.Parameter(torch.empty(5, hidden_size, device=device, dtype=dtype))
+
+    def flat_and_pad_to_seq(self, hidden_states, ref_image_hidden_states):
+        batch_size = len(hidden_states)
+        p = self.patch_size
+
+        img_sizes = [(img.size(1), img.size(2)) for img in hidden_states]
+        l_effective_img_len = [(H // p) * (W // p) for (H, W) in img_sizes]
+
+        if ref_image_hidden_states is not None:
+            ref_image_hidden_states = list(map(lambda ref: comfy.ldm.common_dit.pad_to_patch_size(ref, (p, p)), ref_image_hidden_states))
+            ref_img_sizes = [[(imgs.size(2), imgs.size(3)) if imgs is not None else None for imgs in ref_image_hidden_states]] * batch_size
+            l_effective_ref_img_len = [[(ref_img_size[0] // p) * (ref_img_size[1] // p) for ref_img_size in _ref_img_sizes] if _ref_img_sizes is not None else [0] for _ref_img_sizes in ref_img_sizes]
+        else:
+            ref_img_sizes = [None for _ in range(batch_size)]
+            l_effective_ref_img_len = [[0] for _ in range(batch_size)]
+
+        flat_ref_img_hidden_states = None
+        if ref_image_hidden_states is not None:
+            imgs = []
+            for ref_img in ref_image_hidden_states:
+                B, C, H, W = ref_img.size()
+                ref_img = rearrange(ref_img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
+                imgs.append(ref_img)
+            flat_ref_img_hidden_states = torch.cat(imgs, dim=1)
+
+        img = hidden_states
+        B, C, H, W = img.size()
+        flat_hidden_states = rearrange(img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
+
+        return (
+            flat_hidden_states, flat_ref_img_hidden_states,
+            None, None,
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes,
+        )
+
+    def img_patch_embed_and_refine(self, hidden_states, ref_image_hidden_states, padded_img_mask, padded_ref_img_mask, noise_rotary_emb, ref_img_rotary_emb, l_effective_ref_img_len, l_effective_img_len, temb):
+        batch_size = len(hidden_states)
+
+        hidden_states = self.x_embedder(hidden_states)
+        if ref_image_hidden_states is not None:
+            ref_image_hidden_states = self.ref_image_patch_embedder(ref_image_hidden_states)
+            image_index_embedding = comfy.model_management.cast_to(self.image_index_embedding, dtype=hidden_states.dtype, device=hidden_states.device)
+
+            for i in range(batch_size):
+                shift = 0
+                for j, ref_img_len in enumerate(l_effective_ref_img_len[i]):
+                    ref_image_hidden_states[i, shift:shift + ref_img_len, :] = ref_image_hidden_states[i, shift:shift + ref_img_len, :] + image_index_embedding[j]
+                    shift += ref_img_len
+
+        for layer in self.noise_refiner:
+            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
+
+        if ref_image_hidden_states is not None:
+            for layer in self.ref_image_refiner:
+                ref_image_hidden_states = layer(ref_image_hidden_states, padded_ref_img_mask, ref_img_rotary_emb, temb)
+
+            hidden_states = torch.cat([ref_image_hidden_states, hidden_states], dim=1)
+
+        return hidden_states
+
+    def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, **kwargs):
+        B, C, H, W = x.shape
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        _, _, H_padded, W_padded = hidden_states.shape
+        timestep = 1.0 - timesteps
+        text_hidden_states = context
+        text_attention_mask = attention_mask
+        ref_image_hidden_states = ref_latents
+        device = hidden_states.device
+
+        temb, text_hidden_states = self.time_caption_embed(timestep, text_hidden_states, hidden_states[0].dtype)
+
+        (
+            hidden_states, ref_image_hidden_states,
+            img_mask, ref_img_mask,
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes,
+        ) = self.flat_and_pad_to_seq(hidden_states, ref_image_hidden_states)
+
+        (
+            context_rotary_emb, ref_img_rotary_emb, noise_rotary_emb,
+            rotary_emb, encoder_seq_lengths, seq_lengths,
+        ) = self.rope_embedder(
+            hidden_states.shape[0], text_hidden_states.shape[1], [num_tokens] * text_hidden_states.shape[0],
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes, device,
+        )
+
+        for layer in self.context_refiner:
+            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)
+
+        img_len = hidden_states.shape[1]
+        combined_img_hidden_states = self.img_patch_embed_and_refine(
+            hidden_states, ref_image_hidden_states,
+            img_mask, ref_img_mask,
+            noise_rotary_emb, ref_img_rotary_emb,
+            l_effective_ref_img_len, l_effective_img_len,
+            temb,
+        )
+
+        hidden_states = torch.cat([text_hidden_states, combined_img_hidden_states], dim=1)
+        attention_mask = None
+
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)
+
+        hidden_states = self.norm_out(hidden_states, temb)
+
+        p = self.patch_size
+        output = rearrange(hidden_states[:, -img_len:], 'b (h w) (p1 p2 c) -> b c (h p1) (w p2)',  h=H_padded // p, w=W_padded// p, p1=p, p2=p)[:, :, :H, :W]
+
+        return -output

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/wan/model.py

@@ -9,7 +9,6 @@ from einops import repeat
 from comfy.ldm.modules.attention import optimized_attention
 from comfy.ldm.flux.layers import EmbedND
 from comfy.ldm.flux.math import apply_rope
-from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
 import comfy.ldm.common_dit
 import comfy.model_management
 
@@ -49,8 +48,8 @@ class WanSelfAttention(nn.Module):
         self.k = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.v = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.o = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
-        self.norm_q = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
-        self.norm_k = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_q = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_k = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
 
     def forward(self, x, freqs):
         r"""
@@ -114,7 +113,7 @@ class WanI2VCrossAttention(WanSelfAttention):
         self.k_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.v_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         # self.alpha = nn.Parameter(torch.zeros((1, )))
-        self.norm_k_img = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_k_img = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
 
     def forward(self, x, context, context_img_len):
         r"""
@@ -147,6 +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,
@@ -203,23 +211,108 @@ 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
 
 
+class VaceWanAttentionBlock(WanAttentionBlock):
+    def __init__(
+            self,
+            cross_attn_type,
+            dim,
+            ffn_dim,
+            num_heads,
+            window_size=(-1, -1),
+            qk_norm=True,
+            cross_attn_norm=False,
+            eps=1e-6,
+            block_id=0,
+            operation_settings={}
+    ):
+        super().__init__(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.after_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, c, x, **kwargs):
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+        c = super().forward(c, **kwargs)
+        c_skip = self.after_proj(c)
+        return c_skip, c
+
+
+class WanCamAdapter(nn.Module):
+    def __init__(self, in_dim, out_dim, kernel_size, stride, num_residual_blocks=1, operation_settings={}):
+        super(WanCamAdapter, self).__init__()
+
+        # Pixel Unshuffle: reduce spatial dimensions by a factor of 8
+        self.pixel_unshuffle = nn.PixelUnshuffle(downscale_factor=8)
+
+        # Convolution: reduce spatial dimensions by a factor
+        #  of 2 (without overlap)
+        self.conv = operation_settings.get("operations").Conv2d(in_dim * 64, out_dim, kernel_size=kernel_size, stride=stride, padding=0, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # Residual blocks for feature extraction
+        self.residual_blocks = nn.Sequential(
+            *[WanCamResidualBlock(out_dim, operation_settings = operation_settings) for _ in range(num_residual_blocks)]
+        )
+
+    def forward(self, x):
+        # Reshape to merge the frame dimension into batch
+        bs, c, f, h, w = x.size()
+        x = x.permute(0, 2, 1, 3, 4).contiguous().view(bs * f, c, h, w)
+
+        # Pixel Unshuffle operation
+        x_unshuffled = self.pixel_unshuffle(x)
+
+        # Convolution operation
+        x_conv = self.conv(x_unshuffled)
+
+        # Feature extraction with residual blocks
+        out = self.residual_blocks(x_conv)
+
+        # Reshape to restore original bf dimension
+        out = out.view(bs, f, out.size(1), out.size(2), out.size(3))
+
+        # Permute dimensions to reorder (if needed), e.g., swap channels and feature frames
+        out = out.permute(0, 2, 1, 3, 4)
+
+        return out
+
+
+class WanCamResidualBlock(nn.Module):
+    def __init__(self, dim, operation_settings={}):
+        super(WanCamResidualBlock, self).__init__()
+        self.conv1 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        residual = x
+        out = self.relu(self.conv1(x))
+        out = self.conv2(out)
+        out += residual
+        return out
+
+
 class Head(nn.Module):
 
     def __init__(self, dim, out_dim, patch_size, eps=1e-6, operation_settings={}):
@@ -244,8 +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
 
 
@@ -395,6 +492,7 @@ class WanModel(torch.nn.Module):
         clip_fea=None,
         freqs=None,
         transformer_options={},
+        **kwargs,
     ):
         r"""
         Forward pass through the diffusion model
@@ -422,6 +520,260 @@ class WanModel(torch.nn.Module):
         grid_sizes = x.shape[2:]
         x = x.flatten(2).transpose(1, 2)
 
+        # time embeddings
+        e = self.time_embedding(
+            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))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0])
+
+        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)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
+
+    def unpatchify(self, x, grid_sizes):
+        r"""
+        Reconstruct video tensors from patch embeddings.
+
+        Args:
+            x (List[Tensor]):
+                List of patchified features, each with shape [L, C_out * prod(patch_size)]
+            grid_sizes (Tensor):
+                Original spatial-temporal grid dimensions before patching,
+                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
+
+        Returns:
+            List[Tensor]:
+                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        u = x
+        b = u.shape[0]
+        u = u[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
+        u = torch.einsum('bfhwpqrc->bcfphqwr', u)
+        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
+        return u
+
+
+class VaceWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='vace',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 vace_layers=None,
+                 vace_in_dim=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='t2v', 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}
+
+        # Vace
+        if vace_layers is not None:
+            self.vace_layers = vace_layers
+            self.vace_in_dim = vace_in_dim
+            # vace blocks
+            self.vace_blocks = nn.ModuleList([
+                VaceWanAttentionBlock('t2v_cross_attn', self.dim, self.ffn_dim, self.num_heads, self.window_size, self.qk_norm, self.cross_attn_norm, self.eps, block_id=i, operation_settings=operation_settings)
+                for i in range(self.vace_layers)
+            ])
+
+            self.vace_layers_mapping = {i: n for n, i in enumerate(range(0, self.num_layers, self.num_layers // self.vace_layers))}
+            # vace patch embeddings
+            self.vace_patch_embedding = operations.Conv3d(
+                self.vace_in_dim, self.dim, kernel_size=self.patch_size, stride=self.patch_size, device=device, dtype=torch.float32
+            )
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        vace_context,
+        vace_strength,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # 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))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        orig_shape = list(vace_context.shape)
+        vace_context = vace_context.movedim(0, 1).reshape([-1] + orig_shape[2:])
+        c = self.vace_patch_embedding(vace_context.float()).to(vace_context.dtype)
+        c = c.flatten(2).transpose(1, 2)
+        c = list(c.split(orig_shape[0], dim=0))
+
+        # arguments
+        x_orig = x
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+            ii = self.vace_layers_mapping.get(i, None)
+            if ii is not None:
+                for iii in range(len(c)):
+                    c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                    x += c_skip * vace_strength[iii]
+                del c_skip
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+class CameraWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='camera',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 in_dim_control_adapter=24,
+                 device=None,
+                 dtype=None,
+                 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)
+        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)
+
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+        camera_conditions = None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # 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
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
         # time embeddings
         e = self.time_embedding(
             sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
@@ -456,43 +808,3 @@ class WanModel(torch.nn.Module):
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
-
-    def forward(self, x, timestep, context, clip_fea=None, transformer_options={},**kwargs):
-        bs, c, t, h, w = x.shape
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
-        patch_size = self.patch_size
-        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
-        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
-        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
-        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)
-        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
-        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
-
-        freqs = self.rope_embedder(img_ids).movedim(1, 2)
-        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options)[:, :, :t, :h, :w]
-
-    def unpatchify(self, x, grid_sizes):
-        r"""
-        Reconstruct video tensors from patch embeddings.
-
-        Args:
-            x (List[Tensor]):
-                List of patchified features, each with shape [L, C_out * prod(patch_size)]
-            grid_sizes (Tensor):
-                Original spatial-temporal grid dimensions before patching,
-                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
-
-        Returns:
-            List[Tensor]:
-                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
-        """
-
-        c = self.out_dim
-        u = x
-        b = u.shape[0]
-        u = u[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
-        u = torch.einsum('bfhwpqrc->bcfphqwr', u)
-        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
-        return u

comfy/ldm/wan/vae.py

@@ -52,15 +52,6 @@ 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 +64,11 @@ 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 +148,6 @@ 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):
 
