ComfyUI version v0.8.2

This logic was checking comfy_cast_weights, and going straight to
to the forward_comfy_cast_weights implementation without
attempting to downscale input to fp8 in the event comfy_cast_weights
is set.

The main reason comfy_cast_weights would be set would be for async
offload, which is not a good reason to nix FP8MM.

So instead, and together the underlying exclusions for FP8MM which
are:

* having a weight_function (usually LowVramPatch)
* force_cast_weights (compute dtype override)
* the weight is not Quantized
* the input is already quantized
* the model or layer has MM explictily disabled.

If you get past all of those exclusions, quantize the input tensor.
Then hand the new input, quantized or not off to
forward_comfy_cast_weights to handle it. If the weight is offloaded
but input is quantized you will get an offloaded MM8.

.ci/update_windows/update.py

@@ -53,6 +53,16 @@ try:
     repo.stash(ident)
 except KeyError:
     print("nothing to stash")  # noqa: T201
+except:
+    print("Could not stash, cleaning index and trying again.")  # noqa: T201
+    repo.state_cleanup()
+    repo.index.read_tree(repo.head.peel().tree)
+    repo.index.write()
+    try:
+        repo.stash(ident)
+    except KeyError:
+        print("nothing to stash.")  # noqa: T201
+
 backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S'))
 print("creating backup branch: {}".format(backup_branch_name))  # noqa: T201
 try:
@@ -66,8 +76,10 @@ if branch is None:
     try:
         ref = repo.lookup_reference('refs/remotes/origin/master')
     except:
-        print("pulling.")  # noqa: T201
-        pull(repo)
+        print("fetching.")  # noqa: T201
+        for remote in repo.remotes:
+            if remote.name == "origin":
+                remote.fetch()
         ref = repo.lookup_reference('refs/remotes/origin/master')
     repo.checkout(ref)
     branch = repo.lookup_branch('master')
@@ -149,3 +161,4 @@ try:
         shutil.copy(stable_update_script, stable_update_script_to)
 except:
     pass
+

.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt

@@ -0,0 +1,28 @@
+As of the time of writing this you need this driver for best results:
+https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html
+
+HOW TO RUN:
+
+If you have a AMD gpu:
+
+run_amd_gpu.bat
+
+If you have memory issues you can try disabling the smart memory management by running comfyui with:
+
+run_amd_gpu_disable_smart_memory.bat
+
+IF YOU GET A RED ERROR IN THE UI MAKE SURE YOU HAVE A MODEL/CHECKPOINT IN: ComfyUI\models\checkpoints
+
+You can download the stable diffusion XL one from: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors
+
+
+RECOMMENDED WAY TO UPDATE:
+To update the ComfyUI code: update\update_comfyui.bat
+
+
+TO SHARE MODELS BETWEEN COMFYUI AND ANOTHER UI:
+In the ComfyUI directory you will find a file: extra_model_paths.yaml.example
+Rename this file to: extra_model_paths.yaml and edit it with your favorite text editor.
+
+
+

.ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat

@@ -1,2 +1,2 @@
-.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory
 pause

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

@@ -0,0 +1,3 @@
+..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
+pause

.ci/windows_nvidia_base_files/run_nvidia_gpu.bat

@@ -0,0 +1,3 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
+pause

.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -0,0 +1,3 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
+pause

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -8,13 +8,15 @@ body:
         Before submitting a **Bug Report**, please ensure the following:
 
         - **1:** You are running the latest version of ComfyUI.
-        - **2:** You have looked at the existing bug reports and made sure this isn't already reported.
+        - **2:** You have your ComfyUI logs and relevant workflow on hand and will post them in this bug report.
         - **3:** You confirmed that the bug is not caused by a custom node. You can disable all custom nodes by passing
-        `--disable-all-custom-nodes` command line argument.
+        `--disable-all-custom-nodes` command line argument. If you have custom node try updating them to the latest version.
         - **4:** This is an actual bug in ComfyUI, not just a support question. A bug is when you can specify exact
         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.
+        ## Very Important
+
+        Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored.
   - type: checkboxes
     id: custom-nodes-test
     attributes:

.github/PULL_REQUEST_TEMPLATE/api-node.md

@@ -0,0 +1,21 @@
+<!-- API_NODE_PR_CHECKLIST: do not remove -->
+
+## API Node PR Checklist
+
+### Scope
+- [ ] **Is API Node Change**
+
+### Pricing & Billing
+- [ ] **Need pricing update**
+- [ ] **No pricing update**
+
+If **Need pricing update**:
+- [ ] Metronome rate cards updated
+- [ ] Auto‑billing tests updated and passing
+
+### QA
+- [ ] **QA done**
+- [ ] **QA not required**
+
+### Comms
+- [ ] Informed **Kosinkadink**

.github/workflows/api-node-template.yml

@@ -0,0 +1,58 @@
+name: Append API Node PR template
+
+on:
+  pull_request_target:
+    types: [opened, reopened, synchronize, ready_for_review]
+    paths:
+      - 'comfy_api_nodes/**'   # only run if these files changed
+
+permissions:
+  contents: read
+  pull-requests: write
+
+jobs:
+  inject:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Ensure template exists and append to PR body
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const number = context.payload.pull_request.number;
+            const templatePath = '.github/PULL_REQUEST_TEMPLATE/api-node.md';
+            const marker = '<!-- API_NODE_PR_CHECKLIST: do not remove -->';
+
+            const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number: number });
+
+            let templateText;
+            try {
+              const res = await github.rest.repos.getContent({
+                owner,
+                repo,
+                path: templatePath,
+                ref: pr.base.ref
+              });
+              const buf = Buffer.from(res.data.content, res.data.encoding || 'base64');
+              templateText = buf.toString('utf8');
+            } catch (e) {
+              core.setFailed(`Required PR template not found at "${templatePath}" on ${pr.base.ref}. Please add it to the repo.`);
+              return;
+            }
+
+            // Enforce the presence of the marker inside the template (for idempotence)
+            if (!templateText.includes(marker)) {
+              core.setFailed(`Template at "${templatePath}" does not contain the required marker:\n${marker}\nAdd it so we can detect duplicates safely.`);
+              return;
+            }
+
+            // If the PR already contains the marker, do not append again.
+            const body = pr.body || '';
+            if (body.includes(marker)) {
+              core.info('Template already present in PR body; nothing to inject.');
+              return;
+            }
+
+            const newBody = (body ? body + '\n\n' : '') + templateText + '\n';
+            await github.rest.pulls.update({ owner, repo, pull_number: number, body: newBody });
+            core.notice('API Node template appended to PR description.');

.github/workflows/release-stable-all.yml

@@ -0,0 +1,78 @@
+name: "Release Stable All Portable Versions"
+
+on:
+  workflow_dispatch:
+    inputs:
+      git_tag:
+        description: 'Git tag'
+        required: true
+        type: string
+
+jobs:
+  release_nvidia_default:
+    permissions:
+      contents: "write"
+      packages: "write"
+      pull-requests: "read"
+    name: "Release NVIDIA Default (cu130)"
+    uses: ./.github/workflows/stable-release.yml
+    with:
+      git_tag: ${{ inputs.git_tag }}
+      cache_tag: "cu130"
+      python_minor: "13"
+      python_patch: "9"
+      rel_name: "nvidia"
+      rel_extra_name: ""
+      test_release: true
+    secrets: inherit
+
+  release_nvidia_cu128:
+    permissions:
+      contents: "write"
+      packages: "write"
+      pull-requests: "read"
+    name: "Release NVIDIA cu128"
+    uses: ./.github/workflows/stable-release.yml
+    with:
+      git_tag: ${{ inputs.git_tag }}
+      cache_tag: "cu128"
+      python_minor: "12"
+      python_patch: "10"
+      rel_name: "nvidia"
+      rel_extra_name: "_cu128"
+      test_release: true
+    secrets: inherit
+
+  release_nvidia_cu126:
+    permissions:
+      contents: "write"
+      packages: "write"
+      pull-requests: "read"
+    name: "Release NVIDIA cu126"
+    uses: ./.github/workflows/stable-release.yml
+    with:
+      git_tag: ${{ inputs.git_tag }}
+      cache_tag: "cu126"
+      python_minor: "12"
+      python_patch: "10"
+      rel_name: "nvidia"
+      rel_extra_name: "_cu126"
+      test_release: true
+    secrets: inherit
+
+  release_amd_rocm:
+    permissions:
+      contents: "write"
+      packages: "write"
+      pull-requests: "read"
+    name: "Release AMD ROCm 7.1.1"
+    uses: ./.github/workflows/stable-release.yml
+    with:
+      git_tag: ${{ inputs.git_tag }}
+      cache_tag: "rocm711"
+      python_minor: "12"
+      python_patch: "10"
+      rel_name: "amd"
+      rel_extra_name: ""
+      test_release: false
+    secrets: inherit

.github/workflows/ruff.yml

@@ -21,3 +21,28 @@ jobs:
 
     - name: Run Ruff
       run: ruff check .
+
+  pylint:
+    name: Run Pylint
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+
+    - name: Set up Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.12'
+
+    - name: Install requirements
+      run: |
+        python -m pip install --upgrade pip
+        pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+        pip install -r requirements.txt
+
+    - name: Install Pylint
+      run: pip install pylint
+
+    - name: Run Pylint
+      run: pylint comfy_api_nodes

.github/workflows/stable-release.yml

@@ -2,17 +2,17 @@
 name: "Release Stable Version"
 
 on:
-  workflow_dispatch:
+  workflow_call:
     inputs:
       git_tag:
         description: 'Git tag'
         required: true
         type: string
-      cu:
-        description: 'CUDA version'
+      cache_tag:
+        description: 'Cached dependencies tag'
         required: true
         type: string
-        default: "129"
+        default: "cu129"
       python_minor:
         description: 'Python minor version'
         required: true
@@ -23,7 +23,57 @@ on:
         required: true
         type: string
         default: "6"
-
+      rel_name:
+        description: 'Release name'
+        required: true
+        type: string
+        default: "nvidia"
+      rel_extra_name:
+        description: 'Release extra name'
+        required: false
+        type: string
+        default: ""
+      test_release:
+        description: 'Test Release'
+        required: true
+        type: boolean
+        default: true
+  workflow_dispatch:
+    inputs:
+      git_tag:
+        description: 'Git tag'
+        required: true
+        type: string
+      cache_tag:
+        description: 'Cached dependencies tag'
+        required: true
+        type: string
+        default: "cu129"
+      python_minor:
+        description: 'Python minor version'
+        required: true
+        type: string
+        default: "13"
+      python_patch:
+        description: 'Python patch version'
+        required: true
+        type: string
+        default: "6"
+      rel_name:
+        description: 'Release name'
+        required: true
+        type: string
+        default: "nvidia"
+      rel_extra_name:
+        description: 'Release extra name'
+        required: false
+        type: string
+        default: ""
+      test_release:
+        description: 'Test Release'
+        required: true
+        type: boolean
+        default: true
 
 jobs:
   package_comfy_windows:
@@ -42,15 +92,15 @@ jobs:
         id: cache
         with:
           path: |
-            cu${{ inputs.cu }}_python_deps.tar
+            ${{ inputs.cache_tag }}_python_deps.tar
             update_comfyui_and_python_dependencies.bat
-          key: ${{ runner.os }}-build-cu${{ inputs.cu }}-${{ inputs.python_minor }}
+          key: ${{ runner.os }}-build-${{ inputs.cache_tag }}-${{ inputs.python_minor }}
       - shell: bash
         run: |
-          mv cu${{ inputs.cu }}_python_deps.tar ../
+          mv ${{ inputs.cache_tag }}_python_deps.tar ../
           mv update_comfyui_and_python_dependencies.bat ../
           cd ..
-          tar xf cu${{ inputs.cu }}_python_deps.tar
+          tar xf ${{ inputs.cache_tag }}_python_deps.tar
           pwd
           ls
 
@@ -65,12 +115,19 @@ jobs:
           echo 'import site' >> ./python3${{ inputs.python_minor }}._pth
           curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
           ./python.exe get-pip.py
-          ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
+          ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
+
+          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          ./python.exe -s -m pip install -r requirements_comfyui.txt
+          rm requirements_comfyui.txt
+
           sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
 
-          rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
-          rm ./Lib/site-packages/torch/lib/libprotoc.lib
-          rm ./Lib/site-packages/torch/lib/libprotobuf.lib
+          if test -f ./Lib/site-packages/torch/lib/dnnl.lib; then
+            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
+            rm ./Lib/site-packages/torch/lib/libprotoc.lib
+            rm ./Lib/site-packages/torch/lib/libprotobuf.lib
+          fi
 
           cd ..
 
@@ -85,14 +142,18 @@ jobs:
 
           mkdir update
           cp -r ComfyUI/.ci/update_windows/* ./update/
-          cp -r ComfyUI/.ci/windows_base_files/* ./
+          cp -r ComfyUI/.ci/windows_${{ inputs.rel_name }}_base_files/* ./
           cp ../update_comfyui_and_python_dependencies.bat ./update/
 
           cd ..
 
           "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
-          mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
+          mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_${{ inputs.rel_name }}${{ inputs.rel_extra_name }}.7z
 
+      - shell: bash
+        if: ${{ inputs.test_release }}
+        run: |
+          cd ..
           cd ComfyUI_windows_portable
           python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
@@ -101,10 +162,9 @@ jobs:
           ls
 
       - name: Upload binaries to release
-        uses: svenstaro/upload-release-action@v2
+        uses: softprops/action-gh-release@v2
         with:
-          repo_token: ${{ secrets.GITHUB_TOKEN }}
-          file: ComfyUI_windows_portable_nvidia.7z
-          tag: ${{ inputs.git_tag }}
-          overwrite: true
+          files: ComfyUI_windows_portable_${{ inputs.rel_name }}${{ inputs.rel_extra_name }}.7z
+          tag_name: ${{ inputs.git_tag }}
           draft: true
+          overwrite_files: true

.github/workflows/test-build.yml

@@ -18,7 +18,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
     steps:
       - uses: actions/checkout@v4
       - name: Set up Python ${{ matrix.python-version }}

.github/workflows/test-ci.yml

@@ -5,6 +5,7 @@ on:
   push:
     branches:
       - master
+      - release/**
     paths-ignore:
       - 'app/**'
       - 'input/**'
@@ -19,16 +20,18 @@ jobs:
   test-stable:
     strategy:
       fail-fast: false
+      max-parallel: 1  # This forces sequential execution
       matrix:
         # os: [macos, linux, windows]
-        os: [macos, linux]
-        python_version: ["3.9", "3.10", "3.11", "3.12"]
+        # os: [macos, linux]
+        os: [linux]
+        python_version: ["3.10", "3.11", "3.12"]
         cuda_version: ["12.1"]
         torch_version: ["stable"]
         include:
-          - os: macos
-            runner_label: [self-hosted, macOS]
-            flags: "--use-pytorch-cross-attention"
+          # - os: macos
+          #   runner_label: [self-hosted, macOS]
+          #   flags: "--use-pytorch-cross-attention"
           - os: linux
             runner_label: [self-hosted, Linux]
             flags: ""
@@ -72,15 +75,17 @@ jobs:
   test-unix-nightly:
     strategy:
       fail-fast: false
+      max-parallel: 1  # This forces sequential execution
       matrix:
-        os: [macos, linux]
+        # os: [macos, linux]
+        os: [linux]
         python_version: ["3.11"]
         cuda_version: ["12.1"]
         torch_version: ["nightly"]
         include:
-          - os: macos
-            runner_label: [self-hosted, macOS]
-            flags: "--use-pytorch-cross-attention"
+          # - os: macos
+          #   runner_label: [self-hosted, macOS]
+          #   flags: "--use-pytorch-cross-attention"
           - os: linux
             runner_label: [self-hosted, Linux]
             flags: ""

.github/workflows/test-execution.yml

@@ -0,0 +1,30 @@
+name: Execution Tests
+
+on:
+  push:
+    branches: [ main, master, release/** ]
+  pull_request:
+    branches: [ main, master, release/** ]
+
+jobs:
+  test:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, windows-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
+    continue-on-error: true
+    steps:
+    - uses: actions/checkout@v4
+    - name: Set up Python      
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.12'
+    - name: Install requirements
+      run: |
+        python -m pip install --upgrade pip
+        pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+        pip install -r requirements.txt
+        pip install -r tests-unit/requirements.txt
+    - name: Run Execution Tests
+      run: |
+        python -m pytest tests/execution -v --skip-timing-checks

.github/workflows/test-launch.yml

@@ -2,9 +2,9 @@ name: Test server launches without errors
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:
@@ -32,7 +32,9 @@ jobs:
       working-directory: ComfyUI
     - name: Check for unhandled exceptions in server log
       run: |
-        if grep -qE "Exception|Error" console_output.log; then
+        grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" console_output.log | grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': False, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" > console_output_filtered.log
+        cat console_output_filtered.log
+        if grep -qE "Exception|Error" console_output_filtered.log; then
           echo "Unhandled exception/error found in server log."
           exit 1
         fi

.github/workflows/test-unit.yml

@@ -2,15 +2,15 @@ name: Unit Tests
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:
     strategy:
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
+        os: [ubuntu-latest, windows-2022, macos-latest]
     runs-on: ${{ matrix.os }}
     continue-on-error: true
     steps:

.github/workflows/update-version.yml

@@ -6,6 +6,7 @@ on:
       - "pyproject.toml"
     branches:
       - master
+      - release/**
 
 jobs:
   update-version:

.github/workflows/windows_release_dependencies.yml

@@ -17,7 +17,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "129"
+        default: "130"
 
       python_minor:
         description: 'python minor version'
@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "6"
+        default: "9"
 #  push:
 #    branches:
 #      - master
@@ -56,7 +56,8 @@ jobs:
             ..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio ${{ inputs.xformers }} --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2
             pause" > update_comfyui_and_python_dependencies.bat
 
-            python -m pip wheel --no-cache-dir torch torchvision torchaudio ${{ inputs.xformers }} ${{ inputs.extra_dependencies }} --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cu }} -r requirements.txt pygit2 -w ./temp_wheel_dir
+            grep -v comfyui requirements.txt > requirements_nocomfyui.txt
+            python -m pip wheel --no-cache-dir torch torchvision torchaudio ${{ inputs.xformers }} ${{ inputs.extra_dependencies }} --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cu }} -r requirements_nocomfyui.txt pygit2 -w ./temp_wheel_dir
             python -m pip install --no-cache-dir ./temp_wheel_dir/*
             echo installed basic
             ls -lah temp_wheel_dir

.github/workflows/windows_release_dependencies_manual.yml

@@ -0,0 +1,64 @@
+name: "Windows Release dependencies Manual"
+
+on:
+  workflow_dispatch:
+    inputs:
+      torch_dependencies:
+        description: 'torch dependencies'
+        required: false
+        type: string
+        default: "torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128"
+      cache_tag:
+        description: 'Cached dependencies tag'
+        required: true
+        type: string
+        default: "cu128"
+
+      python_minor:
+        description: 'python minor version'
+        required: true
+        type: string
+        default: "12"
+
+      python_patch:
+        description: 'python patch version'
+        required: true
+        type: string
+        default: "10"
+
+jobs:
+  build_dependencies:
+    runs-on: windows-latest
+    steps:
+        - uses: actions/checkout@v4
+        - uses: actions/setup-python@v5
+          with:
+            python-version: 3.${{ inputs.python_minor }}.${{ inputs.python_patch }}
+
+        - shell: bash
+          run: |
+            echo "@echo off
+            call update_comfyui.bat nopause
+            echo -
+            echo This will try to update pytorch and all python dependencies.
+            echo -
+            echo If you just want to update normally, close this and run update_comfyui.bat instead.
+            echo -
+            pause
+            ..\python_embeded\python.exe -s -m pip install --upgrade ${{ inputs.torch_dependencies }} -r ../ComfyUI/requirements.txt pygit2
+            pause" > update_comfyui_and_python_dependencies.bat
+
+            grep -v comfyui requirements.txt > requirements_nocomfyui.txt
+            python -m pip wheel --no-cache-dir ${{ inputs.torch_dependencies }} -r requirements_nocomfyui.txt pygit2 -w ./temp_wheel_dir
+            python -m pip install --no-cache-dir ./temp_wheel_dir/*
+            echo installed basic
+            ls -lah temp_wheel_dir
+            mv temp_wheel_dir ${{ inputs.cache_tag }}_python_deps
+            tar cf ${{ inputs.cache_tag }}_python_deps.tar ${{ inputs.cache_tag }}_python_deps
+
+        - uses: actions/cache/save@v4
+          with:
+            path: |
+              ${{ inputs.cache_tag }}_python_deps.tar
+              update_comfyui_and_python_dependencies.bat
+            key: ${{ runner.os }}-build-${{ inputs.cache_tag }}-${{ inputs.python_minor }}

.github/workflows/windows_release_nightly_pytorch.yml

@@ -68,7 +68,7 @@ jobs:
 
             mkdir update
             cp -r ComfyUI/.ci/update_windows/* ./update/
-            cp -r ComfyUI/.ci/windows_base_files/* ./
+            cp -r ComfyUI/.ci/windows_nvidia_base_files/* ./
             cp -r ComfyUI/.ci/windows_nightly_base_files/* ./
 
             echo "call update_comfyui.bat nopause

.github/workflows/windows_release_package.yml

@@ -81,7 +81,7 @@ jobs:
 
             mkdir update
             cp -r ComfyUI/.ci/update_windows/* ./update/
-            cp -r ComfyUI/.ci/windows_base_files/* ./
+            cp -r ComfyUI/.ci/windows_nvidia_base_files/* ./
             cp ../update_comfyui_and_python_dependencies.bat ./update/
 
             cd ..

CODEOWNERS

@@ -1,25 +1,2 @@
 # Admins
-* @comfyanonymous
-
-# Note: Github teams syntax cannot be used here as the repo is not owned by Comfy-Org.
-# Inlined the team members for now.
-
-# Maintainers
-*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
-
-# Python web server
-/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
-/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
-/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
-
-# Node developers
-/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
-/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
-/comfy_api_nodes/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+* @comfyanonymous @kosinkadink @guill

QUANTIZATION.md

@@ -0,0 +1,168 @@
+# The Comfy guide to Quantization
+
+
+## How does quantization work?
+
+Quantization aims to map a high-precision value x_f to a lower precision format with minimal loss in accuracy. These smaller formats then serve to reduce the models memory footprint and increase throughput by using specialized hardware.
+
+When simply converting a value from FP16 to FP8 using the round-nearest method we might hit two issues:
+- The dynamic range of FP16 (-65,504, 65,504) far exceeds FP8 formats like E4M3 (-448, 448) or E5M2 (-57,344, 57,344), potentially resulting in clipped values
+- The original values are concentrated in a small range (e.g. -1,1) leaving many FP8-bits "unused"
+
+By using a scaling factor, we aim to map these values into the quantized-dtype range, making use of the full spectrum. One of the easiest approaches, and common, is using per-tensor absolute-maximum scaling.
+
+```
+absmax = max(abs(tensor))
+scale = amax / max_dynamic_range_low_precision
+
+# Quantization
+tensor_q = (tensor / scale).to(low_precision_dtype)
+
+# De-Quantization
+tensor_dq = tensor_q.to(fp16) * scale
+
+tensor_dq ~ tensor
+```
+
+Given that additional information (scaling factor) is needed to "interpret" the quantized values, we describe those as derived datatypes.
+
+
+## Quantization in Comfy
+
+```
+QuantizedTensor (torch.Tensor subclass)
+  ↓ __torch_dispatch__
+Two-Level Registry (generic + layout handlers)
+  ↓
+MixedPrecisionOps + Metadata Detection
+```
+
+### Representation
+
+To represent these derived datatypes, ComfyUI uses a subclass of torch.Tensor to implements these using the `QuantizedTensor` class found in `comfy/quant_ops.py`
+
+A `Layout` class defines how a specific quantization format behaves:
+- Required parameters
+- Quantize method
+- De-Quantize method
+
+```python
+from comfy.quant_ops import QuantizedLayout
+
+class MyLayout(QuantizedLayout):
+    @classmethod
+    def quantize(cls, tensor, **kwargs):
+        # Convert to quantized format
+        qdata = ...
+        params = {'scale': ..., 'orig_dtype': tensor.dtype}
+        return qdata, params
+    
+    @staticmethod
+    def dequantize(qdata, scale, orig_dtype, **kwargs):
+        return qdata.to(orig_dtype) * scale
+```
+
+To then run operations using these QuantizedTensors we use two registry systems to define supported operations. 
+The first is a **generic registry** that handles operations common to all quantized formats (e.g., `.to()`, `.clone()`, `.reshape()`).
+
+The second registry is layout-specific and allows to implement fast-paths like nn.Linear.
+```python
+from comfy.quant_ops import register_layout_op
+
+@register_layout_op(torch.ops.aten.linear.default, MyLayout)
+def my_linear(func, args, kwargs):
+    # Extract tensors, call optimized kernel
+    ...
+```
+When `torch.nn.functional.linear()` is called with QuantizedTensor arguments, `__torch_dispatch__` automatically routes to the registered implementation.
+For any unsupported operation, QuantizedTensor will fallback to call `dequantize` and dispatch using the high-precision implementation.
+
+
+### Mixed Precision
+
+The `MixedPrecisionOps` class (lines 542-648 in `comfy/ops.py`) enables per-layer quantization decisions, allowing different layers in a model to use different precisions. This is activated when a model config contains a `layer_quant_config` dictionary that specifies which layers should be quantized and how.
+
+**Architecture:**
+
+```python
+class MixedPrecisionOps(disable_weight_init):
+    _layer_quant_config = {}  # Maps layer names to quantization configs
+    _compute_dtype = torch.bfloat16  # Default compute / dequantize precision
+```
+
+**Key mechanism:**
+
+The custom `Linear._load_from_state_dict()` method inspects each layer during model loading:
+- If the layer name is **not** in `_layer_quant_config`: load weight as regular tensor in `_compute_dtype`
+- If the layer name **is** in `_layer_quant_config`: 
+  - Load weight as `QuantizedTensor` with the specified layout (e.g., `TensorCoreFP8Layout`)
+  - Load associated quantization parameters (scales, block_size, etc.)
+
+**Why it's needed:**
+
+Not all layers tolerate quantization equally. Sensitive operations like final projections can be kept in higher precision, while compute-heavy matmuls are quantized. This provides most of the performance benefits while maintaining quality.
+
+The system is selected in `pick_operations()` when `model_config.layer_quant_config` is present, making it the highest-priority operation mode.
+
+
+## Checkpoint Format
+
+Quantized checkpoints are stored as standard safetensors files with quantized weight tensors and associated scaling parameters, plus a `_quantization_metadata` JSON entry describing the quantization scheme.
+
+The quantized checkpoint will contain the same layers as the original checkpoint but:
+- The weights are stored as quantized values, sometimes using a different storage datatype. E.g. uint8 container for fp8.
+- For each quantized weight a number of additional scaling parameters are stored alongside depending on the recipe.
+- We store a metadata.json in the metadata of the final safetensor containing the `_quantization_metadata` describing which layers are quantized and what layout has been used.
+
+### Scaling Parameters details
+We define 4 possible scaling parameters that should cover most recipes in the near-future:
+- **weight_scale**: quantization scalers for the weights
+- **weight_scale_2**: global scalers in the context of double scaling
+- **pre_quant_scale**: scalers used for smoothing salient weights
+- **input_scale**: quantization scalers for the activations
+
+| Format | Storage dtype | weight_scale | weight_scale_2 | pre_quant_scale | input_scale |
+|--------|---------------|--------------|----------------|-----------------|-------------|
+| float8_e4m3fn | float32 | float32 (scalar) | - | - | float32 (scalar) |
+
+You can find the defined formats in `comfy/quant_ops.py` (QUANT_ALGOS).
+
+### Quantization Metadata
+
+The metadata stored alongside the checkpoint contains:
+- **format_version**: String to define a version of the standard
+- **layers**: A dictionary mapping layer names to their quantization format. The format string maps to the definitions found in `QUANT_ALGOS`. 
+
+Example:
+```json
+{
+  "_quantization_metadata": {
+    "format_version": "1.0",
+    "layers": {
+      "model.layers.0.mlp.up_proj": "float8_e4m3fn",
+      "model.layers.0.mlp.down_proj": "float8_e4m3fn",
+      "model.layers.1.mlp.up_proj": "float8_e4m3fn"
+    }
+  }
+}
+```
+
+
+## Creating Quantized Checkpoints
+
+To create compatible checkpoints, use any quantization tool provided the output follows the checkpoint format described above and uses a layout defined in `QUANT_ALGOS`.
+
+### Weight Quantization
+
+Weight quantization is straightforward - compute the scaling factor directly from the weight tensor using the absolute maximum method described earlier. Each layer's weights are quantized independently and stored with their corresponding `weight_scale` parameter.
+
+### Calibration (for Activation Quantization)
+
+Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_scale` parameters that cannot be determined from static weights alone. Since activation values depend on actual inputs, we use **post-training calibration (PTQ)**:
+
+1. **Collect statistics**: Run inference on N representative samples
+2. **Track activations**: Record the absolute maximum (`amax`) of inputs to each quantized layer
+3. **Compute scales**: Derive `input_scale` from collected statistics
+4. **Store in checkpoint**: Save `input_scale` parameters alongside weights
+
+The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters.

README.md

@@ -65,8 +65,10 @@ 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/)
    - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/)
+   - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
+   - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
+   - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/)
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
@@ -77,9 +79,9 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [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/) 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/)
+   - [Hunyuan Video 1.5](https://docs.comfy.org/tutorials/video/hunyuan/hunyuan-video-1-5)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
    - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
@@ -113,10 +115,14 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 ## Release Process
 
-ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
+ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
-   - Releases a new stable version (e.g., v0.7.0)
+   - Releases a new stable version (e.g., v0.7.0) roughly every week.
+   - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release.
+   - Minor versions will be used for releases off the master branch.
+   - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense.
+   - Commits outside of the stable release tags may be very unstable and break many custom nodes.
    - Serves as the foundation for the desktop release
 
 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
@@ -173,10 +179,20 @@ There is a portable standalone build for Windows that should work for running on
 
 ### [Direct link to download](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z)
 
-Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints
+Simply download, extract with [7-Zip](https://7-zip.org) or with the windows explorer on recent windows versions and run. For smaller models you normally only need to put the checkpoints (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints but many of the larger models have multiple files. Make sure to follow the instructions to know which subfolder to put them in ComfyUI\models\
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
+Update your Nvidia drivers if it doesn't start.
+
+#### Alternative Downloads:
+
+[Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
+
+[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z).
+
+[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
+
 #### 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.
@@ -192,7 +208,13 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-Python 3.13 is very well supported. If you have trouble with some custom node dependencies you can try 3.12
+Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies.
+
+Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
+
+torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch unless it is less than 2 weeks old.
+
+### Instructions:
 
 Git clone this repo.
 
@@ -201,18 +223,36 @@ Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints
 Put your VAE in: models/vae
 
 
-### AMD GPUs (Linux only)
+### AMD GPUs (Linux)
+
 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/rocm6.4```
 
-This is the command to install the nightly with ROCm 6.4 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.0 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.4```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1```
+
+
+### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only.
+
+These have less hardware support than the builds above but they work on windows. You also need to install the pytorch version specific to your hardware.
+
+RDNA 3 (RX 7000 series):
+
+```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/```
+
+RDNA 3.5 (Strix halo/Ryzen AI Max+ 365):
+
+```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx1151/```
+
+RDNA 4 (RX 9000 series):
+
+```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/```
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
+Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
 
 1. To install PyTorch xpu, use the following command:
 
@@ -222,19 +262,15 @@ This is the command to install the Pytorch xpu nightly which might have some per
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
-
-1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
-
 ### NVIDIA
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130```
 
 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/cu129```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130```
 
 #### Troubleshooting
 
@@ -265,12 +301,6 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve
 
 > **Note**: Remember to add your models, VAE, LoRAs etc. to the corresponding Comfy folders, as discussed in [ComfyUI manual installation](#manual-install-windows-linux).
 