@@ -494,12 +462,6 @@ 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 +507,6 @@ 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, feat_cache[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

@@ -20,6 +20,7 @@ from __future__ import annotations
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
+import comfy.weight_adapter as weight_adapter
 import logging
 import torch
 
@@ -49,139 +50,12 @@ def load_lora(lora, to_load, log_missing=True):
             dora_scale = lora[dora_scale_name]
             loaded_keys.add(dora_scale_name)
 
-        reshape_name = "{}.reshape_weight".format(x)
-        reshape = None
-        if reshape_name in lora.keys():
-            try:
-                reshape = lora[reshape_name].tolist()
-                loaded_keys.add(reshape_name)
-            except:
-                pass
-
-        regular_lora = "{}.lora_up.weight".format(x)
-        diffusers_lora = "{}_lora.up.weight".format(x)
-        diffusers2_lora = "{}.lora_B.weight".format(x)
-        diffusers3_lora = "{}.lora.up.weight".format(x)
-        mochi_lora = "{}.lora_B".format(x)
-        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
-        A_name = None
-
-        if regular_lora in lora.keys():
-            A_name = regular_lora
-            B_name = "{}.lora_down.weight".format(x)
-            mid_name = "{}.lora_mid.weight".format(x)
-        elif diffusers_lora in lora.keys():
-            A_name = diffusers_lora
-            B_name = "{}_lora.down.weight".format(x)
-            mid_name = None
-        elif diffusers2_lora in lora.keys():
-            A_name = diffusers2_lora
-            B_name = "{}.lora_A.weight".format(x)
-            mid_name = None
-        elif diffusers3_lora in lora.keys():
-            A_name = diffusers3_lora
-            B_name = "{}.lora.down.weight".format(x)
-            mid_name = None
-        elif mochi_lora in lora.keys():
-            A_name = mochi_lora
-            B_name = "{}.lora_A".format(x)
-            mid_name = None
-        elif transformers_lora in lora.keys():
-            A_name = transformers_lora
-            B_name ="{}.lora_linear_layer.down.weight".format(x)
-            mid_name = None
-
-        if A_name is not None:
-            mid = None
-            if mid_name is not None and mid_name in lora.keys():
-                mid = lora[mid_name]
-                loaded_keys.add(mid_name)
-            patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale, reshape))
-            loaded_keys.add(A_name)
-            loaded_keys.add(B_name)
-
-
-        ######## loha
-        hada_w1_a_name = "{}.hada_w1_a".format(x)
-        hada_w1_b_name = "{}.hada_w1_b".format(x)
-        hada_w2_a_name = "{}.hada_w2_a".format(x)
-        hada_w2_b_name = "{}.hada_w2_b".format(x)
-        hada_t1_name = "{}.hada_t1".format(x)
-        hada_t2_name = "{}.hada_t2".format(x)
-        if hada_w1_a_name in lora.keys():
-            hada_t1 = None
-            hada_t2 = None
-            if hada_t1_name in lora.keys():
-                hada_t1 = lora[hada_t1_name]
-                hada_t2 = lora[hada_t2_name]
-                loaded_keys.add(hada_t1_name)
-                loaded_keys.add(hada_t2_name)
-
-            patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale))
-            loaded_keys.add(hada_w1_a_name)
-            loaded_keys.add(hada_w1_b_name)
-            loaded_keys.add(hada_w2_a_name)
-            loaded_keys.add(hada_w2_b_name)
-
-
-        ######## lokr
-        lokr_w1_name = "{}.lokr_w1".format(x)
-        lokr_w2_name = "{}.lokr_w2".format(x)
-        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
-        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
-        lokr_t2_name = "{}.lokr_t2".format(x)
-        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
-        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
-
-        lokr_w1 = None
-        if lokr_w1_name in lora.keys():
-            lokr_w1 = lora[lokr_w1_name]
-            loaded_keys.add(lokr_w1_name)
-
-        lokr_w2 = None
-        if lokr_w2_name in lora.keys():
-            lokr_w2 = lora[lokr_w2_name]
-            loaded_keys.add(lokr_w2_name)
-
-        lokr_w1_a = None
-        if lokr_w1_a_name in lora.keys():
-            lokr_w1_a = lora[lokr_w1_a_name]
-            loaded_keys.add(lokr_w1_a_name)
-
-        lokr_w1_b = None
-        if lokr_w1_b_name in lora.keys():
-            lokr_w1_b = lora[lokr_w1_b_name]
-            loaded_keys.add(lokr_w1_b_name)
-
-        lokr_w2_a = None
-        if lokr_w2_a_name in lora.keys():
-            lokr_w2_a = lora[lokr_w2_a_name]
-            loaded_keys.add(lokr_w2_a_name)
-
-        lokr_w2_b = None
-        if lokr_w2_b_name in lora.keys():
-            lokr_w2_b = lora[lokr_w2_b_name]
-            loaded_keys.add(lokr_w2_b_name)
-
-        lokr_t2 = None
-        if lokr_t2_name in lora.keys():
-            lokr_t2 = lora[lokr_t2_name]
-            loaded_keys.add(lokr_t2_name)
-
-        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
-            patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale))
-
-        #glora
-        a1_name = "{}.a1.weight".format(x)
-        a2_name = "{}.a2.weight".format(x)
-        b1_name = "{}.b1.weight".format(x)
-        b2_name = "{}.b2.weight".format(x)
-        if a1_name in lora:
-            patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale))
-            loaded_keys.add(a1_name)
-            loaded_keys.add(a2_name)
-            loaded_keys.add(b1_name)
-            loaded_keys.add(b2_name)
+        for adapter_cls in weight_adapter.adapters:
+            adapter = adapter_cls.load(x, lora, alpha, dora_scale, loaded_keys)
+            if adapter is not None:
+                patch_dict[to_load[x]] = adapter
+                loaded_keys.update(adapter.loaded_keys)
+                continue
 
         w_norm_name = "{}.w_norm".format(x)
         b_norm_name = "{}.b_norm".format(x)
@@ -405,29 +279,23 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["transformer.{}".format(key_lora)] = k
                 key_map["diffusion_model.{}".format(key_lora)] = k  # Old loras
 
+    if isinstance(model, comfy.model_base.HiDream):
+        for k in sdk:
+            if k.startswith("diffusion_model."):
+                if k.endswith(".weight"):
+                    key_lora = k[len("diffusion_model."):-len(".weight")]
+                    key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+                    key_map["transformer.{}".format(key_lora)] = k #SimpleTuner regular format
+
+    if isinstance(model, comfy.model_base.ACEStep):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"): #Official ACE step lora format
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["{}".format(key_lora)] = k
+
     return key_map
 
 
-def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function):
-    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
-    lora_diff *= alpha
-    weight_calc = weight + function(lora_diff).type(weight.dtype)
-    weight_norm = (
-        weight_calc.transpose(0, 1)
-        .reshape(weight_calc.shape[1], -1)
-        .norm(dim=1, keepdim=True)
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
-        .transpose(0, 1)
-    )
-
-    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
-    if strength != 1.0:
-        weight_calc -= weight
-        weight += strength * (weight_calc)
-    else:
-        weight[:] = weight_calc
-    return weight
-
 def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
     """
     Pad a tensor to a new shape with zeros.
@@ -482,6 +350,16 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
         if isinstance(v, list):
             v = (calculate_weight(v[1:], v[0][1](comfy.model_management.cast_to_device(v[0][0], weight.device, intermediate_dtype, copy=True), inplace=True), key, intermediate_dtype=intermediate_dtype), )
 
+        if isinstance(v, weight_adapter.WeightAdapterBase):
+            output = v.calculate_weight(weight, key, strength, strength_model, offset, function, intermediate_dtype, original_weights)
+            if output is None:
+                logging.warning("Calculate Weight Failed: {} {}".format(v.name, key))
+            else:
+                weight = output
+                if old_weight is not None:
+                    weight = old_weight
+            continue
+
         if len(v) == 1:
             patch_type = "diff"
         elif len(v) == 2:
@@ -508,157 +386,6 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             diff_weight = comfy.model_management.cast_to_device(target_weight, weight.device, intermediate_dtype) - \
                           comfy.model_management.cast_to_device(original_weights[key][0][0], weight.device, intermediate_dtype)
             weight += function(strength * comfy.model_management.cast_to_device(diff_weight, weight.device, weight.dtype))
-        elif patch_type == "lora": #lora/locon
-            mat1 = comfy.model_management.cast_to_device(v[0], weight.device, intermediate_dtype)
-            mat2 = comfy.model_management.cast_to_device(v[1], weight.device, intermediate_dtype)
-            dora_scale = v[4]
-            reshape = v[5]
-
-            if reshape is not None:
-                weight = pad_tensor_to_shape(weight, reshape)
-
-            if v[2] is not None:
-                alpha = v[2] / mat2.shape[0]
-            else:
-                alpha = 1.0
-
-            if v[3] is not None:
-                #locon mid weights, hopefully the math is fine because I didn't properly test it
-                mat3 = comfy.model_management.cast_to_device(v[3], weight.device, intermediate_dtype)
-                final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-            try:
-                lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "lokr":
-            w1 = v[0]
-            w2 = v[1]
-            w1_a = v[3]
-            w1_b = v[4]
-            w2_a = v[5]
-            w2_b = v[6]
-            t2 = v[7]
-            dora_scale = v[8]
-            dim = None
-
-            if w1 is None:
-                dim = w1_b.shape[0]
-                w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
-            else:
-                w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
-
-            if w2 is None:
-                dim = w2_b.shape[0]
-                if t2 is None:
-                    w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
-                else:
-                    w2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                        comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
-                                        comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
-                                        comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
-            else:
-                w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
-
-            if len(w2.shape) == 4:
-                w1 = w1.unsqueeze(2).unsqueeze(2)
-            if v[2] is not None and dim is not None:
-                alpha = v[2] / dim
-            else:
-                alpha = 1.0
-
-            try:
-                lora_diff = torch.kron(w1, w2).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "loha":
-            w1a = v[0]
-            w1b = v[1]
-            if v[2] is not None:
-                alpha = v[2] / w1b.shape[0]
-            else:
-                alpha = 1.0
-
-            w2a = v[3]
-            w2b = v[4]
-            dora_scale = v[7]
-            if v[5] is not None: #cp decomposition
-                t1 = v[5]
-                t2 = v[6]
-                m1 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                    comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
-
-                m2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
-            else:
-                m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
-                m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
-
-            try:
-                lora_diff = (m1 * m2).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "glora":
-            dora_scale = v[5]
-
-            old_glora = False
-            if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]:
-                rank = v[0].shape[0]
-                old_glora = True
-
-            if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]:
-                if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]:
-                    pass
-                else:
-                    old_glora = False
-                    rank = v[1].shape[0]
-
-            a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
-            a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
-            b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
-            b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
-
-            if v[4] is not None:
-                alpha = v[4] / rank
-            else:
-                alpha = 1.0
-
-            try:
-                if old_glora:
-                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora
-                else:
-                    if weight.dim() > 2:
-                        lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
-                    else:
-                        lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
-                    lora_diff += torch.mm(b1, b2).reshape(weight.shape)
-
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
         else:
             logging.warning("patch type not recognized {} {}".format(patch_type, key))
 

comfy/model_base.py

@@ -34,10 +34,14 @@ import comfy.ldm.flux.model
 import comfy.ldm.lightricks.model
 import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
+import comfy.ldm.cosmos.predict2
 import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
 import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
+import comfy.ldm.chroma.model
+import comfy.ldm.ace.model
+import comfy.ldm.omnigen.omnigen2
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -46,6 +50,7 @@ import comfy.ops
 from enum import Enum
 from . import utils
 import comfy.latent_formats
+import comfy.model_sampling
 import math
 from typing import TYPE_CHECKING
 if TYPE_CHECKING:
@@ -61,38 +66,39 @@ class ModelType(Enum):
     V_PREDICTION_CONTINUOUS = 7
     FLUX = 8
     IMG_TO_IMG = 9
-
-
-from comfy.model_sampling import EPS, V_PREDICTION, EDM, ModelSamplingDiscrete, ModelSamplingContinuousEDM, StableCascadeSampling, ModelSamplingContinuousV
+    FLOW_COSMOS = 10
 
 
 def model_sampling(model_config, model_type):
-    s = ModelSamplingDiscrete
+    s = comfy.model_sampling.ModelSamplingDiscrete
 
     if model_type == ModelType.EPS:
-        c = EPS
+        c = comfy.model_sampling.EPS
     elif model_type == ModelType.V_PREDICTION:
-        c = V_PREDICTION
+        c = comfy.model_sampling.V_PREDICTION
     elif model_type == ModelType.V_PREDICTION_EDM:
-        c = V_PREDICTION
-        s = ModelSamplingContinuousEDM
+        c = comfy.model_sampling.V_PREDICTION
+        s = comfy.model_sampling.ModelSamplingContinuousEDM
     elif model_type == ModelType.FLOW:
         c = comfy.model_sampling.CONST
         s = comfy.model_sampling.ModelSamplingDiscreteFlow
     elif model_type == ModelType.STABLE_CASCADE:
-        c = EPS
-        s = StableCascadeSampling
+        c = comfy.model_sampling.EPS
+        s = comfy.model_sampling.StableCascadeSampling
     elif model_type == ModelType.EDM:
-        c = EDM
-        s = ModelSamplingContinuousEDM
+        c = comfy.model_sampling.EDM
+        s = comfy.model_sampling.ModelSamplingContinuousEDM
     elif model_type == ModelType.V_PREDICTION_CONTINUOUS:
-        c = V_PREDICTION
-        s = ModelSamplingContinuousV
+        c = comfy.model_sampling.V_PREDICTION
+        s = comfy.model_sampling.ModelSamplingContinuousV
     elif model_type == ModelType.FLUX:
         c = comfy.model_sampling.CONST
         s = comfy.model_sampling.ModelSamplingFlux
     elif model_type == ModelType.IMG_TO_IMG:
         c = comfy.model_sampling.IMG_TO_IMG
+    elif model_type == ModelType.FLOW_COSMOS:
+        c = comfy.model_sampling.COSMOS_RFLOW
+        s = comfy.model_sampling.ModelSamplingCosmosRFlow
 
     class ModelSampling(s, c):
         pass
@@ -100,6 +106,13 @@ def model_sampling(model_config, model_type):
     return ModelSampling(model_config)
 
 
+def convert_tensor(extra, dtype):
+    if hasattr(extra, "dtype"):
+        if extra.dtype != torch.int and extra.dtype != torch.long:
+            extra = extra.to(dtype)
+    return extra
+
+
 class BaseModel(torch.nn.Module):
     def __init__(self, model_config, model_type=ModelType.EPS, device=None, unet_model=UNetModel):
         super().__init__()
@@ -133,6 +146,7 @@ class BaseModel(torch.nn.Module):
         logging.info("model_type {}".format(model_type.name))
         logging.debug("adm {}".format(self.adm_channels))
         self.memory_usage_factor = model_config.memory_usage_factor
+        self.memory_usage_factor_conds = ()
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -162,9 +176,14 @@ class BaseModel(torch.nn.Module):
         extra_conds = {}
         for o in kwargs:
             extra = kwargs[o]
+
             if hasattr(extra, "dtype"):
-                if extra.dtype != torch.int and extra.dtype != torch.long:
-                    extra = extra.to(dtype)
+                extra = convert_tensor(extra, dtype)
+            elif isinstance(extra, list):
+                ex = []
+                for ext in extra:
+                    ex.append(convert_tensor(ext, dtype))
+                extra = ex
             extra_conds[o] = extra
 
         t = self.process_timestep(t, x=x, **extra_conds)
@@ -323,19 +342,28 @@ class BaseModel(torch.nn.Module):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return self.model_sampling.noise_scaling(sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1)), noise, latent_image)
 
-    def memory_required(self, input_shape):
+    def memory_required(self, input_shape, cond_shapes={}):
+        input_shapes = [input_shape]
+        for c in self.memory_usage_factor_conds:
+            shape = cond_shapes.get(c, None)
+            if shape is not None and len(shape) > 0:
+                input_shapes += shape
+
         if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
             dtype = self.get_dtype()
             if self.manual_cast_dtype is not None:
                 dtype = self.manual_cast_dtype
             #TODO: this needs to be tweaked
-            area = input_shape[0] * math.prod(input_shape[2:])
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
         else:
             #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
-            area = input_shape[0] * math.prod(input_shape[2:])
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
 
+    def extra_conds_shapes(self, **kwargs):
+        return {}
+
 
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
     adm_inputs = []
@@ -786,8 +814,9 @@ class PixArt(BaseModel):
         return out
 
 class Flux(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.flux.model.Flux)
+    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 = ("ref_latents",)
 
     def concat_cond(self, **kwargs):
         try:
@@ -848,8 +877,23 @@ class Flux(BaseModel):
         guidance = kwargs.get("guidance", 3.5)
         if guidance is not None:
             out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        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)
         return out
 
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        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 GenmoMochi(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.genmo.joint_model.asymm_models_joint.AsymmDiTJoint)
@@ -922,6 +966,10 @@ class HunyuanVideo(BaseModel):
         if guiding_frame_index is not None:
             out['guiding_frame_index'] = comfy.conds.CONDRegular(torch.FloatTensor([guiding_frame_index]))
 
+        ref_latent = kwargs.get("ref_latent", None)
+        if ref_latent is not None:
+            out['ref_latent'] = comfy.conds.CONDRegular(self.process_latent_in(ref_latent))
+
         return out
 
     def scale_latent_inpaint(self, latent_image, **kwargs):
@@ -970,6 +1018,45 @@ class CosmosVideo(BaseModel):
         latent_image = self.model_sampling.calculate_input(torch.tensor([sigma_noise_augmentation], device=latent_image.device, dtype=latent_image.dtype), latent_image)
         return latent_image * ((sigma ** 2 + self.model_sampling.sigma_data ** 2) ** 0.5)
 
+class CosmosPredict2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW_COSMOS, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.cosmos.predict2.MiniTrainDIT)
+        self.image_to_video = image_to_video
+        if self.image_to_video:
+            self.concat_keys = ("mask_inverted",)
+
+    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)
+
+        out['fps'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", None))
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        if denoise_mask.ndim <= 4:
+            return timestep
+        condition_video_mask_B_1_T_1_1 = denoise_mask.mean(dim=[1, 3, 4], keepdim=True)
+        c_noise_B_1_T_1_1 = 0.0 * (1.0 - condition_video_mask_B_1_T_1_1) + timestep.reshape(timestep.shape[0], 1, 1, 1, 1) * condition_video_mask_B_1_T_1_1
+        out = c_noise_B_1_T_1_1.squeeze(dim=[1, 3, 4])
+        return out
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        sigma = sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1))
+        sigma_noise_augmentation = 0 #TODO
+        if sigma_noise_augmentation != 0:
+            latent_image = latent_image + noise
+        latent_image = self.model_sampling.calculate_input(torch.tensor([sigma_noise_augmentation], device=latent_image.device, dtype=latent_image.dtype), latent_image)
+        sigma = (sigma / (sigma + 1))
+        return latent_image / (1.0 - sigma)
+
 class Lumina2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
@@ -1010,8 +1097,9 @@ 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)]
@@ -1041,8 +1129,85 @@ class WAN21(BaseModel):
         clip_vision_output = kwargs.get("clip_vision_output", None)
         if clip_vision_output is not None:
             out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
+
+        time_dim_concat = kwargs.get("time_dim_concat", None)
+        if time_dim_concat is not None:
+            out['time_dim_concat'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_concat))
+
         return out
 
+
+class WAN21_Vace(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.VaceWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        noise = kwargs.get("noise", None)
+        noise_shape = list(noise.shape)
+        vace_frames = kwargs.get("vace_frames", None)
+        if vace_frames is None:
+            noise_shape[1] = 32
+            vace_frames = [torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)]
+
+        mask = kwargs.get("vace_mask", None)
+        if mask is None:
+            noise_shape[1] = 64
+            mask = [torch.ones(noise_shape, device=noise.device, dtype=noise.dtype)] * len(vace_frames)
+
+        vace_frames_out = []
+        for j in range(len(vace_frames)):
+            vf = vace_frames[j].clone()
+            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)
+            vace_frames_out.append(vf)
+
+        vace_frames = torch.stack(vace_frames_out, dim=1)
+        out['vace_context'] = comfy.conds.CONDRegular(vace_frames)
+
+        vace_strength = kwargs.get("vace_strength", [1.0] * len(vace_frames_out))
+        out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
+        return out
+
+class WAN21_Camera(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.CameraWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        camera_conditions = kwargs.get("camera_conditions", None)
+        if camera_conditions is not None:
+            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)
@@ -1073,4 +1238,68 @@ class HiDream(BaseModel):
         conditioning_llama3 = kwargs.get("conditioning_llama3", None)
         if conditioning_llama3 is not None:
             out['encoder_hidden_states_llama3'] = comfy.conds.CONDRegular(conditioning_llama3)
+        image_cond = kwargs.get("concat_latent_image", None)
+        if image_cond is not None:
+            out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond))
+        return out
+
+class Chroma(Flux):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        guidance = kwargs.get("guidance", 0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out
+
+class ACEStep(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.model.ACEStepTransformer2DModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        noise = kwargs.get("noise", None)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_token_idx'] = comfy.conds.CONDRegular(conditioning_lyrics)
+        out['speaker_embeds'] = comfy.conds.CONDRegular(torch.zeros(noise.shape[0], 512, device=noise.device, dtype=noise.dtype))
+        out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
+        return out
+
+class Omnigen2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+        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)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        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

comfy/model_detection.py

@@ -164,7 +164,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         if in_key in state_dict_keys:
             dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
         dit_config["out_channels"] = 16
-        dit_config["vec_in_dim"] = 768
+        vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
+        if vec_in_key in state_dict_keys:
+            dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072
         dit_config["mlp_ratio"] = 4.0
@@ -174,7 +176,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["axes_dim"] = [16, 56, 56]
         dit_config["theta"] = 10000
         dit_config["qkv_bias"] = True
-        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+            dit_config["image_model"] = "chroma"
+            dit_config["in_channels"] = 64
+            dit_config["out_channels"] = 64
+            dit_config["in_dim"] = 64
+            dit_config["out_dim"] = 3072
+            dit_config["hidden_dim"] = 5120
+            dit_config["n_layers"] = 5
+        else:
+            dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
         return dit_config
 
     if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview
@@ -211,10 +222,39 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv
         dit_config = {}
         dit_config["image_model"] = "ltxv"
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
+        shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape
+        dit_config["attention_head_dim"] = shape[0] // 32
+        dit_config["cross_attention_dim"] = shape[1]
         if metadata is not None and "config" in metadata:
             dit_config.update(json.loads(metadata["config"]).get("transformer", {}))
         return dit_config
 
+    if '{}genre_embedder.weight'.format(key_prefix) in state_dict_keys: #ACE-Step model
+        dit_config = {}
+        dit_config["audio_model"] = "ace"
+        dit_config["attention_head_dim"] = 128
+        dit_config["in_channels"] = 8
+        dit_config["inner_dim"] = 2560
+        dit_config["max_height"] = 16
+        dit_config["max_position"] = 32768
+        dit_config["max_width"] = 32768
+        dit_config["mlp_ratio"] = 2.5
+        dit_config["num_attention_heads"] = 20
+        dit_config["num_layers"] = 24
+        dit_config["out_channels"] = 8
+        dit_config["patch_size"] = [16, 1]
+        dit_config["rope_theta"] = 1000000.0
+        dit_config["speaker_embedding_dim"] = 512
+        dit_config["text_embedding_dim"] = 768
+
+        dit_config["ssl_encoder_depths"] = [8, 8]
+        dit_config["ssl_latent_dims"] = [1024, 768]
+        dit_config["ssl_names"] = ["mert", "m-hubert"]
+        dit_config["lyric_encoder_vocab_size"] = 6693
+        dit_config["lyric_hidden_size"] = 1024
+        return dit_config
+
     if '{}t_block.1.weight'.format(key_prefix) in state_dict_keys: # PixArt
         patch_size = 2
         dit_config = {}
@@ -306,7 +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) + '{}.')
@@ -317,10 +359,17 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["cross_attn_norm"] = True
         dit_config["eps"] = 1e-6
         dit_config["in_dim"] = state_dict['{}patch_embedding.weight'.format(key_prefix)].shape[1]
-        if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
-            dit_config["model_type"] = "i2v"
+        if '{}vace_patch_embedding.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "vace"
+            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"
         else:
-            dit_config["model_type"] = "t2v"
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "i2v"
+            else:
+                dit_config["model_type"] = "t2v"
         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]
@@ -360,6 +409,78 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["text_emb_dim"] = 2048
         return dit_config
 