-#### 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
 
 For models compatible with Ascend Extension for PyTorch (torch_npu). To get started, ensure your environment meets the prerequisites outlined on the [installation](https://ascend.github.io/docs/sources/ascend/quick_install.html) page. Here's a step-by-step guide tailored to your platform and installation method:
@@ -295,6 +325,32 @@ For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step
 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`
 
+
+## [ComfyUI-Manager](https://github.com/Comfy-Org/ComfyUI-Manager/tree/manager-v4)
+
+**ComfyUI-Manager** is an extension that allows you to easily install, update, and manage custom nodes for ComfyUI.
+
+### Setup
+
+1. Install the manager dependencies:
+   ```bash
+   pip install -r manager_requirements.txt
+   ```
+
+2. Enable the manager with the `--enable-manager` flag when running ComfyUI:
+   ```bash
+   python main.py --enable-manager
+   ```
+
+### Command Line Options
+
+| Flag | Description |
+|------|-------------|
+| `--enable-manager` | Enable ComfyUI-Manager |
+| `--enable-manager-legacy-ui` | Use the legacy manager UI instead of the new UI (requires `--enable-manager`) |
+| `--disable-manager-ui` | Disable the manager UI and endpoints while keeping background features like security checks and scheduled installation completion (requires `--enable-manager`) |
+
+
 # Running
 
 ```python main.py```

api_server/routes/internal/internal_routes.py

@@ -58,8 +58,13 @@ class InternalRoutes:
                 return web.json_response({"error": "Invalid directory type"}, status=400)
 
             directory = get_directory_by_type(directory_type)
+
+            def is_visible_file(entry: os.DirEntry) -> bool:
+                """Filter out hidden files (e.g., .DS_Store on macOS)."""
+                return entry.is_file() and not entry.name.startswith('.')
+
             sorted_files = sorted(
-                (entry for entry in os.scandir(directory) if entry.is_file()),
+                (entry for entry in os.scandir(directory) if is_visible_file(entry)),
                 key=lambda entry: -entry.stat().st_mtime
             )
             return web.json_response([entry.name for entry in sorted_files], status=200)

app/frontend_management.py

@@ -10,7 +10,8 @@ import importlib
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import TypedDict, Optional
+from typing import Dict, TypedDict, Optional
+from aiohttp import web
 from importlib.metadata import version
 
 import requests
@@ -42,6 +43,7 @@ def get_installed_frontend_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:
@@ -63,6 +65,7 @@ def get_required_frontend_version():
         logging.error(f"Error reading requirements.txt: {e}")
         return None
 
+
 def check_frontend_version():
     """Check if the frontend version is up to date."""
 
@@ -203,6 +206,37 @@ class FrontendManager:
         """Get the required frontend package version."""
         return get_required_frontend_version()
 
+    @classmethod
+    def get_installed_templates_version(cls) -> str:
+        """Get the currently installed workflow templates package version."""
+        try:
+            templates_version_str = version("comfyui-workflow-templates")
+            return templates_version_str
+        except Exception:
+            return None
+
+    @classmethod
+    def get_required_templates_version(cls) -> str:
+        """Get the required workflow templates 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-workflow-templates=="):
+                        version_str = line.split("==")[-1]
+                        if not is_valid_version(version_str):
+                            logging.error(f"Invalid templates version format in requirements.txt: {version_str}")
+                            return None
+                        return version_str
+                logging.error("comfyui-workflow-templates not found in requirements.txt")
+                return None
+        except FileNotFoundError:
+            logging.error("requirements.txt not found. Cannot determine required templates version.")
+            return None
+        except Exception as e:
+            logging.error(f"Error reading requirements.txt: {e}")
+            return None
+
     @classmethod
     def default_frontend_path(cls) -> str:
         try:
@@ -224,7 +258,54 @@ comfyui-frontend-package is not installed.
             sys.exit(-1)
 
     @classmethod
-    def templates_path(cls) -> str:
+    def template_asset_map(cls) -> Optional[Dict[str, str]]:
+        """Return a mapping of template asset names to their absolute paths."""
+        try:
+            from comfyui_workflow_templates import (
+                get_asset_path,
+                iter_templates,
+            )
+        except ImportError:
+            logging.error(
+                f"""
+********** ERROR ***********
+
+comfyui-workflow-templates is not installed.
+
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
+            return None
+
+        try:
+            template_entries = list(iter_templates())
+        except Exception as exc:
+            logging.error(f"Failed to enumerate workflow templates: {exc}")
+            return None
+
+        asset_map: Dict[str, str] = {}
+        try:
+            for entry in template_entries:
+                for asset in entry.assets:
+                    asset_map[asset.filename] = get_asset_path(
+                        entry.template_id, asset.filename
+                    )
+        except Exception as exc:
+            logging.error(f"Failed to resolve template asset paths: {exc}")
+            return None
+
+        if not asset_map:
+            logging.error("No workflow template assets found. Did the packages install correctly?")
+            return None
+
+        return asset_map
+
+
+    @classmethod
+    def legacy_templates_path(cls) -> Optional[str]:
+        """Return the legacy templates directory shipped inside the meta package."""
         try:
             import comfyui_workflow_templates
 
@@ -243,6 +324,7 @@ comfyui-workflow-templates is not installed.
 ********** ERROR ***********
 """.strip()
             )
+            return None
 
     @classmethod
     def embedded_docs_path(cls) -> str:
@@ -359,3 +441,17 @@ comfyui-workflow-templates is not installed.
             logging.info("Falling back to the default frontend.")
             check_frontend_version()
             return cls.default_frontend_path()
+    @classmethod
+    def template_asset_handler(cls):
+        assets = cls.template_asset_map()
+        if not assets:
+            return None
+
+        async def serve_template(request: web.Request) -> web.StreamResponse:
+            rel_path = request.match_info.get("path", "")
+            target = assets.get(rel_path)
+            if target is None:
+                raise web.HTTPNotFound()
+            return web.FileResponse(target)
+
+        return serve_template

app/model_manager.py

@@ -44,7 +44,7 @@ class ModelFileManager:
         @routes.get("/experiment/models/{folder}")
         async def get_all_models(request):
             folder = request.match_info.get("folder", None)
-            if not folder in folder_paths.folder_names_and_paths:
+            if folder not in folder_paths.folder_names_and_paths:
                 return web.Response(status=404)
             files = self.get_model_file_list(folder)
             return web.json_response(files)
@@ -55,7 +55,7 @@ class ModelFileManager:
             path_index = int(request.match_info.get("path_index", None))
             filename = request.match_info.get("filename", None)
 
-            if not folder_name in folder_paths.folder_names_and_paths:
+            if folder_name not in folder_paths.folder_names_and_paths:
                 return web.Response(status=404)
 
             folders = folder_paths.folder_names_and_paths[folder_name]

app/subgraph_manager.py

@@ -0,0 +1,112 @@
+from __future__ import annotations
+
+from typing import TypedDict
+import os
+import folder_paths
+import glob
+from aiohttp import web
+import hashlib
+
+
+class Source:
+    custom_node = "custom_node"
+
+class SubgraphEntry(TypedDict):
+    source: str
+    """
+    Source of subgraph - custom_nodes vs templates.
+    """
+    path: str
+    """
+    Relative path of the subgraph file.
+    For custom nodes, will be the relative directory like <custom_node_dir>/subgraphs/<name>.json
+    """
+    name: str
+    """
+    Name of subgraph file.
+    """
+    info: CustomNodeSubgraphEntryInfo
+    """
+    Additional info about subgraph; in the case of custom_nodes, will contain nodepack name
+    """
+    data: str
+
+class CustomNodeSubgraphEntryInfo(TypedDict):
+    node_pack: str
+    """Node pack name."""
+
+class SubgraphManager:
+    def __init__(self):
+        self.cached_custom_node_subgraphs: dict[SubgraphEntry] | None = None
+
+    async def load_entry_data(self, entry: SubgraphEntry):
+        with open(entry['path'], 'r') as f:
+            entry['data'] = f.read()
+        return entry
+
+    async def sanitize_entry(self, entry: SubgraphEntry | None, remove_data=False) -> SubgraphEntry | None:
+        if entry is None:
+            return None
+        entry = entry.copy()
+        entry.pop('path', None)
+        if remove_data:
+            entry.pop('data', None)
+        return entry
+
+    async def sanitize_entries(self, entries: dict[str, SubgraphEntry], remove_data=False) -> dict[str, SubgraphEntry]:
+        entries = entries.copy()
+        for key in list(entries.keys()):
+            entries[key] = await self.sanitize_entry(entries[key], remove_data)
+        return entries
+
+    async def get_custom_node_subgraphs(self, loadedModules, force_reload=False):
+        # if not forced to reload and cached, return cache
+        if not force_reload and self.cached_custom_node_subgraphs is not None:
+            return self.cached_custom_node_subgraphs
+        # Load subgraphs from custom nodes
+        subfolder = "subgraphs"
+        subgraphs_dict: dict[SubgraphEntry] = {}
+
+        for folder in folder_paths.get_folder_paths("custom_nodes"):
+            pattern = os.path.join(folder, f"*/{subfolder}/*.json")
+            matched_files = glob.glob(pattern)
+            for file in matched_files:
+                # replace backslashes with forward slashes
+                file = file.replace('\\', '/')
+                info: CustomNodeSubgraphEntryInfo = {
+                    "node_pack": "custom_nodes." + file.split('/')[-3]
+                }
+                source = Source.custom_node
+                # hash source + path to make sure id will be as unique as possible, but
+                # reproducible across backend reloads
+                id = hashlib.sha256(f"{source}{file}".encode()).hexdigest()
+                entry: SubgraphEntry = {
+                    "source": Source.custom_node,
+                    "name": os.path.splitext(os.path.basename(file))[0],
+                    "path": file,
+                    "info": info,
+                }
+                subgraphs_dict[id] = entry
+        self.cached_custom_node_subgraphs = subgraphs_dict
+        return subgraphs_dict
+
+    async def get_custom_node_subgraph(self, id: str, loadedModules):
+        subgraphs = await self.get_custom_node_subgraphs(loadedModules)
+        entry: SubgraphEntry = subgraphs.get(id, None)
+        if entry is not None and entry.get('data', None) is None:
+            await self.load_entry_data(entry)
+        return entry
+
+    def add_routes(self, routes, loadedModules):
+        @routes.get("/global_subgraphs")
+        async def get_global_subgraphs(request):
+            subgraphs_dict = await self.get_custom_node_subgraphs(loadedModules)
+            # NOTE: we may want to include other sources of global subgraphs such as templates in the future;
+            # that's the reasoning for the current implementation
+            return web.json_response(await self.sanitize_entries(subgraphs_dict, remove_data=True))
+
+        @routes.get("/global_subgraphs/{id}")
+        async def get_global_subgraph(request):
+            id = request.match_info.get("id", None)
+            subgraph = await self.get_custom_node_subgraph(id, loadedModules)
+            return web.json_response(await self.sanitize_entry(subgraph))

app/user_manager.py

@@ -59,6 +59,9 @@ class UserManager():
         user = "default"
         if args.multi_user and "comfy-user" in request.headers:
             user = request.headers["comfy-user"]
+            # Block System Users (use same error message to prevent probing)
+            if user.startswith(folder_paths.SYSTEM_USER_PREFIX):
+                raise KeyError("Unknown user: " + user)
 
         if user not in self.users:
             raise KeyError("Unknown user: " + user)
@@ -66,15 +69,16 @@ class UserManager():
         return user
 
     def get_request_user_filepath(self, request, file, type="userdata", create_dir=True):
-        user_directory = folder_paths.get_user_directory()
-
         if type == "userdata":
-            root_dir = user_directory
+            root_dir = folder_paths.get_user_directory()
         else:
             raise KeyError("Unknown filepath type:" + type)
 
         user = self.get_request_user_id(request)
-        path = user_root = os.path.abspath(os.path.join(root_dir, user))
+        user_root = folder_paths.get_public_user_directory(user)
+        if user_root is None:
+            return None
+        path = user_root
 
         # prevent leaving /{type}
         if os.path.commonpath((root_dir, user_root)) != root_dir:
@@ -101,7 +105,11 @@ class UserManager():
         name = name.strip()
         if not name:
             raise ValueError("username not provided")
+        if name.startswith(folder_paths.SYSTEM_USER_PREFIX):
+            raise ValueError("System User prefix not allowed")
         user_id = re.sub("[^a-zA-Z0-9-_]+", '-', name)
+        if user_id.startswith(folder_paths.SYSTEM_USER_PREFIX):
+            raise ValueError("System User prefix not allowed")
         user_id = user_id + "_" + str(uuid.uuid4())
 
         self.users[user_id] = name
@@ -132,7 +140,10 @@ class UserManager():
             if username in self.users.values():
                 return web.json_response({"error": "Duplicate username."}, status=400)
 
-            user_id = self.add_user(username)
+            try:
+                user_id = self.add_user(username)
+            except ValueError as e:
+                return web.json_response({"error": str(e)}, status=400)
             return web.json_response(user_id)
 
         @routes.get("/userdata")
@@ -424,7 +435,7 @@ class UserManager():
                 return source
 
             dest = get_user_data_path(request, check_exists=False, param="dest")
-            if not isinstance(source, str):
+            if not isinstance(dest, str):
                 return dest
 
             overwrite = request.query.get("overwrite", 'true') != "false"

comfy/audio_encoders/audio_encoders.py

@@ -0,0 +1,91 @@
+from .wav2vec2 import Wav2Vec2Model
+from .whisper import WhisperLargeV3
+import comfy.model_management
+import comfy.ops
+import comfy.utils
+import logging
+import torchaudio
+
+
+class AudioEncoderModel():
+    def __init__(self, config):
+        self.load_device = comfy.model_management.text_encoder_device()
+        offload_device = comfy.model_management.text_encoder_offload_device()
+        self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
+        model_type = config.pop("model_type")
+        model_config = dict(config)
+        model_config.update({
+            "dtype": self.dtype,
+            "device": offload_device,
+            "operations": comfy.ops.manual_cast
+        })
+
+        if model_type == "wav2vec2":
+            self.model = Wav2Vec2Model(**model_config)
+        elif model_type == "whisper3":
+            self.model = WhisperLargeV3(**model_config)
+        self.model.eval()
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.model_sample_rate = 16000
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def encode_audio(self, audio, sample_rate):
+        comfy.model_management.load_model_gpu(self.patcher)
+        audio = torchaudio.functional.resample(audio, sample_rate, self.model_sample_rate)
+        out, all_layers = self.model(audio.to(self.load_device))
+        outputs = {}
+        outputs["encoded_audio"] = out
+        outputs["encoded_audio_all_layers"] = all_layers
+        outputs["audio_samples"] = audio.shape[2]
+        return outputs
+
+
+def load_audio_encoder_from_sd(sd, prefix=""):
+    sd = comfy.utils.state_dict_prefix_replace(sd, {"wav2vec2.": ""})
+    if "encoder.layer_norm.bias" in sd: #wav2vec2
+        embed_dim = sd["encoder.layer_norm.bias"].shape[0]
+        if embed_dim == 1024:# large
+            config = {
+                "model_type": "wav2vec2",
+                "embed_dim": 1024,
+                "num_heads": 16,
+                "num_layers": 24,
+                "conv_norm": True,
+                "conv_bias": True,
+                "do_normalize": True,
+                "do_stable_layer_norm": True
+                }
+        elif embed_dim == 768: # base
+            config = {
+                "model_type": "wav2vec2",
+                "embed_dim": 768,
+                "num_heads": 12,
+                "num_layers": 12,
+                "conv_norm": False,
+                "conv_bias": False,
+                "do_normalize": False, # chinese-wav2vec2-base has this False
+                "do_stable_layer_norm": False
+            }
+        else:
+            raise RuntimeError("ERROR: audio encoder file is invalid or unsupported embed_dim: {}".format(embed_dim))
+    elif "model.encoder.embed_positions.weight" in sd:
+        sd = comfy.utils.state_dict_prefix_replace(sd, {"model.": ""})
+        config = {
+            "model_type": "whisper3",
+        }
+    else:
+        raise RuntimeError("ERROR: audio encoder not supported.")
+
+    audio_encoder = AudioEncoderModel(config)
+    m, u = audio_encoder.load_sd(sd)
+    if len(m) > 0:
+        logging.warning("missing audio encoder: {}".format(m))
+    if len(u) > 0:
+        logging.warning("unexpected audio encoder: {}".format(u))
+
+    return audio_encoder

comfy/audio_encoders/wav2vec2.py

@@ -0,0 +1,252 @@
+import torch
+import torch.nn as nn
+from comfy.ldm.modules.attention import optimized_attention_masked
+
+
+class LayerNormConv(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv = operations.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, bias=bias, device=device, dtype=dtype)
+        self.layer_norm = operations.LayerNorm(out_channels, elementwise_affine=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return torch.nn.functional.gelu(self.layer_norm(x.transpose(-2, -1)).transpose(-2, -1))
+
+class LayerGroupNormConv(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv = operations.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, bias=bias, device=device, dtype=dtype)
+        self.layer_norm = operations.GroupNorm(num_groups=out_channels, num_channels=out_channels, affine=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return torch.nn.functional.gelu(self.layer_norm(x))
+
+class ConvNoNorm(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv = operations.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, bias=bias, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return torch.nn.functional.gelu(x)
+
+
+class ConvFeatureEncoder(nn.Module):
+    def __init__(self, conv_dim, conv_bias=False, conv_norm=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        if conv_norm:
+            self.conv_layers = nn.ModuleList([
+                LayerNormConv(1, conv_dim, kernel_size=10, stride=5, bias=True, device=device, dtype=dtype, operations=operations),
+                LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                LayerNormConv(conv_dim, conv_dim, kernel_size=2, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                LayerNormConv(conv_dim, conv_dim, kernel_size=2, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+            ])
+        else:
+            self.conv_layers = nn.ModuleList([
+                LayerGroupNormConv(1, conv_dim, kernel_size=10, stride=5, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                ConvNoNorm(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                ConvNoNorm(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                ConvNoNorm(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                ConvNoNorm(conv_dim, conv_dim, kernel_size=3, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                ConvNoNorm(conv_dim, conv_dim, kernel_size=2, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+                ConvNoNorm(conv_dim, conv_dim, kernel_size=2, stride=2, bias=conv_bias, device=device, dtype=dtype, operations=operations),
+            ])
+
+    def forward(self, x):
+        x = x.unsqueeze(1)
+
+        for conv in self.conv_layers:
+            x = conv(x)
+
+        return x.transpose(1, 2)
+
+
+class FeatureProjection(nn.Module):
+    def __init__(self, conv_dim, embed_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.layer_norm = operations.LayerNorm(conv_dim, eps=1e-05, device=device, dtype=dtype)
+        self.projection = operations.Linear(conv_dim, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.layer_norm(x)
+        x = self.projection(x)
+        return x
+
+
+class PositionalConvEmbedding(nn.Module):
+    def __init__(self, embed_dim=768, kernel_size=128, groups=16):
+        super().__init__()
+        self.conv = nn.Conv1d(
+            embed_dim,
+            embed_dim,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            groups=groups,
+        )
+        self.conv = torch.nn.utils.parametrizations.weight_norm(self.conv, name="weight", dim=2)
+        self.activation = nn.GELU()
+
+    def forward(self, x):
+        x = x.transpose(1, 2)
+        x = self.conv(x)[:, :, :-1]
+        x = self.activation(x)
+        x = x.transpose(1, 2)
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim=768,
+        num_heads=12,
+        num_layers=12,
+        mlp_ratio=4.0,
+        do_stable_layer_norm=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.pos_conv_embed = PositionalConvEmbedding(embed_dim=embed_dim)
+        self.layers = nn.ModuleList([
+            TransformerEncoderLayer(
+                embed_dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                do_stable_layer_norm=do_stable_layer_norm,
+                device=device, dtype=dtype, operations=operations
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.layer_norm = operations.LayerNorm(embed_dim, eps=1e-05, device=device, dtype=dtype)
+        self.do_stable_layer_norm = do_stable_layer_norm
+
+    def forward(self, x, mask=None):
+        x = x + self.pos_conv_embed(x)
+        all_x = ()
+        if not self.do_stable_layer_norm:
+            x = self.layer_norm(x)
+        for layer in self.layers:
+            all_x += (x,)
+            x = layer(x, mask)
+        if self.do_stable_layer_norm:
+            x = self.layer_norm(x)
+        all_x += (x,)
+        return x, all_x
+
+
+class Attention(nn.Module):
+    def __init__(self, embed_dim, num_heads, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+
+        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        assert (mask is None)  # TODO?
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        out = optimized_attention_masked(q, k, v, self.num_heads)
+        return self.out_proj(out)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, embed_dim, mlp_ratio, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.intermediate_dense = operations.Linear(embed_dim, int(embed_dim * mlp_ratio), device=device, dtype=dtype)
+        self.output_dense = operations.Linear(int(embed_dim * mlp_ratio), embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.intermediate_dense(x)
+        x = torch.nn.functional.gelu(x)
+        x = self.output_dense(x)
+        return x
+
+
+class TransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        embed_dim=768,
+        num_heads=12,
+        mlp_ratio=4.0,
+        do_stable_layer_norm=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.attention = Attention(embed_dim, num_heads, device=device, dtype=dtype, operations=operations)
+
+        self.layer_norm = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+        self.feed_forward = FeedForward(embed_dim, mlp_ratio, device=device, dtype=dtype, operations=operations)
+        self.final_layer_norm = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+        self.do_stable_layer_norm = do_stable_layer_norm
+
+    def forward(self, x, mask=None):
+        residual = x
+        if self.do_stable_layer_norm:
+            x = self.layer_norm(x)
+        x = self.attention(x, mask=mask)
+        x = residual + x
+        if not self.do_stable_layer_norm:
+            x = self.layer_norm(x)
+            return self.final_layer_norm(x + self.feed_forward(x))
+        else:
+            return x + self.feed_forward(self.final_layer_norm(x))
+
+
+class Wav2Vec2Model(nn.Module):
+    """Complete Wav2Vec 2.0 model."""
+
+    def __init__(
+        self,
+        embed_dim=1024,
+        final_dim=256,
+        num_heads=16,
+        num_layers=24,
+        conv_norm=True,
+        conv_bias=True,
+        do_normalize=True,
+        do_stable_layer_norm=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        conv_dim = 512
+        self.feature_extractor = ConvFeatureEncoder(conv_dim, conv_norm=conv_norm, conv_bias=conv_bias, device=device, dtype=dtype, operations=operations)
+        self.feature_projection = FeatureProjection(conv_dim, embed_dim, device=device, dtype=dtype, operations=operations)
+
+        self.masked_spec_embed = nn.Parameter(torch.empty(embed_dim, device=device, dtype=dtype))
+        self.do_normalize = do_normalize
+
+        self.encoder = TransformerEncoder(
+            embed_dim=embed_dim,
+            num_heads=num_heads,
+            num_layers=num_layers,
+            do_stable_layer_norm=do_stable_layer_norm,
+            device=device, dtype=dtype, operations=operations
+        )
+
+    def forward(self, x, mask_time_indices=None, return_dict=False):
+        x = torch.mean(x, dim=1)
+
+        if self.do_normalize:
+            x = (x - x.mean()) / torch.sqrt(x.var() + 1e-7)
+
+        features = self.feature_extractor(x)
+        features = self.feature_projection(features)
+        batch_size, seq_len, _ = features.shape
+
+        x, all_x = self.encoder(features)
+        return x, all_x

comfy/audio_encoders/whisper.py

@@ -0,0 +1,186 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchaudio
+from typing import Optional
+from comfy.ldm.modules.attention import optimized_attention_masked
+import comfy.ops
+
+class WhisperFeatureExtractor(nn.Module):
+    def __init__(self, n_mels=128, device=None):
+        super().__init__()
+        self.sample_rate = 16000
+        self.n_fft = 400
+        self.hop_length = 160
+        self.n_mels = n_mels
+        self.chunk_length = 30
+        self.n_samples = 480000
+
+        self.mel_spectrogram = torchaudio.transforms.MelSpectrogram(
+            sample_rate=self.sample_rate,
+            n_fft=self.n_fft,
+            hop_length=self.hop_length,
+            n_mels=self.n_mels,
+            f_min=0,
+            f_max=8000,
+            norm="slaney",
+            mel_scale="slaney",
+        ).to(device)
+
+    def __call__(self, audio):
+        audio = torch.mean(audio, dim=1)
+        batch_size = audio.shape[0]
+        processed_audio = []
+
+        for i in range(batch_size):
+            aud = audio[i]
+            if aud.shape[0] > self.n_samples:
+                aud = aud[:self.n_samples]
+            elif aud.shape[0] < self.n_samples:
+                aud = F.pad(aud, (0, self.n_samples - aud.shape[0]))
+            processed_audio.append(aud)
+
+        audio = torch.stack(processed_audio)
+
+        mel_spec = self.mel_spectrogram(audio.to(self.mel_spectrogram.spectrogram.window.device))[:, :, :-1].to(audio.device)
+
+        log_mel_spec = torch.clamp(mel_spec, min=1e-10).log10()
+        log_mel_spec = torch.maximum(log_mel_spec, log_mel_spec.max() - 8.0)
+        log_mel_spec = (log_mel_spec + 4.0) / 4.0
+
+        return log_mel_spec
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, d_model: int, n_heads: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert d_model % n_heads == 0
+
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.d_k = d_model // n_heads
+
+        self.q_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+        self.k_proj = operations.Linear(d_model, d_model, bias=False, dtype=dtype, device=device)
+        self.v_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+        self.out_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = query.shape
+
+        q = self.q_proj(query)
+        k = self.k_proj(key)
+        v = self.v_proj(value)
+
+        attn_output = optimized_attention_masked(q, k, v, self.n_heads, mask)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, d_model: int, n_heads: int, d_ff: int, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        self.self_attn = MultiHeadAttention(d_model, n_heads, dtype=dtype, device=device, operations=operations)
+        self.self_attn_layer_norm = operations.LayerNorm(d_model, dtype=dtype, device=device)
+
+        self.fc1 = operations.Linear(d_model, d_ff, dtype=dtype, device=device)
+        self.fc2 = operations.Linear(d_ff, d_model, dtype=dtype, device=device)
+        self.final_layer_norm = operations.LayerNorm(d_model, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        residual = x
+        x = self.self_attn_layer_norm(x)
+        x = self.self_attn(x, x, x, attention_mask)
+        x = residual + x
+
+        residual = x
+        x = self.final_layer_norm(x)
+        x = self.fc1(x)
+        x = F.gelu(x)
+        x = self.fc2(x)
+        x = residual + x
+
+        return x
+
+
+class AudioEncoder(nn.Module):
+    def __init__(
+        self,
+        n_mels: int = 128,
+        n_ctx: int = 1500,
+        n_state: int = 1280,
+        n_head: int = 20,
+        n_layer: int = 32,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+
+        self.conv1 = operations.Conv1d(n_mels, n_state, kernel_size=3, padding=1, dtype=dtype, device=device)
+        self.conv2 = operations.Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1, dtype=dtype, device=device)
+
+        self.embed_positions = operations.Embedding(n_ctx, n_state, dtype=dtype, device=device)
+
+        self.layers = nn.ModuleList([
+            EncoderLayer(n_state, n_head, n_state * 4, dtype=dtype, device=device, operations=operations)
+            for _ in range(n_layer)
+        ])
+
+        self.layer_norm = operations.LayerNorm(n_state, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = F.gelu(self.conv1(x))
+        x = F.gelu(self.conv2(x))
+
+        x = x.transpose(1, 2)
+
+        x = x + comfy.ops.cast_to_input(self.embed_positions.weight[:, :x.shape[1]], x)
+
+        all_x = ()
+        for layer in self.layers:
+            all_x += (x,)
+            x = layer(x)
+
+        x = self.layer_norm(x)
+        all_x += (x,)
+        return x, all_x
+
+
+class WhisperLargeV3(nn.Module):
+    def __init__(
+        self,
+        n_mels: int = 128,
+        n_audio_ctx: int = 1500,
+        n_audio_state: int = 1280,
+        n_audio_head: int = 20,
+        n_audio_layer: int = 32,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+
+        self.feature_extractor = WhisperFeatureExtractor(n_mels=n_mels, device=device)
+
+        self.encoder = AudioEncoder(
+            n_mels, n_audio_ctx, n_audio_state, n_audio_head, n_audio_layer,
+            dtype=dtype, device=device, operations=operations
+        )
+
+    def forward(self, audio):
+        mel = self.feature_extractor(audio)
+        x, all_x = self.encoder(mel)
+        return x, all_x

comfy/cldm/cldm.py

@@ -413,7 +413,8 @@ class ControlNet(nn.Module):
         out_middle = []
 
         if self.num_classes is not None:
-            assert y.shape[0] == x.shape[0]
+            if y is None:
+                raise ValueError("y is None, did you try using a controlnet for SDXL on SD1?")
             emb = emb + self.label_emb(y)
 
         h = x

comfy/cli_args.py

@@ -97,6 +97,13 @@ class LatentPreviewMethod(enum.Enum):
     Latent2RGB = "latent2rgb"
     TAESD = "taesd"
 
+    @classmethod
+    def from_string(cls, value: str):
+        for member in cls:
+            if member.value == value:
+                return member
+        return None
+
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
 parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")
@@ -105,6 +112,7 @@ cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
 cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
+cache_group.add_argument("--cache-ram", nargs='?', const=4.0, type=float, default=0, help="Use RAM pressure caching with the specified headroom threshold. If available RAM drops below the threhold the cache remove large items to free RAM. Default 4GB")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
@@ -120,6 +128,12 @@ upcast.add_argument("--force-upcast-attention", action="store_true", help="Force
 upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.")
 
 
+parser.add_argument("--enable-manager", action="store_true", help="Enable the ComfyUI-Manager feature.")
+manager_group = parser.add_mutually_exclusive_group()
+manager_group.add_argument("--disable-manager-ui", action="store_true", help="Disables only the ComfyUI-Manager UI and endpoints. Scheduled installations and similar background tasks will still operate.")
+manager_group.add_argument("--enable-manager-legacy-ui", action="store_true", help="Enables the legacy UI of ComfyUI-Manager")
+
+
 vram_group = parser.add_mutually_exclusive_group()
 vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).")
 vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.")
@@ -130,7 +144,8 @@ 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("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.")
+parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.")
 
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 
@@ -143,8 +158,11 @@ class PerformanceFeature(enum.Enum):
     Fp16Accumulation = "fp16_accumulation"
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
+    AutoTune = "autotune"
 
-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("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
+
+parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.")
 
 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.")
@@ -156,13 +174,14 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win
 parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
 parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
 parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.")
-parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
+parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes. Also prevents the frontend from communicating with the internet.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
 
 parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level')
 parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).")
 
+
 # The default built-in provider hosted under web/
 DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest"
 

comfy/clip_model.py

@@ -2,6 +2,25 @@ import torch
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.ops
 
+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)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        if crop:
+            scale = (size / min(image.shape[2], image.shape[3]))
+            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
+        else:
+            scale_size = (size, size)
+
+        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
+        h = (image.shape[2] - size)//2
+        w = (image.shape[3] - size)//2
+        image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+
 class CLIPAttention(torch.nn.Module):
     def __init__(self, embed_dim, heads, dtype, device, operations):
         super().__init__()
@@ -61,8 +80,12 @@ class CLIPEncoder(torch.nn.Module):
     def forward(self, x, mask=None, intermediate_output=None):
         optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)
 
+        all_intermediate = None
         if intermediate_output is not None:
-            if intermediate_output < 0:
+            if intermediate_output == "all":
+                all_intermediate = []
+                intermediate_output = None
+            elif intermediate_output < 0:
                 intermediate_output = len(self.layers) + intermediate_output
 
         intermediate = None
@@ -70,6 +93,12 @@ class CLIPEncoder(torch.nn.Module):
             x = l(x, mask, optimized_attention)
             if i == intermediate_output:
                 intermediate = x.clone()
+            if all_intermediate is not None:
+                all_intermediate.append(x.unsqueeze(1).clone())
+
+        if all_intermediate is not None:
+            intermediate = torch.cat(all_intermediate, dim=1)
+
         return x, intermediate
 
 class CLIPEmbeddings(torch.nn.Module):

comfy/clip_vision.py

@@ -1,6 +1,5 @@
 from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
 import os
-import torch
 import json
 import logging
 
@@ -17,24 +16,7 @@ class Output:
     def __setitem__(self, key, item):
         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)
-    if not (image.shape[2] == size and image.shape[3] == size):
-        if crop:
-            scale = (size / min(image.shape[2], image.shape[3]))
-            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
-        else:
-            scale_size = (size, size)
-
-        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
-        h = (image.shape[2] - size)//2
-        w = (image.shape[3] - size)//2
-        image = image[:,:,h:h+size,w:w+size]
-    image = torch.clip((255. * image), 0, 255).round() / 255.0
-    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+clip_preprocess = comfy.clip_model.clip_preprocess  # Prevent some stuff from breaking, TODO: remove eventually
 
 IMAGE_ENCODERS = {
     "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
@@ -50,7 +32,13 @@ class ClipVisionModel():
         self.image_size = config.get("image_size", 224)
         self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
         self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
-        model_class = IMAGE_ENCODERS.get(config.get("model_type", "clip_vision_model"))
+        model_type = config.get("model_type", "clip_vision_model")
+        model_class = IMAGE_ENCODERS.get(model_type)
+        if model_type == "siglip_vision_model":
+            self.return_all_hidden_states = True
+        else:
+            self.return_all_hidden_states = False
+
         self.load_device = comfy.model_management.text_encoder_device()
         offload_device = comfy.model_management.text_encoder_offload_device()
         self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
@@ -67,13 +55,19 @@ class ClipVisionModel():
 
     def encode_image(self, image, crop=True):
         comfy.model_management.load_model_gpu(self.patcher)
-        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
-        out = self.model(pixel_values=pixel_values, intermediate_output=-2)
+        pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
+        out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2)
 
         outputs = Output()
         outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
         outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
-        outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device())
+        if self.return_all_hidden_states:
+            all_hs = out[1].to(comfy.model_management.intermediate_device())
+            outputs["penultimate_hidden_states"] = all_hs[:, -2]
+            outputs["all_hidden_states"] = all_hs
+        else:
+            outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device())
+
         outputs["mm_projected"] = out[3]
         return outputs
 
@@ -124,8 +118,12 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336.json")
         else:
             json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
-    elif "embeddings.patch_embeddings.projection.weight" in sd:
+
+    # Dinov2
+    elif 'encoder.layer.39.layer_scale2.lambda1' in sd:
         json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_giant.json")
+    elif 'encoder.layer.23.layer_scale2.lambda1' in sd:
+        json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_large.json")
     else:
         return None
 

comfy/context_windows.py

@@ -51,32 +51,43 @@ class ContextHandlerABC(ABC):
 
 
 class IndexListContextWindow(ContextWindowABC):
-    def __init__(self, index_list: list[int], dim: int=0):
+    def __init__(self, index_list: list[int], dim: int=0, total_frames: int=0):
         self.index_list = index_list
         self.context_length = len(index_list)
         self.dim = dim
+        self.total_frames = total_frames
+        self.center_ratio = (min(index_list) + max(index_list)) / (2 * total_frames)
 
-    def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor:
+    def get_tensor(self, full: torch.Tensor, device=None, dim=None, retain_index_list=[]) -> torch.Tensor:
         if dim is None:
             dim = self.dim
         if dim == 0 and full.shape[dim] == 1:
             return full
-        idx = [slice(None)] * dim + [self.index_list]
-        return full[idx].to(device)
+        idx = tuple([slice(None)] * dim + [self.index_list])
+        window = full[idx]
+        if retain_index_list:
+            idx = tuple([slice(None)] * dim + [retain_index_list])
+            window[idx] = full[idx]
+        return window.to(device)
 
     def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor:
         if dim is None:
             dim = self.dim
-        idx = [slice(None)] * dim + [self.index_list]
+        idx = tuple([slice(None)] * dim + [self.index_list])
         full[idx] += to_add
         return full
 
+    def get_region_index(self, num_regions: int) -> int:
+        region_idx = int(self.center_ratio * num_regions)
+        return min(max(region_idx, 0), num_regions - 1)
+
 
 class IndexListCallbacks:
     EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows"
     COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
     EXECUTE_START = "execute_start"
     EXECUTE_CLEANUP = "execute_cleanup"
+    RESIZE_COND_ITEM = "resize_cond_item"
 
     def init_callbacks(self):
         return {}
@@ -94,7 +105,8 @@ class ContextFuseMethod:
 
 ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
 class IndexListContextHandler(ContextHandlerABC):
-    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, closed_loop=False, dim=0):
+    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1,
+                 closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False):
         self.context_schedule = context_schedule
         self.fuse_method = fuse_method
         self.context_length = context_length
@@ -103,13 +115,18 @@ class IndexListContextHandler(ContextHandlerABC):
         self.closed_loop = closed_loop
         self.dim = dim