+    if '{}blocks.0.mlp.layer1.weight'.format(key_prefix) in state_dict_keys:  # Cosmos predict2
+        dit_config = {}
+        dit_config["image_model"] = "cosmos_predict2"
+        dit_config["max_img_h"] = 240
+        dit_config["max_img_w"] = 240
+        dit_config["max_frames"] = 128
+        concat_padding_mask = True
+        dit_config["in_channels"] = (state_dict['{}x_embedder.proj.1.weight'.format(key_prefix)].shape[1] // 4) - int(concat_padding_mask)
+        dit_config["out_channels"] = 16
+        dit_config["patch_spatial"] = 2
+        dit_config["patch_temporal"] = 1
+        dit_config["model_channels"] = state_dict['{}x_embedder.proj.1.weight'.format(key_prefix)].shape[0]
+        dit_config["concat_padding_mask"] = concat_padding_mask
+        dit_config["crossattn_emb_channels"] = 1024
+        dit_config["pos_emb_cls"] = "rope3d"
+        dit_config["pos_emb_learnable"] = True
+        dit_config["pos_emb_interpolation"] = "crop"
+        dit_config["min_fps"] = 1
+        dit_config["max_fps"] = 30
+
+        dit_config["use_adaln_lora"] = True
+        dit_config["adaln_lora_dim"] = 256
+        if dit_config["model_channels"] == 2048:
+            dit_config["num_blocks"] = 28
+            dit_config["num_heads"] = 16
+        elif dit_config["model_channels"] == 5120:
+            dit_config["num_blocks"] = 36
+            dit_config["num_heads"] = 40
+
+        if dit_config["in_channels"] == 16:
+            dit_config["extra_per_block_abs_pos_emb"] = False
+            dit_config["rope_h_extrapolation_ratio"] = 4.0
+            dit_config["rope_w_extrapolation_ratio"] = 4.0
+            dit_config["rope_t_extrapolation_ratio"] = 1.0
+        elif dit_config["in_channels"] == 17: # img to video
+            if dit_config["model_channels"] == 2048:
+                dit_config["extra_per_block_abs_pos_emb"] = False
+                dit_config["rope_h_extrapolation_ratio"] = 3.0
+                dit_config["rope_w_extrapolation_ratio"] = 3.0
+                dit_config["rope_t_extrapolation_ratio"] = 1.0
+            elif dit_config["model_channels"] == 5120:
+                dit_config["rope_h_extrapolation_ratio"] = 2.0
+                dit_config["rope_w_extrapolation_ratio"] = 2.0
+                dit_config["rope_t_extrapolation_ratio"] = 0.8333333333333334
+
+        dit_config["extra_h_extrapolation_ratio"] = 1.0
+        dit_config["extra_w_extrapolation_ratio"] = 1.0
+        dit_config["extra_t_extrapolation_ratio"] = 1.0
+        dit_config["rope_enable_fps_modulation"] = False
+
+        return dit_config
+
+    if '{}time_caption_embed.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys:  # Omnigen2
+        dit_config = {}
+        dit_config["image_model"] = "omnigen2"
+        dit_config["axes_dim_rope"] = [40, 40, 40]
+        dit_config["axes_lens"] = [1024, 1664, 1664]
+        dit_config["ffn_dim_multiplier"] = None
+        dit_config["hidden_size"] = 2520
+        dit_config["in_channels"] = 16
+        dit_config["multiple_of"] = 256
+        dit_config["norm_eps"] = 1e-05
+        dit_config["num_attention_heads"] = 21
+        dit_config["num_kv_heads"] = 7
+        dit_config["num_layers"] = 32
+        dit_config["num_refiner_layers"] = 2
+        dit_config["out_channels"] = None
+        dit_config["patch_size"] = 2
+        dit_config["text_feat_dim"] = 2048
+        dit_config["timestep_scale"] = 1000.0
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -573,6 +694,9 @@ def convert_config(unet_config):
 
 
 def unet_config_from_diffusers_unet(state_dict, dtype=None):
+    if "conv_in.weight" not in state_dict:
+        return None
+
     match = {}
     transformer_depth = []
 

comfy/model_management.py

@@ -101,7 +101,7 @@ 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
+    import intel_extension_for_pytorch as ipex  # noqa: F401
     _ = torch.xpu.device_count()
     xpu_available = xpu_available or torch.xpu.is_available()
 except:
@@ -128,6 +128,11 @@ try:
 except:
     mlu_available = False
 
+try:
+    ixuca_available = hasattr(torch, "corex")
+except:
+    ixuca_available = False
+
 if args.cpu:
     cpu_state = CPUState.CPU
 
@@ -151,6 +156,12 @@ 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 +197,9 @@ 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']
@@ -288,21 +300,34 @@ 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:
     pass
 
 
+SUPPORT_FP8_OPS = args.supports_fp8_compute
 try:
     if is_amd():
+        try:
+            rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
+        except:
+            rocm_version = (6, -1)
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         logging.info("AMD arch: {}".format(arch))
+        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[0] >= 2 and torch_version_numeric[1] >= 7:  # works on 2.6 but doesn't actually seem to improve much
-                if any((a in arch) for a in ["gfx1100", "gfx1101"]):  # TODO: more arches
+            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: 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
+
 except:
     pass
 
@@ -323,7 +348,7 @@ except:
     pass
 
 try:
-    if torch_version_numeric[0] == 2 and torch_version_numeric[1] >= 5:
+    if torch_version_numeric >= (2, 5):
         torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
 except:
     logging.warning("Warning, could not set allow_fp16_bf16_reduction_math_sdp")
@@ -367,6 +392,8 @@ 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():
@@ -502,6 +529,8 @@ 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
@@ -695,7 +724,7 @@ def unet_inital_load_device(parameters, dtype):
         return torch_dev
 
     cpu_dev = torch.device("cpu")
-    if DISABLE_SMART_MEMORY:
+    if DISABLE_SMART_MEMORY or vram_state == VRAMState.NO_VRAM:
         return cpu_dev
 
     model_size = dtype_size(dtype) * parameters
@@ -725,6 +754,8 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e4m3fn
     if args.fp8_e5m2_unet:
         return torch.float8_e5m2
+    if args.fp8_e8m0fnu_unet:
+        return torch.float8_e8m0fnu
 
     fp8_dtype = None
     if weight_dtype in FLOAT8_TYPES:
@@ -864,6 +895,7 @@ 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
 
@@ -917,7 +949,7 @@ def device_supports_non_blocking(device):
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
     if is_intel_xpu():
-        return False
+        return True
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
     if directml_enabled:
@@ -937,15 +969,74 @@ def force_channels_last():
     #TODO
     return False
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False):
+
+STREAMS = {}
+NUM_STREAMS = 1
+if args.async_offload:
+    NUM_STREAMS = 2
+    logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
+
+stream_counters = {}
+def get_offload_stream(device):
+    stream_counter = stream_counters.get(device, 0)
+    if NUM_STREAMS <= 1:
+        return None
+
+    if device in STREAMS:
+        ss = STREAMS[device]
+        s = ss[stream_counter]
+        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):
+        ss = []
+        for k in range(NUM_STREAMS):
+            ss.append(torch.cuda.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
+    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):
+    if stream is None:
+        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:
         if not copy:
             if dtype is None or weight.dtype == dtype:
                 return weight
+        if stream is not None:
+            with stream:
+                return weight.to(dtype=dtype, copy=copy)
         return weight.to(dtype=dtype, copy=copy)
 
-    r = torch.empty_like(weight, dtype=dtype, device=device)
-    r.copy_(weight, non_blocking=non_blocking)
+    if stream is not None:
+        with stream:
+            r = torch.empty_like(weight, dtype=dtype, device=device)
+            r.copy_(weight, non_blocking=non_blocking)
+    else:
+        r = torch.empty_like(weight, dtype=dtype, device=device)
+        r.copy_(weight, non_blocking=non_blocking)
     return r
 
 def cast_to_device(tensor, device, dtype, copy=False):
@@ -969,6 +1060,8 @@ def xformers_enabled():
         return False
     if is_mlu():
         return False
+    if is_ixuca():
+        return False
     if directml_enabled:
         return False
     return XFORMERS_IS_AVAILABLE
@@ -994,7 +1087,7 @@ def pytorch_attention_flash_attention():
     global ENABLE_PYTORCH_ATTENTION
     if ENABLE_PYTORCH_ATTENTION:
         #TODO: more reliable way of checking for flash attention?
-        if is_nvidia(): #pytorch flash attention only works on Nvidia
+        if is_nvidia():
             return True
         if is_intel_xpu():
             return True
@@ -1004,13 +1097,15 @@ 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():
     upcast = args.force_upcast_attention
 
     macos_version = mac_version()
-    if macos_version is not None and ((14, 5) <= macos_version < (16,)):  # black image bug on recent versions of macOS
+    if macos_version is not None and ((14, 5) <= macos_version):  # black image bug on recent versions of macOS, I don't think it's ever getting fixed
         upcast = True
 
     if upcast:
@@ -1034,8 +1129,8 @@ 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)
@@ -1084,6 +1179,9 @@ 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')
 
@@ -1115,7 +1213,10 @@ 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
@@ -1123,6 +1224,9 @@ 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
 
@@ -1178,11 +1282,17 @@ 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, 6):
+            return True
+        else:
+            return torch.xpu.get_device_capability(device)['has_bfloat16_conversions']
 
     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
@@ -1209,6 +1319,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     return False
 
 def supports_fp8_compute(device=None):
+    if SUPPORT_FP8_OPS:
+        return True
+
     if not is_nvidia():
         return False
 
@@ -1220,15 +1333,22 @@ def supports_fp8_compute(device=None):
     if props.minor < 9:
         return False
 
-    if torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 3):
+    if torch_version_numeric < (2, 3):
         return False
 
     if WINDOWS:
-        if (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 4):
+        if torch_version_numeric < (2, 4):
             return False
 
     return True
 
+def extended_fp16_support():
+    # TODO: check why some models work with fp16 on newer torch versions but not on older
+    if torch_version_numeric < (2, 7):
+        return False
+
+    return True
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:

comfy/model_patcher.py

@@ -17,23 +17,26 @@
 """
 
 from __future__ import annotations
-from typing import Optional, Callable
-import torch
+
+import collections
 import copy
 import inspect
 import logging
-import uuid
-import collections
 import math
+import uuid
+from typing import Callable, Optional
+
+import torch
 
-import comfy.utils
 import comfy.float
-import comfy.model_management
-import comfy.lora
 import comfy.hooks
+import comfy.lora
+import comfy.model_management
 import comfy.patcher_extension
-from comfy.patcher_extension import CallbacksMP, WrappersMP, PatcherInjection
+import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
+from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP
+
 
 def string_to_seed(data):
     crc = 0xFFFFFFFF
@@ -376,6 +379,9 @@ 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
 

comfy/model_sampling.py

@@ -77,6 +77,25 @@ class IMG_TO_IMG(X0):
     def calculate_input(self, sigma, noise):
         return noise
 
+class COSMOS_RFLOW:
+    def calculate_input(self, sigma, noise):
+        sigma = (sigma / (sigma + 1))
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
+        return noise * (1.0 - sigma)
+
+    def calculate_denoised(self, sigma, model_output, model_input):
+        sigma = (sigma / (sigma + 1))
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
+        return model_input * (1.0 - sigma) - model_output * sigma
+
+    def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
+        noise = noise * sigma
+        noise += latent_image
+        return noise
+
+    def inverse_noise_scaling(self, sigma, latent):
+        return latent
 
 class ModelSamplingDiscrete(torch.nn.Module):
     def __init__(self, model_config=None, zsnr=None):
@@ -111,13 +130,14 @@ class ModelSamplingDiscrete(torch.nn.Module):
         self.num_timesteps = int(timesteps)
         self.linear_start = linear_start
         self.linear_end = linear_end
+        self.zsnr = zsnr
 