         self._step = 0
+        self.freenoise = freenoise
+        self.cond_retain_index_list = [int(x.strip()) for x in cond_retain_index_list.split(",")] if cond_retain_index_list else []
+        self.split_conds_to_windows = split_conds_to_windows
 
         self.callbacks = {}
 
     def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
         # for now, assume first dim is batch - should have stored on BaseModel in actual implementation
         if x_in.size(self.dim) > self.context_length:
-            logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.")
+            logging.info(f"Using context windows {self.context_length} with overlap {self.context_overlap} for {x_in.size(self.dim)} frames.")
+            if self.cond_retain_index_list:
+                logging.info(f"Retaining original cond for indexes: {self.cond_retain_index_list}")
             return True
         return False
 
@@ -123,6 +140,11 @@ class IndexListContextHandler(ContextHandlerABC):
             return None
         # reuse or resize cond items to match context requirements
         resized_cond = []
+        # if multiple conds, split based on primary region
+        if self.split_conds_to_windows and len(cond_in) > 1:
+            region = window.get_region_index(len(cond_in))
+            logging.info(f"Splitting conds to windows; using region {region} for window {window.index_list[0]}-{window.index_list[-1]} with center ratio {window.center_ratio:.3f}")
+            cond_in = [cond_in[region]]
         # cond object is a list containing a dict - outer list is irrelevant, so just loop through it
         for actual_cond in cond_in:
             resized_actual_cond = actual_cond.copy()
@@ -145,13 +167,38 @@ class IndexListContextHandler(ContextHandlerABC):
                         new_cond_item = cond_item.copy()
                         # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
                         for cond_key, cond_value in new_cond_item.items():
+                            # Allow callbacks to handle custom conditioning items
+                            handled = False
+                            for callback in comfy.patcher_extension.get_all_callbacks(
+                                IndexListCallbacks.RESIZE_COND_ITEM, self.callbacks
+                            ):
+                                result = callback(cond_key, cond_value, window, x_in, device, new_cond_item)
+                                if result is not None:
+                                    new_cond_item[cond_key] = result
+                                    handled = True
+                                    break
+                            if handled:
+                                continue
                             if isinstance(cond_value, torch.Tensor):
-                                if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim):
+                                if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
+                                   (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):
                                     new_cond_item[cond_key] = window.get_tensor(cond_value, device)
+                            # Handle audio_embed (temporal dim is 1)
+                            elif cond_key == "audio_embed" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                audio_cond = cond_value.cond
+                                if audio_cond.ndim > 1 and audio_cond.size(1) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(audio_cond, device, dim=1))
+                            # Handle vace_context (temporal dim is 3)
+                            elif cond_key == "vace_context" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                vace_cond = cond_value.cond
+                                if vace_cond.ndim >= 4 and vace_cond.size(3) == x_in.size(self.dim):
+                                    sliced_vace = window.get_tensor(vace_cond, device, dim=3, retain_index_list=self.cond_retain_index_list)
+                                    new_cond_item[cond_key] = cond_value._copy_with(sliced_vace)
                             # if has cond that is a Tensor, check if needs to be subset
                             elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
-                                if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim):
-                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device))
+                                if  (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \
+                                    (cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim)):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device, retain_index_list=self.cond_retain_index_list))
                             elif cond_key == "num_video_frames": # for SVD
                                 new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond)
                                 new_cond_item[cond_key].cond = window.context_length
@@ -164,7 +211,7 @@ class IndexListContextHandler(ContextHandlerABC):
         return resized_cond
 
     def set_step(self, timestep: torch.Tensor, model_options: dict[str]):
-        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001)
+        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep[0], rtol=0.0001)
         matches = torch.nonzero(mask)
         if torch.numel(matches) == 0:
             raise Exception("No sample_sigmas matched current timestep; something went wrong.")
@@ -173,7 +220,7 @@ class IndexListContextHandler(ContextHandlerABC):
     def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:
         full_length = x_in.size(self.dim) # TODO: choose dim based on model
         context_windows = self.context_schedule.func(full_length, self, model_options)
-        context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows]
+        context_windows = [IndexListContextWindow(window, dim=self.dim, total_frames=full_length) for window in context_windows]
         return context_windows
 
     def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
@@ -250,8 +297,8 @@ class IndexListContextHandler(ContextHandlerABC):
                     prev_weight = (bias_total / (bias_total + bias))
                     new_weight = (bias / (bias_total + bias))
                     # account for dims of tensors
-                    idx_window = [slice(None)] * self.dim + [idx]
-                    pos_window = [slice(None)] * self.dim + [pos]
+                    idx_window = tuple([slice(None)] * self.dim + [idx])
+                    pos_window = tuple([slice(None)] * self.dim + [pos])
                     # apply new values
                     conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight
                     biases_final[i][idx] = bias_total + bias
@@ -287,6 +334,28 @@ def create_prepare_sampling_wrapper(model: ModelPatcher):
     )
 
 
+def _sampler_sample_wrapper(executor, guider, sigmas, extra_args, callback, noise, *args, **kwargs):
+    model_options = extra_args.get("model_options", None)
+    if model_options is None:
+        raise Exception("model_options not found in sampler_sample_wrapper; this should never happen, something went wrong.")
+    handler: IndexListContextHandler = model_options.get("context_handler", None)
+    if handler is None:
+        raise Exception("context_handler not found in sampler_sample_wrapper; this should never happen, something went wrong.")
+    if not handler.freenoise:
+        return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs)
+    noise = apply_freenoise(noise, handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"])
+
+    return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs)
+
+
+def create_sampler_sample_wrapper(model: ModelPatcher):
+    model.add_wrapper_with_key(
+        comfy.patcher_extension.WrappersMP.SAMPLER_SAMPLE,
+        "ContextWindows_sampler_sample",
+        _sampler_sample_wrapper
+    )
+
+
 def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor:
     total_dims = len(x_in.shape)
     weights_tensor = torch.Tensor(weights).to(device=device)
@@ -538,3 +607,29 @@ def shift_window_to_end(window: list[int], num_frames: int):
     for i in range(len(window)):
         # 2) add end_delta to each val to slide windows to end
         window[i] = window[i] + end_delta
+
+
+# https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved/blob/90fb1331201a4b29488089e4fbffc0d82cc6d0a9/animatediff/sample_settings.py#L465
+def apply_freenoise(noise: torch.Tensor, dim: int, context_length: int, context_overlap: int, seed: int):
+    logging.info("Context windows: Applying FreeNoise")
+    generator = torch.Generator(device='cpu').manual_seed(seed)
+    latent_video_length = noise.shape[dim]
+    delta = context_length - context_overlap
+
+    for start_idx in range(0, latent_video_length - context_length, delta):
+        place_idx = start_idx + context_length
+
+        actual_delta = min(delta, latent_video_length - place_idx)
+        if actual_delta <= 0:
+            break
+
+        list_idx = torch.randperm(actual_delta, generator=generator, device='cpu') + start_idx
+
+        source_slice = [slice(None)] * noise.ndim
+        source_slice[dim] = list_idx
+        target_slice = [slice(None)] * noise.ndim
+        target_slice[dim] = slice(place_idx, place_idx + actual_delta)
+
+        noise[tuple(target_slice)] = noise[tuple(source_slice)]
+
+    return noise

comfy/controlnet.py

@@ -253,7 +253,10 @@ class ControlNet(ControlBase):
                 to_concat = []
                 for c in self.extra_concat_orig:
                     c = c.to(self.cond_hint.device)
-                    c = comfy.utils.common_upscale(c, self.cond_hint.shape[3], self.cond_hint.shape[2], self.upscale_algorithm, "center")
+                    c = comfy.utils.common_upscale(c, self.cond_hint.shape[-1], self.cond_hint.shape[-2], self.upscale_algorithm, "center")
+                    if c.ndim < self.cond_hint.ndim:
+                        c = c.unsqueeze(2)
+                        c = comfy.utils.repeat_to_batch_size(c, self.cond_hint.shape[2], dim=2)
                     to_concat.append(comfy.utils.repeat_to_batch_size(c, self.cond_hint.shape[0]))
                 self.cond_hint = torch.cat([self.cond_hint] + to_concat, dim=1)
 
@@ -307,11 +310,13 @@ class ControlLoraOps:
             self.bias = None
 
         def forward(self, input):
-            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
             if self.up is not None:
-                return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
+                x = torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
             else:
-                return torch.nn.functional.linear(input, weight, bias)
+                x = torch.nn.functional.linear(input, weight, bias)
+            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
     class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp):
         def __init__(
@@ -347,12 +352,13 @@ class ControlLoraOps:
 
 
         def forward(self, input):
-            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
             if self.up is not None:
-                return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
+                x = torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
             else:
-                return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
-
+                x = torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
+            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
 class ControlLora(ControlNet):
     def __init__(self, control_weights, global_average_pooling=False, model_options={}): #TODO? model_options
@@ -585,11 +591,18 @@ def load_controlnet_flux_instantx(sd, model_options={}):
 
 def load_controlnet_qwen_instantx(sd, model_options={}):
     model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options)
-    control_model = comfy.ldm.qwen_image.controlnet.QwenImageControlNetModel(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_latent_channels = sd.get("controlnet_x_embedder.weight").shape[1]
+
+    extra_condition_channels = 0
+    concat_mask = False
+    if control_latent_channels == 68: #inpaint controlnet
+        extra_condition_channels = control_latent_channels - 64
+        concat_mask = True
+    control_model = comfy.ldm.qwen_image.controlnet.QwenImageControlNetModel(extra_condition_channels=extra_condition_channels, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
     control_model = controlnet_load_state_dict(control_model, sd)
     latent_format = comfy.latent_formats.Wan21()
     extra_conds = []
-    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
+    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
     return control
 
 def convert_mistoline(sd):

comfy/hooks.py

@@ -527,7 +527,8 @@ class HookKeyframeGroup:
                         if self._current_keyframe.get_effective_guarantee_steps(max_sigma) > 0:
                             break
                     # if eval_c is outside the percent range, stop looking further
-                    else: break
+                    else:
+                        break
         # update steps current context is used
         self._current_used_steps += 1
         # update current timestep this was performed on

comfy/image_encoders/dino2.py

@@ -31,6 +31,20 @@ class LayerScale(torch.nn.Module):
     def forward(self, x):
         return x * comfy.model_management.cast_to_device(self.lambda1, x.device, x.dtype)
 
+class Dinov2MLP(torch.nn.Module):
+    def __init__(self, hidden_size: int, dtype, device, operations):
+        super().__init__()
+
+        mlp_ratio = 4
+        hidden_features = int(hidden_size * mlp_ratio)
+        self.fc1 = operations.Linear(hidden_size, hidden_features, bias = True, device=device, dtype=dtype)
+        self.fc2 = operations.Linear(hidden_features, hidden_size, bias = True, device=device, dtype=dtype)
+
+    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
+        hidden_state = self.fc1(hidden_state)
+        hidden_state = torch.nn.functional.gelu(hidden_state)
+        hidden_state = self.fc2(hidden_state)
+        return hidden_state
 
 class SwiGLUFFN(torch.nn.Module):
     def __init__(self, dim, dtype, device, operations):
@@ -50,12 +64,15 @@ class SwiGLUFFN(torch.nn.Module):
 
 
 class Dino2Block(torch.nn.Module):
-    def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations):
+    def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations, use_swiglu_ffn):
         super().__init__()
         self.attention = Dino2AttentionBlock(dim, num_heads, layer_norm_eps, dtype, device, operations)
         self.layer_scale1 = LayerScale(dim, dtype, device, operations)
         self.layer_scale2 = LayerScale(dim, dtype, device, operations)
-        self.mlp = SwiGLUFFN(dim, dtype, device, operations)
+        if use_swiglu_ffn:
+            self.mlp = SwiGLUFFN(dim, dtype, device, operations)
+        else:
+            self.mlp = Dinov2MLP(dim, dtype, device, operations)
         self.norm1 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
         self.norm2 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
 
@@ -66,9 +83,10 @@ class Dino2Block(torch.nn.Module):
 
 
 class Dino2Encoder(torch.nn.Module):
-    def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations):
+    def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations, use_swiglu_ffn):
         super().__init__()
-        self.layer = torch.nn.ModuleList([Dino2Block(dim, num_heads, layer_norm_eps, dtype, device, operations) for _ in range(num_layers)])
+        self.layer = torch.nn.ModuleList([Dino2Block(dim, num_heads, layer_norm_eps, dtype, device, operations, use_swiglu_ffn = use_swiglu_ffn)
+                                          for _ in range(num_layers)])
 
     def forward(self, x, intermediate_output=None):
         optimized_attention = optimized_attention_for_device(x.device, False, small_input=True)
@@ -78,8 +96,8 @@ class Dino2Encoder(torch.nn.Module):
                 intermediate_output = len(self.layer) + intermediate_output
 
         intermediate = None
-        for i, l in enumerate(self.layer):
-            x = l(x, optimized_attention)
+        for i, layer in enumerate(self.layer):
+            x = layer(x, optimized_attention)
             if i == intermediate_output:
                 intermediate = x.clone()
         return x, intermediate
@@ -128,9 +146,10 @@ class Dinov2Model(torch.nn.Module):
         dim = config_dict["hidden_size"]
         heads = config_dict["num_attention_heads"]
         layer_norm_eps = config_dict["layer_norm_eps"]
+        use_swiglu_ffn = config_dict["use_swiglu_ffn"]
 
         self.embeddings = Dino2Embeddings(dim, dtype, device, operations)
-        self.encoder = Dino2Encoder(dim, heads, layer_norm_eps, num_layers, dtype, device, operations)
+        self.encoder = Dino2Encoder(dim, heads, layer_norm_eps, num_layers, dtype, device, operations, use_swiglu_ffn = use_swiglu_ffn)
         self.layernorm = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
 
     def forward(self, pixel_values, attention_mask=None, intermediate_output=None):

comfy/image_encoders/dino2_large.json

@@ -0,0 +1,22 @@
+{
+  "hidden_size": 1024,
+  "use_mask_token": true,
+  "patch_size": 14,
+  "image_size": 518,
+  "num_channels": 3,
+  "num_attention_heads": 16,
+  "initializer_range": 0.02,
+  "attention_probs_dropout_prob": 0.0,
+  "hidden_dropout_prob": 0.0,
+  "hidden_act": "gelu",
+  "mlp_ratio": 4,
+  "model_type": "dinov2",
+  "num_hidden_layers": 24,
+  "layer_norm_eps": 1e-6,
+  "qkv_bias": true,
+  "use_swiglu_ffn": false,
+  "layerscale_value": 1.0,
+  "drop_path_rate": 0.0,
+  "image_mean": [0.485, 0.456, 0.406],
+  "image_std": [0.229, 0.224, 0.225]
+}

comfy/k_diffusion/sampling.py

@@ -74,6 +74,9 @@ def get_ancestral_step(sigma_from, sigma_to, eta=1.):
 
 def default_noise_sampler(x, seed=None):
     if seed is not None:
+        if x.device == torch.device("cpu"):
+            seed += 1
+
         generator = torch.Generator(device=x.device)
         generator.manual_seed(seed)
     else:
@@ -86,24 +89,24 @@ class BatchedBrownianTree:
     """A wrapper around torchsde.BrownianTree that enables batches of entropy."""
 
     def __init__(self, x, t0, t1, seed=None, **kwargs):
-        self.cpu_tree = True
-        if "cpu" in kwargs:
-            self.cpu_tree = kwargs.pop("cpu")
+        self.cpu_tree = kwargs.pop("cpu", True)
         t0, t1, self.sign = self.sort(t0, t1)
-        w0 = kwargs.get('w0', torch.zeros_like(x))
+        w0 = kwargs.pop('w0', None)
+        if w0 is None:
+            w0 = torch.zeros_like(x)
+        self.batched = False
         if seed is None:
-            seed = torch.randint(0, 2 ** 63 - 1, []).item()
-        self.batched = True
-        try:
-            assert len(seed) == x.shape[0]
+            seed = (torch.randint(0, 2 ** 63 - 1, ()).item(),)
+        elif isinstance(seed, (tuple, list)):
+            if len(seed) != x.shape[0]:
+                raise ValueError("Passing a list or tuple of seeds to BatchedBrownianTree requires a length matching the batch size.")
+            self.batched = True
             w0 = w0[0]
-        except TypeError:
-            seed = [seed]
-            self.batched = False
-        if self.cpu_tree:
-            self.trees = [torchsde.BrownianTree(t0.cpu(), w0.cpu(), t1.cpu(), entropy=s, **kwargs) for s in seed]
         else:
-            self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
+            seed = (seed,)
+        if self.cpu_tree:
+            t0, w0, t1 = t0.detach().cpu(), w0.detach().cpu(), t1.detach().cpu()
+        self.trees = tuple(torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed)
 
     @staticmethod
     def sort(a, b):
@@ -111,11 +114,10 @@ class BatchedBrownianTree:
 
     def __call__(self, t0, t1):
         t0, t1, sign = self.sort(t0, t1)
+        device, dtype = t0.device, t0.dtype
         if self.cpu_tree:
-            w = torch.stack([tree(t0.cpu().float(), t1.cpu().float()).to(t0.dtype).to(t0.device) for tree in self.trees]) * (self.sign * sign)
-        else:
-            w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
-
+            t0, t1 = t0.detach().cpu().float(), t1.detach().cpu().float()
+        w = torch.stack([tree(t0, t1) for tree in self.trees]).to(device=device, dtype=dtype) * (self.sign * sign)
         return w if self.batched else w[0]
 
 
@@ -171,6 +173,16 @@ def offset_first_sigma_for_snr(sigmas, model_sampling, percent_offset=1e-4):
     return sigmas
 
 
+def ei_h_phi_1(h: torch.Tensor) -> torch.Tensor:
+    """Compute the result of h*phi_1(h) in exponential integrator methods."""
+    return torch.expm1(h)
+
+
+def ei_h_phi_2(h: torch.Tensor) -> torch.Tensor:
+    """Compute the result of h*phi_2(h) in exponential integrator methods."""
+    return (torch.expm1(h) - h) / h
+
+
 @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)."""
@@ -853,6 +865,11 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     return x
 
 
+@torch.no_grad()
+def sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
+    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_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """DPM-Solver++(3M) SDE."""
@@ -925,6 +942,16 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     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_heun_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
+    if len(sigmas) <= 1:
+        return x
+    extra_args = {} if extra_args is None else extra_args
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    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_heun(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_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:
@@ -1533,15 +1560,17 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
 
 
 @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):
+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, solver_type="phi_1"):
     """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
-    arXiv: https://arxiv.org/abs/2305.14267
+    arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023)
     """
+    if solver_type not in {"phi_1", "phi_2"}:
+        raise ValueError("solver_type must be 'phi_1' or 'phi_2'")
+
     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')
@@ -1549,55 +1578,70 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
     lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
     sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
 
+    fac = 1 / (2 * r)
+
     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 * h
-            fac = 1 / (2 * r)
-            sigma_s_1 = sigma_fn(lambda_s_1)
+            continue
 
-            # 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()
+        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 = torch.lerp(lambda_s, lambda_t, r)
+        sigma_s_1 = sigma_fn(lambda_s_1)
 
-            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 = (-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])
+        alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+        alpha_t = sigmas[i + 1] * lambda_t.exp()
 
-            # Step 1
-            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 1
+        x_2 = sigma_s_1 / sigmas[i] * (-r * h * eta).exp() * x - alpha_s_1 * ei_h_phi_1(-r * h_eta) * denoised
+        if inject_noise:
+            sde_noise = (-2 * r * h * eta).expm1().neg().sqrt() * noise_sampler(sigmas[i], sigma_s_1)
+            x_2 = x_2 + sde_noise * sigma_s_1 * s_noise
+        denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
-            # Step 2
-            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
+        # Step 2
+        if solver_type == "phi_1":
+            denoised_d = torch.lerp(denoised, denoised_2, fac)
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        elif solver_type == "phi_2":
+            b2 = ei_h_phi_2(-h_eta) / r
+            b1 = ei_h_phi_1(-h_eta) - b2
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2)
+
+        if inject_noise:
+            segment_factor = (r - 1) * h * eta
+            sde_noise = sde_noise * segment_factor.exp()
+            sde_noise = sde_noise + segment_factor.mul(2).expm1().neg().sqrt() * noise_sampler(sigma_s_1, sigmas[i + 1])
+            x = x + sde_noise * sigmas[i + 1] * s_noise
     return x
 
+@torch.no_grad()
+def sample_exp_heun_2_x0(model, x, sigmas, extra_args=None, callback=None, disable=None, solver_type="phi_2"):
+    """Deterministic exponential Heun second order method in data prediction (x0) and logSNR time."""
+    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None, r=1.0, solver_type=solver_type)
+
+
+@torch.no_grad()
+def sample_exp_heun_2_x0_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type="phi_2"):
+    """Stochastic exponential Heun second order method in data prediction (x0) and logSNR time."""
+    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=1.0, solver_type=solver_type)
+
 
 @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
+    arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023)
     """
     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')
@@ -1609,45 +1653,49 @@ def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=Non
         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)
+            continue
 
-            # 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()
+        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 = torch.lerp(lambda_s, lambda_t, r_1)
+        lambda_s_2 = torch.lerp(lambda_s, lambda_t, r_2)
+        sigma_s_1, sigma_s_2 = sigma_fn(lambda_s_1), sigma_fn(lambda_s_2)
 
-            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])
+        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()
 
-            # 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 1
+        x_2 = sigma_s_1 / sigmas[i] * (-r_1 * h * eta).exp() * x - alpha_s_1 * ei_h_phi_1(-r_1 * h_eta) * denoised
+        if inject_noise:
+            sde_noise = (-2 * r_1 * h * eta).expm1().neg().sqrt() * noise_sampler(sigmas[i], sigma_s_1)
+            x_2 = x_2 + sde_noise * sigma_s_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 2
+        a3_2 = r_2 / r_1 * ei_h_phi_2(-r_2 * h_eta)
+        a3_1 = ei_h_phi_1(-r_2 * h_eta) - a3_2
+        x_3 = sigma_s_2 / sigmas[i] * (-r_2 * h * eta).exp() * x - alpha_s_2 * (a3_1 * denoised + a3_2 * denoised_2)
+        if inject_noise:
+            segment_factor = (r_1 - r_2) * h * eta
+            sde_noise = sde_noise * segment_factor.exp()
+            sde_noise = sde_noise + segment_factor.mul(2).expm1().neg().sqrt() * noise_sampler(sigma_s_1, sigma_s_2)
+            x_3 = x_3 + sde_noise * sigma_s_2 * s_noise
+        denoised_3 = model(x_3, sigma_s_2 * s_in, **extra_args)
 
-            # Step 3
-            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
+        # Step 3
+        b3 = ei_h_phi_2(-h_eta) / r_2
+        b1 = ei_h_phi_1(-h_eta) - b3
+        x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b3 * denoised_3)
+        if inject_noise:
+            segment_factor = (r_2 - 1) * h * eta
+            sde_noise = sde_noise * segment_factor.exp()
+            sde_noise = sde_noise + segment_factor.mul(2).expm1().neg().sqrt() * noise_sampler(sigma_s_2, sigmas[i + 1])
+            x = x + sde_noise * sigmas[i + 1] * s_noise
     return x
 
 
@@ -1731,7 +1779,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
         # Predictor
         if sigmas[i + 1] == 0:
             # Denoising step
-            x = denoised
+            x_pred = denoised
         else:
             tau_t = tau_func(sigmas[i + 1])
             curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
@@ -1752,7 +1800,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
             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
+    return x_pred
 
 
 @torch.no_grad()

comfy/latent_formats.py

@@ -6,6 +6,7 @@ class LatentFormat:
     latent_dimensions = 2
     latent_rgb_factors = None
     latent_rgb_factors_bias = None
+    latent_rgb_factors_reshape = None
     taesd_decoder_name = None
 
     def process_in(self, latent):
@@ -178,6 +179,54 @@ class Flux(SD3):
     def process_out(self, latent):
         return (latent / self.scale_factor) + self.shift_factor
 
+class Flux2(LatentFormat):
+    latent_channels = 128
+
+    def __init__(self):
+        self.latent_rgb_factors =[
+            [0.0058, 0.0113, 0.0073],
+            [0.0495, 0.0443, 0.0836],
+            [-0.0099, 0.0096, 0.0644],
+            [0.2144, 0.3009, 0.3652],
+            [0.0166, -0.0039, -0.0054],
+            [0.0157, 0.0103, -0.0160],
+            [-0.0398, 0.0902, -0.0235],
+            [-0.0052, 0.0095, 0.0109],
+            [-0.3527, -0.2712, -0.1666],
+            [-0.0301, -0.0356, -0.0180],
+            [-0.0107, 0.0078, 0.0013],
+            [0.0746, 0.0090, -0.0941],
+            [0.0156, 0.0169, 0.0070],
+            [-0.0034, -0.0040, -0.0114],
+            [0.0032, 0.0181, 0.0080],
+            [-0.0939, -0.0008, 0.0186],
+            [0.0018, 0.0043, 0.0104],
+            [0.0284, 0.0056, -0.0127],
+            [-0.0024, -0.0022, -0.0030],
+            [0.1207, -0.0026, 0.0065],
+            [0.0128, 0.0101, 0.0142],
+            [0.0137, -0.0072, -0.0007],
+            [0.0095, 0.0092, -0.0059],
+            [0.0000, -0.0077, -0.0049],
+            [-0.0465, -0.0204, -0.0312],
+            [0.0095, 0.0012, -0.0066],
+            [0.0290, -0.0034, 0.0025],
+            [0.0220, 0.0169, -0.0048],
+            [-0.0332, -0.0457, -0.0468],
+            [-0.0085, 0.0389, 0.0609],
+            [-0.0076, 0.0003, -0.0043],
+            [-0.0111, -0.0460, -0.0614],
+        ]
+
+        self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851]
+        self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2)
+
+    def process_in(self, latent):
+        return latent
+
+    def process_out(self, latent):
+        return latent
+
 class Mochi(LatentFormat):
     latent_channels = 12
     latent_dimensions = 3
@@ -358,6 +407,11 @@ class LTXV(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512]
 
+class LTXAV(LTXV):
+    def __init__(self):
+        self.latent_rgb_factors = None
+        self.latent_rgb_factors_bias = None
+
 class HunyuanVideo(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3
@@ -382,6 +436,7 @@ class HunyuanVideo(LatentFormat):
     ]
 
     latent_rgb_factors_bias = [ 0.0259, -0.0192, -0.0761]
+    taesd_decoder_name = "taehv"
 
 class Cosmos1CV8x8x8(LatentFormat):
     latent_channels = 16
@@ -445,7 +500,7 @@ class Wan21(LatentFormat):
         ]).view(1, self.latent_channels, 1, 1, 1)
 
 
-        self.taesd_decoder_name = None #TODO
+        self.taesd_decoder_name = "lighttaew2_1"
 
     def process_in(self, latent):
         latents_mean = self.latents_mean.to(latent.device, latent.dtype)
@@ -516,6 +571,7 @@ class Wan22(Wan21):
 
     def __init__(self):
         self.scale_factor = 1.0
+        self.taesd_decoder_name = "lighttaew2_2"
         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,
@@ -533,11 +589,155 @@ class Wan22(Wan21):
                 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 HunyuanImage21(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 2
+    scale_factor = 0.75289
+
+    latent_rgb_factors = [
+        [-0.0154, -0.0397, -0.0521],
+        [ 0.0005,  0.0093,  0.0006],
+        [-0.0805, -0.0773, -0.0586],
+        [-0.0494, -0.0487, -0.0498],
+        [-0.0212, -0.0076, -0.0261],
+        [-0.0179, -0.0417, -0.0505],
+        [ 0.0158,  0.0310,  0.0239],
+        [ 0.0409,  0.0516,  0.0201],
+        [ 0.0350,  0.0553,  0.0036],
+        [-0.0447, -0.0327, -0.0479],
+        [-0.0038, -0.0221, -0.0365],
+        [-0.0423, -0.0718, -0.0654],
+        [ 0.0039,  0.0368,  0.0104],
+        [ 0.0655,  0.0217,  0.0122],
+        [ 0.0490,  0.1638,  0.2053],
+        [ 0.0932,  0.0829,  0.0650],
+        [-0.0186, -0.0209, -0.0135],
+        [-0.0080, -0.0076, -0.0148],
+        [-0.0284, -0.0201,  0.0011],
+        [-0.0642, -0.0294, -0.0777],
+        [-0.0035,  0.0076, -0.0140],
+        [ 0.0519,  0.0731,  0.0887],
+        [-0.0102,  0.0095,  0.0704],
+        [ 0.0068,  0.0218, -0.0023],
+        [-0.0726, -0.0486, -0.0519],
+        [ 0.0260,  0.0295,  0.0263],
+        [ 0.0250,  0.0333,  0.0341],
+        [ 0.0168, -0.0120, -0.0174],
+        [ 0.0226,  0.1037,  0.0114],
+        [ 0.2577,  0.1906,  0.1604],
+        [-0.0646, -0.0137, -0.0018],
+        [-0.0112,  0.0309,  0.0358],
+        [-0.0347,  0.0146, -0.0481],
+        [ 0.0234,  0.0179,  0.0201],
+        [ 0.0157,  0.0313,  0.0225],
+        [ 0.0423,  0.0675,  0.0524],
+        [-0.0031,  0.0027, -0.0255],
+        [ 0.0447,  0.0555,  0.0330],
+        [-0.0152,  0.0103,  0.0299],
+        [-0.0755, -0.0489, -0.0635],
+        [ 0.0853,  0.0788,  0.1017],
+        [-0.0272, -0.0294, -0.0471],
+        [ 0.0440,  0.0400, -0.0137],
+        [ 0.0335,  0.0317, -0.0036],
+        [-0.0344, -0.0621, -0.0984],
+        [-0.0127, -0.0630, -0.0620],
+        [-0.0648,  0.0360,  0.0924],
+        [-0.0781, -0.0801, -0.0409],
+        [ 0.0363,  0.0613,  0.0499],
+        [ 0.0238,  0.0034,  0.0041],
+        [-0.0135,  0.0258,  0.0310],
+        [ 0.0614,  0.1086,  0.0589],
+        [ 0.0428,  0.0350,  0.0205],
+        [ 0.0153,  0.0173, -0.0018],
+        [-0.0288, -0.0455, -0.0091],
+        [ 0.0344,  0.0109, -0.0157],
+        [-0.0205, -0.0247, -0.0187],
+        [ 0.0487,  0.0126,  0.0064],
+        [-0.0220, -0.0013,  0.0074],
+        [-0.0203, -0.0094, -0.0048],
+        [-0.0719,  0.0429, -0.0442],
+        [ 0.1042,  0.0497,  0.0356],
+        [-0.0659, -0.0578, -0.0280],
+        [-0.0060, -0.0322, -0.0234]]
+
+    latent_rgb_factors_bias = [0.0007, -0.0256, -0.0206]
+
+class HunyuanImage21Refiner(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 3
+    scale_factor = 1.03682
+
+    def process_in(self, latent):
+        out = latent * self.scale_factor
+        out = torch.cat((out[:, :, :1], out), dim=2)
+        out = out.permute(0, 2, 1, 3, 4)
+        b, f_times_2, c, h, w = out.shape
+        out = out.reshape(b, f_times_2 // 2, 2 * c, h, w)
+        out = out.permute(0, 2, 1, 3, 4).contiguous()
+        return out
+
+    def process_out(self, latent):
+        z = latent / self.scale_factor
+        z = z.permute(0, 2, 1, 3, 4)
+        b, f, c, h, w = z.shape
+        z = z.reshape(b, f, 2, c // 2, h, w)
+        z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w)
+        z = z.permute(0, 2, 1, 3, 4)
+        z = z[:, :, 1:]
+        return z
+
+class HunyuanVideo15(LatentFormat):
+    latent_rgb_factors = [
+        [ 0.0568, -0.0521, -0.0131],
+        [ 0.0014,  0.0735,  0.0326],
+        [ 0.0186,  0.0531, -0.0138],
+        [-0.0031,  0.0051,  0.0288],
+        [ 0.0110,  0.0556,  0.0432],
+        [-0.0041, -0.0023, -0.0485],
+        [ 0.0530,  0.0413,  0.0253],
+        [ 0.0283,  0.0251,  0.0339],
+        [ 0.0277, -0.0372, -0.0093],
+        [ 0.0393,  0.0944,  0.1131],
+        [ 0.0020,  0.0251,  0.0037],
+        [-0.0017,  0.0012,  0.0234],
+        [ 0.0468,  0.0436,  0.0203],
+        [ 0.0354,  0.0439, -0.0233],
+        [ 0.0090,  0.0123,  0.0346],
+        [ 0.0382,  0.0029,  0.0217],
+        [ 0.0261, -0.0300,  0.0030],
+        [-0.0088, -0.0220, -0.0283],
+        [-0.0272, -0.0121, -0.0363],
+        [-0.0664, -0.0622,  0.0144],
+        [ 0.0414,  0.0479,  0.0529],
+        [ 0.0355,  0.0612, -0.0247],
+        [ 0.0147,  0.0264,  0.0174],
+        [ 0.0438,  0.0038,  0.0542],
+        [ 0.0431, -0.0573, -0.0033],
+        [-0.0162, -0.0211, -0.0406],
+        [-0.0487, -0.0295, -0.0393],
+        [ 0.0005, -0.0109,  0.0253],
+        [ 0.0296,  0.0591,  0.0353],
+        [ 0.0119,  0.0181, -0.0306],
+        [-0.0085, -0.0362,  0.0229],
+        [ 0.0005, -0.0106,  0.0242]
+    ]
+
+    latent_rgb_factors_bias = [ 0.0456, -0.0202, -0.0644]
+    latent_channels = 32
+    latent_dimensions = 3
+    scale_factor = 1.03682
+    taesd_decoder_name = "lighttaehy1_5"
+
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1
     scale_factor = 0.9990943042622529
 
+class Hunyuan3Dv2_1(LatentFormat):
+    scale_factor = 1.0039506158752403
+    latent_channels = 64
+    latent_dimensions = 1
+
 class Hunyuan3Dv2mini(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1
@@ -546,3 +746,20 @@ class Hunyuan3Dv2mini(LatentFormat):
 class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
+
+class ChromaRadiance(LatentFormat):
+    latent_channels = 3
+
+    def __init__(self):
+        self.latent_rgb_factors = [
+            # R    G    B
+            [ 1.0, 0.0, 0.0 ],
+            [ 0.0, 1.0, 0.0 ],
+            [ 0.0, 0.0, 1.0 ]
+        ]
+
+    def process_in(self, latent):
+        return latent
+
+    def process_out(self, latent):
+        return latent

comfy/ldm/ace/attention.py

@@ -133,6 +133,7 @@ class Attention(nn.Module):
         hidden_states: torch.Tensor,
         encoder_hidden_states: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
+        transformer_options={},
         **cross_attention_kwargs,
     ) -> torch.Tensor:
         return self.processor(
@@ -140,6 +141,7 @@ class Attention(nn.Module):
             hidden_states,
             encoder_hidden_states=encoder_hidden_states,
             attention_mask=attention_mask,
+            transformer_options=transformer_options,
             **cross_attention_kwargs,
         )
 
@@ -366,6 +368,7 @@ class CustomerAttnProcessor2_0:
         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,
+        transformer_options={},
         *args,
         **kwargs,
     ) -> torch.Tensor:
@@ -433,7 +436,7 @@ class CustomerAttnProcessor2_0:
 
         # 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,
+            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True, transformer_options=transformer_options,
         ).to(query.dtype)
 
         # linear proj
@@ -697,6 +700,7 @@ class LinearTransformerBlock(nn.Module):
         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,
+        transformer_options={},
     ):
 
         N = hidden_states.shape[0]
@@ -720,6 +724,7 @@ class LinearTransformerBlock(nn.Module):
                 encoder_attention_mask=encoder_attention_mask,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+                transformer_options=transformer_options,
             )
         else:
             attn_output, _ = self.attn(
@@ -729,6 +734,7 @@ class LinearTransformerBlock(nn.Module):
                 encoder_attention_mask=None,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=None,
+                transformer_options=transformer_options,
             )
 
         if self.use_adaln_single:
@@ -743,6 +749,7 @@ class LinearTransformerBlock(nn.Module):
                 encoder_attention_mask=encoder_attention_mask,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+                transformer_options=transformer_options,
             )
             hidden_states = attn_output + hidden_states
 

comfy/ldm/ace/model.py

@@ -314,6 +314,7 @@ class ACEStepTransformer2DModel(nn.Module):
         output_length: int = 0,
         block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
         controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        transformer_options={},
     ):
         embedded_timestep = self.timestep_embedder(self.time_proj(timestep).to(dtype=hidden_states.dtype))
         temb = self.t_block(embedded_timestep)
@@ -339,6 +340,7 @@ class ACEStepTransformer2DModel(nn.Module):
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=encoder_rotary_freqs_cis,
                 temb=temb,
+                transformer_options=transformer_options,
             )
 
         output = self.final_layer(hidden_states, embedded_timestep, output_length)
@@ -393,6 +395,7 @@ class ACEStepTransformer2DModel(nn.Module):
 
         output_length = hidden_states.shape[-1]
 
+        transformer_options = kwargs.get("transformer_options", {})
         output = self.decode(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
@@ -402,6 +405,7 @@ class ACEStepTransformer2DModel(nn.Module):
             output_length=output_length,
             block_controlnet_hidden_states=block_controlnet_hidden_states,
             controlnet_scale=controlnet_scale,
+            transformer_options=transformer_options,
         )
 
         return output

comfy/ldm/ace/vae/music_dcae_pipeline.py

@@ -23,8 +23,6 @@ class MusicDCAE(torch.nn.Module):
         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),
         ])
@@ -37,10 +35,6 @@ class MusicDCAE(torch.nn.Module):
         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)):
@@ -73,10 +67,8 @@ class MusicDCAE(torch.nn.Module):
             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):
@@ -91,9 +83,7 @@ class MusicDCAE(torch.nn.Module):
             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)
@@ -101,7 +91,6 @@ class MusicDCAE(torch.nn.Module):
         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)

comfy/ldm/audio/dit.py

@@ -298,7 +298,8 @@ class Attention(nn.Module):
         mask = None,
         context_mask = None,
         rotary_pos_emb = None,
-        causal = None
+        causal = None,
+        transformer_options={},
     ):
         h, kv_h, has_context = self.num_heads, self.kv_heads, context is not None
 
@@ -363,7 +364,7 @@ class Attention(nn.Module):
             heads_per_kv_head = h // kv_h
             k, v = map(lambda t: t.repeat_interleave(heads_per_kv_head, dim = 1), (k, v))
 
-        out = optimized_attention(q, k, v, h, skip_reshape=True)
+        out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options)
         out = self.to_out(out)
 
         if mask is not None:
@@ -488,7 +489,8 @@ class TransformerBlock(nn.Module):
         global_cond=None,
         mask = None,
         context_mask = None,
-        rotary_pos_emb = None
+        rotary_pos_emb = None,
+        transformer_options={}
     ):
         if self.global_cond_dim is not None and self.global_cond_dim > 0 and global_cond is not None:
 
@@ -498,12 +500,12 @@ class TransformerBlock(nn.Module):
             residual = x
             x = self.pre_norm(x)
             x = x * (1 + scale_self) + shift_self
-            x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb)
+            x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb, transformer_options=transformer_options)
             x = x * torch.sigmoid(1 - gate_self)
             x = x + residual
 
             if context is not None:
-                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
+                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask, transformer_options=transformer_options)
 
             if self.conformer is not None:
                 x = x + self.conformer(x)
@@ -517,10 +519,10 @@ class TransformerBlock(nn.Module):
             x = x + residual
 
         else:
-            x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb)
+            x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb, transformer_options=transformer_options)
 
             if context is not None:
-                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
+                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask, transformer_options=transformer_options)
 
             if self.conformer is not None:
                 x = x + self.conformer(x)
@@ -606,7 +608,8 @@ class ContinuousTransformer(nn.Module):
         return_info = False,
         **kwargs
     ):
-        patches_replace = kwargs.get("transformer_options", {}).get("patches_replace", {})
+        transformer_options = kwargs.get("transformer_options", {})
+        patches_replace = transformer_options.get("patches_replace", {})
         batch, seq, device = *x.shape[:2], x.device
         context = kwargs["context"]
 
@@ -632,7 +635,7 @@ class ContinuousTransformer(nn.Module):
         # Attention layers
 
         if self.rotary_pos_emb is not None:
-            rotary_pos_emb = self.rotary_pos_emb.forward_from_seq_len(x.shape[1], dtype=x.dtype, device=x.device)
+            rotary_pos_emb = self.rotary_pos_emb.forward_from_seq_len(x.shape[1], dtype=torch.float, device=x.device)
         else:
             rotary_pos_emb = None
 
@@ -645,13 +648,13 @@ class ContinuousTransformer(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = layer(args["img"], rotary_pos_emb=args["pe"], global_cond=args["vec"], context=args["txt"])
+                    out["img"] = layer(args["img"], rotary_pos_emb=args["pe"], global_cond=args["vec"], context=args["txt"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": global_cond, "pe": rotary_pos_emb}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": global_cond, "pe": rotary_pos_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, context=context)
+                x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, context=context, transformer_options=transformer_options)
             # x = checkpoint(layer, x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, **kwargs)
 
             if return_info:

comfy/ldm/aura/mmdit.py

@@ -85,7 +85,7 @@ class SingleAttention(nn.Module):
         )
 
     #@torch.compile()
-    def forward(self, c):
+    def forward(self, c, transformer_options={}):
 
         bsz, seqlen1, _ = c.shape
 
@@ -95,7 +95,7 @@ class SingleAttention(nn.Module):
         v = v.view(bsz, seqlen1, self.n_heads, self.head_dim)
         q, k = self.q_norm1(q), self.k_norm1(k)
 
-        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
+        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True, transformer_options=transformer_options)
         c = self.w1o(output)
         return c
 
@@ -144,7 +144,7 @@ class DoubleAttention(nn.Module):
 
 
     #@torch.compile()
-    def forward(self, c, x):
+    def forward(self, c, x, transformer_options={}):
 
         bsz, seqlen1, _ = c.shape
         bsz, seqlen2, _ = x.shape
@@ -168,7 +168,7 @@ class DoubleAttention(nn.Module):
             torch.cat([cv, xv], dim=1),
         )
 
-        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
+        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True, transformer_options=transformer_options)
 
         c, x = output.split([seqlen1, seqlen2], dim=1)
         c = self.w1o(c)
@@ -207,7 +207,7 @@ class MMDiTBlock(nn.Module):
         self.is_last = is_last
 
     #@torch.compile()
-    def forward(self, c, x, global_cond, **kwargs):
+    def forward(self, c, x, global_cond, transformer_options={}, **kwargs):
 
         cres, xres = c, x
 
@@ -225,7 +225,7 @@ class MMDiTBlock(nn.Module):
         x = modulate(self.normX1(x), xshift_msa, xscale_msa)
 
         # attention
-        c, x = self.attn(c, x)
+        c, x = self.attn(c, x, transformer_options=transformer_options)
 
 
         c = self.normC2(cres + cgate_msa.unsqueeze(1) * c)
@@ -255,13 +255,13 @@ class DiTBlock(nn.Module):
         self.mlp = MLP(dim, hidden_dim=dim * 4, dtype=dtype, device=device, operations=operations)
 
     #@torch.compile()
-    def forward(self, cx, global_cond, **kwargs):
+    def forward(self, cx, global_cond, transformer_options={}, **kwargs):
         cxres = cx
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.modCX(
             global_cond
         ).chunk(6, dim=1)
         cx = modulate(self.norm1(cx), shift_msa, scale_msa)
-        cx = self.attn(cx)
+        cx = self.attn(cx, transformer_options=transformer_options)
         cx = self.norm2(cxres + gate_msa.unsqueeze(1) * cx)
         mlpout = self.mlp(modulate(cx, shift_mlp, scale_mlp))
         cx = gate_mlp.unsqueeze(1) * mlpout
@@ -473,13 +473,14 @@ class MMDiT(nn.Module):
                         out = {}
                         out["txt"], out["img"] = layer(args["txt"],
                                                        args["img"],
-                                                       args["vec"])
+                                                       args["vec"],
+                                                       transformer_options=args["transformer_options"])
                         return out
-                    out = blocks_replace[("double_block", i)]({"img": x, "txt": c, "vec": global_cond}, {"original_block": block_wrap})
+                    out = blocks_replace[("double_block", i)]({"img": x, "txt": c, "vec": global_cond, "transformer_options": transformer_options}, {"original_block": block_wrap})
                     c = out["txt"]
                     x = out["img"]
                 else:
-                    c, x = layer(c, x, global_cond, **kwargs)
+                    c, x = layer(c, x, global_cond, transformer_options=transformer_options, **kwargs)
 
         if len(self.single_layers) > 0:
             c_len = c.size(1)
@@ -488,13 +489,13 @@ class MMDiT(nn.Module):
                 if ("single_block", i) in blocks_replace:
                     def block_wrap(args):
                         out = {}
-                        out["img"] = layer(args["img"], args["vec"])
+                        out["img"] = layer(args["img"], args["vec"], transformer_options=args["transformer_options"])
                         return out
 
-                    out = blocks_replace[("single_block", i)]({"img": cx, "vec": global_cond}, {"original_block": block_wrap})
+                    out = blocks_replace[("single_block", i)]({"img": cx, "vec": global_cond, "transformer_options": transformer_options}, {"original_block": block_wrap})
                     cx = out["img"]
                 else:
-                    cx = layer(cx, global_cond, **kwargs)
+                    cx = layer(cx, global_cond, transformer_options=transformer_options, **kwargs)
 
             x = cx[:, c_len:]
 

comfy/ldm/cascade/common.py

@@ -32,12 +32,12 @@ class OptimizedAttention(nn.Module):
 
         self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
 
-    def forward(self, q, k, v):
+    def forward(self, q, k, v, transformer_options={}):
         q = self.to_q(q)
         k = self.to_k(k)
         v = self.to_v(v)
 
-        out = optimized_attention(q, k, v, self.heads)
+        out = optimized_attention(q, k, v, self.heads, transformer_options=transformer_options)
 
         return self.out_proj(out)
 
@@ -47,13 +47,13 @@ class Attention2D(nn.Module):
         self.attn = OptimizedAttention(c, nhead, dtype=dtype, device=device, operations=operations)
         # self.attn = nn.MultiheadAttention(c, nhead, dropout=dropout, bias=True, batch_first=True, dtype=dtype, device=device)
 
-    def forward(self, x, kv, self_attn=False):
+    def forward(self, x, kv, self_attn=False, transformer_options={}):
         orig_shape = x.shape
         x = x.view(x.size(0), x.size(1), -1).permute(0, 2, 1)  # Bx4xHxW -> Bx(HxW)x4
         if self_attn:
             kv = torch.cat([x, kv], dim=1)
         # x = self.attn(x, kv, kv, need_weights=False)[0]
-        x = self.attn(x, kv, kv)
+        x = self.attn(x, kv, kv, transformer_options=transformer_options)
         x = x.permute(0, 2, 1).view(*orig_shape)
         return x
 
@@ -114,9 +114,9 @@ class AttnBlock(nn.Module):
             operations.Linear(c_cond, c, dtype=dtype, device=device)
         )
 
-    def forward(self, x, kv):
+    def forward(self, x, kv, transformer_options={}):
         kv = self.kv_mapper(kv)
-        x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn)
+        x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn, transformer_options=transformer_options)
         return x
 
 

comfy/ldm/cascade/stage_b.py

@@ -173,7 +173,7 @@ class StageB(nn.Module):
         clip = self.clip_norm(clip)
         return clip
 
-    def _down_encode(self, x, r_embed, clip):
+    def _down_encode(self, x, r_embed, clip, transformer_options={}):
         level_outputs = []
         block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
         for down_block, downscaler, repmap in block_group:
@@ -187,7 +187,7 @@ class StageB(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -199,7 +199,7 @@ class StageB(nn.Module):
             level_outputs.insert(0, x)
         return level_outputs
 
-    def _up_decode(self, level_outputs, r_embed, clip):
+    def _up_decode(self, level_outputs, r_embed, clip, transformer_options={}):
         x = level_outputs[0]
         block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
         for i, (up_block, upscaler, repmap) in enumerate(block_group):
@@ -216,7 +216,7 @@ class StageB(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -228,7 +228,7 @@ class StageB(nn.Module):
             x = upscaler(x)
         return x
 
-    def forward(self, x, r, effnet, clip, pixels=None, **kwargs):
+    def forward(self, x, r, effnet, clip, pixels=None, transformer_options={}, **kwargs):
         if pixels is None:
             pixels = x.new_zeros(x.size(0), 3, 8, 8)
 
@@ -245,8 +245,8 @@ class StageB(nn.Module):
             nn.functional.interpolate(effnet, size=x.shape[-2:], mode='bilinear', align_corners=True))
         x = x + nn.functional.interpolate(self.pixels_mapper(pixels), size=x.shape[-2:], mode='bilinear',
                                           align_corners=True)
-        level_outputs = self._down_encode(x, r_embed, clip)
-        x = self._up_decode(level_outputs, r_embed, clip)
+        level_outputs = self._down_encode(x, r_embed, clip, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, transformer_options=transformer_options)
         return self.clf(x)
 
     def update_weights_ema(self, src_model, beta=0.999):

comfy/ldm/cascade/stage_c.py

@@ -182,7 +182,7 @@ class StageC(nn.Module):
         clip = self.clip_norm(clip)
         return clip
 
-    def _down_encode(self, x, r_embed, clip, cnet=None):
+    def _down_encode(self, x, r_embed, clip, cnet=None, transformer_options={}):
         level_outputs = []
         block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
         for down_block, downscaler, repmap in block_group:
@@ -201,7 +201,7 @@ class StageC(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -213,7 +213,7 @@ class StageC(nn.Module):
             level_outputs.insert(0, x)
         return level_outputs
 
-    def _up_decode(self, level_outputs, r_embed, clip, cnet=None):
+    def _up_decode(self, level_outputs, r_embed, clip, cnet=None, transformer_options={}):
         x = level_outputs[0]
         block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
         for i, (up_block, upscaler, repmap) in enumerate(block_group):
@@ -235,7 +235,7 @@ class StageC(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -247,7 +247,7 @@ class StageC(nn.Module):
             x = upscaler(x)
         return x
 
-    def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, **kwargs):
+    def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, transformer_options={}, **kwargs):
         # Process the conditioning embeddings
         r_embed = self.gen_r_embedding(r).to(dtype=x.dtype)
         for c in self.t_conds:
@@ -262,8 +262,8 @@ class StageC(nn.Module):
 
         # Model Blocks
         x = self.embedding(x)
-        level_outputs = self._down_encode(x, r_embed, clip, cnet)
-        x = self._up_decode(level_outputs, r_embed, clip, cnet)
+        level_outputs = self._down_encode(x, r_embed, clip, cnet, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, cnet, transformer_options=transformer_options)
         return self.clf(x)
 
     def update_weights_ema(self, src_model, beta=0.999):

comfy/ldm/chroma/layers.py

@@ -1,15 +1,15 @@
 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,
 )
 
+# TODO: remove this in a few months
+SingleStreamBlock = None
+DoubleStreamBlock = None
 
 
 class ChromaModulationOut(ModulationOut):
@@ -48,124 +48,6 @@ class Approximator(nn.Module):
         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__()

comfy/ldm/chroma/model.py

@@ -11,12 +11,12 @@ import comfy.ldm.common_dit
 from comfy.ldm.flux.layers import (
     EmbedND,
     timestep_embedding,
+    DoubleStreamBlock,
+    SingleStreamBlock,
 )
 
 from .layers import (
-    DoubleStreamBlock,
     LastLayer,
-    SingleStreamBlock,
     Approximator,
     ChromaModulationOut,
 )
@@ -40,7 +40,8 @@ class ChromaParams:
     out_dim: int
     hidden_dim: int
     n_layers: int
-
+    txt_ids_dims: list
+    vec_in_dim: int
 
 
 
@@ -90,6 +91,7 @@ class Chroma(nn.Module):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
+                    modulation=False,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -98,7 +100,7 @@ class Chroma(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=False, dtype=dtype, device=device, operations=operations)
                 for _ in range(params.depth_single_blocks)
             ]
         )
@@ -151,8 +153,6 @@ class Chroma(nn.Module):
         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)
@@ -180,7 +180,10 @@ class Chroma(nn.Module):
         pe = self.pe_embedder(ids)
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if i not in self.skip_mmdit:
                 double_mod = (
                     self.get_modulations(mod_vectors, "double_img", idx=i),
@@ -193,14 +196,16 @@ class Chroma(nn.Module):
                                                        txt=args["txt"],
                                                        vec=args["vec"],
                                                        pe=args["pe"],
-                                                       attn_mask=args.get("attn_mask"))
+                                                       attn_mask=args.get("attn_mask"),
+                                                       transformer_options=args.get("transformer_options"))
                         return out
 
                     out = blocks_replace[("double_block", i)]({"img": img,
                                                                "txt": txt,
                                                                "vec": double_mod,
                                                                "pe": pe,
-                                                               "attn_mask": attn_mask},
+                                                               "attn_mask": attn_mask,
+                                                               "transformer_options": transformer_options},
                                                               {"original_block": block_wrap})
                     txt = out["txt"]
                     img = out["img"]
@@ -209,7 +214,8 @@ class Chroma(nn.Module):
                                      txt=txt,
                                      vec=double_mod,
                                      pe=pe,
-                                     attn_mask=attn_mask)
+                                     attn_mask=attn_mask,
+                                     transformer_options=transformer_options)
 
                 if control is not None: # Controlnet
                     control_i = control.get("input")
@@ -220,7 +226,10 @@ class Chroma(nn.Module):
 
         img = torch.cat((txt, img), 1)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if i not in self.skip_dit:
                 single_mod = self.get_modulations(mod_vectors, "single", idx=i)
                 if ("single_block", i) in blocks_replace:
@@ -229,17 +238,19 @@ class Chroma(nn.Module):
                         out["img"] = block(args["img"],
                                            vec=args["vec"],
                                            pe=args["pe"],
-                                           attn_mask=args.get("attn_mask"))
+                                           attn_mask=args.get("attn_mask"),
+                                           transformer_options=args.get("transformer_options"))
                         return out
 
                     out = blocks_replace[("single_block", i)]({"img": img,
                                                                "vec": single_mod,
                                                                "pe": pe,
-                                                               "attn_mask": attn_mask},
+                                                               "attn_mask": attn_mask,
+                                                               "transformer_options": transformer_options},
                                                               {"original_block": block_wrap})
                     img = out["img"]
                 else:
-                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask)
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
 
                 if control is not None: # Controlnet
                     control_o = control.get("output")
@@ -249,8 +260,9 @@ class Chroma(nn.Module):
                             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)
+        if hasattr(self, "final_layer"):
+            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):
@@ -266,6 +278,9 @@ class Chroma(nn.Module):
 
         img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=self.patch_size, pw=self.patch_size)
 
+        if img.ndim != 3 or context.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
         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)

comfy/ldm/chroma_radiance/layers.py

@@ -0,0 +1,206 @@
+# Adapted from https://github.com/lodestone-rock/flow
+from functools import lru_cache
+
+import torch
+from torch import nn
+
+from comfy.ldm.flux.layers import RMSNorm
+
+
+class NerfEmbedder(nn.Module):
+    """
+    An embedder module that combines input features with a 2D positional
+    encoding that mimics the Discrete Cosine Transform (DCT).
+
+    This module takes an input tensor of shape (B, P^2, C), where P is the
+    patch size, and enriches it with positional information before projecting
+    it to a new hidden size.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        hidden_size_input: int,
+        max_freqs: int,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        """
+        Initializes the NerfEmbedder.
+
+        Args:
+            in_channels (int): The number of channels in the input tensor.
+            hidden_size_input (int): The desired dimension of the output embedding.
+            max_freqs (int): The number of frequency components to use for both
+                             the x and y dimensions of the positional encoding.
+                             The total number of positional features will be max_freqs^2.
+        """
+        super().__init__()
+        self.dtype = dtype
+        self.max_freqs = max_freqs
+        self.hidden_size_input = hidden_size_input
+
+        # A linear layer to project the concatenated input features and
+        # positional encodings to the final output dimension.
+        self.embedder = nn.Sequential(
+            operations.Linear(in_channels + max_freqs**2, hidden_size_input, dtype=dtype, device=device)
+        )
+
+    @lru_cache(maxsize=4)
+    def fetch_pos(self, patch_size: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
+        """
+        Generates and caches 2D DCT-like positional embeddings for a given patch size.
+
+        The LRU cache is a performance optimization that avoids recomputing the
+        same positional grid on every forward pass.
+
+        Args:
+            patch_size (int): The side length of the square input patch.
+            device: The torch device to create the tensors on.
+            dtype: The torch dtype for the tensors.
+
+        Returns:
+            A tensor of shape (1, patch_size^2, max_freqs^2) containing the
+            positional embeddings.
+        """
+        # Create normalized 1D coordinate grids from 0 to 1.
+        pos_x = torch.linspace(0, 1, patch_size, device=device, dtype=dtype)
+        pos_y = torch.linspace(0, 1, patch_size, device=device, dtype=dtype)
+
+        # Create a 2D meshgrid of coordinates.
+        pos_y, pos_x = torch.meshgrid(pos_y, pos_x, indexing="ij")
+
+        # Reshape positions to be broadcastable with frequencies.
+        # Shape becomes (patch_size^2, 1, 1).
+        pos_x = pos_x.reshape(-1, 1, 1)
+        pos_y = pos_y.reshape(-1, 1, 1)
+
+        # Create a 1D tensor of frequency values from 0 to max_freqs-1.
+        freqs = torch.linspace(0, self.max_freqs - 1, self.max_freqs, dtype=dtype, device=device)
+
+        # Reshape frequencies to be broadcastable for creating 2D basis functions.
+        # freqs_x shape: (1, max_freqs, 1)
+        # freqs_y shape: (1, 1, max_freqs)
+        freqs_x = freqs[None, :, None]
+        freqs_y = freqs[None, None, :]
+
+        # A custom weighting coefficient, not part of standard DCT.
+        # This seems to down-weight the contribution of higher-frequency interactions.
+        coeffs = (1 + freqs_x * freqs_y) ** -1
+
+        # Calculate the 1D cosine basis functions for x and y coordinates.
+        # This is the core of the DCT formulation.
+        dct_x = torch.cos(pos_x * freqs_x * torch.pi)
+        dct_y = torch.cos(pos_y * freqs_y * torch.pi)
+
+        # Combine the 1D basis functions to create 2D basis functions by element-wise
+        # multiplication, and apply the custom coefficients. Broadcasting handles the
+        # combination of all (pos_x, freqs_x) with all (pos_y, freqs_y).
+        # The result is flattened into a feature vector for each position.
+        dct = (dct_x * dct_y * coeffs).view(1, -1, self.max_freqs ** 2)
+
+        return dct
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass for the embedder.
+
+        Args:
+            inputs (Tensor): The input tensor of shape (B, P^2, C).
+
+        Returns:
+            Tensor: The output tensor of shape (B, P^2, hidden_size_input).
+        """
+        # Get the batch size, number of pixels, and number of channels.
+        B, P2, C = inputs.shape
+
+        # Infer the patch side length from the number of pixels (P^2).
+        patch_size = int(P2 ** 0.5)
+
+        input_dtype = inputs.dtype
+        inputs = inputs.to(dtype=self.dtype)
+
+        # Fetch the pre-computed or cached positional embeddings.
+        dct = self.fetch_pos(patch_size, inputs.device, self.dtype)
+
+        # Repeat the positional embeddings for each item in the batch.
+        dct = dct.repeat(B, 1, 1)
+
+        # Concatenate the original input features with the positional embeddings
+        # along the feature dimension.
+        inputs = torch.cat((inputs, dct), dim=-1)
+
+        # Project the combined tensor to the target hidden size.
+        return self.embedder(inputs).to(dtype=input_dtype)
+
+
+class NerfGLUBlock(nn.Module):
+    """
+    A NerfBlock using a Gated Linear Unit (GLU) like MLP.
+    """
+    def __init__(self, hidden_size_s: int, hidden_size_x: int, mlp_ratio, dtype=None, device=None, operations=None):
+        super().__init__()
+        # The total number of parameters for the MLP is increased to accommodate
+        # the gate, value, and output projection matrices.
+        # We now need to generate parameters for 3 matrices.
+        total_params = 3 * hidden_size_x**2 * mlp_ratio
+        self.param_generator = operations.Linear(hidden_size_s, total_params, dtype=dtype, device=device)
+        self.norm = RMSNorm(hidden_size_x, dtype=dtype, device=device, operations=operations)
+        self.mlp_ratio = mlp_ratio
+
+
+    def forward(self, x: torch.Tensor, s: torch.Tensor) -> torch.Tensor:
+        batch_size, num_x, hidden_size_x = x.shape
+        mlp_params = self.param_generator(s)
+
+        # Split the generated parameters into three parts for the gate, value, and output projection.
+        fc1_gate_params, fc1_value_params, fc2_params = mlp_params.chunk(3, dim=-1)
+
+        # Reshape the parameters into matrices for batch matrix multiplication.
+        fc1_gate = fc1_gate_params.view(batch_size, hidden_size_x, hidden_size_x * self.mlp_ratio)
+        fc1_value = fc1_value_params.view(batch_size, hidden_size_x, hidden_size_x * self.mlp_ratio)
+        fc2 = fc2_params.view(batch_size, hidden_size_x * self.mlp_ratio, hidden_size_x)
+
+        # Normalize the generated weight matrices as in the original implementation.
+        fc1_gate = torch.nn.functional.normalize(fc1_gate, dim=-2)
+        fc1_value = torch.nn.functional.normalize(fc1_value, dim=-2)
+        fc2 = torch.nn.functional.normalize(fc2, dim=-2)
+
+        res_x = x
+        x = self.norm(x)
+
+        # Apply the final output projection.
+        x = torch.bmm(torch.nn.functional.silu(torch.bmm(x, fc1_gate)) * torch.bmm(x, fc1_value), fc2)
+
+        return x + res_x
+
+
+class NerfFinalLayer(nn.Module):
+    def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
+        self.linear = operations.Linear(hidden_size, out_channels, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # RMSNorm normalizes over the last dimension, but our channel dim (C) is at dim=1.
+        # So we temporarily move the channel dimension to the end for the norm operation.
+        return self.linear(self.norm(x.movedim(1, -1))).movedim(-1, 1)
+
+
+class NerfFinalLayerConv(nn.Module):
+    def __init__(self, hidden_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
+        self.conv = operations.Conv2d(
+            in_channels=hidden_size,
+            out_channels=out_channels,
+            kernel_size=3,
+            padding=1,
+            dtype=dtype,
+            device=device,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # RMSNorm normalizes over the last dimension, but our channel dim (C) is at dim=1.
+        # So we temporarily move the channel dimension to the end for the norm operation.
+        return self.conv(self.norm(x.movedim(1, -1)).movedim(-1, 1))

comfy/ldm/chroma_radiance/model.py

@@ -0,0 +1,335 @@
+# Credits:
+# Original Flux code can be found on: https://github.com/black-forest-labs/flux
+# Chroma Radiance adaption referenced from https://github.com/lodestone-rock/flow
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from torch import Tensor, nn
+from einops import repeat
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import EmbedND, DoubleStreamBlock, SingleStreamBlock
+
+from comfy.ldm.chroma.model import Chroma, ChromaParams
+from comfy.ldm.chroma.layers import (
+    Approximator,
+)
+from .layers import (
+    NerfEmbedder,
+    NerfGLUBlock,
+    NerfFinalLayer,
+    NerfFinalLayerConv,
+)
+
+
+@dataclass
+class ChromaRadianceParams(ChromaParams):
+    patch_size: int
+    nerf_hidden_size: int
+    nerf_mlp_ratio: int
+    nerf_depth: int
+    nerf_max_freqs: int
+    # Setting nerf_tile_size to 0 disables tiling.
+    nerf_tile_size: int
+    # Currently one of linear (legacy) or conv.
+    nerf_final_head_type: str
+    # None means use the same dtype as the model.
+    nerf_embedder_dtype: Optional[torch.dtype]
+    use_x0: bool
+
+class ChromaRadiance(Chroma):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        if operations is None:
+            raise RuntimeError("Attempt to create ChromaRadiance object without setting operations")
+        nn.Module.__init__(self)
+        self.dtype = dtype
+        params = ChromaRadianceParams(**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_patch = operations.Conv2d(
+            params.in_channels,
+            params.hidden_size,
+            kernel_size=params.patch_size,
+            stride=params.patch_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,
+                    modulation=False,
+                    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,
+                    modulation=False,
+                    dtype=dtype, device=device, operations=operations,
+                )
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        # pixel channel concat with DCT
+        self.nerf_image_embedder = NerfEmbedder(
+            in_channels=params.in_channels,
+            hidden_size_input=params.nerf_hidden_size,
+            max_freqs=params.nerf_max_freqs,
+            dtype=params.nerf_embedder_dtype or dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.nerf_blocks = nn.ModuleList([
+            NerfGLUBlock(
+                hidden_size_s=params.hidden_size,
+                hidden_size_x=params.nerf_hidden_size,
+                mlp_ratio=params.nerf_mlp_ratio,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            ) for _ in range(params.nerf_depth)
+        ])
+
+        if params.nerf_final_head_type == "linear":
+            self.nerf_final_layer = NerfFinalLayer(
+                params.nerf_hidden_size,
+                out_channels=params.in_channels,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+        elif params.nerf_final_head_type == "conv":
+            self.nerf_final_layer_conv = NerfFinalLayerConv(
+                params.nerf_hidden_size,
+                out_channels=params.in_channels,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+        else:
+            errstr = f"Unsupported nerf_final_head_type {params.nerf_final_head_type}"
+            raise ValueError(errstr)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+        if params.use_x0:
+            self.register_buffer("__x0__", torch.tensor([]))
+
+    @property
+    def _nerf_final_layer(self) -> nn.Module:
+        if self.params.nerf_final_head_type == "linear":
+            return self.nerf_final_layer
+        if self.params.nerf_final_head_type == "conv":
+            return self.nerf_final_layer_conv
+        # Impossible to get here as we raise an error on unexpected types on initialization.
+        raise NotImplementedError
+
+    def img_in(self, img: Tensor) -> Tensor:
+        img = self.img_in_patch(img) # -> [B, Hidden, H/P, W/P]
+        # flatten into a sequence for the transformer.
+        return img.flatten(2).transpose(1, 2) # -> [B, NumPatches, Hidden]
+
+    def forward_nerf(
+        self,
+        img_orig: Tensor,
+        img_out: Tensor,
+        params: ChromaRadianceParams,
+    ) -> Tensor:
+        B, C, H, W = img_orig.shape
+        num_patches = img_out.shape[1]
+        patch_size = params.patch_size
+
+        # Store the raw pixel values of each patch for the NeRF head later.
+        # unfold creates patches: [B, C * P * P, NumPatches]
+        nerf_pixels = nn.functional.unfold(img_orig, kernel_size=patch_size, stride=patch_size)
+        nerf_pixels = nerf_pixels.transpose(1, 2) # -> [B, NumPatches, C * P * P]
+
+        # Reshape for per-patch processing
+        nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
+        nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)
+
+        if params.nerf_tile_size > 0 and num_patches > params.nerf_tile_size:
+            # Enable tiling if nerf_tile_size isn't 0 and we actually have more patches than
+            # the tile size.
+            img_dct = self.forward_tiled_nerf(nerf_hidden, nerf_pixels, B, C, num_patches, patch_size, params)
+        else:
+            # Get DCT-encoded pixel embeddings [pixel-dct]
+            img_dct = self.nerf_image_embedder(nerf_pixels)
+
+            # Pass through the dynamic MLP blocks (the NeRF)
+            for block in self.nerf_blocks:
+                img_dct = block(img_dct, nerf_hidden)
+
+        # Reassemble the patches into the final image.
+        img_dct = img_dct.transpose(1, 2) # -> [B*NumPatches, C, P*P]
+        # Reshape to combine with batch dimension for fold
+        img_dct = img_dct.reshape(B, num_patches, -1) # -> [B, NumPatches, C*P*P]
+        img_dct = img_dct.transpose(1, 2) # -> [B, C*P*P, NumPatches]
+        img_dct = nn.functional.fold(
+            img_dct,
+            output_size=(H, W),
+            kernel_size=patch_size,
+            stride=patch_size,
+        )
+        return self._nerf_final_layer(img_dct)
+
+    def forward_tiled_nerf(
+        self,
+        nerf_hidden: Tensor,
+        nerf_pixels: Tensor,
+        batch: int,
+        channels: int,
+        num_patches: int,
+        patch_size: int,
+        params: ChromaRadianceParams,
+    ) -> Tensor:
+        """
+        Processes the NeRF head in tiles to save memory.
+        nerf_hidden has shape [B, L, D]
+        nerf_pixels has shape [B, L, C * P * P]
+        """
+        tile_size = params.nerf_tile_size
+        output_tiles = []
+        # Iterate over the patches in tiles. The dimension L (num_patches) is at index 1.
+        for i in range(0, num_patches, tile_size):
+            end = min(i + tile_size, num_patches)
+
+            # Slice the current tile from the input tensors
+            nerf_hidden_tile = nerf_hidden[i * batch:end * batch]
+            nerf_pixels_tile = nerf_pixels[i * batch:end * batch]
+
+            # get DCT-encoded pixel embeddings [pixel-dct]
+            img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile)
+
+            # pass through the dynamic MLP blocks (the NeRF)
+            for block in self.nerf_blocks:
+                img_dct_tile = block(img_dct_tile, nerf_hidden_tile)
+
+            output_tiles.append(img_dct_tile)
+
+        # Concatenate the processed tiles along the patch dimension
+        return torch.cat(output_tiles, dim=0)
+
+    def radiance_get_override_params(self, overrides: dict) -> ChromaRadianceParams:
+        params = self.params
+        if not overrides:
+            return params
+        params_dict = {k: getattr(params, k) for k in params.__dataclass_fields__}
+        nullable_keys = frozenset(("nerf_embedder_dtype",))
+        bad_keys = tuple(k for k in overrides if k not in params_dict)
+        if bad_keys:
+            e = f"Unknown key(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"
+            raise ValueError(e)
+        bad_keys = tuple(
+            k
+            for k, v in overrides.items()
+            if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys)
+        )
+        if bad_keys:
+            e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"
+            raise ValueError(e)
+        # At this point it's all valid keys and values so we can merge with the existing params.
+        params_dict |= overrides
+        return params.__class__(**params_dict)
+
+    def _apply_x0_residual(self, predicted, noisy, timesteps):
+
+        # non zero during training to prevent 0 div
+        eps = 0.0
+        return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps)
+
+    def _forward(
+        self,
+        x: Tensor,
+        timestep: Tensor,
+        context: Tensor,
+        guidance: Optional[Tensor],
+        control: Optional[dict]=None,
+        transformer_options: dict={},
+        **kwargs: dict,
+    ) -> Tensor:
+        bs, c, h, w = x.shape
+        img = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+
+        if img.ndim != 4:
+            raise ValueError("Input img tensor must be in [B, C, H, W] format.")
+        if context.ndim != 3:
+            raise ValueError("Input txt tensors must have 3 dimensions.")
+
+        params = self.radiance_get_override_params(transformer_options.get("chroma_radiance_options", {}))
+
+        h_len = (img.shape[-2] // self.patch_size)
+        w_len = (img.shape[-1] // 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)
+
+        img_out = self.forward_orig(
+            img,
+            img_ids,
+            context,
+            txt_ids,
+            timestep,
+            guidance,
+            control,
+            transformer_options,
+            attn_mask=kwargs.get("attention_mask", None),
+        )
+
+        out = self.forward_nerf(img, img_out, params)[:, :, :h, :w]
+
+        # If x0 variant → v-pred, just return this instead
+        if hasattr(self, "__x0__"):
+            out = self._apply_x0_residual(out, img, timestep)
+        return out
+

comfy/ldm/cosmos/blocks.py

@@ -176,6 +176,7 @@ class Attention(nn.Module):
         context=None,
         mask=None,
         rope_emb=None,
+        transformer_options={},
         **kwargs,
     ):
         """
@@ -184,7 +185,7 @@ class Attention(nn.Module):
             context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
         """
         q, k, v = self.cal_qkv(x, context, mask, rope_emb=rope_emb, **kwargs)
-        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True)
+        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True, transformer_options=transformer_options)
         del q, k, v
         out = rearrange(out, " b n s c -> s b (n c)")
         return self.to_out(out)
@@ -546,6 +547,7 @@ class VideoAttn(nn.Module):
         context: Optional[torch.Tensor] = None,
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         """
         Forward pass for video attention.
@@ -571,6 +573,7 @@ class VideoAttn(nn.Module):
             context_M_B_D,
             crossattn_mask,
             rope_emb=rope_emb_L_1_1_D,
+            transformer_options=transformer_options,
         )
         x_T_H_W_B_D = rearrange(x_THW_B_D, "(t h w) b d -> t h w b d", h=H, w=W)
         return x_T_H_W_B_D
@@ -665,6 +668,7 @@ class DITBuildingBlock(nn.Module):
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_3D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         """
         Forward pass for dynamically configured blocks with adaptive normalization.
@@ -702,6 +706,7 @@ class DITBuildingBlock(nn.Module):
                 adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
                 context=None,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
             )
         elif self.block_type in ["cross_attn", "ca"]:
             x = x + gate_1_1_1_B_D * self.block(
@@ -709,6 +714,7 @@ class DITBuildingBlock(nn.Module):
                 context=crossattn_emb,
                 crossattn_mask=crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
             )
         else:
             raise ValueError(f"Unknown block type: {self.block_type}")
@@ -784,6 +790,7 @@ class GeneralDITTransformerBlock(nn.Module):
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_3D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         for block in self.blocks:
             x = block(
@@ -793,5 +800,6 @@ class GeneralDITTransformerBlock(nn.Module):
                 crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                 adaln_lora_B_3D=adaln_lora_B_3D,
+                transformer_options=transformer_options,
             )
         return x

comfy/ldm/cosmos/model.py

@@ -520,6 +520,7 @@ class GeneralDIT(nn.Module):
                 x.shape == extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape
             ), f"{x.shape} != {extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape} {original_shape}"
 
+        transformer_options = kwargs.get("transformer_options", {})
         for _, block in self.blocks.items():
             assert (
                 self.blocks["block0"].x_format == block.x_format
@@ -534,6 +535,7 @@ class GeneralDIT(nn.Module):
                 crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                 adaln_lora_B_3D=adaln_lora_B_3D,
+                transformer_options=transformer_options,
             )
 
         x_B_T_H_W_D = rearrange(x, "T H W B D -> B T H W D")

comfy/ldm/cosmos/predict2.py

@@ -44,7 +44,7 @@ class GPT2FeedForward(nn.Module):
         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:
+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, transformer_options: Optional[dict] = {}) -> torch.Tensor:
     """Computes multi-head attention using PyTorch's native implementation.
 
     This function provides a PyTorch backend alternative to Transformer Engine's attention operation.
@@ -71,7 +71,7 @@ def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H
     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)
+    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, transformer_options=transformer_options)
 
 
 class Attention(nn.Module):
@@ -180,8 +180,8 @@ class Attention(nn.Module):
 
         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]
+    def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, transformer_options: Optional[dict] = {}) -> torch.Tensor:
+        result = self.attn_op(q, k, v, transformer_options=transformer_options)  # [B, S, H, D]
         return self.output_dropout(self.output_proj(result))
 
     def forward(
@@ -189,6 +189,7 @@ class Attention(nn.Module):
         x: torch.Tensor,
         context: Optional[torch.Tensor] = None,
         rope_emb: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         """
         Args:
@@ -196,7 +197,7 @@ class Attention(nn.Module):
             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)
+        return self.compute_attention(q, k, v, transformer_options=transformer_options)
 
 
 class Timesteps(nn.Module):
@@ -459,6 +460,7 @@ class Block(nn.Module):
         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,
+        transformer_options: Optional[dict] = {},
     ) -> 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
@@ -512,6 +514,7 @@ class Block(nn.Module):
                 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,
+                transformer_options=transformer_options,
             ),
             "b (t h w) d -> b t h w d",
             t=T,
@@ -525,6 +528,7 @@ class Block(nn.Module):
             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,
+            transformer_options: Optional[dict] = {},
         ) -> 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
@@ -534,6 +538,7 @@ class Block(nn.Module):
                     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,
+                    transformer_options=transformer_options,
                 ),
                 "b (t h w) d -> b t h w d",
                 t=T,
@@ -547,6 +552,7 @@ class Block(nn.Module):
             self.layer_norm_cross_attn,
             scale_cross_attn_B_T_1_1_D,
             shift_cross_attn_B_T_1_1_D,
+            transformer_options=transformer_options,
         )
         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
 
@@ -865,6 +871,7 @@ class MiniTrainDIT(nn.Module):
             "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,
+            "transformer_options": kwargs.get("transformer_options", {}),
         }
         for block in self.blocks:
             x_B_T_H_W_D = block(

comfy/ldm/flux/layers.py

@@ -48,15 +48,44 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
     return embedding
 
 class MLPEmbedder(nn.Module):
-    def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
+    def __init__(self, in_dim: int, hidden_dim: int, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
-        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
         self.silu = nn.SiLU()
-        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
 
     def forward(self, x: Tensor) -> Tensor:
         return self.out_layer(self.silu(self.in_layer(x)))
 
+class YakMLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
+        self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
+        self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dtype=None, device=None, operations=None):