         # self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32))
         # self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32))
         # self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32))
 
         sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
-        if zsnr:
+        if self.zsnr:
             sigmas = rescale_zero_terminal_snr_sigmas(sigmas)
 
         self.set_sigmas(sigmas)
@@ -349,3 +369,15 @@ class ModelSamplingFlux(torch.nn.Module):
         if percent >= 1.0:
             return 0.0
         return flux_time_shift(self.shift, 1.0, 1.0 - percent)
+
+
+class ModelSamplingCosmosRFlow(ModelSamplingContinuousEDM):
+    def timestep(self, sigma):
+        return sigma / (sigma + 1)
+
+    def sigma(self, timestep):
+        sigma_max = self.sigma_max
+        if timestep >= (sigma_max / (sigma_max + 1)):
+            return sigma_max
+
+        return timestep / (1 - timestep)

comfy/ops.py

@@ -22,6 +22,7 @@ import comfy.model_management
 from comfy.cli_args import args, PerformanceFeature
 import comfy.float
 import comfy.rmsnorm
+import contextlib
 
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
@@ -37,20 +38,31 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
         if device is None:
             device = input.device
 
+    offload_stream = comfy.model_management.get_offload_stream(device)
+    if offload_stream is not None:
+        wf_context = offload_stream
+    else:
+        wf_context = contextlib.nullcontext()
+
     bias = None
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
         has_function = len(s.bias_function) > 0
-        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
+        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
+
         if has_function:
-            for f in s.bias_function:
-                bias = f(bias)
+            with wf_context:
+                for f in s.bias_function:
+                    bias = f(bias)
 
     has_function = len(s.weight_function) > 0
-    weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
+    weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
     if has_function:
-        for f in s.weight_function:
-            weight = f(weight)
+        with wf_context:
+            for f in s.weight_function:
+                weight = f(weight)
+
+    comfy.model_management.sync_stream(device, offload_stream)
     return weight, bias
 
 class CastWeightBiasOp:
@@ -296,10 +308,10 @@ def fp8_linear(self, input):
         if scale_input is None:
             scale_input = torch.ones((), device=input.device, dtype=torch.float32)
             input = torch.clamp(input, min=-448, max=448, out=input)
-            input = input.reshape(-1, input_shape[2]).to(dtype)
+            input = input.reshape(-1, input_shape[2]).to(dtype).contiguous()
         else:
             scale_input = scale_input.to(input.device)
-            input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype)
+            input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype).contiguous()
 
         if bias is not None:
             o = torch._scaled_mm(input, w, out_dtype=input_dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
@@ -324,9 +336,12 @@ 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

@@ -30,7 +30,7 @@ if RMSNorm is None:
         def __init__(
             self,
             normalized_shape,
-            eps=None,
+            eps=1e-6,
             elementwise_affine=True,
             device=None,
             dtype=None,

comfy/sampler_helpers.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 import uuid
+import math
+import collections
 import comfy.model_management
 import comfy.conds
 import comfy.utils
@@ -104,6 +106,21 @@ def cleanup_additional_models(models):
         if hasattr(m, 'cleanup'):
             m.cleanup()
 
+def estimate_memory(model, noise_shape, conds):
+    cond_shapes = collections.defaultdict(list)
+    cond_shapes_min = {}
+    for _, cs in conds.items():
+        for cond in cs:
+            for k, v in model.model.extra_conds_shapes(**cond).items():
+                cond_shapes[k].append(v)
+                if cond_shapes_min.get(k, None) is None:
+                    cond_shapes_min[k] = [v]
+                elif math.prod(v) > math.prod(cond_shapes_min[k][0]):
+                    cond_shapes_min[k] = [v]
+
+    memory_required = model.model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:]), cond_shapes=cond_shapes)
+    minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
+    return memory_required, minimum_memory_required
 
 def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
@@ -117,9 +134,8 @@ def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=Non
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
-    memory_required = model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory
-    minimum_memory_required = model.memory_required([noise_shape[0]] + list(noise_shape[1:])) + inference_memory
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required)
+    memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory)
     real_model = model.model
 
     return real_model, conds, models

comfy/samplers.py

@@ -256,7 +256,13 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
             for i in range(1, len(to_batch_temp) + 1):
                 batch_amount = to_batch_temp[:len(to_batch_temp)//i]
                 input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
-                if model.memory_required(input_shape) * 1.5 < free_memory:
+                cond_shapes = collections.defaultdict(list)
+                for tt in batch_amount:
+                    cond = {k: v.size() for k, v in to_run[tt][0].conditioning.items()}
+                    for k, v in to_run[tt][0].conditioning.items():
+                        cond_shapes[k].append(v.size())
+
+                if model.memory_required(input_shape, cond_shapes=cond_shapes) * 1.5 < free_memory:
                     to_batch = batch_amount
                     break
 
@@ -367,7 +373,11 @@ 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,
@@ -710,7 +720,7 @@ 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", "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={}):
@@ -1033,13 +1043,13 @@ 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

@@ -14,9 +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
 
@@ -42,6 +45,8 @@ import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
 import comfy.text_encoders.wan
 import comfy.text_encoders.hidream
+import comfy.text_encoders.ace
+import comfy.text_encoders.omnigen2
 
 import comfy.model_patcher
 import comfy.lora
@@ -120,6 +125,7 @@ class CLIP:
         self.layer_idx = None
         self.use_clip_schedule = False
         logging.info("CLIP/text encoder model load device: {}, offload device: {}, current: {}, dtype: {}".format(load_device, offload_device, params['device'], dtype))
+        self.tokenizer_options = {}
 
     def clone(self):
         n = CLIP(no_init=True)
@@ -127,6 +133,7 @@ class CLIP:
         n.cond_stage_model = self.cond_stage_model
         n.tokenizer = self.tokenizer
         n.layer_idx = self.layer_idx
+        n.tokenizer_options = self.tokenizer_options.copy()
         n.use_clip_schedule = self.use_clip_schedule
         n.apply_hooks_to_conds = self.apply_hooks_to_conds
         return n
@@ -134,10 +141,18 @@ class CLIP:
     def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
         return self.patcher.add_patches(patches, strength_patch, strength_model)
 
+    def set_tokenizer_option(self, option_name, value):
+        self.tokenizer_options[option_name] = value
+
     def clip_layer(self, layer_idx):
         self.layer_idx = layer_idx
 
     def tokenize(self, text, return_word_ids=False, **kwargs):
+        tokenizer_options = kwargs.get("tokenizer_options", {})
+        if len(self.tokenizer_options) > 0:
+            tokenizer_options = {**self.tokenizer_options, **tokenizer_options}
+        if len(tokenizer_options) > 0:
+            kwargs["tokenizer_options"] = tokenizer_options
         return self.tokenizer.tokenize_with_weights(text, return_word_ids, **kwargs)
 
     def add_hooks_to_dict(self, pooled_dict: dict[str]):
@@ -270,6 +285,7 @@ class VAE:
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
+        self.extra_1d_channel = None
 
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
@@ -405,17 +421,30 @@ 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
@@ -427,6 +456,20 @@ class VAE:
                 ddconfig = {"embed_dim": 64, "num_freqs": 8, "include_pi": False, "heads": 16, "width": 1024, "num_decoder_layers": 16, "qkv_bias": False, "qk_norm": True, "geo_decoder_mlp_expand_ratio": mlp_expand, "geo_decoder_downsample_ratio": downsample_ratio, "geo_decoder_ln_post": ln_post}
                 self.first_stage_model = comfy.ldm.hunyuan3d.vae.ShapeVAE(**ddconfig)
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+            elif "vocoder.backbone.channel_layers.0.0.bias" in sd: #Ace Step Audio
+                self.first_stage_model = comfy.ldm.ace.vae.music_dcae_pipeline.MusicDCAE(source_sample_rate=44100)
+                self.memory_used_encode = lambda shape, dtype: (shape[2] * 330) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype)
+                self.latent_channels = 8
+                self.output_channels = 2
+                self.upscale_ratio = 4096
+                self.downscale_ratio = 4096
+                self.latent_dim = 2
+                self.process_output = lambda audio: audio
+                self.process_input = lambda audio: audio
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                self.disable_offload = True
+                self.extra_1d_channel = 16
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -485,7 +528,13 @@ class VAE:
         return output
 
     def decode_tiled_1d(self, samples, tile_x=128, overlap=32):
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        if samples.ndim == 3:
+            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        else:
+            og_shape = samples.shape
+            samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1))
+            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float()
+
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device))
 
     def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)):
@@ -505,9 +554,24 @@ class VAE:
         samples /= 3.0
         return samples
 
-    def encode_tiled_1d(self, samples, tile_x=128 * 2048, overlap=32 * 2048):
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
-        return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=(1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device)
+    def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048):
+        if self.latent_dim == 1:
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+            out_channels = self.latent_channels
+            upscale_amount = 1 / self.downscale_ratio
+        else:
+            extra_channel_size = self.extra_1d_channel
+            out_channels = self.latent_channels * extra_channel_size
+            tile_x = tile_x // extra_channel_size
+            overlap = overlap // extra_channel_size
+            upscale_amount = 1 / self.downscale_ratio
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float()
+
+        out = comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device)
+        if self.latent_dim == 1:
+            return out
+        else:
+            return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1)
 
     def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)):
         encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
@@ -532,7 +596,7 @@ class VAE:
         except model_management.OOM_EXCEPTION:
             logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
             dims = samples_in.ndim - 2
-            if dims == 1:
+            if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
             elif dims == 2:
                 pixel_samples = self.decode_tiled_(samples_in)
@@ -599,7 +663,7 @@ class VAE:
                 tile = 256
                 overlap = tile // 4
                 samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
-            elif self.latent_dim == 1:
+            elif self.latent_dim == 1 or self.extra_1d_channel is not None:
                 samples = self.encode_tiled_1d(pixel_samples)
             else:
                 samples = self.encode_tiled_(pixel_samples)
@@ -703,6 +767,10 @@ class CLIPType(Enum):
     COSMOS = 11
     LUMINA2 = 12
     WAN = 13
+    HIDREAM = 14
+    CHROMA = 15
+    ACE = 16
+    OMNIGEN2 = 17
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -722,6 +790,7 @@ class TEModel(Enum):
     LLAMA3_8 = 7
     T5_XXL_OLD = 8
     GEMMA_2_2B = 9
+    QWEN25_3B = 10
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -742,6 +811,8 @@ def detect_te_model(sd):
         return TEModel.T5_BASE
     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
     if "model.layers.0.post_attention_layernorm.weight" in sd:
         return TEModel.LLAMA3_8
     return None
@@ -791,6 +862,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.SD3:
                 clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False)
                 clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
+            elif clip_type == CLIPType.HIDREAM:
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
+                clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else:
                 clip_target.clip = sdxl_clip.SDXLRefinerClipModel
                 clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -804,13 +878,17 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.LTXV:
                 clip_target.clip = comfy.text_encoders.lt.ltxv_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.lt.LTXVT5Tokenizer
-            elif clip_type == CLIPType.PIXART:
+            elif clip_type == CLIPType.PIXART or clip_type == CLIPType.CHROMA:
                 clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer
             elif clip_type == CLIPType.WAN:
                 clip_target.clip = comfy.text_encoders.wan.te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.wan.WanT5Tokenizer
                 tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+            elif clip_type == CLIPType.HIDREAM:
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
+                                                                        clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None)
+                clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer
@@ -821,16 +899,32 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.clip = comfy.text_encoders.aura_t5.AuraT5Model
             clip_target.tokenizer = comfy.text_encoders.aura_t5.AuraT5Tokenizer
         elif te_model == TEModel.T5_BASE:
-            clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
-            clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
+            if clip_type == CLIPType.ACE or "spiece_model" in clip_data[0]:
+                clip_target.clip = comfy.text_encoders.ace.AceT5Model
+                clip_target.tokenizer = comfy.text_encoders.ace.AceT5Tokenizer
+                tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+            else:
+                clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
+                clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
         elif te_model == TEModel.GEMMA_2_2B:
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
             tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+        elif te_model == TEModel.LLAMA3_8:
+            clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
+                                                                        clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)
+            clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
+        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
         else:
+            # clip_l
             if clip_type == CLIPType.SD3:
                 clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False)
                 clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
+            elif clip_type == CLIPType.HIDREAM:
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
+                clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else:
                 clip_target.clip = sd1_clip.SD1ClipModel
                 clip_target.tokenizer = sd1_clip.SD1Tokenizer
@@ -848,6 +942,24 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.HUNYUAN_VIDEO:
             clip_target.clip = comfy.text_encoders.hunyuan_video.hunyuan_video_clip(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideoTokenizer
+        elif clip_type == CLIPType.HIDREAM:
+            # Detect
+            hidream_dualclip_classes = []
+            for hidream_te in clip_data:
+                te_model = detect_te_model(hidream_te)
+                hidream_dualclip_classes.append(te_model)
+
+            clip_l = TEModel.CLIP_L in hidream_dualclip_classes
+            clip_g = TEModel.CLIP_G in hidream_dualclip_classes
+            t5 = TEModel.T5_XXL in hidream_dualclip_classes
+            llama = TEModel.LLAMA3_8 in hidream_dualclip_classes
+
+            # Initialize t5xxl_detect and llama_detect kwargs if needed
+            t5_kwargs = t5xxl_detect(clip_data) if t5 else {}
+            llama_kwargs = llama_detect(clip_data) if llama else {}
+
+            clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, **t5_kwargs, **llama_kwargs)
+            clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -880,6 +992,12 @@ 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)
@@ -908,7 +1026,7 @@ 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):
@@ -992,7 +1110,28 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     return (model_patcher, clip, vae, clipvision)
 