+    if yak_mlp:
+        return YakMLP(hidden_size, mlp_hidden_dim, dtype=dtype, device=device, operations=operations)
+    if mlp_silu_act:
+        return nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+            SiLUActivation(),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+        )
+    else:
+        return 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),
+        )
 
 class RMSNorm(torch.nn.Module):
     def __init__(self, dim: int, dtype=None, device=None, operations=None):
@@ -80,14 +109,14 @@ class QKNorm(torch.nn.Module):
 
 
 class SelfAttention(nn.Module):
-    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, proj_bias: bool = True, dtype=None, device=None, operations=None):
         super().__init__()
         self.num_heads = num_heads
         head_dim = dim // num_heads
 
         self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
-        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, bias=proj_bias, dtype=dtype, device=device)
 
 
 @dataclass
@@ -98,11 +127,11 @@ class ModulationOut:
 
 
 class Modulation(nn.Module):
-    def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
+    def __init__(self, dim: int, double: bool, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
         self.is_double = double
         self.multiplier = 6 if double else 3
-        self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
+        self.lin = operations.Linear(dim, self.multiplier * dim, bias=bias, dtype=dtype, device=device)
 
     def forward(self, vec: Tensor) -> tuple:
         if vec.ndim == 2:
@@ -129,77 +158,107 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
         return tensor
 
 
+class SiLUActivation(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.gate_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        return self.gate_fn(x1) * x2
+
+
 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):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=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_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.modulation = modulation
+
+        if self.modulation:
+            self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+
         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_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_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_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.img_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+
+        if self.modulation:
+            self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+
         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_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_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.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+
         self.flipped_img_txt = flipped_img_txt
 
-    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None):
-        img_mod1, img_mod2 = self.img_mod(vec)
-        txt_mod1, txt_mod2 = self.txt_mod(vec)
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
+        if self.modulation:
+            img_mod1, img_mod2 = self.img_mod(vec)
+            txt_mod1, txt_mod2 = self.txt_mod(vec)
+        else:
+            (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
         # prepare image for attention
         img_modulated = self.img_norm1(img)
         img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
         img_qkv = self.img_attn.qkv(img_modulated)
+        del 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)
+        del img_qkv
         img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
 
         # prepare txt for attention
         txt_modulated = self.txt_norm1(txt)
         txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt)
         txt_qkv = self.txt_attn.qkv(txt_modulated)
+        del 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)
+        del txt_qkv
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
 
         if self.flipped_img_txt:
+            q = torch.cat((img_q, txt_q), dim=2)
+            del img_q, txt_q
+            k = torch.cat((img_k, txt_k), dim=2)
+            del img_k, txt_k
+            v = torch.cat((img_v, txt_v), dim=2)
+            del img_v, txt_v
             # run actual attention
-            attn = attention(torch.cat((img_q, txt_q), dim=2),
-                             torch.cat((img_k, txt_k), dim=2),
-                             torch.cat((img_v, txt_v), dim=2),
-                             pe=pe, mask=attn_mask)
+            attn = attention(q, k, v,
+                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
+            del q, k, v
 
             img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
         else:
+            q = torch.cat((txt_q, img_q), dim=2)
+            del txt_q, img_q
+            k = torch.cat((txt_k, img_k), dim=2)
+            del txt_k, img_k
+            v = torch.cat((txt_v, img_v), dim=2)
+            del txt_v, img_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)
+            attn = attention(q, k, v,
+                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
+            del q, k, v
 
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
-        img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
-        img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
+        img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        del img_attn
+        img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
 
         # calculate the txt bloks
         txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)
+        del txt_attn
         txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt)
 
         if txt.dtype == torch.float16:
@@ -220,6 +279,10 @@ class SingleStreamBlock(nn.Module):
         num_heads: int,
         mlp_ratio: float = 4.0,
         qk_scale: float = None,
+        modulation=True,
+        mlp_silu_act=False,
+        bias=True,
+        yak_mlp=False,
         dtype=None,
         device=None,
         operations=None
@@ -231,30 +294,55 @@ class SingleStreamBlock(nn.Module):
         self.scale = qk_scale or head_dim**-0.5
 
         self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp_hidden_dim_first = self.mlp_hidden_dim
+        self.yak_mlp = yak_mlp
+        if mlp_silu_act:
+            self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2)
+            self.mlp_act = SiLUActivation()
+        else:
+            self.mlp_act = nn.GELU(approximate="tanh")
+
+        if self.yak_mlp:
+            self.mlp_hidden_dim_first *= 2
+            self.mlp_act = nn.SiLU()
+
         # qkv and mlp_in
-        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, 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.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=bias, 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")
-        self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+        if modulation:
+            self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+        else:
+            self.modulation = None
 
-    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None) -> Tensor:
-        mod, _ = self.modulation(vec)
-        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor:
+        if self.modulation:
+            mod, _ = self.modulation(vec)
+        else:
+            mod = vec
+
+        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        del qkv
         q, k = self.norm(q, k, v)
 
         # compute attention
-        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
+        del q, k, v
         # compute activation in mlp stream, cat again and run second linear layer
-        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        if self.yak_mlp:
+            mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]
+        else:
+            mlp = self.mlp_act(mlp)
+        output = self.linear2(torch.cat((attn, mlp), 2))
         x += apply_mod(output, mod.gate, None, modulation_dims)
         if x.dtype == torch.float16:
             x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
@@ -262,11 +350,11 @@ class SingleStreamBlock(nn.Module):
 
 
 class LastLayer(nn.Module):
-    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, bias=True, 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, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
-        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=bias, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=bias, dtype=dtype, device=device))
 
     def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor:
         if vec.ndim == 2:

comfy/ldm/flux/math.py

@@ -4,23 +4,16 @@ from torch import Tensor
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
+import logging
 
 
-def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
-    q_shape = q.shape
-    k_shape = k.shape
-
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
     if pe is not None:
-        q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
-        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
-        q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
-        k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
-
+        q, k = apply_rope(q, k, pe)
     heads = q.shape[1]
-    x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask)
+    x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
 
-
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
     if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
@@ -36,10 +29,19 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     return out.to(dtype=torch.float32, device=pos.device)
 
 
-def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-    xq_ = xq.to(dtype=freqs_cis.dtype).reshape(*xq.shape[:-1], -1, 1, 2)
-    xk_ = xk.to(dtype=freqs_cis.dtype).reshape(*xk.shape[:-1], -1, 1, 2)
-    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
-    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
-    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
+try:
+    import comfy.quant_ops
+    apply_rope = comfy.quant_ops.ck.apply_rope
+    apply_rope1 = comfy.quant_ops.ck.apply_rope1
+except:
+    logging.warning("No comfy kitchen, using old apply_rope functions.")
+    def apply_rope1(x: Tensor, freqs_cis: Tensor):
+        x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
 
+        x_out = freqs_cis[..., 0] * x_[..., 0]
+        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+
+        return x_out.reshape(*x.shape).type_as(x)
+
+    def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+        return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)

comfy/ldm/flux/model.py

@@ -15,6 +15,8 @@ from .layers import (
     MLPEmbedder,
     SingleStreamBlock,
     timestep_embedding,
+    Modulation,
+    RMSNorm
 )
 
 @dataclass
@@ -33,6 +35,14 @@ class FluxParams:
     patch_size: int
     qkv_bias: bool
     guidance_embed: bool
+    txt_ids_dims: list
+    global_modulation: bool = False
+    mlp_silu_act: bool = False
+    ops_bias: bool = True
+    default_ref_method: str = "offset"
+    ref_index_scale: float = 1.0
+    yak_mlp: bool = False
+    txt_norm: bool = False
 
 
 class Flux(nn.Module):
@@ -58,13 +68,22 @@ class Flux(nn.Module):
         self.hidden_size = params.hidden_size
         self.num_heads = params.num_heads
         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.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
-        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+        if params.vec_in_dim is not None:
+            self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.vector_in = None
+
         self.guidance_in = (
-            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
         )
-        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
+
+        if params.txt_norm:
+            self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations)
+        else:
+            self.txt_norm = None
 
         self.double_blocks = nn.ModuleList(
             [
@@ -73,6 +92,10 @@ class Flux(nn.Module):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
+                    modulation=params.global_modulation is False,
+                    mlp_silu_act=params.mlp_silu_act,
+                    proj_bias=params.ops_bias,
+                    yak_mlp=params.yak_mlp,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -81,13 +104,30 @@ class Flux(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, yak_mlp=params.yak_mlp, 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.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+
+        if params.global_modulation:
+            self.double_stream_modulation_img = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.double_stream_modulation_txt = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.single_stream_modulation = Modulation(
+                self.hidden_size, double=False, bias=False, dtype=dtype, device=device, operations=operations
+            )
 
     def forward_orig(
         self,
@@ -103,9 +143,7 @@ class Flux(nn.Module):
         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 = transformer_options.get("patches", {})
         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.")
@@ -117,9 +155,27 @@ class Flux(nn.Module):
             if guidance is not None:
                 vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
 
-        vec = vec + self.vector_in(y[:,:self.params.vec_in_dim])
+        if self.vector_in is not None:
+            if y is None:
+                y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+
+        if self.txt_norm is not None:
+            txt = self.txt_norm(txt)
         txt = self.txt_in(txt)
 
+        vec_orig = vec
+        if self.params.global_modulation:
+            vec = (self.double_stream_modulation_img(vec_orig), self.double_stream_modulation_txt(vec_orig))
+
+        if "post_input" in patches:
+            for p in patches["post_input"]:
+                out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids})
+                img = out["img"]
+                txt = out["txt"]
+                img_ids = out["img_ids"]
+                txt_ids = out["txt_ids"]
+
         if img_ids is not None:
             ids = torch.cat((txt_ids, img_ids), dim=1)
             pe = self.pe_embedder(ids)
@@ -127,7 +183,10 @@ class Flux(nn.Module):
             pe = None
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -135,14 +194,16 @@ class Flux(nn.Module):
                                                    txt=args["txt"],
                                                    vec=args["vec"],
                                                    pe=args["pe"],
-                                                   attn_mask=args.get("attn_mask"))
+                                                   attn_mask=args.get("attn_mask"),
+                                                   transformer_options=args.get("transformer_options"))
                     return out
 
                 out = blocks_replace[("double_block", i)]({"img": img,
                                                            "txt": txt,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
@@ -151,52 +212,61 @@ class Flux(nn.Module):
                                  txt=txt,
                                  vec=vec,
                                  pe=pe,
-                                 attn_mask=attn_mask)
+                                 attn_mask=attn_mask,
+                                 transformer_options=transformer_options)
 
             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[:, :add.shape[1]] += 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)
 
+        if self.params.global_modulation:
+            vec, _ = self.single_stream_modulation(vec_orig)
+
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = 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"))
+                                       attn_mask=args.get("attn_mask"),
+                                       transformer_options=args.get("transformer_options"))
                     return out
 
                 out = blocks_replace[("single_block", i)]({"img": img,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
 
             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[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
 
         img = img[:, txt.shape[1] :, ...]
 
-        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        img = self.final_layer(img, vec_orig)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def process_img(self, x, index=0, h_offset=0, w_offset=0):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}):
         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))
@@ -208,10 +278,22 @@ class Flux(nn.Module):
         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)
+        steps_h = h_len
+        steps_w = w_len
+
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        img_ids = torch.zeros((steps_h, steps_w, len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
         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)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=torch.float32).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=torch.float32).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):
@@ -227,18 +309,24 @@ class Flux(nn.Module):
 
         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, img_ids = self.process_img(x, transformer_options=transformer_options)
         img_tokens = img.shape[1]
         if ref_latents is not None:
             h = 0
             w = 0
             index = 0
-            index_ref_method = kwargs.get("ref_latents_method", "offset") == "index"
+            ref_latents_method = kwargs.get("ref_latents_method", self.params.default_ref_method)
             for ref in ref_latents:
-                if index_ref_method:
-                    index += 1
+                if ref_latents_method == "index":
+                    index += self.params.ref_index_scale
                     h_offset = 0
                     w_offset = 0
+                elif ref_latents_method == "uxo":
+                    index = 0
+                    h_offset = h_len * patch_size + h
+                    w_offset = w_len * patch_size + w
+                    h += ref.shape[-2]
+                    w += ref.shape[-1]
                 else:
                     index = 1
                     h_offset = 0
@@ -254,7 +342,12 @@ class Flux(nn.Module):
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
 
-        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
+
+        if len(self.params.txt_ids_dims) > 0:
+            for i in self.params.txt_ids_dims:
+                txt_ids[:, :, i] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32)
+
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
         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]
+        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_orig,:w_orig]

comfy/ldm/genmo/joint_model/asymm_models_joint.py

@@ -109,6 +109,7 @@ class AsymmetricAttention(nn.Module):
         scale_x: torch.Tensor,  # (B, dim_x), modulation for pre-RMSNorm.
         scale_y: torch.Tensor,  # (B, dim_y), modulation for pre-RMSNorm.
         crop_y,
+        transformer_options={},
         **rope_rotation,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         rope_cos = rope_rotation.get("rope_cos")
@@ -143,7 +144,7 @@ class AsymmetricAttention(nn.Module):
 
         xy = optimized_attention(q,
                                  k,
-                                 v, self.num_heads, skip_reshape=True)
+                                 v, self.num_heads, skip_reshape=True, transformer_options=transformer_options)
 
         x, y = torch.tensor_split(xy, (q_x.shape[1],), dim=1)
         x = self.proj_x(x)
@@ -224,6 +225,7 @@ class AsymmetricJointBlock(nn.Module):
         x: torch.Tensor,
         c: torch.Tensor,
         y: torch.Tensor,
+        transformer_options={},
         **attn_kwargs,
     ):
         """Forward pass of a block.
@@ -256,6 +258,7 @@ class AsymmetricJointBlock(nn.Module):
             y,
             scale_x=scale_msa_x,
             scale_y=scale_msa_y,
+            transformer_options=transformer_options,
             **attn_kwargs,
         )
 
@@ -524,10 +527,11 @@ class AsymmDiTJoint(nn.Module):
                                                     args["txt"],
                                                     rope_cos=args["rope_cos"],
                                                     rope_sin=args["rope_sin"],
-                                                    crop_y=args["num_tokens"]
+                                                    crop_y=args["num_tokens"],
+                                                    transformer_options=args["transformer_options"]
                                                     )
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 y_feat = out["txt"]
                 x = out["img"]
             else:
@@ -538,6 +542,7 @@ class AsymmDiTJoint(nn.Module):
                     rope_cos=rope_cos,
                     rope_sin=rope_sin,
                     crop_y=num_tokens,
+                    transformer_options=transformer_options,
                 )  # (B, M, D), (B, L, D)
         del y_feat  # Final layers don't use dense text features.
 

comfy/ldm/hidream/model.py

@@ -72,8 +72,8 @@ class TimestepEmbed(nn.Module):
         return t_emb
 
 
-def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor):
-    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2])
+def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, transformer_options={}):
+    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2], transformer_options=transformer_options)
 
 
 class HiDreamAttnProcessor_flashattn:
@@ -86,6 +86,7 @@ class HiDreamAttnProcessor_flashattn:
         image_tokens_masks: Optional[torch.FloatTensor] = None,
         text_tokens: Optional[torch.FloatTensor] = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
         *args,
         **kwargs,
     ) -> torch.FloatTensor:
@@ -133,7 +134,7 @@ class HiDreamAttnProcessor_flashattn:
             query = torch.cat([query_1, query_2], dim=-1)
             key = torch.cat([key_1, key_2], dim=-1)
 
-        hidden_states = attention(query, key, value)
+        hidden_states = attention(query, key, value, transformer_options=transformer_options)
 
         if not attn.single:
             hidden_states_i, hidden_states_t = torch.split(hidden_states, [num_image_tokens, num_text_tokens], dim=1)
@@ -199,6 +200,7 @@ class HiDreamAttention(nn.Module):
         image_tokens_masks: torch.FloatTensor = None,
         norm_text_tokens: torch.FloatTensor = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
     ) -> torch.Tensor:
         return self.processor(
             self,
@@ -206,6 +208,7 @@ class HiDreamAttention(nn.Module):
             image_tokens_masks = image_tokens_masks,
             text_tokens = norm_text_tokens,
             rope = rope,
+            transformer_options=transformer_options,
         )
 
 
@@ -406,7 +409,7 @@ class HiDreamImageSingleTransformerBlock(nn.Module):
         text_tokens: Optional[torch.FloatTensor] = None,
         adaln_input: Optional[torch.FloatTensor] = None,
         rope: torch.FloatTensor = None,
-
+        transformer_options={},
     ) -> torch.FloatTensor:
         wtype = image_tokens.dtype
         shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i = \
@@ -419,6 +422,7 @@ class HiDreamImageSingleTransformerBlock(nn.Module):
             norm_image_tokens,
             image_tokens_masks,
             rope = rope,
+            transformer_options=transformer_options,
         )
         image_tokens = gate_msa_i * attn_output_i + image_tokens
 
@@ -483,6 +487,7 @@ class HiDreamImageTransformerBlock(nn.Module):
         text_tokens: Optional[torch.FloatTensor] = None,
         adaln_input: Optional[torch.FloatTensor] = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
     ) -> torch.FloatTensor:
         wtype = image_tokens.dtype
         shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i, \
@@ -500,6 +505,7 @@ class HiDreamImageTransformerBlock(nn.Module):
             image_tokens_masks,
             norm_text_tokens,
             rope = rope,
+            transformer_options=transformer_options,
         )
 
         image_tokens = gate_msa_i * attn_output_i + image_tokens
@@ -550,6 +556,7 @@ class HiDreamImageBlock(nn.Module):
         text_tokens: Optional[torch.FloatTensor] = None,
         adaln_input: torch.FloatTensor = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
     ) -> torch.FloatTensor:
         return self.block(
             image_tokens,
@@ -557,6 +564,7 @@ class HiDreamImageBlock(nn.Module):
             text_tokens,
             adaln_input,
             rope,
+            transformer_options=transformer_options,
         )
 
 
@@ -786,6 +794,7 @@ class HiDreamImageTransformer2DModel(nn.Module):
                 text_tokens = cur_encoder_hidden_states,
                 adaln_input = adaln_input,
                 rope = rope,
+                transformer_options=transformer_options,
             )
             initial_encoder_hidden_states = initial_encoder_hidden_states[:, :initial_encoder_hidden_states_seq_len]
             block_id += 1
@@ -809,6 +818,7 @@ class HiDreamImageTransformer2DModel(nn.Module):
                 text_tokens=None,
                 adaln_input=adaln_input,
                 rope=rope,
+                transformer_options=transformer_options,
             )
             hidden_states = hidden_states[:, :hidden_states_seq_len]
             block_id += 1

comfy/ldm/hunyuan3d/model.py

@@ -99,14 +99,16 @@ class Hunyuan3Dv2(nn.Module):
                                                    txt=args["txt"],
                                                    vec=args["vec"],
                                                    pe=args["pe"],
-                                                   attn_mask=args.get("attn_mask"))
+                                                   attn_mask=args.get("attn_mask"),
+                                                   transformer_options=args["transformer_options"])
                     return out
 
                 out = blocks_replace[("double_block", i)]({"img": img,
                                                            "txt": txt,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
@@ -115,7 +117,8 @@ class Hunyuan3Dv2(nn.Module):
                                  txt=txt,
                                  vec=vec,
                                  pe=pe,
-                                 attn_mask=attn_mask)
+                                 attn_mask=attn_mask,
+                                 transformer_options=transformer_options)
 
         img = torch.cat((txt, img), 1)
 
@@ -126,17 +129,19 @@ class Hunyuan3Dv2(nn.Module):
                     out["img"] = block(args["img"],
                                        vec=args["vec"],
                                        pe=args["pe"],
-                                       attn_mask=args.get("attn_mask"))
+                                       attn_mask=args.get("attn_mask"),
+                                       transformer_options=args["transformer_options"])
                     return out
 
                 out = blocks_replace[("single_block", i)]({"img": img,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
 
         img = img[:, txt.shape[1]:, ...]
         img = self.final_layer(img, vec)

comfy/ldm/hunyuan3d/vae.py

@@ -4,81 +4,458 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-
-
-from typing import Union, Tuple, List, Callable, Optional
-
 import numpy as np
-from einops import repeat, rearrange
+import math
 from tqdm import tqdm
+
+from typing import Optional
+
 import logging
 
 import comfy.ops
 ops = comfy.ops.disable_weight_init
 
-def generate_dense_grid_points(
-    bbox_min: np.ndarray,
-    bbox_max: np.ndarray,
-    octree_resolution: int,
-    indexing: str = "ij",
-):
-    length = bbox_max - bbox_min
-    num_cells = octree_resolution
+def fps(src: torch.Tensor, batch: torch.Tensor, sampling_ratio: float, start_random: bool = True):
 
-    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
-    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
-    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
-    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
-    xyz = np.stack((xs, ys, zs), axis=-1)
-    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+    # manually create the pointer vector
+    assert src.size(0) == batch.numel()
 
-    return xyz, grid_size, length
+    batch_size = int(batch.max()) + 1
+    deg = src.new_zeros(batch_size, dtype = torch.long)
+
+    deg.scatter_add_(0, batch, torch.ones_like(batch))
+
+    ptr_vec = deg.new_zeros(batch_size + 1)
+    torch.cumsum(deg, 0, out=ptr_vec[1:])
+
+    #return fps_sampling(src, ptr_vec, ratio)
+    sampled_indicies = []
+
+    for b in range(batch_size):
+        # start and the end of each batch
+        start, end = ptr_vec[b].item(), ptr_vec[b + 1].item()
+        # points from the point cloud
+        points = src[start:end]
+
+        num_points = points.size(0)
+        num_samples = max(1, math.ceil(num_points * sampling_ratio))
+
+        selected = torch.zeros(num_samples, device = src.device, dtype = torch.long)
+        distances = torch.full((num_points,), float("inf"), device = src.device)
+
+        # select a random start point
+        if start_random:
+            farthest = torch.randint(0, num_points, (1,), device = src.device)
+        else:
+            farthest = torch.tensor([0], device = src.device, dtype = torch.long)
+
+        for i in range(num_samples):
+            selected[i] = farthest
+            centroid = points[farthest].squeeze(0)
+            dist = torch.norm(points - centroid, dim = 1) # compute euclidean distance
+            distances = torch.minimum(distances, dist)
+            farthest = torch.argmax(distances)
+
+        sampled_indicies.append(torch.arange(start, end)[selected])
+
+    return torch.cat(sampled_indicies, dim = 0)
+class PointCrossAttention(nn.Module):
+    def __init__(self,
+        num_latents: int,
+        downsample_ratio: float,
+        pc_size: int,
+        pc_sharpedge_size: int,
+        point_feats: int,
+        width: int,
+        heads: int,
+        layers: int,
+        fourier_embedder,
+        normal_pe: bool = False,
+        qkv_bias: bool = False,
+        use_ln_post: bool = True,
+        qk_norm: bool = True):
+
+        super().__init__()
+
+        self.fourier_embedder = fourier_embedder
+
+        self.pc_size = pc_size
+        self.normal_pe = normal_pe
+        self.downsample_ratio = downsample_ratio
+        self.pc_sharpedge_size = pc_sharpedge_size
+        self.num_latents = num_latents
+        self.point_feats = point_feats
+
+        self.input_proj = nn.Linear(self.fourier_embedder.out_dim + point_feats, width)
+
+        self.cross_attn = ResidualCrossAttentionBlock(
+            width = width,
+            heads = heads,
+            qkv_bias = qkv_bias,
+            qk_norm = qk_norm
+        )
+
+        self.self_attn = None
+        if layers > 0:
+            self.self_attn = Transformer(
+                width = width,
+                heads = heads,
+                qkv_bias = qkv_bias,
+                qk_norm = qk_norm,
+                layers = layers
+            )
+
+        if use_ln_post:
+            self.ln_post = nn.LayerNorm(width)
+        else:
+            self.ln_post = None
+
+    def sample_points_and_latents(self, point_cloud: torch.Tensor, features: torch.Tensor):
+
+        """
+        Subsample points randomly from the point cloud (input_pc)
+        Further sample the subsampled points to get query_pc
+        take the fourier embeddings for both input and query pc
+
+        Mental Note: FPS-sampled points (query_pc) act as latent tokens that attend to and learn from the broader context in input_pc.
+        Goal: get a smaller represenation (query_pc) to represent the entire scence structure by learning from a broader subset (input_pc).
+        More computationally efficient.
+
+        Features are additional information for each point in the cloud
+        """
+
+        B, _, D = point_cloud.shape
+
+        num_latents = int(self.num_latents)
+
+        num_random_query = self.pc_size / (self.pc_size + self.pc_sharpedge_size) * num_latents
+        num_sharpedge_query = num_latents - num_random_query
+
+        # Split random and sharpedge surface points
+        random_pc, sharpedge_pc = torch.split(point_cloud, [self.pc_size, self.pc_sharpedge_size], dim=1)
+
+        # assert statements
+        assert random_pc.shape[1] <= self.pc_size, "Random surface points size must be less than or equal to pc_size"
+        assert sharpedge_pc.shape[1] <= self.pc_sharpedge_size, "Sharpedge surface points size must be less than or equal to pc_sharpedge_size"
+
+        input_random_pc_size = int(num_random_query * self.downsample_ratio)
+        random_query_pc, random_input_pc, random_idx_pc, random_idx_query = \
+            self.subsample(pc = random_pc, num_query = num_random_query, input_pc_size = input_random_pc_size)
+
+        input_sharpedge_pc_size = int(num_sharpedge_query * self.downsample_ratio)
+
+        if input_sharpedge_pc_size == 0:
+            sharpedge_input_pc = torch.zeros(B, 0, D, dtype = random_input_pc.dtype).to(point_cloud.device)
+            sharpedge_query_pc = torch.zeros(B, 0, D, dtype= random_query_pc.dtype).to(point_cloud.device)
+
+        else:
+            sharpedge_query_pc, sharpedge_input_pc, sharpedge_idx_pc, sharpedge_idx_query = \
+            self.subsample(pc = sharpedge_pc, num_query = num_sharpedge_query, input_pc_size = input_sharpedge_pc_size)
+
+        # concat the random and sharpedges
+        query_pc = torch.cat([random_query_pc, sharpedge_query_pc], dim = 1)
+        input_pc = torch.cat([random_input_pc, sharpedge_input_pc], dim = 1)
+
+        query = self.fourier_embedder(query_pc)
+        data = self.fourier_embedder(input_pc)
+
+        if self.point_feats > 0:
+            random_surface_features, sharpedge_surface_features = torch.split(features, [self.pc_size, self.pc_sharpedge_size], dim = 1)
+
+            input_random_surface_features, query_random_features = \
+                self.handle_features(features = random_surface_features, idx_pc = random_idx_pc, batch_size = B,
+                                     input_pc_size = input_random_pc_size, idx_query = random_idx_query)
+
+            if input_sharpedge_pc_size == 0:
+                input_sharpedge_surface_features = torch.zeros(B, 0, self.point_feats,
+                                                               dtype = input_random_surface_features.dtype, device = point_cloud.device)
+
+                query_sharpedge_features = torch.zeros(B, 0, self.point_feats,
+                                                       dtype = query_random_features.dtype, device = point_cloud.device)
+            else:
+
+                input_sharpedge_surface_features, query_sharpedge_features = \
+                    self.handle_features(idx_pc = sharpedge_idx_pc, features = sharpedge_surface_features,
+                                         batch_size = B, idx_query = sharpedge_idx_query, input_pc_size = input_sharpedge_pc_size)
+
+            query_features = torch.cat([query_random_features, query_sharpedge_features], dim = 1)
+            input_features = torch.cat([input_random_surface_features, input_sharpedge_surface_features], dim = 1)
+
+            if self.normal_pe:
+                # apply the fourier embeddings on the first 3 dims (xyz)
+                input_features_pe = self.fourier_embedder(input_features[..., :3])
+                query_features_pe = self.fourier_embedder(query_features[..., :3])
+                # replace the first 3 dims with the new PE ones
+                input_features = torch.cat([input_features_pe, input_features[..., :3]], dim = -1)
+                query_features = torch.cat([query_features_pe, query_features[..., :3]], dim = -1)
+
+            # concat at the channels dim
+            query = torch.cat([query, query_features], dim = -1)
+            data = torch.cat([data, input_features], dim = -1)
+
+        # don't return pc_info to avoid unnecessary memory usuage
+        return query.view(B, -1, query.shape[-1]), data.view(B, -1, data.shape[-1])
+
+    def forward(self, point_cloud: torch.Tensor, features: torch.Tensor):
+
+        query, data = self.sample_points_and_latents(point_cloud = point_cloud, features = features)
+
+        # apply projections
+        query = self.input_proj(query)
+        data = self.input_proj(data)
+
+        # apply cross attention between query and data
+        latents = self.cross_attn(query, data)
+
+        if self.self_attn is not None:
+            latents = self.self_attn(latents)
+
+        if self.ln_post is not None:
+            latents = self.ln_post(latents)
+
+        return latents
 
 
-class VanillaVolumeDecoder:
+    def subsample(self, pc, num_query, input_pc_size: int):
+
+        """
+        num_query: number of points to keep after FPS
+        input_pc_size: number of points to select before FPS
+        """
+
+        B, _, D = pc.shape
+        query_ratio = num_query / input_pc_size
+
+        # random subsampling of points inside the point cloud
+        idx_pc = torch.randperm(pc.shape[1], device = pc.device)[:input_pc_size]
+        input_pc = pc[:, idx_pc, :]
+
+        # flatten to allow applying fps across the whole batch
+        flattent_input_pc = input_pc.view(B * input_pc_size, D)
+
+        # construct a batch_down tensor to tell fps
+        # which points belong to which batch
+        N_down = int(flattent_input_pc.shape[0] / B)
+        batch_down = torch.arange(B).to(pc.device)
+        batch_down = torch.repeat_interleave(batch_down, N_down)
+
+        idx_query = fps(flattent_input_pc, batch_down, sampling_ratio = query_ratio)
+        query_pc = flattent_input_pc[idx_query].view(B, -1, D)
+
+        return query_pc, input_pc, idx_pc, idx_query
+
+    def handle_features(self, features, idx_pc, input_pc_size, batch_size: int, idx_query):
+
+        B = batch_size
+
+        input_surface_features = features[:, idx_pc, :]
+        flattent_input_features = input_surface_features.view(B * input_pc_size, -1)
+        query_features = flattent_input_features[idx_query].view(B, -1,
+                                                                 flattent_input_features.shape[-1])
+
+        return input_surface_features, query_features
+
+def normalize_mesh(mesh, scale = 0.9999):
+    """Normalize mesh to fit in [-scale, scale]. Translate mesh so its center is [0,0,0]"""
+
+    bbox = mesh.bounds
+    center = (bbox[1] + bbox[0]) / 2
+
+    max_extent = (bbox[1] - bbox[0]).max()
+    mesh.apply_translation(-center)
+    mesh.apply_scale((2 * scale) / max_extent)
+
+    return mesh
+
+def sample_pointcloud(mesh, num = 200000):
+    """ Uniformly sample points from the surface of the mesh """
+
+    points, face_idx = mesh.sample(num, return_index = True)
+    normals = mesh.face_normals[face_idx]
+    return torch.from_numpy(points.astype(np.float32)), torch.from_numpy(normals.astype(np.float32))
+
+def detect_sharp_edges(mesh, threshold=0.985):
+    """Return edge indices (a, b) that lie on sharp boundaries of the mesh."""
+
+    V, F = mesh.vertices, mesh.faces
+    VN, FN = mesh.vertex_normals, mesh.face_normals
+
+    sharp_mask = np.ones(V.shape[0])
+    for i in range(3):
+        indices = F[:, i]
+        alignment = np.einsum('ij,ij->i', VN[indices], FN)
+        dot_stack = np.stack((sharp_mask[indices], alignment), axis=-1)
+        sharp_mask[indices] = np.min(dot_stack, axis=-1)
+
+    edge_a = np.concatenate([F[:, 0], F[:, 1], F[:, 2]])
+    edge_b = np.concatenate([F[:, 1], F[:, 2], F[:, 0]])
+    sharp_edges = (sharp_mask[edge_a] < threshold) & (sharp_mask[edge_b] < threshold)
+
+    return edge_a[sharp_edges], edge_b[sharp_edges]
+
+
+def sharp_sample_pointcloud(mesh, num = 16384):
+    """ Sample points preferentially from sharp edges in the mesh. """
+
+    edge_a, edge_b = detect_sharp_edges(mesh)
+    V, VN = mesh.vertices, mesh.vertex_normals
+
+    va, vb = V[edge_a], V[edge_b]
+    na, nb = VN[edge_a], VN[edge_b]
+
+    edge_lengths = np.linalg.norm(vb - va, axis=-1)
+    weights = edge_lengths / edge_lengths.sum()
+
+    indices = np.searchsorted(np.cumsum(weights), np.random.rand(num))
+    t = np.random.rand(num, 1)
+
+    samples = t * va[indices] + (1 - t) * vb[indices]
+    normals = t * na[indices] + (1 - t) * nb[indices]
+
+    return samples.astype(np.float32), normals.astype(np.float32)
+
+def load_surface_sharpedge(mesh, num_points=4096, num_sharp_points=4096, sharpedge_flag = True, device = "cuda"):
+    """Load a surface with optional sharp-edge annotations from a trimesh mesh."""
+
+    import trimesh
+
+    try:
+        mesh_full = trimesh.util.concatenate(mesh.dump())
+    except Exception:
+        mesh_full = trimesh.util.concatenate(mesh)
+
+    mesh_full = normalize_mesh(mesh_full)
+
+    faces = mesh_full.faces
+    vertices = mesh_full.vertices
+    origin_face_count = faces.shape[0]
+
+    mesh_surface = trimesh.Trimesh(vertices=vertices, faces=faces[:origin_face_count])
+    mesh_fill = trimesh.Trimesh(vertices=vertices, faces=faces[origin_face_count:])
+
+    area_surface = mesh_surface.area
+    area_fill = mesh_fill.area
+    total_area = area_surface + area_fill
+
+    sample_num = 499712 // 2
+    fill_ratio = area_fill / total_area if total_area > 0 else 0
+
+    num_fill = int(sample_num * fill_ratio)
+    num_surface = sample_num - num_fill
+
+    surf_pts, surf_normals = sample_pointcloud(mesh_surface, num_surface)
+    fill_pts, fill_normals = (torch.zeros(0, 3), torch.zeros(0, 3)) if num_fill == 0 else sample_pointcloud(mesh_fill, num_fill)
+
+    sharp_pts, sharp_normals = sharp_sample_pointcloud(mesh_surface, sample_num)
+
+    def assemble_tensor(points, normals, label=None):
+
+        data = torch.cat([points, normals], dim=1).half().to(device)
+
+        if label is not None:
+            label_tensor = torch.full((data.shape[0], 1), float(label), dtype=torch.float16).to(device)
+            data = torch.cat([data, label_tensor], dim=1)
+
+        return data
+
+    surface = assemble_tensor(torch.cat([surf_pts.to(device), fill_pts.to(device)], dim=0),
+                              torch.cat([surf_normals.to(device), fill_normals.to(device)], dim=0),
+                              label = 0 if sharpedge_flag else None)
+
+    sharp_surface = assemble_tensor(torch.from_numpy(sharp_pts), torch.from_numpy(sharp_normals),
+                                    label = 1 if sharpedge_flag else None)
+
+    rng = np.random.default_rng()
+
+    surface = surface[rng.choice(surface.shape[0], num_points, replace = False)]
+    sharp_surface = sharp_surface[rng.choice(sharp_surface.shape[0], num_sharp_points, replace = False)]
+
+    full = torch.cat([surface, sharp_surface], dim = 0).unsqueeze(0)
+
+    return full
+
+class SharpEdgeSurfaceLoader:
+    """ Load mesh surface and sharp edge samples. """
+
+    def __init__(self, num_uniform_points = 8192, num_sharp_points = 8192):
+
+        self.num_uniform_points = num_uniform_points
+        self.num_sharp_points = num_sharp_points
+        self.total_points = num_uniform_points + num_sharp_points
+
+    def __call__(self, mesh_input, device = "cuda"):
+        mesh = self._load_mesh(mesh_input)
+        return load_surface_sharpedge(mesh, self.num_uniform_points, self.num_sharp_points, device = device)
+
+    @staticmethod
+    def _load_mesh(mesh_input):
+        import trimesh
+
+        if isinstance(mesh_input, str):
+            mesh = trimesh.load(mesh_input, force="mesh", merge_primitives = True)
+        else:
+            mesh = mesh_input
+
+        if isinstance(mesh, trimesh.Scene):
+            combined = None
+            for obj in mesh.geometry.values():
+                combined = obj if combined is None else combined + obj
+            return combined
+
+        return mesh
+
+class DiagonalGaussianDistribution:
+    def __init__(self, params: torch.Tensor, feature_dim: int = -1):
+
+        # divide quant channels (8) into mean and log variance
+        self.mean, self.logvar = torch.chunk(params, 2, dim = feature_dim)
+
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.std = torch.exp(0.5 * self.logvar)
+
+    def sample(self):
+
+        eps = torch.randn_like(self.std)
+        z = self.mean + eps * self.std
+
+        return z
+
+################################################
+# Volume Decoder
+################################################
+
+class VanillaVolumeDecoder():
     @torch.no_grad()
-    def __call__(
-        self,
-        latents: torch.FloatTensor,
-        geo_decoder: Callable,
-        bounds: Union[Tuple[float], List[float], float] = 1.01,
-        num_chunks: int = 10000,
-        octree_resolution: int = None,
-        enable_pbar: bool = True,
-        **kwargs,
-    ):
-        device = latents.device
-        dtype = latents.dtype
-        batch_size = latents.shape[0]
+    def __call__(self, latents: torch.Tensor, geo_decoder: callable, octree_resolution: int, bounds = 1.01,
+                 num_chunks: int = 10_000, enable_pbar: bool = True, **kwargs):
 
-        # 1. generate query points
         if isinstance(bounds, float):
             bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
 
-        bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6])
-        xyz_samples, grid_size, length = generate_dense_grid_points(
-            bbox_min=bbox_min,
-            bbox_max=bbox_max,
-            octree_resolution=octree_resolution,
-            indexing="ij"
-        )
-        xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3)
+        bbox_min, bbox_max = torch.tensor(bounds[:3]), torch.tensor(bounds[3:])
+
+        x = torch.linspace(bbox_min[0], bbox_max[0], int(octree_resolution) + 1, dtype = torch.float32)
+        y = torch.linspace(bbox_min[1], bbox_max[1], int(octree_resolution) + 1, dtype = torch.float32)
+        z = torch.linspace(bbox_min[2], bbox_max[2], int(octree_resolution) + 1, dtype = torch.float32)
+
+        [xs, ys, zs] = torch.meshgrid(x, y, z, indexing = "ij")
+        xyz = torch.stack((xs, ys, zs), axis=-1).to(latents.device, dtype = latents.dtype).contiguous().reshape(-1, 3)
+        grid_size = [int(octree_resolution) + 1, int(octree_resolution) + 1, int(octree_resolution) + 1]
 
-        # 2. latents to 3d volume
         batch_logits = []
-        for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), desc="Volume Decoding",
+        for start in tqdm(range(0, xyz.shape[0], num_chunks), desc="Volume Decoding",
                           disable=not enable_pbar):
-            chunk_queries = xyz_samples[start: start + num_chunks, :]
-            chunk_queries = repeat(chunk_queries, "p c -> b p c", b=batch_size)
-            logits = geo_decoder(queries=chunk_queries, latents=latents)
+
+            chunk_queries = xyz[start: start + num_chunks, :]
+            chunk_queries = chunk_queries.unsqueeze(0).repeat(latents.shape[0], 1, 1)
+            logits = geo_decoder(queries = chunk_queries, latents = latents)
             batch_logits.append(logits)
 
-        grid_logits = torch.cat(batch_logits, dim=1)
-        grid_logits = grid_logits.view((batch_size, *grid_size)).float()
+        grid_logits = torch.cat(batch_logits, dim = 1)
+        grid_logits = grid_logits.view((latents.shape[0], *grid_size)).float()
 
         return grid_logits
 
-
 class FourierEmbedder(nn.Module):
     """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
     each feature dimension of `x[..., i]` into:
@@ -175,13 +552,11 @@ class FourierEmbedder(nn.Module):
         else:
             return x
 
-
 class CrossAttentionProcessor:
     def __call__(self, attn, q, k, v):
         out = comfy.ops.scaled_dot_product_attention(q, k, v)
         return out
 
-
 class DropPath(nn.Module):
     """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
     """
@@ -232,38 +607,41 @@ class MLP(nn.Module):
     def forward(self, x):
         return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
 
-
 class QKVMultiheadCrossAttention(nn.Module):
     def __init__(
         self,
-        *,
         heads: int,
+        n_data = None,
         width=None,
         qk_norm=False,
         norm_layer=ops.LayerNorm
     ):
         super().__init__()
         self.heads = heads
+        self.n_data = n_data
         self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
         self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
 
-        self.attn_processor = CrossAttentionProcessor()
-
     def forward(self, q, kv):
+
         _, n_ctx, _ = q.shape
         bs, n_data, width = kv.shape
+
         attn_ch = width // self.heads // 2
         q = q.view(bs, n_ctx, self.heads, -1)
+
         kv = kv.view(bs, n_data, self.heads, -1)
         k, v = torch.split(kv, attn_ch, dim=-1)
 
         q = self.q_norm(q)
         k = self.k_norm(k)
-        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
-        out = self.attn_processor(self, q, k, v)
-        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
-        return out
 
+        q, k, v = [t.permute(0, 2, 1, 3) for t in (q, k, v)]
+        out = F.scaled_dot_product_attention(q, k, v)
+
+        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
+
+        return out
 
 class MultiheadCrossAttention(nn.Module):
     def __init__(
@@ -306,7 +684,6 @@ class MultiheadCrossAttention(nn.Module):
         x = self.c_proj(x)
         return x
 
-
 class ResidualCrossAttentionBlock(nn.Module):
     def __init__(
         self,
@@ -366,7 +743,7 @@ class QKVMultiheadAttention(nn.Module):
         q = self.q_norm(q)
         k = self.k_norm(k)
 
-        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        q, k, v = [t.permute(0, 2, 1, 3) for t in (q, k, v)]
         out = F.scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
         return out
 
@@ -383,8 +760,7 @@ class MultiheadAttention(nn.Module):
         drop_path_rate: float = 0.0
     ):
         super().__init__()
-        self.width = width
-        self.heads = heads
+
         self.c_qkv = ops.Linear(width, width * 3, bias=qkv_bias)
         self.c_proj = ops.Linear(width, width)
         self.attention = QKVMultiheadAttention(
@@ -491,7 +867,7 @@ class CrossAttentionDecoder(nn.Module):
         self.query_proj = ops.Linear(self.fourier_embedder.out_dim, width)
         if self.downsample_ratio != 1:
             self.latents_proj = ops.Linear(width * downsample_ratio, width)
-        if self.enable_ln_post == False:
+        if not self.enable_ln_post:
             qk_norm = False
         self.cross_attn_decoder = ResidualCrossAttentionBlock(
             width=width,
@@ -522,28 +898,44 @@ class CrossAttentionDecoder(nn.Module):
 
 class ShapeVAE(nn.Module):
     def __init__(
-        self,
-        *,
-        embed_dim: int,
-        width: int,
-        heads: int,
-        num_decoder_layers: int,
-        geo_decoder_downsample_ratio: int = 1,
-        geo_decoder_mlp_expand_ratio: int = 4,
-        geo_decoder_ln_post: bool = True,
-        num_freqs: int = 8,
-        include_pi: bool = True,
-        qkv_bias: bool = True,
-        qk_norm: bool = False,
-        label_type: str = "binary",
-        drop_path_rate: float = 0.0,
-        scale_factor: float = 1.0,
+            self,
+            *,
+            num_latents: int = 4096,
+            embed_dim: int = 64,
+            width: int = 1024,
+            heads: int = 16,
+            num_decoder_layers: int = 16,
+            num_encoder_layers: int = 8,
+            pc_size: int = 81920,
+            pc_sharpedge_size: int = 0,
+            point_feats: int = 4,
+            downsample_ratio: int = 20,
+            geo_decoder_downsample_ratio: int = 1,
+            geo_decoder_mlp_expand_ratio: int = 4,
+            geo_decoder_ln_post: bool = True,
+            num_freqs: int = 8,
+            qkv_bias: bool = False,
+            qk_norm: bool = True,
+            drop_path_rate: float = 0.0,
+            include_pi: bool = False,
+            scale_factor: float = 1.0039506158752403,
+            label_type: str = "binary",
     ):
         super().__init__()
         self.geo_decoder_ln_post = geo_decoder_ln_post
 
         self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
 
+        self.encoder = PointCrossAttention(layers = num_encoder_layers,
+                                    num_latents = num_latents,
+                                    downsample_ratio = downsample_ratio,
+                                    heads = heads,
+                                    pc_size = pc_size,
+                                    width = width,
+                                    point_feats = point_feats,
+                                    fourier_embedder = self.fourier_embedder,
+                                    pc_sharpedge_size = pc_sharpedge_size)
+
         self.post_kl = ops.Linear(embed_dim, width)
 
         self.transformer = Transformer(
@@ -583,5 +975,14 @@ class ShapeVAE(nn.Module):
         grid_logits = self.volume_decoder(latents, self.geo_decoder, bounds=bounds, num_chunks=num_chunks, octree_resolution=octree_resolution, enable_pbar=enable_pbar)
         return grid_logits.movedim(-2, -1)
 
-    def encode(self, x):
-        return None
+    def encode(self, surface):
+
+        pc, feats = surface[:, :, :3], surface[:, :, 3:]
+        latents = self.encoder(pc, feats)
+
+        moments = self.pre_kl(latents)
+        posterior = DiagonalGaussianDistribution(moments, feature_dim = -1)
+
+        latents = posterior.sample()
+
+        return latents

comfy/ldm/hunyuan3dv2_1/hunyuandit.py

@@ -0,0 +1,659 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+class GELU(nn.Module):
+
+    def __init__(self, dim_in: int, dim_out: int, operations, device, dtype):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out, device = device, dtype = dtype)
+
+    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+
+        if gate.device.type == "mps":
+            return F.gelu(gate.to(dtype = torch.float32)).to(dtype = gate.dtype)
+
+        return F.gelu(gate)
+
+    def forward(self, hidden_states):
+
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+
+        return hidden_states
+
+class FeedForward(nn.Module):
+
+    def __init__(self, dim: int, dim_out = None, mult: int = 4,
+                dropout: float = 0.0, inner_dim = None, operations = None, device = None, dtype = None):
+
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+
+        dim_out = dim_out if dim_out is not None else dim
+
+        act_fn = GELU(dim, inner_dim, operations = operations, device = device, dtype = dtype)
+
+        self.net = nn.ModuleList([])
+        self.net.append(act_fn)
+
+        self.net.append(nn.Dropout(dropout))
+        self.net.append(operations.Linear(inner_dim, dim_out, device = device, dtype = dtype))
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+class AddAuxLoss(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x, loss):
+        # do nothing in forward (no computation)
+        ctx.requires_aux_loss = loss.requires_grad
+        ctx.dtype = loss.dtype
+
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        # add the aux loss gradients
+        grad_loss = None
+        # put the aux grad the same as the main grad loss
+        # aux grad contributes equally
+        if ctx.requires_aux_loss:
+            grad_loss = torch.ones(1, dtype = ctx.dtype, device = grad_output.device)
+
+        return grad_output, grad_loss
+
+class MoEGate(nn.Module):
+
+    def __init__(self, embed_dim, num_experts=16, num_experts_per_tok=2, aux_loss_alpha=0.01, device = None, dtype = None):
+
+        super().__init__()
+        self.top_k = num_experts_per_tok
+        self.n_routed_experts = num_experts
+
+        self.alpha = aux_loss_alpha
+
+        self.gating_dim = embed_dim
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim), device = device, dtype = dtype))
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+
+        # flatten hidden states
+        hidden_states = hidden_states.view(-1, hidden_states.size(-1))
+
+        # get logits and pass it to softmax
+        logits = F.linear(hidden_states, comfy.model_management.cast_to(self.weight, dtype=hidden_states.dtype, device=hidden_states.device), bias = None)
+        scores = logits.softmax(dim = -1)
+
+        topk_weight, topk_idx = torch.topk(scores, k = self.top_k, dim = -1, sorted = False)
+
+        if self.training and self.alpha > 0.0:
+            scores_for_aux = scores
+
+            # used bincount instead of one hot encoding
+            counts = torch.bincount(topk_idx.view(-1), minlength = self.n_routed_experts).float()
+            ce = counts / topk_idx.numel()  # normalized expert usage
+
+            # mean expert score
+            Pi = scores_for_aux.mean(0)
+
+            # expert balance loss
+            aux_loss = (Pi * ce * self.n_routed_experts).sum() * self.alpha
+        else:
+            aux_loss = None
+
+        return topk_idx, topk_weight, aux_loss
+
+class MoEBlock(nn.Module):
+    def __init__(self, dim, num_experts: int = 6, moe_top_k: int = 2, dropout: float = 0.0,
+                 ff_inner_dim: int = None, operations = None, device = None, dtype = None):
+        super().__init__()
+
+        self.moe_top_k = moe_top_k
+        self.num_experts = num_experts
+
+        self.experts = nn.ModuleList([
+            FeedForward(dim, dropout = dropout, inner_dim = ff_inner_dim, operations = operations, device = device, dtype = dtype)
+            for _ in range(num_experts)
+        ])
+
+        self.gate = MoEGate(dim, num_experts = num_experts, num_experts_per_tok = moe_top_k, device = device, dtype = dtype)
+        self.shared_experts = FeedForward(dim, dropout = dropout, inner_dim = ff_inner_dim, operations = operations, device = device, dtype = dtype)
+
+    def forward(self, hidden_states) -> torch.Tensor:
+
+        identity = hidden_states
+        orig_shape = hidden_states.shape
+        topk_idx, topk_weight, aux_loss = self.gate(hidden_states)
+
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+
+        if self.training:
+
+            hidden_states = hidden_states.repeat_interleave(self.moe_top_k, dim = 0)
+            y = torch.empty_like(hidden_states, dtype = hidden_states.dtype)
+
+            for i, expert in enumerate(self.experts):
+                tmp = expert(hidden_states[flat_topk_idx == i])
+                y[flat_topk_idx == i] = tmp.to(hidden_states.dtype)
+
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim = 1)
+            y =  y.view(*orig_shape)
+
+            y = AddAuxLoss.apply(y, aux_loss)
+        else:
+            y = self.moe_infer(hidden_states, flat_expert_indices = flat_topk_idx,flat_expert_weights = topk_weight.view(-1, 1)).view(*orig_shape)
+
+        y = y + self.shared_experts(identity)
+
+        return y
+
+    @torch.no_grad()
+    def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
+
+        expert_cache = torch.zeros_like(x)
+        idxs = flat_expert_indices.argsort()
+
+        # no need for .numpy().cpu() here
+        tokens_per_expert = flat_expert_indices.bincount().cumsum(0)
+        token_idxs = idxs // self.moe_top_k
+
+        for i, end_idx in enumerate(tokens_per_expert):
+
+            start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+
+            if start_idx == end_idx:
+                continue
+
+            expert = self.experts[i]
+            exp_token_idx = token_idxs[start_idx:end_idx]
+
+            expert_tokens = x[exp_token_idx]
+            expert_out = expert(expert_tokens)
+
+            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
+
+            # use index_add_ with a 1-D index tensor directly avoids building a large [N, D] index map and extra memcopy required by scatter_reduce_
+            # + avoid dtype conversion
+            expert_cache.index_add_(0, exp_token_idx, expert_out)
+
+        return expert_cache
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, downscale_freq_shift: float = 0.0,
+                 scale: float = 1.0, max_period: int = 10000):
+        super().__init__()
+
+        self.num_channels = num_channels
+        half_dim = num_channels // 2
+
+        # precompute the “inv_freq” vector once
+        exponent = -math.log(max_period) * torch.arange(
+            half_dim, dtype=torch.float32
+        ) / (half_dim - downscale_freq_shift)
+
+        inv_freq = torch.exp(exponent)
+
+        # pad
+        if num_channels % 2 == 1:
+            # we’ll pad a zero at the end of the cos-half
+            inv_freq = torch.cat([inv_freq, inv_freq.new_zeros(1)])
+
+        # register to buffer so it moves with the device
+        self.register_buffer("inv_freq", inv_freq, persistent = False)
+        self.scale = scale
+
+    def forward(self, timesteps: torch.Tensor):
+
+        x = timesteps.float().unsqueeze(1) * self.inv_freq.to(timesteps.device).unsqueeze(0)
+
+
+        # fused CUDA kernels for sin and cos
+        sin_emb = x.sin()
+        cos_emb = x.cos()
+
+        emb = torch.cat([sin_emb, cos_emb], dim = 1)
+
+        # scale factor
+        if self.scale != 1.0:
+            emb = emb * self.scale
+
+        # If we padded inv_freq for odd, emb is already wide enough; otherwise:
+        if emb.shape[1] > self.num_channels:
+            emb = emb[:, :self.num_channels]
+
+        return emb
+
+class TimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size = 256, cond_proj_dim = None, operations = None, device = None, dtype = None):
+        super().__init__()
+
+        self.mlp = nn.Sequential(
+            operations.Linear(hidden_size, frequency_embedding_size, bias=True, device = device, dtype = dtype),
+            nn.GELU(),
+            operations.Linear(frequency_embedding_size, hidden_size, bias=True, device = device, dtype = dtype),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+        if cond_proj_dim is not None:
+            self.cond_proj = operations.Linear(cond_proj_dim, frequency_embedding_size, bias=False, device = device, dtype = dtype)
+
+        self.time_embed = Timesteps(hidden_size)
+
+    def forward(self, timesteps, condition):
+
+        timestep_embed = self.time_embed(timesteps).type(self.mlp[0].weight.dtype)
+
+        if condition is not None:
+            cond_embed = self.cond_proj(condition)
+            timestep_embed = timestep_embed + cond_embed
+
+        time_conditioned = self.mlp(timestep_embed)
+
+        # for broadcasting with image tokens
+        return time_conditioned.unsqueeze(1)
+
+class MLP(nn.Module):
+    def __init__(self, *, width: int, operations = None, device = None, dtype = None):
+        super().__init__()
+        self.width = width
+        self.fc1 = operations.Linear(width, width * 4, device = device, dtype = dtype)
+        self.fc2 = operations.Linear(width * 4, width, device = device, dtype = dtype)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        return self.fc2(self.gelu(self.fc1(x)))
+
+class CrossAttention(nn.Module):
+    def __init__(
+        self,
+        qdim,
+        kdim,
+        num_heads,
+        qkv_bias=True,
+        qk_norm=False,
+        norm_layer=nn.LayerNorm,
+        use_fp16: bool = False,
+        operations = None,
+        dtype = None,
+        device = None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.qdim = qdim
+        self.kdim = kdim
+
+        self.num_heads = num_heads
+        self.head_dim = self.qdim // num_heads
+
+        self.scale = self.head_dim ** -0.5
+
+        self.to_q = operations.Linear(qdim, qdim, bias=qkv_bias, device = device, dtype = dtype)
+        self.to_k = operations.Linear(kdim, qdim, bias=qkv_bias, device = device, dtype = dtype)
+        self.to_v = operations.Linear(kdim, qdim, bias=qkv_bias, device = device, dtype = dtype)
+
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        if norm_layer == nn.LayerNorm:
+            norm_layer = operations.LayerNorm
+        else:
+            norm_layer = operations.RMSNorm
+
+        self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.out_proj = operations.Linear(qdim, qdim, bias=True, device = device, dtype = dtype)
+
+    def forward(self, x, y):
+
+        b, s1, _ = x.shape
+        _, s2, _ = y.shape
+
+        y = y.to(next(self.to_k.parameters()).dtype)
+
+        q = self.to_q(x)
+        k = self.to_k(y)
+        v = self.to_v(y)
+
+        kv = torch.cat((k, v), dim=-1)
+        split_size = kv.shape[-1] // self.num_heads // 2
+
+        kv = kv.view(1, -1, self.num_heads, split_size * 2)
+        k, v = torch.split(kv, split_size, dim=-1)
+
+        q = q.view(b, s1, self.num_heads, self.head_dim)
+        k = k.view(b, s2, self.num_heads, self.head_dim)
+        v = v.reshape(b, s2, self.num_heads * self.head_dim)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        x = optimized_attention(
+            q.reshape(b, s1, self.num_heads * self.head_dim),
+            k.reshape(b, s2, self.num_heads * self.head_dim),
+            v,
+            heads=self.num_heads,
+        )
+
+        out = self.out_proj(x)
+
+        return out
+
+class Attention(nn.Module):
+
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        qkv_bias = True,
+        qk_norm = False,
+        norm_layer = nn.LayerNorm,
+        use_fp16: bool = False,
+        operations = None,
+        device = None,
+        dtype = None
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = self.dim // num_heads
+        self.scale = self.head_dim ** -0.5
+
+        self.to_q = operations.Linear(dim, dim, bias = qkv_bias, device = device, dtype = dtype)
+        self.to_k = operations.Linear(dim, dim, bias = qkv_bias, device = device, dtype = dtype)
+        self.to_v = operations.Linear(dim, dim, bias = qkv_bias, device = device, dtype = dtype)
+
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        if norm_layer == nn.LayerNorm:
+            norm_layer = operations.LayerNorm
+        else:
+            norm_layer = operations.RMSNorm
+
+        self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.out_proj = operations.Linear(dim, dim, device = device, dtype = dtype)
+
+    def forward(self, x):
+        B, N, _ = x.shape
+
+        query = self.to_q(x)
+        key = self.to_k(x)
+        value = self.to_v(x)
+
+        qkv_combined = torch.cat((query, key, value), dim=-1)
+        split_size = qkv_combined.shape[-1] // self.num_heads // 3
+
+        qkv = qkv_combined.view(1, -1, self.num_heads, split_size * 3)
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        query = query.reshape(B, N, self.num_heads, self.head_dim)
+        key = key.reshape(B, N, self.num_heads, self.head_dim)
+        value = value.reshape(B, N, self.num_heads * self.head_dim)
+
+        query = self.q_norm(query)
+        key = self.k_norm(key)
+
+        x = optimized_attention(
+            query.reshape(B, N, self.num_heads * self.head_dim),
+            key.reshape(B, N, self.num_heads * self.head_dim),
+            value,
+            heads=self.num_heads,
+        )
+
+        x = self.out_proj(x)
+        return x
+
+class HunYuanDiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        c_emb_size,
+        num_heads,
+        text_states_dim=1024,
+        qk_norm=False,
+        norm_layer=nn.LayerNorm,
+        qk_norm_layer=True,
+        qkv_bias=True,
+        skip_connection=True,
+        timested_modulate=False,
+        use_moe: bool = False,
+        num_experts: int = 8,
+        moe_top_k: int = 2,
+        use_fp16: bool = False,
+        operations = None,
+        device = None, dtype = None
+    ):
+        super().__init__()
+
+        # eps can't be 1e-6 in fp16 mode because of numerical stability issues
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        self.norm1 = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+
+        self.attn1 = Attention(hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, qk_norm=qk_norm,
+                               norm_layer=qk_norm_layer, use_fp16 = use_fp16, device = device, dtype = dtype, operations = operations)
+
+        self.norm2 = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+
+        self.timested_modulate = timested_modulate
+        if self.timested_modulate:
+            self.default_modulation = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(c_emb_size, hidden_size, bias=True, device = device, dtype = dtype)
+            )
+
+        self.attn2 = CrossAttention(hidden_size, text_states_dim, num_heads=num_heads, qkv_bias=qkv_bias,
+                                    qk_norm=qk_norm, norm_layer=qk_norm_layer, use_fp16 = use_fp16,
+                                    device = device, dtype = dtype, operations = operations)
+
+        self.norm3 = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+
+        if skip_connection:
+            self.skip_norm = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+            self.skip_linear = operations.Linear(2 * hidden_size, hidden_size, device = device, dtype = dtype)
+        else:
+            self.skip_linear = None
+
+        self.use_moe = use_moe
+
+        if self.use_moe:
+            self.moe = MoEBlock(
+                hidden_size,
+                num_experts = num_experts,
+                moe_top_k = moe_top_k,
+                dropout = 0.0,
+                ff_inner_dim = int(hidden_size * 4.0),
+                device = device, dtype = dtype,
+                operations = operations
+            )
+        else:
+            self.mlp = MLP(width=hidden_size, operations=operations, device = device, dtype = dtype)
+
+    def forward(self, hidden_states, conditioning=None, text_states=None, skip_tensor=None):
+
+        if self.skip_linear is not None:
+            combined = torch.cat([skip_tensor, hidden_states], dim=-1)
+            hidden_states = self.skip_linear(combined)
+            hidden_states = self.skip_norm(hidden_states)
+
+        # self attention
+        if self.timested_modulate:
+            modulation_shift = self.default_modulation(conditioning).unsqueeze(dim=1)
+            hidden_states = hidden_states + modulation_shift
+
+        self_attn_out = self.attn1(self.norm1(hidden_states))
+        hidden_states = hidden_states + self_attn_out
+
+        # cross attention
+        hidden_states = hidden_states + self.attn2(self.norm2(hidden_states), text_states)
+
+        # MLP Layer
+        mlp_input = self.norm3(hidden_states)
+
+        if self.use_moe:
+            hidden_states = hidden_states + self.moe(mlp_input)
+        else:
+            hidden_states = hidden_states + self.mlp(mlp_input)
+
+        return hidden_states
+
+class FinalLayer(nn.Module):
+
+    def __init__(self, final_hidden_size, out_channels, operations, use_fp16: bool = False, device = None, dtype = None):
+        super().__init__()
+
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        self.norm_final = operations.LayerNorm(final_hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+        self.linear = operations.Linear(final_hidden_size, out_channels, bias = True, device = device, dtype = dtype)
+
+    def forward(self, x):
+        x = self.norm_final(x)
+        x = x[:, 1:]
+        x = self.linear(x)
+        return x
+
+class HunYuanDiTPlain(nn.Module):
+
+    # init with the defaults values from https://huggingface.co/tencent/Hunyuan3D-2.1/blob/main/hunyuan3d-dit-v2-1/config.yaml
+    def __init__(
+        self,
+        in_channels: int = 64,
+        hidden_size: int = 2048,
+        context_dim: int = 1024,
+        depth: int = 21,
+        num_heads: int = 16,
+        qk_norm: bool = True,
+        qkv_bias: bool = False,
+        num_moe_layers: int = 6,
+        guidance_cond_proj_dim = 2048,
+        norm_type = 'layer',
+        num_experts: int = 8,
+        moe_top_k: int = 2,
+        use_fp16: bool = False,
+        dtype = None,
+        device = None,
+        operations = None,
+        **kwargs
+        ):
+
+        self.dtype = dtype
+
+        super().__init__()
+
+        self.depth = depth
+
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+
+        norm = operations.LayerNorm if norm_type == 'layer' else operations.RMSNorm
+        qk_norm = operations.RMSNorm
+
+        self.context_dim = context_dim
+        self.guidance_cond_proj_dim = guidance_cond_proj_dim
+
+        self.x_embedder = operations.Linear(in_channels, hidden_size, bias = True, device = device, dtype = dtype)
+        self.t_embedder = TimestepEmbedder(hidden_size, hidden_size * 4, cond_proj_dim = guidance_cond_proj_dim, device = device, dtype = dtype, operations = operations)
+
+
+        # HUnYuanDiT Blocks
+        self.blocks = nn.ModuleList([
+            HunYuanDiTBlock(hidden_size=hidden_size,
+                            c_emb_size=hidden_size,
+                            num_heads=num_heads,
+                            text_states_dim=context_dim,
+                            qk_norm=qk_norm,
+                            norm_layer = norm,
+                            qk_norm_layer = qk_norm,
+                            skip_connection=layer > depth // 2,
+                            qkv_bias=qkv_bias,
+                            use_moe=True if depth - layer <= num_moe_layers else False,
+                            num_experts=num_experts,
+                            moe_top_k=moe_top_k,
+                            use_fp16 = use_fp16,
+                            device = device, dtype = dtype, operations = operations)
+            for layer in range(depth)
+        ])
+
+        self.depth = depth
+
+        self.final_layer = FinalLayer(hidden_size, self.out_channels, use_fp16 = use_fp16, operations = operations, device = device, dtype = dtype)
+
+    def forward(self, x, t, context, transformer_options = {}, **kwargs):
+
+        x = x.movedim(-1, -2)
+        uncond_emb, cond_emb = context.chunk(2, dim = 0)
+
+        context = torch.cat([cond_emb, uncond_emb], dim = 0)
+        main_condition = context
+
+        t = 1.0 - t
+
+        time_embedded = self.t_embedder(t, condition = kwargs.get('guidance_cond'))
+
+        x = x.to(dtype = next(self.x_embedder.parameters()).dtype)
+        x_embedded = self.x_embedder(x)
+
+        combined = torch.cat([time_embedded, x_embedded], dim=1)
+
+        def block_wrap(args):
+            return block(
+                args["x"],
+                args["t"],
+                args["cond"],
+                skip_tensor=args.get("skip"),)
+
+        skip_stack = []
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for idx, block in enumerate(self.blocks):
+            if idx <= self.depth // 2:
+                skip_input = None
+            else:
+                skip_input = skip_stack.pop()
+
+            if ("block", idx) in blocks_replace:
+
+                combined = blocks_replace[("block", idx)](
+                    {
+                        "x": combined,
+                        "t": time_embedded,
+                        "cond": main_condition,
+                        "skip": skip_input,
+                    },
+                    {"original_block": block_wrap},
+                )
+            else:
+                combined = block(combined, time_embedded, main_condition, skip_tensor=skip_input)
+
+            if idx < self.depth // 2:
+                skip_stack.append(combined)
+
+        output = self.final_layer(combined)
+        output =  output.movedim(-2, -1) * (-1.0)
+
+        cond_emb, uncond_emb = output.chunk(2, dim = 0)
+        return torch.cat([uncond_emb, cond_emb])

comfy/ldm/hunyuan_video/model.py

@@ -6,7 +6,6 @@ import comfy.ldm.flux.layers
 import comfy.ldm.modules.diffusionmodules.mmdit
 from comfy.ldm.modules.attention import optimized_attention
 
-
 from dataclasses import dataclass
 from einops import repeat
 
@@ -40,6 +39,11 @@ class HunyuanVideoParams:
     patch_size: list
     qkv_bias: bool
     guidance_embed: bool
+    byt5: bool
+    meanflow: bool
+    use_cond_type_embedding: bool
+    vision_in_dim: int
+    meanflow_sum: bool
 
 
 class SelfAttentionRef(nn.Module):
@@ -78,13 +82,13 @@ class TokenRefinerBlock(nn.Module):
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
 
-    def forward(self, x, c, mask):
+    def forward(self, x, c, mask, transformer_options={}):
         mod1, mod2 = self.adaLN_modulation(c).chunk(2, dim=1)
 
         norm_x = self.norm1(x)
         qkv = self.self_attn.qkv(norm_x)
         q, k, v = qkv.reshape(qkv.shape[0], qkv.shape[1], 3, self.heads, -1).permute(2, 0, 3, 1, 4)
-        attn = optimized_attention(q, k, v, self.heads, mask=mask, skip_reshape=True)
+        attn = optimized_attention(q, k, v, self.heads, mask=mask, skip_reshape=True, transformer_options=transformer_options)
 
         x = x + self.self_attn.proj(attn) * mod1.unsqueeze(1)
         x = x + self.mlp(self.norm2(x)) * mod2.unsqueeze(1)
@@ -115,14 +119,14 @@ class IndividualTokenRefiner(nn.Module):
             ]
         )
 
-    def forward(self, x, c, mask):
+    def forward(self, x, c, mask, transformer_options={}):
         m = None
         if mask is not None:
             m = mask.view(mask.shape[0], 1, 1, mask.shape[1]).repeat(1, 1, mask.shape[1], 1)
             m = m + m.transpose(2, 3)
 
         for block in self.blocks:
-            x = block(x, c, m)
+            x = block(x, c, m, transformer_options=transformer_options)
         return x
 
 
@@ -150,17 +154,45 @@ class TokenRefiner(nn.Module):
         x,
         timesteps,
         mask,
+        transformer_options={},
     ):
         t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype))
         # m = mask.float().unsqueeze(-1)
         # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise
-        c = x.sum(dim=1) / x.shape[1]
+        if x.dtype == torch.float16:
+            c = x.float().sum(dim=1) / x.shape[1]
+        else:
+            c = x.sum(dim=1) / x.shape[1]
 
         c = t + self.c_embedder(c.to(x.dtype))
         x = self.input_embedder(x)
-        x = self.individual_token_refiner(x, c, mask)
+        x = self.individual_token_refiner(x, c, mask, transformer_options=transformer_options)
         return x
 
+
+class ByT5Mapper(nn.Module):
+    def __init__(self, in_dim, out_dim, hidden_dim, out_dim1, use_res=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.layernorm = operations.LayerNorm(in_dim, dtype=dtype, device=device)
+        self.fc1 = operations.Linear(in_dim, hidden_dim, dtype=dtype, device=device)
+        self.fc2 = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+        self.fc3 = operations.Linear(out_dim, out_dim1, dtype=dtype, device=device)
+        self.use_res = use_res
+        self.act_fn = nn.GELU()
+
+    def forward(self, x):
+        if self.use_res:
+            res = x
+        x = self.layernorm(x)
+        x = self.fc1(x)
+        x = self.act_fn(x)
+        x = self.fc2(x)
+        x2 = self.act_fn(x)
+        x2 = self.fc3(x2)
+        if self.use_res:
+            x2 = x2 + res
+        return x2
+
 class HunyuanVideo(nn.Module):
     """
     Transformer model for flow matching on sequences.
@@ -169,11 +201,15 @@ class HunyuanVideo(nn.Module):
     def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
         params = HunyuanVideoParams(**kwargs)
         self.params = params
         self.patch_size = params.patch_size
         self.in_channels = params.in_channels
         self.out_channels = params.out_channels
+        self.use_cond_type_embedding = params.use_cond_type_embedding
+        self.vision_in_dim = params.vision_in_dim
         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}"
@@ -185,9 +221,13 @@ class HunyuanVideo(nn.Module):
         self.num_heads = params.num_heads
         self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
 
-        self.img_in = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(None, self.patch_size, self.in_channels, self.hidden_size, conv3d=True, dtype=dtype, device=device, operations=operations)
+        self.img_in = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(None, self.patch_size, self.in_channels, self.hidden_size, conv3d=len(self.patch_size) == 3, dtype=dtype, device=device, operations=operations)
         self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
-        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        if params.vec_in_dim is not None:
+            self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.vector_in = None
+
         self.guidance_in = (
             MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
         )
@@ -215,9 +255,38 @@ class HunyuanVideo(nn.Module):
             ]
         )
 
+        if params.byt5:
+            self.byt5_in = ByT5Mapper(
+                in_dim=1472,
+                out_dim=2048,
+                hidden_dim=2048,
+                out_dim1=self.hidden_size,
+                use_res=False,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.byt5_in = None
+
+        if params.meanflow:
+            self.time_r_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.time_r_in = None
+
         if final_layer:
             self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations)
 
+        # HunyuanVideo 1.5 specific modules
+        if self.vision_in_dim is not None:
+            from comfy.ldm.wan.model import MLPProj
+            self.vision_in = MLPProj(in_dim=self.vision_in_dim, out_dim=self.hidden_size, operation_settings=operation_settings)
+        else:
+            self.vision_in = None
+        if self.use_cond_type_embedding:
+            # 0: text_encoder feature 1: byt5 feature 2: vision_encoder feature
+            self.cond_type_embedding = nn.Embedding(3, self.hidden_size)
+        else:
+            self.cond_type_embedding = None
+
     def forward_orig(
         self,
         img: Tensor,
@@ -226,10 +295,13 @@ class HunyuanVideo(nn.Module):
         txt_ids: Tensor,
         txt_mask: Tensor,
         timesteps: Tensor,
-        y: Tensor,
+        y: Tensor = None,
+        txt_byt5=None,
+        clip_fea=None,
         guidance: Tensor = None,
         guiding_frame_index=None,
         ref_latent=None,
+        disable_time_r=False,
         control=None,
         transformer_options={},
     ) -> Tensor:
@@ -240,6 +312,14 @@ 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 (self.time_r_in is not None) and (not disable_time_r):
+            w = torch.where(transformer_options['sigmas'][0] == transformer_options['sample_sigmas'])[0]  # This most likely could be improved
+            if len(w) > 0:
+                timesteps_r = transformer_options['sample_sigmas'][w[0] + 1]
+                timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype)
+                vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype))
+                vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2
+
         if ref_latent is not None:
             ref_latent_ids = self.img_ids(ref_latent)
             ref_latent = self.img_in(ref_latent)
@@ -250,13 +330,17 @@ class HunyuanVideo(nn.Module):
 
         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])
-            vec = torch.cat([(vec_ + token_replace_vec).unsqueeze(1), (vec_ + vec).unsqueeze(1)], dim=1)
+            if self.vector_in is not None:
+                vec_ = self.vector_in(y[:, :self.params.vec_in_dim])
+                vec = torch.cat([(vec_ + token_replace_vec).unsqueeze(1), (vec_ + vec).unsqueeze(1)], dim=1)
+            else:
+                vec = torch.cat([(token_replace_vec).unsqueeze(1), (vec).unsqueeze(1)], dim=1)
             frame_tokens = (initial_shape[-1] // self.patch_size[-1]) * (initial_shape[-2] // self.patch_size[-2])
             modulation_dims = [(0, frame_tokens, 0), (frame_tokens, None, 1)]
             modulation_dims_txt = [(0, None, 1)]
         else:
-            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+            if self.vector_in is not None:
+                vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
             modulation_dims = None
             modulation_dims_txt = None
 
@@ -267,7 +351,32 @@ class HunyuanVideo(nn.Module):
         if txt_mask is not None and not torch.is_floating_point(txt_mask):
             txt_mask = (txt_mask - 1).to(img.dtype) * torch.finfo(img.dtype).max
 
-        txt = self.txt_in(txt, timesteps, txt_mask)
+        txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options)
+
+        if self.cond_type_embedding is not None:
+            self.cond_type_embedding.to(txt.device)
+            cond_emb = self.cond_type_embedding(torch.zeros_like(txt[:, :, 0], device=txt.device, dtype=torch.long))
+            txt = txt + cond_emb.to(txt.dtype)
+
+        if self.byt5_in is not None and txt_byt5 is not None:
+            txt_byt5 = self.byt5_in(txt_byt5)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(torch.ones_like(txt_byt5[:, :, 0], device=txt_byt5.device, dtype=torch.long))
+                txt_byt5 = txt_byt5 + cond_emb.to(txt_byt5.dtype)
+                txt = torch.cat((txt_byt5, txt), dim=1) # byt5 first for HunyuanVideo1.5
+            else:
+                txt = torch.cat((txt, txt_byt5), dim=1)
+            txt_byt5_ids = torch.zeros((txt_ids.shape[0], txt_byt5.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
+            txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1)
+
+        if clip_fea is not None:
+            txt_vision_states = self.vision_in(clip_fea)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(2 * torch.ones_like(txt_vision_states[:, :, 0], dtype=torch.long, device=txt_vision_states.device))
+                txt_vision_states = txt_vision_states + cond_emb
+            txt = torch.cat((txt_vision_states.to(txt.dtype), txt), dim=1)
+            extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
+            txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1)
 
         ids = torch.cat((img_ids, txt_ids), dim=1)
         pe = self.pe_embedder(ids)
@@ -281,18 +390,21 @@ class HunyuanVideo(nn.Module):
             attn_mask = None
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = 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["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"])
+                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt, 'transformer_options': transformer_options}, {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
             else:
-                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt)
+                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt, transformer_options=transformer_options)
 
             if control is not None: # Controlnet