 
-def load_diffusion_model_state_dict(sd, model_options={}): #load unet in diffusers or regular format
+def load_diffusion_model_state_dict(sd, model_options={}):
+    """
+    Loads a UNet diffusion model from a state dictionary, supporting both diffusers and regular formats.
+
+    Args:
+        sd (dict): State dictionary containing model weights and configuration
+        model_options (dict, optional): Additional options for model loading. Supports:
+            - dtype: Override model data type
+            - custom_operations: Custom model operations
+            - fp8_optimizations: Enable FP8 optimizations
+
+    Returns:
+        ModelPatcher: A wrapped model instance that handles device management and weight loading.
+        Returns None if the model configuration cannot be detected.
+
+    The function:
+    1. Detects and handles different model formats (regular, diffusers, mmdit)
+    2. Configures model dtype based on parameters and device capabilities
+    3. Handles weight conversion and device placement
+    4. Manages model optimization settings
+    5. Loads weights and returns a device-managed model instance
+    """
     dtype = model_options.get("dtype", None)
 
     #Allow loading unets from checkpoint files
@@ -1050,7 +1189,7 @@ def load_diffusion_model_state_dict(sd, model_options={}): #load unet in diffuse
     model.load_model_weights(new_sd, "")
     left_over = sd.keys()
     if len(left_over) > 0:
-        logging.info("left over keys in unet: {}".format(left_over))
+        logging.info("left over keys in diffusion model: {}".format(left_over))
     return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
 
 
@@ -1058,8 +1197,8 @@ def load_diffusion_model(unet_path, model_options={}):
     sd = comfy.utils.load_torch_file(unet_path)
     model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
-        logging.error("ERROR UNSUPPORTED UNET {}".format(unet_path))
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
+        logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".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_clip.py

@@ -457,13 +457,14 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
     return embed_out
 
 class SDTokenizer:
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, tokenizer_data={}, tokenizer_args={}):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, tokenizer_data={}, tokenizer_args={}):
         if tokenizer_path is None:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
         self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
         self.max_length = tokenizer_data.get("{}_max_length".format(embedding_key), max_length)
-        self.min_length = min_length
+        self.min_length = tokenizer_data.get("{}_min_length".format(embedding_key), min_length)
         self.end_token = None
+        self.min_padding = min_padding
 
         empty = self.tokenizer('')["input_ids"]
         self.tokenizer_adds_end_token = has_end_token
@@ -481,7 +482,8 @@ class SDTokenizer:
             if end_token is not None:
                 self.end_token = end_token
             else:
-                self.end_token = empty[0]
+                if has_end_token:
+                    self.end_token = empty[0]
 
         if pad_token is not None:
             self.pad_token = pad_token
@@ -518,13 +520,15 @@ class SDTokenizer:
         return (embed, leftover)
 
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+    def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_options={}, **kwargs):
         '''
         Takes a prompt and converts it to a list of (token, weight, word id) elements.
         Tokens can both be integer tokens and pre computed CLIP tensors.
         Word id values are unique per word and embedding, where the id 0 is reserved for non word tokens.
         Returned list has the dimensions NxM where M is the input size of CLIP
         '''
+        min_length = tokenizer_options.get("{}_min_length".format(self.embedding_key), self.min_length)
+        min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
         parsed_weights = token_weights(text, 1.0)
@@ -603,10 +607,12 @@ class SDTokenizer:
         #fill last batch
         if self.end_token is not None:
             batch.append((self.end_token, 1.0, 0))
-        if self.pad_to_max_length:
+        if min_padding is not None:
+            batch.extend([(self.pad_token, 1.0, 0)] * min_padding)
+        if self.pad_to_max_length and len(batch) < self.max_length:
             batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch)))
-        if self.min_length is not None and len(batch) < self.min_length:
-            batch.extend([(self.pad_token, 1.0, 0)] * (self.min_length - len(batch)))
+        if min_length is not None and len(batch) < min_length:
+            batch.extend([(self.pad_token, 1.0, 0)] * (min_length - len(batch)))
 
         if not return_word_ids:
             batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens]
@@ -634,7 +640,7 @@ class SD1Tokenizer:
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out[self.clip_name] = getattr(self, self.clip).tokenize_with_weights(text, return_word_ids)
+        out[self.clip_name] = getattr(self, self.clip).tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):

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/sdxl_clip.py

@@ -28,8 +28,8 @@ class SDXLTokenizer:
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids)
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):

comfy/supported_models.py

@@ -17,6 +17,8 @@ import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
 import comfy.text_encoders.wan
+import comfy.text_encoders.ace
+import comfy.text_encoders.omnigen2
 
 from . import supported_models_base
 from . import latent_formats
@@ -785,6 +787,10 @@ class LTXV(supported_models_base.BASE):
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = (unet_config.get("cross_attention_dim", 2048) / 2048) * 5.5
+
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.LTXV(self, device=device)
         return out
@@ -903,6 +909,48 @@ class CosmosI2V(CosmosT2V):
         out = model_base.CosmosVideo(self, image_to_video=True, device=device)
         return out
 
+class CosmosT2IPredict2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "cosmos_predict2",
+        "in_channels": 16,
+    }
+
+    sampling_settings = {
+        "sigma_data": 1.0,
+        "sigma_max": 80.0,
+        "sigma_min": 0.002,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    memory_usage_factor = 1.0
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.9
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.CosmosPredict2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.cosmos.CosmosT5Tokenizer, comfy.text_encoders.cosmos.te(**t5_detect))
+
+class CosmosI2VPredict2(CosmosT2IPredict2):
+    unet_config = {
+        "image_model": "cosmos_predict2",
+        "in_channels": 17,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.CosmosPredict2(self, image_to_video=True, device=device)
+        return out
+
 class Lumina2(supported_models_base.BASE):
     unet_config = {
         "image_model": "lumina2",
@@ -987,6 +1035,43 @@ class WAN21_FunControl2V(WAN21_T2V):
         out = model_base.WAN21(self, image_to_video=False, device=device)
         return out
 
+class WAN21_Camera(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "camera",
+        "in_dim": 32,
+    }
+
+    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",
+        "model_type": "vace",
+    }
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = 1.2 * self.memory_usage_factor
+
+    def get_model(self, state_dict, prefix="", device=None):
+        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",
@@ -1054,7 +1139,97 @@ class HiDream(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return None #  TODO
 
+class Chroma(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "chroma",
+    }
 
-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, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream]
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+    }
+
+    latent_format = comfy.latent_formats.Flux
+
+    memory_usage_factor = 3.2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Chroma(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.pixart_t5.PixArtTokenizer, comfy.text_encoders.pixart_t5.pixart_te(**t5_detect))
+
+class ACEStep(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio
+
+    memory_usage_factor = 0.5
+
+    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.ACEStep(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace.AceT5Tokenizer, comfy.text_encoders.ace.AceT5Model)
+
+class Omnigen2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "omnigen2",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 2.6,
+    }
+
+    memory_usage_factor = 1.65 #TODO
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        if comfy.model_management.extended_fp16_support():
+            self.supported_inference_dtypes = [torch.float16] + self.supported_inference_dtypes
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Omnigen2(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_3b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.Omnigen2Tokenizer, comfy.text_encoders.omnigen2.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, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace.py

@@ -0,0 +1,153 @@
+from comfy import sd1_clip
+from .spiece_tokenizer import SPieceTokenizer
+import comfy.text_encoders.t5
+import os
+import re
+import torch
+import logging
+
+from tokenizers import Tokenizer
+from .ace_text_cleaners import multilingual_cleaners, japanese_to_romaji
+
+SUPPORT_LANGUAGES = {
+    "en": 259, "de": 260, "fr": 262, "es": 284, "it": 285,
+    "pt": 286, "pl": 294, "tr": 295, "ru": 267, "cs": 293,
+    "nl": 297, "ar": 5022, "zh": 5023, "ja": 5412, "hu": 5753,
+    "ko": 6152, "hi": 6680
+}
+
+structure_pattern = re.compile(r"\[.*?\]")
+
+DEFAULT_VOCAB_FILE = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "ace_lyrics_tokenizer"), "vocab.json")
+
+
+class VoiceBpeTokenizer:
+    def __init__(self, vocab_file=DEFAULT_VOCAB_FILE):
+        self.tokenizer = None
+        if vocab_file is not None:
+            self.tokenizer = Tokenizer.from_file(vocab_file)
+
+    def preprocess_text(self, txt, lang):
+        txt = multilingual_cleaners(txt, lang)
+        return txt
+
+    def encode(self, txt, lang='en'):
+        # lang = lang.split("-")[0]  # remove the region
+        # self.check_input_length(txt, lang)
+        txt = self.preprocess_text(txt, lang)
+        lang = "zh-cn" if lang == "zh" else lang
+        txt = f"[{lang}]{txt}"
+        txt = txt.replace(" ", "[SPACE]")
+        return self.tokenizer.encode(txt).ids
+
+    def get_lang(self, line):
+        if line.startswith("[") and line[3:4] == ']':
+            lang = line[1:3].lower()
+            if lang in SUPPORT_LANGUAGES:
+                return lang, line[4:]
+        return "en", line
+
+    def __call__(self, string):
+        lines = string.split("\n")
+        lyric_token_idx = [261]
+        for line in lines:
+            line = line.strip()
+            if not line:
+                lyric_token_idx += [2]
+                continue
+
+            lang, line = self.get_lang(line)
+
+            if lang not in SUPPORT_LANGUAGES:
+                lang = "en"
+            if "zh" in lang:
+                lang = "zh"
+            if "spa" in lang:
+                lang = "es"
+
+            try:
+                line_out = japanese_to_romaji(line)
+                if line_out != line:
+                    lang = "ja"
+                line = line_out
+            except:
+                pass
+
+            try:
+                if structure_pattern.match(line):
+                    token_idx = self.encode(line, "en")
+                else:
+                    token_idx = self.encode(line, lang)
+                lyric_token_idx = lyric_token_idx + token_idx + [2]
+            except Exception as e:
+                logging.warning("tokenize error {} for line {} major_language {}".format(e, line, lang))
+        return {"input_ids": lyric_token_idx}
+
+    @staticmethod
+    def from_pretrained(path, **kwargs):
+        return VoiceBpeTokenizer(path, **kwargs)
+
+    def get_vocab(self):
+        return {}
+
+
+class UMT5BaseModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "umt5_config_base.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=False, model_options=model_options)
+
+class UMT5BaseTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=768, embedding_key='umt5base', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=0, tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+class LyricsTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "ace_lyrics_tokenizer"), "vocab.json")
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='lyrics', tokenizer_class=VoiceBpeTokenizer, has_start_token=True, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=2, has_end_token=False, tokenizer_data=tokenizer_data)
+
+class AceT5Tokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.voicebpe = LyricsTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.umt5base = UMT5BaseTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = {}
+        out["lyrics"] = self.voicebpe.tokenize_with_weights(kwargs.get("lyrics", ""), return_word_ids, **kwargs)
+        out["umt5base"] = self.umt5base.tokenize_with_weights(text, return_word_ids, **kwargs)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.umt5base.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return self.umt5base.state_dict()
+
+class AceT5Model(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__()
+        self.umt5base = UMT5BaseModel(device=device, dtype=dtype, model_options=model_options)
+        self.dtypes = set()
+        if dtype is not None:
+            self.dtypes.add(dtype)
+
+    def set_clip_options(self, options):
+        self.umt5base.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.umt5base.reset_clip_options()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_umt5base = token_weight_pairs["umt5base"]
+        token_weight_pairs_lyrics = token_weight_pairs["lyrics"]
+
+        t5_out, t5_pooled = self.umt5base.encode_token_weights(token_weight_pairs_umt5base)
+
+        lyrics_embeds = torch.tensor(list(map(lambda a: a[0], token_weight_pairs_lyrics[0]))).unsqueeze(0)
+        return t5_out, None, {"conditioning_lyrics": lyrics_embeds}
+
+    def load_sd(self, sd):
+        return self.umt5base.load_sd(sd)