                 control_i = control.get("input")
@@ -303,17 +415,20 @@ class HunyuanVideo(nn.Module):
 
         img = torch.cat((img, txt), 1)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = 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["attention_mask"], modulation_dims=args["modulation_dims"])
+                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims=args["modulation_dims"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims}, {"original_block": block_wrap})
+                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims, 'transformer_options': transformer_options}, {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims, transformer_options=transformer_options)
 
             if control is not None: # Controlnet
                 control_o = control.get("output")
@@ -328,12 +443,16 @@ class HunyuanVideo(nn.Module):
 
         img = self.final_layer(img, vec, modulation_dims=modulation_dims)  # (N, T, patch_size ** 2 * out_channels)
 
-        shape = initial_shape[-3:]
+        shape = initial_shape[-len(self.patch_size):]
         for i in range(len(shape)):
             shape[i] = shape[i] // self.patch_size[i]
         img = img.reshape([img.shape[0]] + shape + [self.out_channels] + self.patch_size)
-        img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
-        img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
+        if img.ndim == 8:
+            img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
+            img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
+        else:
+            img = img.permute(0, 3, 1, 4, 2, 5)
+            img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3])
         return img
 
     def img_ids(self, x):
@@ -348,16 +467,30 @@ class HunyuanVideo(nn.Module):
         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)
         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):
+    def img_ids_2d(self, x):
+        bs, c, h, w = x.shape
+        patch_size = self.patch_size
+        h_len = ((h + (patch_size[0] // 2)) // patch_size[0])
+        w_len = ((w + (patch_size[1] // 2)) // patch_size[1])
+        img_ids = torch.zeros((h_len, w_len, 2), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 0] = img_ids[:, :, 0] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        return repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
             self,
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, y, guidance, attention_mask, guiding_frame_index, ref_latent, control, transformer_options, **kwargs)
+        ).execute(x, timestep, context, y, txt_byt5, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
 
-    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, ref_latent, control=control, transformer_options=transformer_options)
+    def _forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+        bs = x.shape[0]
+        if len(self.patch_size) == 3:
+            img_ids = self.img_ids(x)
+            txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        else:
+            img_ids = self.img_ids_2d(x)
+            txt_ids = torch.zeros((bs, context.shape[1], 2), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, clip_fea, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options)
         return out

comfy/ldm/hunyuan_video/upsampler.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
+from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
+import model_management
+import model_patcher
+
+class SRResidualCausalBlock3D(nn.Module):
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block = nn.Sequential(
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+class SRModel3DV2(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        hidden_channels: int = 64,
+        num_blocks: int = 6,
+        global_residual: bool = False,
+    ):
+        super().__init__()
+        self.in_conv = VideoConv3d(in_channels, hidden_channels, kernel_size=3)
+        self.blocks = nn.ModuleList([SRResidualCausalBlock3D(hidden_channels) for _ in range(num_blocks)])
+        self.out_conv = VideoConv3d(hidden_channels, out_channels, kernel_size=3)
+        self.global_residual = bool(global_residual)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        y = self.in_conv(x)
+        for blk in self.blocks:
+            y = blk(y)
+        y = self.out_conv(y)
+        if self.global_residual and (y.shape == residual.shape):
+            y = y + residual
+        return y
+
+
+class Upsampler(nn.Module):
+    def __init__(
+        self,
+        z_channels: int,
+        out_channels: int,
+        block_out_channels: tuple[int, ...],
+        num_res_blocks: int = 2,
+    ):
+        super().__init__()
+        self.num_res_blocks = num_res_blocks
+        self.block_out_channels = block_out_channels
+        self.z_channels = z_channels
+
+        ch = block_out_channels[0]
+        self.conv_in = VideoConv3d(z_channels, ch, kernel_size=3)
+
+        self.up = nn.ModuleList()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                    out_channels=tgt,
+                                                    temb_channels=0,
+                                                    conv_shortcut=False,
+                                                    conv_op=VideoConv3d, norm_op=RMS_norm)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            self.up.append(stage)
+
+        self.norm_out = RMS_norm(ch)
+        self.conv_out = VideoConv3d(ch, out_channels, kernel_size=3)
+
+    def forward(self, z):
+        """
+        Args:
+            z: (B, C, T, H, W)
+            target_shape: (H, W)
+        """
+        # z to block_in
+        repeats = self.block_out_channels[0] // (self.z_channels)
+        x = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1)
+
+        # upsampling
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+
+        out = self.conv_out(F.silu(self.norm_out(x)))
+        return out
+
+UPSAMPLERS = {
+    "720p": SRModel3DV2,
+    "1080p": Upsampler,
+}
+
+class HunyuanVideo15SRModel():
+    def __init__(self, model_type, config):
+        self.load_device = model_management.vae_device()
+        offload_device = model_management.vae_offload_device()
+        self.dtype = model_management.vae_dtype(self.load_device)
+        self.model_class = UPSAMPLERS.get(model_type)
+        self.model = self.model_class(**config).eval()
+
+        self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=True)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def resample_latent(self, latent):
+        model_management.load_model_gpu(self.patcher)
+        return self.model(latent.to(self.load_device))

comfy/ldm/hunyuan_video/vae.py

@@ -0,0 +1,136 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class PixelShuffle2D(nn.Module):
+    def __init__(self, in_dim, out_dim, op=ops.Conv2d):
+        super().__init__()
+        self.conv = op(in_dim, out_dim >> 2, 3, 1, 1)
+        self.ratio = (in_dim << 2) // out_dim
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        h2, w2 = h >> 1, w >> 1
+        y = self.conv(x).view(b, -1, h2, 2, w2, 2).permute(0, 3, 5, 1, 2, 4).reshape(b, -1, h2, w2)
+        r = x.view(b, c, h2, 2, w2, 2).permute(0, 3, 5, 1, 2, 4).reshape(b, c << 2, h2, w2)
+        return y + r.view(b, y.shape[1], self.ratio, h2, w2).mean(2)
+
+
+class PixelUnshuffle2D(nn.Module):
+    def __init__(self, in_dim, out_dim, op=ops.Conv2d):
+        super().__init__()
+        self.conv = op(in_dim, out_dim << 2, 3, 1, 1)
+        self.scale = (out_dim << 2) // in_dim
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        h2, w2 = h << 1, w << 1
+        y = self.conv(x).view(b, 2, 2, -1, h, w).permute(0, 3, 4, 1, 5, 2).reshape(b, -1, h2, w2)
+        r = x.repeat_interleave(self.scale, 1).view(b, 2, 2, -1, h, w).permute(0, 3, 4, 1, 5, 2).reshape(b, -1, h2, w2)
+        return y + r
+
+
+class Encoder(nn.Module):
+    def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, downsample_match_channel=True, **_):
+        super().__init__()
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+        self.conv_in = ops.Conv2d(in_channels, block_out_channels[0], 3, 1, 1)
+
+        self.down = nn.ModuleList()
+        ch = block_out_channels[0]
+        depth = (ffactor_spatial >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=ops.Conv2d)
+                                        for j in range(num_res_blocks)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and downsample_match_channel else ch
+                stage.downsample = PixelShuffle2D(ch, nxt, ops.Conv2d)
+                ch = nxt
+            self.down.append(stage)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv2d)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+
+        self.norm_out = ops.GroupNorm(32, ch, 1e-6, True)
+        self.conv_out = ops.Conv2d(ch, z_channels << 1, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv_in(x)
+
+        for stage in self.down:
+            for blk in stage.block:
+                x = blk(x)
+            if hasattr(stage, 'downsample'):
+                x = stage.downsample(x)
+
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+
+        b, c, h, w = x.shape
+        grp = c // (self.z_channels << 1)
+        skip = x.view(b, c // grp, grp, h, w).mean(2)
+
+        return self.conv_out(F.silu(self.norm_out(x))) + skip
+
+
+class Decoder(nn.Module):
+    def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, upsample_match_channel=True, **_):
+        super().__init__()
+        block_out_channels = block_out_channels[::-1]
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+
+        ch = block_out_channels[0]
+        self.conv_in = ops.Conv2d(z_channels, ch, 3, 1, 1)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv2d)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=ops.Conv2d)
+
+        self.up = nn.ModuleList()
+        depth = (ffactor_spatial >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=ops.Conv2d)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and upsample_match_channel else ch
+                stage.upsample = PixelUnshuffle2D(ch, nxt, ops.Conv2d)
+                ch = nxt
+            self.up.append(stage)
+
+        self.norm_out = ops.GroupNorm(32, ch, 1e-6, True)
+        self.conv_out = ops.Conv2d(ch, out_channels, 3, 1, 1)
+
+    def forward(self, z):
+        x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+            if hasattr(stage, 'upsample'):
+                x = stage.upsample(x)
+
+        return self.conv_out(F.silu(self.norm_out(x)))

comfy/ldm/hunyuan_video/vae_refiner.py

@@ -0,0 +1,313 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed
+import comfy.ops
+import comfy.ldm.models.autoencoder
+import comfy.model_management
+ops = comfy.ops.disable_weight_init
+
+
+class RMS_norm(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        shape = (dim, 1, 1, 1)
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.empty(shape))
+
+    def forward(self, x):
+        return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
+
+class DnSmpl(nn.Module):
+    def __init__(self, ic, oc, tds, refiner_vae, op):
+        super().__init__()
+        fct = 2 * 2 * 2 if tds else 1 * 2 * 2
+        assert oc % fct == 0
+        self.conv = op(ic, oc // fct, kernel_size=3, stride=1, padding=1)
+        self.refiner_vae = refiner_vae
+
+        self.tds = tds
+        self.gs = fct * ic // oc
+
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+        r1 = 2 if self.tds else 1
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
+
+            hf = h[:, :, :1, :, :]
+            b, c, f, ht, wd = hf.shape
+            hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2)
+            hf = hf.permute(0, 4, 6, 1, 2, 3, 5)
+            hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2)
+            hf = torch.cat([hf, hf], dim=1)
+
+            h = h[:, :, 1:, :, :]
+
+            xf = x[:, :, :1, :, :]
+            b, ci, f, ht, wd = xf.shape
+            xf = xf.reshape(b, ci, f, ht // 2, 2, wd // 2, 2)
+            xf = xf.permute(0, 4, 6, 1, 2, 3, 5)
+            xf = xf.reshape(b, 2 * 2 * ci, f, ht // 2, wd // 2)
+            B, C, T, H, W = xf.shape
+            xf = xf.view(B, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2)
+
+            x = x[:, :, 1:, :, :]
+
+        if h.shape[2] == 0:
+            return hf + xf
+
+        b, c, frms, ht, wd = h.shape
+        nf = frms // r1
+        h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
+        h = h.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
+
+        b, ci, frms, ht, wd = x.shape
+        nf = frms // r1
+        x = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
+        x = x.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        x = x.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
+        B, C, T, H, W = x.shape
+        x = x.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
+
+
+class UpSmpl(nn.Module):
+    def __init__(self, ic, oc, tus, refiner_vae, op):
+        super().__init__()
+        fct = 2 * 2 * 2 if tus else 1 * 2 * 2
+        self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1)
+        self.refiner_vae = refiner_vae
+
+        self.tus = tus
+        self.rp = fct * oc // ic
+
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+        r1 = 2 if self.tus else 1
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+
+        if self.tus and self.refiner_vae and conv_carry_in is None:
+            hf = h[:, :, :1, :, :]
+            b, c, f, ht, wd = hf.shape
+            nc = c // (2 * 2)
+            hf = hf.reshape(b, 2, 2, nc, f, ht, wd)
+            hf = hf.permute(0, 3, 4, 5, 1, 6, 2)
+            hf = hf.reshape(b, nc, f, ht * 2, wd * 2)
+            hf = hf[:, : hf.shape[1] // 2]
+
+            h = h[:, :, 1:, :, :]
+
+            xf = x[:, :, :1, :, :]
+            b, ci, f, ht, wd = xf.shape
+            xf = xf.repeat_interleave(repeats=self.rp // 2, dim=1)
+            b, c, f, ht, wd = xf.shape
+            nc = c // (2 * 2)
+            xf = xf.reshape(b, 2, 2, nc, f, ht, wd)
+            xf = xf.permute(0, 3, 4, 5, 1, 6, 2)
+            xf = xf.reshape(b, nc, f, ht * 2, wd * 2)
+
+            x = x[:, :, 1:, :, :]
+
+        b, c, frms, ht, wd = h.shape
+        nc = c // (r1 * 2 * 2)
+        h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        h = h.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+
+        x = x.repeat_interleave(repeats=self.rp, dim=1)
+        b, c, frms, ht, wd = x.shape
+        nc = c // (r1 * 2 * 2)
+        x = x.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        x = x.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        x = x.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+
+        if self.tus and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
+
+class Encoder(nn.Module):
+    def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_):
+        super().__init__()
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+        self.ffactor_temporal = ffactor_temporal
+
+        self.refiner_vae = refiner_vae
+        if self.refiner_vae:
+            conv_op = CarriedConv3d
+            norm_op = RMS_norm
+        else:
+            conv_op = ops.Conv3d
+            norm_op = Normalize
+
+        self.conv_in = conv_op(in_channels, block_out_channels[0], 3, 1, 1)
+
+        self.down = nn.ModuleList()
+        ch = block_out_channels[0]
+        depth = (ffactor_spatial >> 1).bit_length()
+        depth_temporal = ((ffactor_spatial // self.ffactor_temporal) >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=conv_op, norm_op=norm_op)
+                                        for j in range(num_res_blocks)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and downsample_match_channel else ch
+                stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
+                ch = nxt
+            self.down.append(stage)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+
+        self.norm_out = norm_op(ch)
+        self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
+
+        self.regul = comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer()
+
+    def forward(self, x):
+        if not self.refiner_vae and x.shape[2] == 1:
+            x = x.expand(-1, -1, self.ffactor_temporal, -1, -1)
+
+        if self.refiner_vae:
+            xl = [x[:, :, :1, :, :]]
+            if x.shape[2] > self.ffactor_temporal:
+                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // self.ffactor_temporal) * self.ffactor_temporal, :, :], self.ffactor_temporal * 2, dim=2)
+            x = xl
+        else:
+            x = [x]
+        out = []
+
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+
+            x1 = [ x1 ]
+            x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
+
+            for stage in self.down:
+                for blk in stage.block:
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'downsample'):
+                    x1 = stage.downsample(x1, conv_carry_in, conv_carry_out)
+
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+
+        out = torch_cat_if_needed(out, dim=2)
+
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
+        del out
+
+        b, c, t, h, w = x.shape
+        grp = c // (self.z_channels << 1)
+        skip = x.view(b, c // grp, grp, t, h, w).mean(2)
+
+        out = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip
+
+        if self.refiner_vae:
+            out = self.regul(out)[0]
+
+        return out
+
+class Decoder(nn.Module):
+    def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
+                 ffactor_spatial, ffactor_temporal, upsample_match_channel=True, refiner_vae=True, **_):
+        super().__init__()
+        block_out_channels = block_out_channels[::-1]
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+
+        self.refiner_vae = refiner_vae
+        if self.refiner_vae:
+            conv_op = CarriedConv3d
+            norm_op = RMS_norm
+        else:
+            conv_op = ops.Conv3d
+            norm_op = Normalize
+
+        ch = block_out_channels[0]
+        self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
+
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
+
+        self.up = nn.ModuleList()
+        depth = (ffactor_spatial >> 1).bit_length()
+        depth_temporal = (ffactor_temporal >> 1).bit_length()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=conv_op, norm_op=norm_op)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            if i < depth:
+                nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and upsample_match_channel else ch
+                stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
+                ch = nxt
+            self.up.append(stage)
+
+        self.norm_out = norm_op(ch)
+        self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1)
+
+    def forward(self, z):
+        x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+
+        if self.refiner_vae:
+            x = torch.split(x, 2, dim=2)
+        else:
+            x = [ x ]
+        out = []
+
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+            for stage in self.up:
+                for blk in stage.block:
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'upsample'):
+                    x1 = stage.upsample(x1, conv_carry_in, conv_carry_out)
+
+            x1 = [ F.silu(self.norm_out(x1)) ]
+            x1 = conv_carry_causal_3d(x1, self.conv_out, conv_carry_in, conv_carry_out)
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+        del x
+
+        out = torch_cat_if_needed(out, dim=2)
+
+        if not self.refiner_vae:
+            if z.shape[-3] == 1:
+                out = out[:, :, -1:]
+
+        return out
+

comfy/ldm/kandinsky5/model.py

@@ -0,0 +1,413 @@
+import torch
+from torch import nn
+import math
+
+import comfy.ldm.common_dit
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope1
+from comfy.ldm.flux.layers import EmbedND
+
+def attention(q, k, v, heads, transformer_options={}):
+    return optimized_attention(
+        q.transpose(1, 2),
+        k.transpose(1, 2),
+        v.transpose(1, 2),
+        heads=heads,
+        skip_reshape=True,
+        transformer_options=transformer_options
+    )
+
+def apply_scale_shift_norm(norm, x, scale, shift):
+    return torch.addcmul(shift, norm(x), scale + 1.0)
+
+def apply_gate_sum(x, out, gate):
+    return torch.addcmul(x, gate, out)
+
+def get_shift_scale_gate(params):
+    shift, scale, gate = torch.chunk(params, 3, dim=-1)
+    return tuple(x.unsqueeze(1) for x in (shift, scale, gate))
+
+def get_freqs(dim, max_period=10000.0):
+    return torch.exp(-math.log(max_period) * torch.arange(start=0, end=dim, dtype=torch.float32) / dim)
+
+
+class TimeEmbeddings(nn.Module):
+    def __init__(self, model_dim, time_dim, max_period=10000.0, operation_settings=None):
+        super().__init__()
+        assert model_dim % 2 == 0
+        self.model_dim = model_dim
+        self.max_period = max_period
+        self.register_buffer("freqs", get_freqs(model_dim // 2, max_period), persistent=False)
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.activation = nn.SiLU()
+        self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, timestep, dtype):
+        args = torch.outer(timestep, self.freqs.to(device=timestep.device))
+        time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(dtype)
+        time_embed = self.out_layer(self.activation(self.in_layer(time_embed)))
+        return time_embed
+
+
+class TextEmbeddings(nn.Module):
+    def __init__(self, text_dim, model_dim, operation_settings=None):
+        super().__init__()
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(text_dim, model_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm = operations.LayerNorm(model_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, text_embed):
+        text_embed = self.in_layer(text_embed)
+        return self.norm(text_embed).type_as(text_embed)
+
+
+class VisualEmbeddings(nn.Module):
+    def __init__(self, visual_dim, model_dim, patch_size, operation_settings=None):
+        super().__init__()
+        self.patch_size = patch_size
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(visual_dim, model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        x = x.movedim(1, -1)  # B C T H W -> B T H W C
+        B, T, H, W, dim = x.shape
+        pt, ph, pw = self.patch_size
+
+        x = x.view(
+            B,
+            T // pt, pt,
+            H // ph, ph,
+            W // pw, pw,
+            dim,
+        ).permute(0, 1, 3, 5, 2, 4, 6, 7).flatten(4, 7)
+
+        return self.in_layer(x)
+
+
+class Modulation(nn.Module):
+    def __init__(self, time_dim, model_dim, num_params, operation_settings=None):
+        super().__init__()
+        self.activation = nn.SiLU()
+        self.out_layer = operation_settings.get("operations").Linear(time_dim, num_params * model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        return self.out_layer(self.activation(x))
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, num_channels, head_dim, operation_settings=None):
+        super().__init__()
+        assert num_channels % head_dim == 0
+        self.num_heads = num_channels // head_dim
+        self.head_dim = head_dim
+
+        operations = operation_settings.get("operations")
+        self.to_query = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_key = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_value = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.query_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.key_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.out_layer = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.num_chunks = 2
+
+    def _compute_qk(self, x, freqs, proj_fn, norm_fn):
+        result = proj_fn(x).view(*x.shape[:-1], self.num_heads, -1)
+        return apply_rope1(norm_fn(result), freqs)
+
+    def _forward(self, x, freqs, transformer_options={}):
+        q = self._compute_qk(x, freqs, self.to_query, self.query_norm)
+        k = self._compute_qk(x, freqs, self.to_key, self.key_norm)
+        v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1)
+        out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        return self.out_layer(out)
+
+    def _forward_chunked(self, x, freqs, transformer_options={}):
+        def process_chunks(proj_fn, norm_fn):
+            x_chunks = torch.chunk(x, self.num_chunks, dim=1)
+            freqs_chunks = torch.chunk(freqs, self.num_chunks, dim=1)
+            chunks = []
+            for x_chunk, freqs_chunk in zip(x_chunks, freqs_chunks):
+                chunks.append(self._compute_qk(x_chunk, freqs_chunk, proj_fn, norm_fn))
+            return torch.cat(chunks, dim=1)
+
+        q = process_chunks(self.to_query, self.query_norm)
+        k = process_chunks(self.to_key, self.key_norm)
+        v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1)
+        out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        return self.out_layer(out)
+
+    def forward(self, x, freqs, transformer_options={}):
+        if x.shape[1] > 8192:
+            return self._forward_chunked(x, freqs, transformer_options=transformer_options)
+        else:
+            return self._forward(x, freqs, transformer_options=transformer_options)
+
+
+class CrossAttention(SelfAttention):
+    def get_qkv(self, x, context):
+        q = self.to_query(x).view(*x.shape[:-1], self.num_heads, -1)
+        k = self.to_key(context).view(*context.shape[:-1], self.num_heads, -1)
+        v = self.to_value(context).view(*context.shape[:-1], self.num_heads, -1)
+        return q, k, v
+
+    def forward(self, x, context, transformer_options={}):
+        q, k, v = self.get_qkv(x, context)
+        out = attention(self.query_norm(q), self.key_norm(k), v, self.num_heads, transformer_options=transformer_options)
+        return self.out_layer(out)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, ff_dim, operation_settings=None):
+        super().__init__()
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(dim, ff_dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.activation = nn.GELU()
+        self.out_layer = operations.Linear(ff_dim, dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.num_chunks = 4
+
+    def _forward(self, x):
+        return self.out_layer(self.activation(self.in_layer(x)))
+
+    def _forward_chunked(self, x):
+        chunks = torch.chunk(x, self.num_chunks, dim=1)
+        output_chunks = []
+        for chunk in chunks:
+            output_chunks.append(self._forward(chunk))
+        return torch.cat(output_chunks, dim=1)
+
+    def forward(self, x):
+        if x.shape[1] > 8192:
+            return self._forward_chunked(x)
+        else:
+            return self._forward(x)
+
+
+class OutLayer(nn.Module):
+    def __init__(self, model_dim, time_dim, visual_dim, patch_size, operation_settings=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.modulation = Modulation(time_dim, model_dim, 2, operation_settings=operation_settings)
+        operations = operation_settings.get("operations")
+        self.norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.out_layer = operations.Linear(model_dim, math.prod(patch_size) * visual_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, visual_embed, time_embed):
+        B, T, H, W, _ = visual_embed.shape
+        shift, scale = torch.chunk(self.modulation(time_embed), 2, dim=-1)
+        scale = scale[:, None, None, None, :]
+        shift = shift[:, None, None, None, :]
+        visual_embed = apply_scale_shift_norm(self.norm, visual_embed, scale, shift)
+        x = self.out_layer(visual_embed)
+
+        out_dim = x.shape[-1] // (self.patch_size[0] * self.patch_size[1] * self.patch_size[2])
+        x = x.view(
+            B, T, H, W,
+            out_dim,
+            self.patch_size[0], self.patch_size[1], self.patch_size[2]
+        )
+        return x.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(2, 3).flatten(3, 4).flatten(4, 5)
+
+
+class TransformerEncoderBlock(nn.Module):
+    def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None):
+        super().__init__()
+        self.text_modulation = Modulation(time_dim, model_dim, 6, operation_settings=operation_settings)
+        operations = operation_settings.get("operations")
+
+        self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
+
+    def forward(self, x, time_embed, freqs, transformer_options={}):
+        self_attn_params, ff_params = torch.chunk(self.text_modulation(time_embed), 2, dim=-1)
+        shift, scale, gate = get_shift_scale_gate(self_attn_params)
+        out = apply_scale_shift_norm(self.self_attention_norm, x, scale, shift)
+        out = self.self_attention(out, freqs, transformer_options=transformer_options)
+        x = apply_gate_sum(x, out, gate)
+
+        shift, scale, gate = get_shift_scale_gate(ff_params)
+        out = apply_scale_shift_norm(self.feed_forward_norm, x, scale, shift)
+        out = self.feed_forward(out)
+        x = apply_gate_sum(x, out, gate)
+        return x
+
+
+class TransformerDecoderBlock(nn.Module):
+    def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None):
+        super().__init__()
+        self.visual_modulation = Modulation(time_dim, model_dim, 9, operation_settings=operation_settings)
+
+        operations = operation_settings.get("operations")
+        self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.cross_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.cross_attention = CrossAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
+
+    def forward(self, visual_embed, text_embed, time_embed, freqs, transformer_options={}):
+        self_attn_params, cross_attn_params, ff_params = torch.chunk(self.visual_modulation(time_embed), 3, dim=-1)
+        # self attention
+        shift, scale, gate = get_shift_scale_gate(self_attn_params)
+        visual_out = apply_scale_shift_norm(self.self_attention_norm, visual_embed, scale, shift)
+        visual_out = self.self_attention(visual_out, freqs, transformer_options=transformer_options)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        # cross attention
+        shift, scale, gate = get_shift_scale_gate(cross_attn_params)
+        visual_out = apply_scale_shift_norm(self.cross_attention_norm, visual_embed, scale, shift)
+        visual_out = self.cross_attention(visual_out, text_embed, transformer_options=transformer_options)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        # feed forward
+        shift, scale, gate = get_shift_scale_gate(ff_params)
+        visual_out = apply_scale_shift_norm(self.feed_forward_norm, visual_embed, scale, shift)
+        visual_out = self.feed_forward(visual_out)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        return visual_embed
+
+
+class Kandinsky5(nn.Module):
+    def __init__(
+        self,
+        in_visual_dim=16, out_visual_dim=16, in_text_dim=3584, in_text_dim2=768, time_dim=512,
+        model_dim=1792, ff_dim=7168, visual_embed_dim=132, patch_size=(1, 2, 2), num_text_blocks=2, num_visual_blocks=32,
+        axes_dims=(16, 24, 24), rope_scale_factor=(1.0, 2.0, 2.0),
+        dtype=None, device=None, operations=None, **kwargs
+    ):
+        super().__init__()
+        head_dim = sum(axes_dims)
+        self.rope_scale_factor = rope_scale_factor
+        self.in_visual_dim = in_visual_dim
+        self.model_dim = model_dim
+        self.patch_size = patch_size
+        self.visual_embed_dim = visual_embed_dim
+        self.dtype = dtype
+        self.device = device
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.time_embeddings = TimeEmbeddings(model_dim, time_dim, operation_settings=operation_settings)
+        self.text_embeddings = TextEmbeddings(in_text_dim, model_dim, operation_settings=operation_settings)
+        self.pooled_text_embeddings = TextEmbeddings(in_text_dim2, time_dim, operation_settings=operation_settings)
+        self.visual_embeddings = VisualEmbeddings(visual_embed_dim, model_dim, patch_size, operation_settings=operation_settings)
+
+        self.text_transformer_blocks = nn.ModuleList(
+            [TransformerEncoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_text_blocks)]
+        )
+
+        self.visual_transformer_blocks = nn.ModuleList(
+            [TransformerDecoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_visual_blocks)]
+        )
+
+        self.out_layer = OutLayer(model_dim, time_dim, out_visual_dim, patch_size, operation_settings=operation_settings)
+
+        self.rope_embedder_3d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=axes_dims)
+        self.rope_embedder_1d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=[head_dim])
+
+    def rope_encode_1d(self, seq_len, seq_start=0, steps=None, device=None, dtype=None, transformer_options={}):
+        steps = seq_len if steps is None else steps
+        seq_ids = torch.linspace(seq_start, seq_start + (seq_len - 1), steps=steps, device=device, dtype=dtype)
+        seq_ids = seq_ids.reshape(-1, 1).unsqueeze(0)  # Shape: (1, steps, 1)
+        freqs = self.rope_embedder_1d(seq_ids).movedim(1, 2)
+        return freqs
+
+    def rope_encode_3d(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}):
+
+        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 steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        h_start = 0
+        w_start = 0
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            t_start += rope_options.get("shift_t", 0.0)
+            h_start += rope_options.get("shift_y", 0.0)
+            w_start += rope_options.get("shift_x", 0.0)
+        else:
+            rope_scale_factor = self.rope_scale_factor
+            if self.model_dim == 4096: # pro video model uses different rope scaling at higher resolutions
+                if h * w >= 14080:
+                    rope_scale_factor = (1.0, 3.16, 3.16)
+
+            t_len = (t_len - 1.0) / rope_scale_factor[0] + 1.0
+            h_len = (h_len - 1.0) / rope_scale_factor[1] + 1.0
+            w_len = (w_len - 1.0) / rope_scale_factor[2] + 1.0
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder_3d(img_ids).movedim(1, 2)
+        return freqs
+
+    def forward_orig(self, x, timestep, context, y, freqs, freqs_text, transformer_options={}, **kwargs):
+        patches_replace = transformer_options.get("patches_replace", {})
+        context = self.text_embeddings(context)
+        time_embed = self.time_embeddings(timestep, x.dtype) + self.pooled_text_embeddings(y)
+
+        for block in self.text_transformer_blocks:
+            context = block(context, time_embed, freqs_text, transformer_options=transformer_options)
+
+        visual_embed = self.visual_embeddings(x)
+        visual_shape = visual_embed.shape[:-1]
+        visual_embed = visual_embed.flatten(1, -2)
+
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.visual_transformer_blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.visual_transformer_blocks):
+            transformer_options["block_index"] = i
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    return block(x=args["x"], context=args["context"], time_embed=args["time_embed"], freqs=args["freqs"], transformer_options=args.get("transformer_options"))
+                visual_embed = blocks_replace[("double_block", i)]({"x": visual_embed, "context": context, "time_embed": time_embed, "freqs": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})["x"]
+            else:
+                visual_embed = block(visual_embed, context, time_embed, freqs=freqs, transformer_options=transformer_options)
+
+        visual_embed = visual_embed.reshape(*visual_shape, -1)
+        return self.out_layer(visual_embed, time_embed)
+
+    def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
+        original_dims = x.ndim
+        if original_dims == 4:
+            x = x.unsqueeze(2)
+        bs, c, t_len, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        if time_dim_replace is not None:
+            time_dim_replace = comfy.ldm.common_dit.pad_to_patch_size(time_dim_replace, self.patch_size)
+            x[:, :time_dim_replace.shape[1], :time_dim_replace.shape[2]] = time_dim_replace
+
+        freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+        freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+
+        out = self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs)
+        if original_dims == 4:
+            out = out.squeeze(2)
+        return out
+
+    def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, time_dim_replace=time_dim_replace, transformer_options=transformer_options, **kwargs)

comfy/ldm/lightricks/av_model.py

@@ -0,0 +1,837 @@
+from typing import Tuple
+import torch
+import torch.nn as nn
+from comfy.ldm.lightricks.model import (
+    CrossAttention,
+    FeedForward,
+    AdaLayerNormSingle,
+    PixArtAlphaTextProjection,
+    LTXVModel,
+)
+from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
+import comfy.ldm.common_dit
+
+class BasicAVTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        v_dim,
+        a_dim,
+        v_heads,
+        a_heads,
+        vd_head,
+        ad_head,
+        v_context_dim=None,
+        a_context_dim=None,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        self.attn_precision = attn_precision
+
+        self.attn1 = CrossAttention(
+            query_dim=v_dim,
+            heads=v_heads,
+            dim_head=vd_head,
+            context_dim=None,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.audio_attn1 = CrossAttention(
+            query_dim=a_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            context_dim=None,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.attn2 = CrossAttention(
+            query_dim=v_dim,
+            context_dim=v_context_dim,
+            heads=v_heads,
+            dim_head=vd_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.audio_attn2 = CrossAttention(
+            query_dim=a_dim,
+            context_dim=a_context_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Q: Video, K,V: Audio
+        self.audio_to_video_attn = CrossAttention(
+            query_dim=v_dim,
+            context_dim=a_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Q: Audio, K,V: Video
+        self.video_to_audio_attn = CrossAttention(
+            query_dim=a_dim,
+            context_dim=v_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.ff = FeedForward(
+            v_dim, dim_out=v_dim, glu=True, dtype=dtype, device=device, operations=operations
+        )
+        self.audio_ff = FeedForward(
+            a_dim, dim_out=a_dim, glu=True, dtype=dtype, device=device, operations=operations
+        )
+
+        self.scale_shift_table = nn.Parameter(torch.empty(6, v_dim, device=device, dtype=dtype))
+        self.audio_scale_shift_table = nn.Parameter(
+            torch.empty(6, a_dim, device=device, dtype=dtype)
+        )
+
+        self.scale_shift_table_a2v_ca_audio = nn.Parameter(
+            torch.empty(5, a_dim, device=device, dtype=dtype)
+        )
+        self.scale_shift_table_a2v_ca_video = nn.Parameter(
+            torch.empty(5, v_dim, device=device, dtype=dtype)
+        )
+
+    def get_ada_values(
+        self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None)
+    ):
+        num_ada_params = scale_shift_table.shape[0]
+
+        ada_values = (
+            scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype)
+            + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :]
+        ).unbind(dim=2)
+        return ada_values
+
+    def get_av_ca_ada_values(
+        self,
+        scale_shift_table: torch.Tensor,
+        batch_size: int,
+        scale_shift_timestep: torch.Tensor,
+        gate_timestep: torch.Tensor,
+        num_scale_shift_values: int = 4,
+    ):
+        scale_shift_ada_values = self.get_ada_values(
+            scale_shift_table[:num_scale_shift_values, :],
+            batch_size,
+            scale_shift_timestep,
+        )
+        gate_ada_values = self.get_ada_values(
+            scale_shift_table[num_scale_shift_values:, :],
+            batch_size,
+            gate_timestep,
+        )
+
+        scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values]
+        gate_ada_values = [t.squeeze(2) for t in gate_ada_values]
+
+        return (*scale_shift_chunks, *gate_ada_values)
+
+    def forward(
+        self,
+        x: Tuple[torch.Tensor, torch.Tensor],
+        v_context=None,
+        a_context=None,
+        attention_mask=None,
+        v_timestep=None,
+        a_timestep=None,
+        v_pe=None,
+        a_pe=None,
+        v_cross_pe=None,
+        a_cross_pe=None,
+        v_cross_scale_shift_timestep=None,
+        a_cross_scale_shift_timestep=None,
+        v_cross_gate_timestep=None,
+        a_cross_gate_timestep=None,
+        transformer_options=None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        run_vx = transformer_options.get("run_vx", True)
+        run_ax = transformer_options.get("run_ax", True)
+
+        vx, ax = x
+        run_ax = run_ax and ax.numel() > 0
+        run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0
+        run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True)
+
+        if run_vx:
+            vshift_msa, vscale_msa, vgate_msa = (
+                self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 3))
+            )
+
+            norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa
+            vx += self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) * vgate_msa
+            vx += self.attn2(
+                comfy.ldm.common_dit.rms_norm(vx),
+                context=v_context,
+                mask=attention_mask,
+                transformer_options=transformer_options,
+            )
+
+            del vshift_msa, vscale_msa, vgate_msa
+
+        if run_ax:
+            ashift_msa, ascale_msa, agate_msa = (
+                self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 3))
+            )
+
+            norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa
+            ax += (
+                self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options)
+                * agate_msa
+            )
+            ax += self.audio_attn2(
+                comfy.ldm.common_dit.rms_norm(ax),
+                context=a_context,
+                mask=attention_mask,
+                transformer_options=transformer_options,
+            )
+
+            del ashift_msa, ascale_msa, agate_msa
+
+        # Audio - Video cross attention.
+        if run_a2v or run_v2a:
+            # norm3
+            vx_norm3 = comfy.ldm.common_dit.rms_norm(vx)
+            ax_norm3 = comfy.ldm.common_dit.rms_norm(ax)
+
+            (
+                scale_ca_audio_hidden_states_a2v,
+                shift_ca_audio_hidden_states_a2v,
+                scale_ca_audio_hidden_states_v2a,
+                shift_ca_audio_hidden_states_v2a,
+                gate_out_v2a,
+            ) = self.get_av_ca_ada_values(
+                self.scale_shift_table_a2v_ca_audio,
+                ax.shape[0],
+                a_cross_scale_shift_timestep,
+                a_cross_gate_timestep,
+            )
+
+            (
+                scale_ca_video_hidden_states_a2v,
+                shift_ca_video_hidden_states_a2v,
+                scale_ca_video_hidden_states_v2a,
+                shift_ca_video_hidden_states_v2a,