comfy/text_encoders/ace_text_cleaners.py

@@ -0,0 +1,395 @@
+# basic text cleaners for the ACE step model
+# I didn't copy the ones from the reference code because I didn't want to deal with the dependencies
+# TODO: more languages than english?
+
+import re
+
+def japanese_to_romaji(japanese_text):
+    """
+    Convert Japanese hiragana and katakana to romaji (Latin alphabet representation).
+
+    Args:
+        japanese_text (str): Text containing hiragana and/or katakana characters
+
+    Returns:
+        str: The romaji (Latin alphabet) equivalent
+    """
+    # Dictionary mapping kana characters to their romaji equivalents
+    kana_map = {
+        # Katakana characters
+        'ア': 'a', 'イ': 'i', 'ウ': 'u', 'エ': 'e', 'オ': 'o',
+        'カ': 'ka', 'キ': 'ki', 'ク': 'ku', 'ケ': 'ke', 'コ': 'ko',
+        'サ': 'sa', 'シ': 'shi', 'ス': 'su', 'セ': 'se', 'ソ': 'so',
+        'タ': 'ta', 'チ': 'chi', 'ツ': 'tsu', 'テ': 'te', 'ト': 'to',
+        'ナ': 'na', 'ニ': 'ni', 'ヌ': 'nu', 'ネ': 'ne', 'ノ': 'no',
+        'ハ': 'ha', 'ヒ': 'hi', 'フ': 'fu', 'ヘ': 'he', 'ホ': 'ho',
+        'マ': 'ma', 'ミ': 'mi', 'ム': 'mu', 'メ': 'me', 'モ': 'mo',
+        'ヤ': 'ya', 'ユ': 'yu', 'ヨ': 'yo',
+        'ラ': 'ra', 'リ': 'ri', 'ル': 'ru', 'レ': 're', 'ロ': 'ro',
+        'ワ': 'wa', 'ヲ': 'wo', 'ン': 'n',
+
+        # Katakana voiced consonants
+        'ガ': 'ga', 'ギ': 'gi', 'グ': 'gu', 'ゲ': 'ge', 'ゴ': 'go',
+        'ザ': 'za', 'ジ': 'ji', 'ズ': 'zu', 'ゼ': 'ze', 'ゾ': 'zo',
+        'ダ': 'da', 'ヂ': 'ji', 'ヅ': 'zu', 'デ': 'de', 'ド': 'do',
+        'バ': 'ba', 'ビ': 'bi', 'ブ': 'bu', 'ベ': 'be', 'ボ': 'bo',
+        'パ': 'pa', 'ピ': 'pi', 'プ': 'pu', 'ペ': 'pe', 'ポ': 'po',
+
+        # Katakana combinations
+        'キャ': 'kya', 'キュ': 'kyu', 'キョ': 'kyo',
+        'シャ': 'sha', 'シュ': 'shu', 'ショ': 'sho',
+        'チャ': 'cha', 'チュ': 'chu', 'チョ': 'cho',
+        'ニャ': 'nya', 'ニュ': 'nyu', 'ニョ': 'nyo',
+        'ヒャ': 'hya', 'ヒュ': 'hyu', 'ヒョ': 'hyo',
+        'ミャ': 'mya', 'ミュ': 'myu', 'ミョ': 'myo',
+        'リャ': 'rya', 'リュ': 'ryu', 'リョ': 'ryo',
+        'ギャ': 'gya', 'ギュ': 'gyu', 'ギョ': 'gyo',
+        'ジャ': 'ja', 'ジュ': 'ju', 'ジョ': 'jo',
+        'ビャ': 'bya', 'ビュ': 'byu', 'ビョ': 'byo',
+        'ピャ': 'pya', 'ピュ': 'pyu', 'ピョ': 'pyo',
+
+        # Katakana small characters and special cases
+        'ッ': '', # Small tsu (doubles the following consonant)
+        'ャ': 'ya', 'ュ': 'yu', 'ョ': 'yo',
+
+        # Katakana extras
+        'ヴ': 'vu', 'ファ': 'fa', 'フィ': 'fi', 'フェ': 'fe', 'フォ': 'fo',
+        'ウィ': 'wi', 'ウェ': 'we', 'ウォ': 'wo',
+
+        # Hiragana characters
+        'あ': 'a', 'い': 'i', 'う': 'u', 'え': 'e', 'お': 'o',
+        'か': 'ka', 'き': 'ki', 'く': 'ku', 'け': 'ke', 'こ': 'ko',
+        'さ': 'sa', 'し': 'shi', 'す': 'su', 'せ': 'se', 'そ': 'so',
+        'た': 'ta', 'ち': 'chi', 'つ': 'tsu', 'て': 'te', 'と': 'to',
+        'な': 'na', 'に': 'ni', 'ぬ': 'nu', 'ね': 'ne', 'の': 'no',
+        'は': 'ha', 'ひ': 'hi', 'ふ': 'fu', 'へ': 'he', 'ほ': 'ho',
+        'ま': 'ma', 'み': 'mi', 'む': 'mu', 'め': 'me', 'も': 'mo',
+        'や': 'ya', 'ゆ': 'yu', 'よ': 'yo',
+        'ら': 'ra', 'り': 'ri', 'る': 'ru', 'れ': 're', 'ろ': 'ro',
+        'わ': 'wa', 'を': 'wo', 'ん': 'n',
+
+        # Hiragana voiced consonants
+        'が': 'ga', 'ぎ': 'gi', 'ぐ': 'gu', 'げ': 'ge', 'ご': 'go',
+        'ざ': 'za', 'じ': 'ji', 'ず': 'zu', 'ぜ': 'ze', 'ぞ': 'zo',
+        'だ': 'da', 'ぢ': 'ji', 'づ': 'zu', 'で': 'de', 'ど': 'do',
+        'ば': 'ba', 'び': 'bi', 'ぶ': 'bu', 'べ': 'be', 'ぼ': 'bo',
+        'ぱ': 'pa', 'ぴ': 'pi', 'ぷ': 'pu', 'ぺ': 'pe', 'ぽ': 'po',
+
+        # Hiragana combinations
+        'きゃ': 'kya', 'きゅ': 'kyu', 'きょ': 'kyo',
+        'しゃ': 'sha', 'しゅ': 'shu', 'しょ': 'sho',
+        'ちゃ': 'cha', 'ちゅ': 'chu', 'ちょ': 'cho',
+        'にゃ': 'nya', 'にゅ': 'nyu', 'にょ': 'nyo',
+        'ひゃ': 'hya', 'ひゅ': 'hyu', 'ひょ': 'hyo',
+        'みゃ': 'mya', 'みゅ': 'myu', 'みょ': 'myo',
+        'りゃ': 'rya', 'りゅ': 'ryu', 'りょ': 'ryo',
+        'ぎゃ': 'gya', 'ぎゅ': 'gyu', 'ぎょ': 'gyo',
+        'じゃ': 'ja', 'じゅ': 'ju', 'じょ': 'jo',
+        'びゃ': 'bya', 'びゅ': 'byu', 'びょ': 'byo',
+        'ぴゃ': 'pya', 'ぴゅ': 'pyu', 'ぴょ': 'pyo',
+
+        # Hiragana small characters and special cases
+        'っ': '', # Small tsu (doubles the following consonant)
+        'ゃ': 'ya', 'ゅ': 'yu', 'ょ': 'yo',
+
+        # Common punctuation and spaces
+        '　': ' ', # Japanese space
+        '、': ', ', '。': '. ',
+    }
+
+    result = []
+    i = 0
+
+    while i < len(japanese_text):
+        # Check for small tsu (doubling the following consonant)
+        if i < len(japanese_text) - 1 and (japanese_text[i] == 'っ' or japanese_text[i] == 'ッ'):
+            if i < len(japanese_text) - 1 and japanese_text[i+1] in kana_map:
+                next_romaji = kana_map[japanese_text[i+1]]
+                if next_romaji and next_romaji[0] not in 'aiueon':
+                    result.append(next_romaji[0])  # Double the consonant
+            i += 1
+            continue
+
+        # Check for combinations with small ya, yu, yo
+        if i < len(japanese_text) - 1 and japanese_text[i+1] in ('ゃ', 'ゅ', 'ょ', 'ャ', 'ュ', 'ョ'):
+            combo = japanese_text[i:i+2]
+            if combo in kana_map:
+                result.append(kana_map[combo])
+                i += 2
+                continue
+
+        # Regular character
+        if japanese_text[i] in kana_map:
+            result.append(kana_map[japanese_text[i]])
+        else:
+            # If it's not in our map, keep it as is (might be kanji, romaji, etc.)
+            result.append(japanese_text[i])
+
+        i += 1
+
+    return ''.join(result)
+
+def number_to_text(num, ordinal=False):
+    """
+    Convert a number (int or float) to its text representation.
+
+    Args:
+        num: The number to convert
+
+    Returns:
+        str: Text representation of the number
+    """
+
+    if not isinstance(num, (int, float)):
+        return "Input must be a number"
+
+    # Handle special case of zero
+    if num == 0:
+        return "zero"
+
+    # Handle negative numbers
+    negative = num < 0
+    num = abs(num)
+
+    # Handle floats
+    if isinstance(num, float):
+        # Split into integer and decimal parts
+        int_part = int(num)
+
+        # Convert both parts
+        int_text = _int_to_text(int_part)
+
+        # Handle decimal part (convert to string and remove '0.')
+        decimal_str = str(num).split('.')[1]
+        decimal_text = " point " + " ".join(_digit_to_text(int(digit)) for digit in decimal_str)
+
+        result = int_text + decimal_text
+    else:
+        # Handle integers
+        result = _int_to_text(num)
+
+    # Add 'negative' prefix for negative numbers
+    if negative:
+        result = "negative " + result
+
+    return result
+
+
+def _int_to_text(num):
+    """Helper function to convert an integer to text"""
+
+    ones = ["", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine",
+            "ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen",
+            "seventeen", "eighteen", "nineteen"]
+
+    tens = ["", "", "twenty", "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety"]
+
+    if num < 20:
+        return ones[num]
+
+    if num < 100:
+        return tens[num // 10] + (" " + ones[num % 10] if num % 10 != 0 else "")
+
+    if num < 1000:
+        return ones[num // 100] + " hundred" + (" " + _int_to_text(num % 100) if num % 100 != 0 else "")
+
+    if num < 1000000:
+        return _int_to_text(num // 1000) + " thousand" + (" " + _int_to_text(num % 1000) if num % 1000 != 0 else "")
+
+    if num < 1000000000:
+        return _int_to_text(num // 1000000) + " million" + (" " + _int_to_text(num % 1000000) if num % 1000000 != 0 else "")
+
+    return _int_to_text(num // 1000000000) + " billion" + (" " + _int_to_text(num % 1000000000) if num % 1000000000 != 0 else "")
+
+
+def _digit_to_text(digit):
+    """Convert a single digit to text"""
+    digits = ["zero", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine"]
+    return digits[digit]
+
+
+_whitespace_re = re.compile(r"\s+")
+
+
+# List of (regular expression, replacement) pairs for abbreviations:
+_abbreviations = {
+    "en": [
+        (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+        for x in [
+            ("mrs", "misess"),
+            ("mr", "mister"),
+            ("dr", "doctor"),
+            ("st", "saint"),
+            ("co", "company"),
+            ("jr", "junior"),
+            ("maj", "major"),
+            ("gen", "general"),
+            ("drs", "doctors"),
+            ("rev", "reverend"),
+            ("lt", "lieutenant"),
+            ("hon", "honorable"),
+            ("sgt", "sergeant"),
+            ("capt", "captain"),
+            ("esq", "esquire"),
+            ("ltd", "limited"),
+            ("col", "colonel"),
+            ("ft", "fort"),
+        ]
+    ],
+}
+
+
+def expand_abbreviations_multilingual(text, lang="en"):
+    for regex, replacement in _abbreviations[lang]:
+        text = re.sub(regex, replacement, text)
+    return text
+
+
+_symbols_multilingual = {
+    "en": [
+        (re.compile(r"%s" % re.escape(x[0]), re.IGNORECASE), x[1])
+        for x in [
+            ("&", " and "),
+            ("@", " at "),
+            ("%", " percent "),
+            ("#", " hash "),
+            ("$", " dollar "),
+            ("£", " pound "),
+            ("°", " degree "),
+        ]
+    ],
+}
+
+
+def expand_symbols_multilingual(text, lang="en"):
+    for regex, replacement in _symbols_multilingual[lang]:
+        text = re.sub(regex, replacement, text)
+        text = text.replace("  ", " ")  # Ensure there are no double spaces
+    return text.strip()
+
+
+_ordinal_re = {
+    "en": re.compile(r"([0-9]+)(st|nd|rd|th)"),
+}
+_number_re = re.compile(r"[0-9]+")
+_currency_re = {
+    "USD": re.compile(r"((\$[0-9\.\,]*[0-9]+)|([0-9\.\,]*[0-9]+\$))"),
+    "GBP": re.compile(r"((£[0-9\.\,]*[0-9]+)|([0-9\.\,]*[0-9]+£))"),
+    "EUR": re.compile(r"(([0-9\.\,]*[0-9]+€)|((€[0-9\.\,]*[0-9]+)))"),
+}
+
+_comma_number_re = re.compile(r"\b\d{1,3}(,\d{3})*(\.\d+)?\b")
+_dot_number_re = re.compile(r"\b\d{1,3}(.\d{3})*(\,\d+)?\b")
+_decimal_number_re = re.compile(r"([0-9]+[.,][0-9]+)")
+
+
+def _remove_commas(m):
+    text = m.group(0)
+    if "," in text:
+        text = text.replace(",", "")
+    return text
+
+
+def _remove_dots(m):
+    text = m.group(0)
+    if "." in text:
+        text = text.replace(".", "")
+    return text
+
+
+def _expand_decimal_point(m, lang="en"):
+    amount = m.group(1).replace(",", ".")
+    return number_to_text(float(amount))
+
+
+def _expand_currency(m, lang="en", currency="USD"):
+    amount = float((re.sub(r"[^\d.]", "", m.group(0).replace(",", "."))))
+    full_amount = number_to_text(amount)
+
+    and_equivalents = {
+        "en": ", ",
+        "es": " con ",
+        "fr": " et ",
+        "de": " und ",
+        "pt": " e ",
+        "it": " e ",
+        "pl": ", ",
+        "cs": ", ",
+        "ru": ", ",
+        "nl": ", ",
+        "ar": ", ",
+        "tr": ", ",
+        "hu": ", ",
+        "ko": ", ",
+    }
+
+    if amount.is_integer():
+        last_and = full_amount.rfind(and_equivalents[lang])
+        if last_and != -1:
+            full_amount = full_amount[:last_and]
+
+    return full_amount
+
+
+def _expand_ordinal(m, lang="en"):
+    return number_to_text(int(m.group(1)), ordinal=True)
+
+
+def _expand_number(m, lang="en"):
+    return number_to_text(int(m.group(0)))
+
+
+def expand_numbers_multilingual(text, lang="en"):
+    if lang in ["en", "ru"]:
+        text = re.sub(_comma_number_re, _remove_commas, text)
+    else:
+        text = re.sub(_dot_number_re, _remove_dots, text)
+    try:
+        text = re.sub(_currency_re["GBP"], lambda m: _expand_currency(m, lang, "GBP"), text)
+        text = re.sub(_currency_re["USD"], lambda m: _expand_currency(m, lang, "USD"), text)
+        text = re.sub(_currency_re["EUR"], lambda m: _expand_currency(m, lang, "EUR"), text)
+    except:
+        pass
+
+    text = re.sub(_decimal_number_re, lambda m: _expand_decimal_point(m, lang), text)
+    text = re.sub(_ordinal_re[lang], lambda m: _expand_ordinal(m, lang), text)
+    text = re.sub(_number_re, lambda m: _expand_number(m, lang), text)
+    return text
+
+
+def lowercase(text):
+    return text.lower()
+
+
+def collapse_whitespace(text):
+    return re.sub(_whitespace_re, " ", text)
+
+
+def multilingual_cleaners(text, lang):
+    text = text.replace('"', "")
+    if lang == "tr":
+        text = text.replace("İ", "i")
+        text = text.replace("Ö", "ö")
+        text = text.replace("Ü", "ü")
+    text = lowercase(text)
+    try:
+        text = expand_numbers_multilingual(text, lang)
+    except:
+        pass
+    try:
+        text = expand_abbreviations_multilingual(text, lang)
+    except:
+        pass
+    try:
+        text = expand_symbols_multilingual(text, lang=lang)
+    except:
+        pass
+    text = collapse_whitespace(text)
+    return text
+
+
+def basic_cleaners(text):
+    """Basic pipeline that lowercases and collapses whitespace without transliteration."""
+    text = lowercase(text)
+    text = collapse_whitespace(text)
+    return text

comfy/text_encoders/flux.py

@@ -19,8 +19,8 @@ class FluxTokenizer:
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
-        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):

comfy/text_encoders/hidream.py

@@ -16,11 +16,11 @@ class HiDreamTokenizer:
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids)
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
-        t5xxl = self.t5xxl.tokenize_with_weights(text, return_word_ids)
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
+        t5xxl = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         out["t5xxl"] = [t5xxl[0]]  # Use only first 128 tokens
-        out["llama"] = self.llama.tokenize_with_weights(text, return_word_ids)
+        out["llama"] = self.llama.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):
@@ -109,14 +109,18 @@ class HiDreamTEModel(torch.nn.Module):
         if self.t5xxl is not None:
             t5_output = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
             t5_out, t5_pooled = t5_output[:2]
+        else:
+            t5_out = None
 
         if self.llama is not None:
             ll_output = self.llama.encode_token_weights(token_weight_pairs_llama)
             ll_out, ll_pooled = ll_output[:2]
             ll_out = ll_out[:, 1:]
+        else:
+            ll_out = None
 
         if t5_out is None:
-            t5_out = torch.zeros((1, 1, 4096), device=comfy.model_management.intermediate_device())
+            t5_out = torch.zeros((1, 128, 4096), device=comfy.model_management.intermediate_device())
 
         if ll_out is None:
             ll_out = torch.zeros((1, 32, 1, 4096), device=comfy.model_management.intermediate_device())

comfy/text_encoders/hunyuan_video.py

@@ -49,13 +49,13 @@ class HunyuanVideoTokenizer:
 
     def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, image_embeds=None, image_interleave=1, **kwargs):
         out = {}
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
 
         if llama_template is None:
             llama_text = self.llama_template.format(text)
         else:
             llama_text = llama_template.format(text)
-        llama_text_tokens = self.llama.tokenize_with_weights(llama_text, return_word_ids)
+        llama_text_tokens = self.llama.tokenize_with_weights(llama_text, return_word_ids, **kwargs)
         embed_count = 0
         for r in llama_text_tokens:
             for i in range(len(r)):

comfy/text_encoders/hydit.py

@@ -41,8 +41,8 @@ class HyditTokenizer:
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids)
-        out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids)
+        out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):

comfy/text_encoders/llama.py

@@ -24,6 +24,24 @@ class Llama2Config:
     head_dim = 128
     rms_norm_add = False
     mlp_activation = "silu"
+    qkv_bias = False
+
+@dataclass
+class Qwen25_3BConfig:
+    vocab_size: int = 151936
+    hidden_size: int = 2048
+    intermediate_size: int = 11008
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 2
+    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:
@@ -40,6 +58,7 @@ class Gemma2_2B_Config:
     head_dim = 256
     rms_norm_add = True
     mlp_activation = "gelu_pytorch_tanh"
+    qkv_bias = False
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -98,9 +117,9 @@ class Attention(nn.Module):
         self.inner_size = self.num_heads * self.head_dim
 
         ops = ops or nn
-        self.q_proj = ops.Linear(config.hidden_size, self.inner_size, bias=False, device=device, dtype=dtype)
-        self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False, device=device, dtype=dtype)
-        self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False, device=device, dtype=dtype)
+        self.q_proj = ops.Linear(config.hidden_size, self.inner_size, bias=config.qkv_bias, device=device, dtype=dtype)
+        self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.qkv_bias, device=device, dtype=dtype)
+        self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.qkv_bias, device=device, dtype=dtype)
         self.o_proj = ops.Linear(self.inner_size, config.hidden_size, bias=False, device=device, dtype=dtype)
 
     def forward(
@@ -320,6 +339,14 @@ class Llama2(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen25_3B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen25_3BConfig(**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):

comfy/text_encoders/long_clipl.json

@@ -1,25 +0,0 @@
-{
-  "_name_or_path": "openai/clip-vit-large-patch14",
-  "architectures": [
-    "CLIPTextModel"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 49407,
-  "hidden_act": "quick_gelu",
-  "hidden_size": 768,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 3072,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 248,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 12,
-  "num_hidden_layers": 12,
-  "pad_token_id": 1,
-  "projection_dim": 768,
-  "torch_dtype": "float32",
-  "transformers_version": "4.24.0",
-  "vocab_size": 49408
-}

comfy/text_encoders/omnigen2.py

@@ -0,0 +1,44 @@
+from transformers import Qwen2Tokenizer
+from comfy import sd1_clip
+import comfy.text_encoders.llama
+import os
+
+
+class Qwen25_3BTokenizer(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=2048, embedding_key='qwen25_3b', 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 Omnigen2Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_3b", tokenizer=Qwen25_3BTokenizer)
+        self.llama_template = '<|im_start|>system\nYou are a helpful assistant that generates high-quality images based on user instructions.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\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_3BModel(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_3B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class Omnigen2Model(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="qwen25_3b", clip_model=Qwen25_3BModel, model_options=model_options)
+
+
+def te(dtype_llama=None, llama_scaled_fp8=None):
+    class Omnigen2TEModel_(Omnigen2Model):
+        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 Omnigen2TEModel_

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/qwen25_tokenizer/tokenizer_config.json

@@ -0,0 +1,241 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<|img|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151666": {
+      "content": "<|endofimg|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151667": {
+      "content": "<|meta|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151668": {
+      "content": "<|endofmeta|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen2_5_VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

comfy/text_encoders/sd3_clip.py

@@ -45,9 +45,9 @@ class SD3Tokenizer:
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids)
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
-        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids)
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):

comfy/text_encoders/spiece_tokenizer.py

@@ -1,4 +1,5 @@
 import torch
+import os
 
 class SPieceTokenizer:
     @staticmethod
@@ -15,6 +16,8 @@ class SPieceTokenizer:
         if isinstance(tokenizer_path, bytes):
             self.tokenizer = sentencepiece.SentencePieceProcessor(model_proto=tokenizer_path, add_bos=self.add_bos, add_eos=self.add_eos)
         else:
+            if not os.path.isfile(tokenizer_path):
+                raise ValueError("invalid tokenizer")
             self.tokenizer = sentencepiece.SentencePieceProcessor(model_file=tokenizer_path, add_bos=self.add_bos, add_eos=self.add_eos)
 
     def get_vocab(self):

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)