+                gate_out_a2v,
+            ) = self.get_av_ca_ada_values(
+                self.scale_shift_table_a2v_ca_video,
+                vx.shape[0],
+                v_cross_scale_shift_timestep,
+                v_cross_gate_timestep,
+            )
+
+            if run_a2v:
+                vx_scaled = (
+                    vx_norm3 * (1 + scale_ca_video_hidden_states_a2v)
+                    + shift_ca_video_hidden_states_a2v
+                )
+                ax_scaled = (
+                    ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v)
+                    + shift_ca_audio_hidden_states_a2v
+                )
+                vx += (
+                    self.audio_to_video_attn(
+                        vx_scaled,
+                        context=ax_scaled,
+                        pe=v_cross_pe,
+                        k_pe=a_cross_pe,
+                        transformer_options=transformer_options,
+                    )
+                    * gate_out_a2v
+                )
+
+                del gate_out_a2v
+                del scale_ca_video_hidden_states_a2v,\
+                    shift_ca_video_hidden_states_a2v,\
+                    scale_ca_audio_hidden_states_a2v,\
+                    shift_ca_audio_hidden_states_a2v,\
+
+            if run_v2a:
+                ax_scaled = (
+                    ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a)
+                    + shift_ca_audio_hidden_states_v2a
+                )
+                vx_scaled = (
+                    vx_norm3 * (1 + scale_ca_video_hidden_states_v2a)
+                    + shift_ca_video_hidden_states_v2a
+                )
+                ax += (
+                    self.video_to_audio_attn(
+                        ax_scaled,
+                        context=vx_scaled,
+                        pe=a_cross_pe,
+                        k_pe=v_cross_pe,
+                        transformer_options=transformer_options,
+                    )
+                    * gate_out_v2a
+                )
+
+                del gate_out_v2a
+                del scale_ca_video_hidden_states_v2a,\
+                    shift_ca_video_hidden_states_v2a,\
+                    scale_ca_audio_hidden_states_v2a,\
+                    shift_ca_audio_hidden_states_v2a
+
+        if run_vx:
+            vshift_mlp, vscale_mlp, vgate_mlp = (
+                self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, None))
+            )
+
+            vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp
+            vx += self.ff(vx_scaled) * vgate_mlp
+            del vshift_mlp, vscale_mlp, vgate_mlp
+
+        if run_ax:
+            ashift_mlp, ascale_mlp, agate_mlp = (
+                self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, None))
+            )
+
+            ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp
+            ax += self.audio_ff(ax_scaled) * agate_mlp
+
+            del ashift_mlp, ascale_mlp, agate_mlp
+
+
+        return vx, ax
+
+
+class LTXAVModel(LTXVModel):
+    """LTXAV model for audio-video generation."""
+
+    def __init__(
+        self,
+        in_channels=128,
+        audio_in_channels=128,
+        cross_attention_dim=4096,
+        audio_cross_attention_dim=2048,
+        attention_head_dim=128,
+        audio_attention_head_dim=64,
+        num_attention_heads=32,
+        audio_num_attention_heads=32,
+        caption_channels=3840,
+        num_layers=48,
+        positional_embedding_theta=10000.0,
+        positional_embedding_max_pos=[20, 2048, 2048],
+        audio_positional_embedding_max_pos=[20],
+        causal_temporal_positioning=False,
+        vae_scale_factors=(8, 32, 32),
+        use_middle_indices_grid=False,
+        timestep_scale_multiplier=1000.0,
+        av_ca_timestep_scale_multiplier=1.0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        # Store audio-specific parameters
+        self.audio_in_channels = audio_in_channels
+        self.audio_cross_attention_dim = audio_cross_attention_dim
+        self.audio_attention_head_dim = audio_attention_head_dim
+        self.audio_num_attention_heads = audio_num_attention_heads
+        self.audio_positional_embedding_max_pos = audio_positional_embedding_max_pos
+
+        # Calculate audio dimensions
+        self.audio_inner_dim = audio_num_attention_heads * audio_attention_head_dim
+        self.audio_out_channels = audio_in_channels
+
+        # Audio-specific constants
+        self.num_audio_channels = 8
+        self.audio_frequency_bins = 16
+
+        self.av_ca_timestep_scale_multiplier = av_ca_timestep_scale_multiplier
+
+        super().__init__(
+            in_channels=in_channels,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+            caption_channels=caption_channels,
+            num_layers=num_layers,
+            positional_embedding_theta=positional_embedding_theta,
+            positional_embedding_max_pos=positional_embedding_max_pos,
+            causal_temporal_positioning=causal_temporal_positioning,
+            vae_scale_factors=vae_scale_factors,
+            use_middle_indices_grid=use_middle_indices_grid,
+            timestep_scale_multiplier=timestep_scale_multiplier,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+            **kwargs,
+        )
+
+    def _init_model_components(self, device, dtype, **kwargs):
+        """Initialize LTXAV-specific components."""
+        # Audio-specific projections
+        self.audio_patchify_proj = self.operations.Linear(
+            self.audio_in_channels, self.audio_inner_dim, bias=True, dtype=dtype, device=device
+        )
+
+        # Audio-specific AdaLN
+        self.audio_adaln_single = AdaLayerNormSingle(
+            self.audio_inner_dim,
+            use_additional_conditions=False,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+        num_scale_shift_values = 4
+        self.av_ca_video_scale_shift_adaln_single = AdaLayerNormSingle(
+            self.inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=num_scale_shift_values,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+        self.av_ca_a2v_gate_adaln_single = AdaLayerNormSingle(
+            self.inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=1,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+        self.av_ca_audio_scale_shift_adaln_single = AdaLayerNormSingle(
+            self.audio_inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=num_scale_shift_values,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+        self.av_ca_v2a_gate_adaln_single = AdaLayerNormSingle(
+            self.audio_inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=1,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+        # Audio caption projection
+        self.audio_caption_projection = PixArtAlphaTextProjection(
+            in_features=self.caption_channels,
+            hidden_size=self.audio_inner_dim,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+    def _init_transformer_blocks(self, device, dtype, **kwargs):
+        """Initialize transformer blocks for LTXAV."""
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicAVTransformerBlock(
+                    v_dim=self.inner_dim,
+                    a_dim=self.audio_inner_dim,
+                    v_heads=self.num_attention_heads,
+                    a_heads=self.audio_num_attention_heads,
+                    vd_head=self.attention_head_dim,
+                    ad_head=self.audio_attention_head_dim,
+                    v_context_dim=self.cross_attention_dim,
+                    a_context_dim=self.audio_cross_attention_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=self.operations,
+                )
+                for _ in range(self.num_layers)
+            ]
+        )
+
+    def _init_output_components(self, device, dtype):
+        """Initialize output components for LTXAV."""
+        # Video output components
+        super()._init_output_components(device, dtype)
+        # Audio output components
+        self.audio_scale_shift_table = nn.Parameter(
+            torch.empty(2, self.audio_inner_dim, dtype=dtype, device=device)
+        )
+        self.audio_norm_out = self.operations.LayerNorm(
+            self.audio_inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
+        )
+        self.audio_proj_out = self.operations.Linear(
+            self.audio_inner_dim, self.audio_out_channels, dtype=dtype, device=device
+        )
+        self.a_patchifier = AudioPatchifier(1, start_end=True)
+
+    def separate_audio_and_video_latents(self, x, audio_length):
+        """Separate audio and video latents from combined input."""
+        # vx = x[:, : self.in_channels]
+        # ax = x[:, self.in_channels :]
+        #
+        # ax = ax.reshape(ax.shape[0], -1)
+        # ax = ax[:, : audio_length * self.num_audio_channels * self.audio_frequency_bins]
+        #
+        # ax = ax.reshape(
+        #     ax.shape[0], self.num_audio_channels, audio_length, self.audio_frequency_bins
+        # )
+
+        vx = x[0]
+        ax = x[1] if len(x) > 1 else torch.zeros(
+            (vx.shape[0], self.num_audio_channels, 0, self.audio_frequency_bins),
+            device=vx.device, dtype=vx.dtype
+        )
+        return vx, ax
+
+    def recombine_audio_and_video_latents(self, vx, ax, target_shape=None):
+        if ax.numel() == 0:
+            return vx
+        else:
+            return [vx, ax]
+        """Recombine audio and video latents for output."""
+        # if ax.device != vx.device or ax.dtype != vx.dtype:
+        #     logging.warning("Audio and video latents are on different devices or dtypes.")
+        #     ax = ax.to(device=vx.device, dtype=vx.dtype)
+        #     logging.warning(f"Audio audio latent moved to device: {ax.device}, dtype: {ax.dtype}")
+        #
+        # ax = ax.reshape(ax.shape[0], -1)
+        # # pad to f x h x w of the video latents
+        # divisor = vx.shape[-1] * vx.shape[-2] * vx.shape[-3]
+        # if target_shape is None:
+        #     repetitions = math.ceil(ax.shape[-1] / divisor)
+        # else:
+        #     repetitions = target_shape[1] - vx.shape[1]
+        # padded_len = repetitions * divisor
+        # ax = F.pad(ax, (0, padded_len - ax.shape[-1]))
+        # ax = ax.reshape(ax.shape[0], -1, vx.shape[-3], vx.shape[-2], vx.shape[-1])
+        # return torch.cat([vx, ax], dim=1)
+
+    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
+        """Process input for LTXAV - separate audio and video, then patchify."""
+        audio_length = kwargs.get("audio_length", 0)
+        # Separate audio and video latents
+        vx, ax = self.separate_audio_and_video_latents(x, audio_length)
+        [vx, v_pixel_coords, additional_args] = super()._process_input(
+            vx, keyframe_idxs, denoise_mask, **kwargs
+        )
+
+        ax, a_latent_coords = self.a_patchifier.patchify(ax)
+        ax = self.audio_patchify_proj(ax)
+
+        # additional_args.update({"av_orig_shape": list(x.shape)})
+        return [vx, ax], [v_pixel_coords, a_latent_coords], additional_args
+
+    def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs):
+        """Prepare timestep embeddings."""
+        # TODO: some code reuse is needed here.
+        grid_mask = kwargs.get("grid_mask", None)
+        if grid_mask is not None:
+            timestep = timestep[:, grid_mask]
+
+        timestep = timestep * self.timestep_scale_multiplier
+        v_timestep, v_embedded_timestep = self.adaln_single(
+            timestep.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=hidden_dtype,
+        )
+
+        # Second dimension is 1 or number of tokens (if timestep_per_token)
+        v_timestep = v_timestep.view(batch_size, -1, v_timestep.shape[-1])
+        v_embedded_timestep = v_embedded_timestep.view(
+            batch_size, -1, v_embedded_timestep.shape[-1]
+        )
+
+        # Prepare audio timestep
+        a_timestep = kwargs.get("a_timestep")
+        if a_timestep is not None:
+            a_timestep = a_timestep * self.timestep_scale_multiplier
+            av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier
+
+            av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single(
+                a_timestep.flatten(),
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single(
+                timestep.flatten(),
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single(
+                timestep.flatten() * av_ca_factor,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single(
+                a_timestep.flatten() * av_ca_factor,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+
+            a_timestep, a_embedded_timestep = self.audio_adaln_single(
+                a_timestep.flatten(),
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1])
+            a_embedded_timestep = a_embedded_timestep.view(
+                batch_size, -1, a_embedded_timestep.shape[-1]
+            )
+            cross_av_timestep_ss = [
+                av_ca_audio_scale_shift_timestep,
+                av_ca_video_scale_shift_timestep,
+                av_ca_a2v_gate_noise_timestep,
+                av_ca_v2a_gate_noise_timestep,
+            ]
+            cross_av_timestep_ss = list(
+                [t.view(batch_size, -1, t.shape[-1]) for t in cross_av_timestep_ss]
+            )
+        else:
+            a_timestep = timestep
+            a_embedded_timestep = kwargs.get("embedded_timestep")
+            cross_av_timestep_ss = []
+
+        return [v_timestep, a_timestep, cross_av_timestep_ss], [
+            v_embedded_timestep,
+            a_embedded_timestep,
+        ]
+
+    def _prepare_context(self, context, batch_size, x, attention_mask=None):
+        vx = x[0]
+        ax = x[1]
+        v_context, a_context = torch.split(
+            context, int(context.shape[-1] / 2), len(context.shape) - 1
+        )
+
+        v_context, attention_mask = super()._prepare_context(
+            v_context, batch_size, vx, attention_mask
+        )
+        if self.audio_caption_projection is not None:
+            a_context = self.audio_caption_projection(a_context)
+            a_context = a_context.view(batch_size, -1, ax.shape[-1])
+
+        return [v_context, a_context], attention_mask
+
+    def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype):
+        v_pixel_coords = pixel_coords[0]
+        v_pe = super()._prepare_positional_embeddings(v_pixel_coords, frame_rate, x_dtype)
+
+        a_latent_coords = pixel_coords[1]
+        a_pe = self._precompute_freqs_cis(
+            a_latent_coords,
+            dim=self.audio_inner_dim,
+            out_dtype=x_dtype,
+            max_pos=self.audio_positional_embedding_max_pos,
+            use_middle_indices_grid=self.use_middle_indices_grid,
+            num_attention_heads=self.audio_num_attention_heads,
+        )
+
+        # calculate positional embeddings for the middle of the token duration, to use in av cross attention layers.
+        max_pos = max(
+            self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0]
+        )
+        v_pixel_coords = v_pixel_coords.to(torch.float32)
+        v_pixel_coords[:, 0] = v_pixel_coords[:, 0] * (1.0 / frame_rate)
+        av_cross_video_freq_cis = self._precompute_freqs_cis(
+            v_pixel_coords[:, 0:1, :],
+            dim=self.audio_cross_attention_dim,
+            out_dtype=x_dtype,
+            max_pos=[max_pos],
+            use_middle_indices_grid=True,
+            num_attention_heads=self.audio_num_attention_heads,
+        )
+        av_cross_audio_freq_cis = self._precompute_freqs_cis(
+            a_latent_coords[:, 0:1, :],
+            dim=self.audio_cross_attention_dim,
+            out_dtype=x_dtype,
+            max_pos=[max_pos],
+            use_middle_indices_grid=True,
+            num_attention_heads=self.audio_num_attention_heads,
+        )
+
+        return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)]
+
+    def _process_transformer_blocks(
+        self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs
+    ):
+        vx = x[0]
+        ax = x[1]
+        v_context = context[0]
+        a_context = context[1]
+        v_timestep = timestep[0]
+        a_timestep = timestep[1]
+        v_pe, av_cross_video_freq_cis = pe[0]
+        a_pe, av_cross_audio_freq_cis = pe[1]
+
+        (
+            av_ca_audio_scale_shift_timestep,
+            av_ca_video_scale_shift_timestep,
+            av_ca_a2v_gate_noise_timestep,
+            av_ca_v2a_gate_noise_timestep,
+        ) = timestep[2]
+
+        """Process transformer blocks for LTXAV."""
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+
+        # Process transformer blocks
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(
+                        args["img"],
+                        v_context=args["v_context"],
+                        a_context=args["a_context"],
+                        attention_mask=args["attention_mask"],
+                        v_timestep=args["v_timestep"],
+                        a_timestep=args["a_timestep"],
+                        v_pe=args["v_pe"],
+                        a_pe=args["a_pe"],
+                        v_cross_pe=args["v_cross_pe"],
+                        a_cross_pe=args["a_cross_pe"],
+                        v_cross_scale_shift_timestep=args["v_cross_scale_shift_timestep"],
+                        a_cross_scale_shift_timestep=args["a_cross_scale_shift_timestep"],
+                        v_cross_gate_timestep=args["v_cross_gate_timestep"],
+                        a_cross_gate_timestep=args["a_cross_gate_timestep"],
+                        transformer_options=args["transformer_options"],
+                    )
+                    return out
+
+                out = blocks_replace[("double_block", i)](
+                    {
+                        "img": (vx, ax),
+                        "v_context": v_context,
+                        "a_context": a_context,
+                        "attention_mask": attention_mask,
+                        "v_timestep": v_timestep,
+                        "a_timestep": a_timestep,
+                        "v_pe": v_pe,
+                        "a_pe": a_pe,
+                        "v_cross_pe": av_cross_video_freq_cis,
+                        "a_cross_pe": av_cross_audio_freq_cis,
+                        "v_cross_scale_shift_timestep": av_ca_video_scale_shift_timestep,
+                        "a_cross_scale_shift_timestep": av_ca_audio_scale_shift_timestep,
+                        "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep,
+                        "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep,
+                        "transformer_options": transformer_options,
+                    },
+                    {"original_block": block_wrap},
+                )
+                vx, ax = out["img"]
+            else:
+                vx, ax = block(
+                    (vx, ax),
+                    v_context=v_context,
+                    a_context=a_context,
+                    attention_mask=attention_mask,
+                    v_timestep=v_timestep,
+                    a_timestep=a_timestep,
+                    v_pe=v_pe,
+                    a_pe=a_pe,
+                    v_cross_pe=av_cross_video_freq_cis,
+                    a_cross_pe=av_cross_audio_freq_cis,
+                    v_cross_scale_shift_timestep=av_ca_video_scale_shift_timestep,
+                    a_cross_scale_shift_timestep=av_ca_audio_scale_shift_timestep,
+                    v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep,
+                    a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep,
+                    transformer_options=transformer_options,
+                )
+
+        return [vx, ax]
+
+    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
+        vx = x[0]
+        ax = x[1]
+        v_embedded_timestep = embedded_timestep[0]
+        a_embedded_timestep = embedded_timestep[1]
+        vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs)
+
+        # Process audio output
+        a_scale_shift_values = (
+            self.audio_scale_shift_table[None, None].to(device=a_embedded_timestep.device, dtype=a_embedded_timestep.dtype)
+            + a_embedded_timestep[:, :, None]
+        )
+        a_shift, a_scale = a_scale_shift_values[:, :, 0], a_scale_shift_values[:, :, 1]
+
+        ax = self.audio_norm_out(ax)
+        ax = ax * (1 + a_scale) + a_shift
+        ax = self.audio_proj_out(ax)
+
+        # Unpatchify audio
+        ax = self.a_patchifier.unpatchify(
+            ax, channels=self.num_audio_channels, freq=self.audio_frequency_bins
+        )
+
+        # Recombine audio and video
+        original_shape = kwargs.get("av_orig_shape")
+        return self.recombine_audio_and_video_latents(vx, ax, original_shape)
+
+    def forward(
+        self,
+        x,
+        timestep,
+        context,
+        attention_mask=None,
+        frame_rate=25,
+        transformer_options={},
+        keyframe_idxs=None,
+        **kwargs,
+    ):
+        """
+        Forward pass for LTXAV model.
+
+        Args:
+            x: Combined audio-video input tensor
+            timestep: Tuple of (video_timestep, audio_timestep) or single timestep
+            context: Context tensor (e.g., text embeddings)
+            attention_mask: Attention mask tensor
+            frame_rate: Frame rate for temporal processing
+            transformer_options: Additional options for transformer blocks
+            keyframe_idxs: Keyframe indices for temporal processing
+            **kwargs: Additional keyword arguments including audio_length
+
+        Returns:
+            Combined audio-video output tensor
+        """
+        # Handle timestep format
+        if isinstance(timestep, (tuple, list)) and len(timestep) == 2:
+            v_timestep, a_timestep = timestep
+            kwargs["a_timestep"] = a_timestep
+            timestep = v_timestep
+        else:
+            kwargs["a_timestep"] = timestep
+
+        # Call parent forward method
+        return super().forward(
+            x,
+            timestep,
+            context,
+            attention_mask,
+            frame_rate,
+            transformer_options,
+            keyframe_idxs,
+            **kwargs,
+        )

comfy/ldm/lightricks/embeddings_connector.py

@@ -0,0 +1,305 @@
+import math
+from typing import Optional
+
+import comfy.ldm.common_dit
+import torch
+from comfy.ldm.lightricks.model import (
+    CrossAttention,
+    FeedForward,
+    generate_freq_grid_np,
+    interleaved_freqs_cis,
+    split_freqs_cis,
+)
+from torch import nn
+
+
+class BasicTransformerBlock1D(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        standardization_norm (`str`, *optional*, defaults to `"layer_norm"`): The type of pre-normalization to use. Can be `"layer_norm"` or `"rms_norm"`.
+        norm_eps (`float`, *optional*, defaults to 1e-5): Epsilon value for normalization layers.
+        qk_norm (`str`, *optional*, defaults to None):
+            Set to 'layer_norm' or `rms_norm` to perform query and key normalization.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        ff_inner_dim (`int`, *optional*): Dimension of the inner feed-forward layer. If not provided, defaults to `dim * 4`.
+        ff_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the feed-forward layer.
+        attention_out_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the attention output layer.
+        use_rope (`bool`, *optional*, defaults to `False`): Whether to use Rotary Position Embeddings (RoPE).
+        ffn_dim_mult (`int`, *optional*, defaults to 4): Multiplier for the inner dimension of the feed-forward layer.
+    """
+
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        context_dim=None,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.attn1 = CrossAttention(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            context_dim=None,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # 3. Feed-forward
+        self.ff = FeedForward(
+            dim,
+            dim_out=dim,
+            glu=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def forward(self, hidden_states, attention_mask=None, pe=None) -> torch.FloatTensor:
+
+        # Notice that normalization is always applied before the real computation in the following blocks.
+
+        # 1. Normalization Before Self-Attention
+        norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states)
+
+        norm_hidden_states = norm_hidden_states.squeeze(1)
+
+        # 2. Self-Attention
+        attn_output = self.attn1(norm_hidden_states, mask=attention_mask, pe=pe)
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 3. Normalization before Feed-Forward
+        norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states)
+
+        # 4. Feed-forward
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class Embeddings1DConnector(nn.Module):
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels=128,
+        cross_attention_dim=2048,
+        attention_head_dim=128,
+        num_attention_heads=30,
+        num_layers=2,
+        positional_embedding_theta=10000.0,
+        positional_embedding_max_pos=[4096],
+        causal_temporal_positioning=False,
+        num_learnable_registers: Optional[int] = 128,
+        dtype=None,
+        device=None,
+        operations=None,
+        split_rope=False,
+        double_precision_rope=False,
+        **kwargs,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.out_channels = in_channels
+        self.num_attention_heads = num_attention_heads
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.causal_temporal_positioning = causal_temporal_positioning
+        self.positional_embedding_theta = positional_embedding_theta
+        self.positional_embedding_max_pos = positional_embedding_max_pos
+        self.split_rope = split_rope
+        self.double_precision_rope = double_precision_rope
+        self.transformer_1d_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock1D(
+                    self.inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    context_dim=cross_attention_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        inner_dim = num_attention_heads * attention_head_dim
+        self.num_learnable_registers = num_learnable_registers
+        if self.num_learnable_registers:
+            self.learnable_registers = nn.Parameter(
+                torch.rand(
+                    self.num_learnable_registers, inner_dim, dtype=dtype, device=device
+                )
+                * 2.0
+                - 1.0
+            )
+
+    def get_fractional_positions(self, indices_grid):
+        fractional_positions = torch.stack(
+            [
+                indices_grid[:, i] / self.positional_embedding_max_pos[i]
+                for i in range(1)
+            ],
+            dim=-1,
+        )
+        return fractional_positions
+
+    def precompute_freqs(self, indices_grid, spacing):
+        source_dtype = indices_grid.dtype
+        dtype = (
+            torch.float32
+            if source_dtype in (torch.bfloat16, torch.float16)
+            else source_dtype
+        )
+
+        fractional_positions = self.get_fractional_positions(indices_grid)
+        indices = (
+            generate_freq_grid_np(
+                self.positional_embedding_theta,
+                indices_grid.shape[1],
+                self.inner_dim,
+            )
+            if self.double_precision_rope
+            else self.generate_freq_grid(spacing, dtype, fractional_positions.device)
+        ).to(device=fractional_positions.device)
+
+        if spacing == "exp_2":
+            freqs = (
+                (indices * fractional_positions.unsqueeze(-1))
+                .transpose(-1, -2)
+                .flatten(2)
+            )
+        else:
+            freqs = (
+                (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
+                .transpose(-1, -2)
+                .flatten(2)
+            )
+        return freqs
+
+    def generate_freq_grid(self, spacing, dtype, device):
+        dim = self.inner_dim
+        theta = self.positional_embedding_theta
+        n_pos_dims = 1
+        n_elem = 2 * n_pos_dims  # 2 for cos and sin e.g. x 3 = 6
+        start = 1
+        end = theta
+
+        if spacing == "exp":
+            indices = theta ** (torch.arange(0, dim, n_elem, device="cpu", dtype=torch.float32) / (dim - n_elem))
+            indices = indices.to(dtype=dtype, device=device)
+        elif spacing == "exp_2":
+            indices = 1.0 / theta ** (torch.arange(0, dim, n_elem, device=device) / dim)
+            indices = indices.to(dtype=dtype)
+        elif spacing == "linear":
+            indices = torch.linspace(
+                start, end, dim // n_elem, device=device, dtype=dtype
+            )
+        elif spacing == "sqrt":
+            indices = torch.linspace(
+                start**2, end**2, dim // n_elem, device=device, dtype=dtype
+            ).sqrt()
+
+        indices = indices * math.pi / 2
+
+        return indices
+
+    def precompute_freqs_cis(self, indices_grid, spacing="exp"):
+        dim = self.inner_dim
+        n_elem = 2  # 2 because of cos and sin
+        freqs = self.precompute_freqs(indices_grid, spacing)
+        if self.split_rope:
+            expected_freqs = dim // 2
+            current_freqs = freqs.shape[-1]
+            pad_size = expected_freqs - current_freqs
+            cos_freq, sin_freq = split_freqs_cis(
+                freqs, pad_size, self.num_attention_heads
+            )
+        else:
+            cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem)
+        return cos_freq.to(self.dtype), sin_freq.to(self.dtype), self.split_rope
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            indices_grid (`torch.LongTensor` of shape `(batch size, 3, num latent pixels)`):
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # 1. Input
+
+        if self.num_learnable_registers:
+            num_registers_duplications = math.ceil(
+                max(1024, hidden_states.shape[1]) / self.num_learnable_registers
+            )
+            learnable_registers = torch.tile(
+                self.learnable_registers.to(hidden_states), (num_registers_duplications, 1)
+            )
+
+            hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1)
+
+            if attention_mask is not None:
+                attention_mask = torch.zeros([1, 1, 1, hidden_states.shape[1]], dtype=attention_mask.dtype, device=attention_mask.device)
+
+        indices_grid = torch.arange(
+            hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device
+        )
+        indices_grid = indices_grid[None, None, :]
+        freqs_cis = self.precompute_freqs_cis(indices_grid)
+
+        # 2. Blocks
+        for block_idx, block in enumerate(self.transformer_1d_blocks):
+            hidden_states = block(
+                hidden_states, attention_mask=attention_mask, pe=freqs_cis
+            )
+
+        # 3. Output
+        # if self.output_scale is not None:
+        #     hidden_states = hidden_states / self.output_scale
+
+        hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states)
+
+        return hidden_states, attention_mask

comfy/ldm/lightricks/latent_upsampler.py

@@ -0,0 +1,292 @@
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+def _rational_for_scale(scale: float) -> Tuple[int, int]:
+    mapping = {0.75: (3, 4), 1.5: (3, 2), 2.0: (2, 1), 4.0: (4, 1)}
+    if float(scale) not in mapping:
+        raise ValueError(
+            f"Unsupported spatial_scale {scale}. Choose from {list(mapping.keys())}"
+        )
+    return mapping[float(scale)]
+
+
+class PixelShuffleND(nn.Module):
+    def __init__(self, dims, upscale_factors=(2, 2, 2)):
+        super().__init__()
+        assert dims in [1, 2, 3], "dims must be 1, 2, or 3"
+        self.dims = dims
+        self.upscale_factors = upscale_factors
+
+    def forward(self, x):
+        if self.dims == 3:
+            return rearrange(
+                x,
+                "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+                p1=self.upscale_factors[0],
+                p2=self.upscale_factors[1],
+                p3=self.upscale_factors[2],
+            )
+        elif self.dims == 2:
+            return rearrange(
+                x,
+                "b (c p1 p2) h w -> b c (h p1) (w p2)",
+                p1=self.upscale_factors[0],
+                p2=self.upscale_factors[1],
+            )
+        elif self.dims == 1:
+            return rearrange(
+                x,
+                "b (c p1) f h w -> b c (f p1) h w",
+                p1=self.upscale_factors[0],
+            )
+
+
+class BlurDownsample(nn.Module):
+    """
+    Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel.
+    Applies only on H,W. Works for dims=2 or dims=3 (per-frame).
+    """
+
+    def __init__(self, dims: int, stride: int):
+        super().__init__()
+        assert dims in (2, 3)
+        assert stride >= 1 and isinstance(stride, int)
+        self.dims = dims
+        self.stride = stride
+
+        # 5x5 separable binomial kernel [1,4,6,4,1] (outer product), normalized
+        k = torch.tensor([1.0, 4.0, 6.0, 4.0, 1.0])
+        k2d = k[:, None] @ k[None, :]
+        k2d = (k2d / k2d.sum()).float()  # shape (5,5)
+        self.register_buffer("kernel", k2d[None, None, :, :])  # (1,1,5,5)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.stride == 1:
+            return x
+
+        def _apply_2d(x2d: torch.Tensor) -> torch.Tensor:
+            # x2d: (B, C, H, W)
+            B, C, H, W = x2d.shape
+            weight = self.kernel.expand(C, 1, 5, 5)  # depthwise
+            x2d = F.conv2d(
+                x2d, weight=weight, bias=None, stride=self.stride, padding=2, groups=C
+            )
+            return x2d
+
+        if self.dims == 2:
+            return _apply_2d(x)
+        else:
+            # dims == 3: apply per-frame on H,W
+            b, c, f, h, w = x.shape
+            x = rearrange(x, "b c f h w -> (b f) c h w")
+            x = _apply_2d(x)
+            h2, w2 = x.shape[-2:]
+            x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f, h=h2, w=w2)
+            return x
+
+
+class SpatialRationalResampler(nn.Module):
+    """
+    Fully-learned rational spatial scaling: up by 'num' via PixelShuffle, then anti-aliased
+    downsample by 'den' using fixed blur + stride. Operates on H,W only.
+
+    For dims==3, work per-frame for spatial scaling (temporal axis untouched).
+    """
+
+    def __init__(self, mid_channels: int, scale: float):
+        super().__init__()
+        self.scale = float(scale)
+        self.num, self.den = _rational_for_scale(self.scale)
+        self.conv = nn.Conv2d(
+            mid_channels, (self.num**2) * mid_channels, kernel_size=3, padding=1
+        )
+        self.pixel_shuffle = PixelShuffleND(2, upscale_factors=(self.num, self.num))
+        self.blur_down = BlurDownsample(dims=2, stride=self.den)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        b, c, f, h, w = x.shape
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = self.conv(x)
+        x = self.pixel_shuffle(x)
+        x = self.blur_down(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f)
+        return x
+
+
+class ResBlock(nn.Module):
+    def __init__(
+        self, channels: int, mid_channels: Optional[int] = None, dims: int = 3
+    ):
+        super().__init__()
+        if mid_channels is None:
+            mid_channels = channels
+
+        Conv = nn.Conv2d if dims == 2 else nn.Conv3d
+
+        self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1)
+        self.norm1 = nn.GroupNorm(32, mid_channels)
+        self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1)
+        self.norm2 = nn.GroupNorm(32, channels)
+        self.activation = nn.SiLU()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.activation(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        x = self.activation(x + residual)
+        return x
+
+
+class LatentUpsampler(nn.Module):
+    """
+    Model to spatially upsample VAE latents.
+
+    Args:
+        in_channels (`int`): Number of channels in the input latent
+        mid_channels (`int`): Number of channels in the middle layers
+        num_blocks_per_stage (`int`): Number of ResBlocks to use in each stage (pre/post upsampling)
+        dims (`int`): Number of dimensions for convolutions (2 or 3)
+        spatial_upsample (`bool`): Whether to spatially upsample the latent
+        temporal_upsample (`bool`): Whether to temporally upsample the latent
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 128,
+        mid_channels: int = 512,
+        num_blocks_per_stage: int = 4,
+        dims: int = 3,
+        spatial_upsample: bool = True,
+        temporal_upsample: bool = False,
+        spatial_scale: float = 2.0,
+        rational_resampler: bool = False,
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.mid_channels = mid_channels
+        self.num_blocks_per_stage = num_blocks_per_stage
+        self.dims = dims
+        self.spatial_upsample = spatial_upsample
+        self.temporal_upsample = temporal_upsample
+        self.spatial_scale = float(spatial_scale)
+        self.rational_resampler = rational_resampler
+
+        Conv = nn.Conv2d if dims == 2 else nn.Conv3d
+
+        self.initial_conv = Conv(in_channels, mid_channels, kernel_size=3, padding=1)
+        self.initial_norm = nn.GroupNorm(32, mid_channels)
+        self.initial_activation = nn.SiLU()
+
+        self.res_blocks = nn.ModuleList(
+            [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]
+        )
+
+        if spatial_upsample and temporal_upsample:
+            self.upsampler = nn.Sequential(
+                nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1),
+                PixelShuffleND(3),
+            )
+        elif spatial_upsample:
+            if rational_resampler:
+                self.upsampler = SpatialRationalResampler(
+                    mid_channels=mid_channels, scale=self.spatial_scale
+                )
+            else:
+                self.upsampler = nn.Sequential(
+                    nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1),
+                    PixelShuffleND(2),
+                )
+        elif temporal_upsample:
+            self.upsampler = nn.Sequential(
+                nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1),
+                PixelShuffleND(1),
+            )
+        else:
+            raise ValueError(
+                "Either spatial_upsample or temporal_upsample must be True"
+            )
+
+        self.post_upsample_res_blocks = nn.ModuleList(
+            [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]
+        )
+
+        self.final_conv = Conv(mid_channels, in_channels, kernel_size=3, padding=1)
+
+    def forward(self, latent: torch.Tensor) -> torch.Tensor:
+        b, c, f, h, w = latent.shape
+
+        if self.dims == 2:
+            x = rearrange(latent, "b c f h w -> (b f) c h w")
+            x = self.initial_conv(x)
+            x = self.initial_norm(x)
+            x = self.initial_activation(x)
+
+            for block in self.res_blocks:
+                x = block(x)
+
+            x = self.upsampler(x)
+
+            for block in self.post_upsample_res_blocks:
+                x = block(x)
+
+            x = self.final_conv(x)
+            x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f)
+        else:
+            x = self.initial_conv(latent)
+            x = self.initial_norm(x)
+            x = self.initial_activation(x)
+
+            for block in self.res_blocks:
+                x = block(x)
+
+            if self.temporal_upsample:
+                x = self.upsampler(x)
+                x = x[:, :, 1:, :, :]
+            else:
+                if isinstance(self.upsampler, SpatialRationalResampler):
+                    x = self.upsampler(x)
+                else:
+                    x = rearrange(x, "b c f h w -> (b f) c h w")
+                    x = self.upsampler(x)
+                    x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f)
+
+            for block in self.post_upsample_res_blocks:
+                x = block(x)
+
+            x = self.final_conv(x)
+
+        return x
+
+    @classmethod
+    def from_config(cls, config):
+        return cls(
+            in_channels=config.get("in_channels", 4),
+            mid_channels=config.get("mid_channels", 128),
+            num_blocks_per_stage=config.get("num_blocks_per_stage", 4),
+            dims=config.get("dims", 2),
+            spatial_upsample=config.get("spatial_upsample", True),
+            temporal_upsample=config.get("temporal_upsample", False),
+            spatial_scale=config.get("spatial_scale", 2.0),
+            rational_resampler=config.get("rational_resampler", False),
+        )
+
+    def config(self):
+        return {
+            "_class_name": "LatentUpsampler",
+            "in_channels": self.in_channels,
+            "mid_channels": self.mid_channels,
+            "num_blocks_per_stage": self.num_blocks_per_stage,
+            "dims": self.dims,
+            "spatial_upsample": self.spatial_upsample,
+            "temporal_upsample": self.temporal_upsample,
+            "spatial_scale": self.spatial_scale,
+            "rational_resampler": self.rational_resampler,
+        }

comfy/ldm/lightricks/model.py

@@ -1,14 +1,47 @@
+from abc import ABC, abstractmethod
+from enum import Enum
+import functools
+import math
+from typing import Dict, Optional, Tuple
+
+from einops import rearrange
+import numpy as np
 import torch
 from torch import nn
 import comfy.patcher_extension
 import comfy.ldm.modules.attention
 import comfy.ldm.common_dit
-from einops import rearrange
-import math
-from typing import Dict, Optional, Tuple
 
 from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
 
+def _log_base(x, base):
+    return np.log(x) / np.log(base)
+
+class LTXRopeType(str, Enum):
+    INTERLEAVED = "interleaved"
+    SPLIT = "split"
+
+    KEY = "rope_type"
+
+    @classmethod
+    def from_dict(cls, kwargs, default=None):
+        if default is None:
+            default = cls.INTERLEAVED
+        return cls(kwargs.get(cls.KEY, default))
+
+
+class LTXFrequenciesPrecision(str, Enum):
+    FLOAT32 = "float32"
+    FLOAT64 = "float64"
+
+    KEY = "frequencies_precision"
+
+    @classmethod
+    def from_dict(cls, kwargs, default=None):
+        if default is None:
+            default = cls.FLOAT32
+        return cls(kwargs.get(cls.KEY, default))
+
 
 def get_timestep_embedding(
     timesteps: torch.Tensor,
@@ -40,9 +73,7 @@ def get_timestep_embedding(
     assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
 
     half_dim = embedding_dim // 2
-    exponent = -math.log(max_period) * torch.arange(
-        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
-    )
+    exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device)
     exponent = exponent / (half_dim - downscale_freq_shift)
 
     emb = torch.exp(exponent)
@@ -74,7 +105,9 @@ class TimestepEmbedding(nn.Module):
         post_act_fn: Optional[str] = None,
         cond_proj_dim=None,
         sample_proj_bias=True,
-        dtype=None, device=None, operations=None,
+        dtype=None,
+        device=None,
+        operations=None,
     ):
         super().__init__()
 
@@ -91,7 +124,9 @@ class TimestepEmbedding(nn.Module):
             time_embed_dim_out = out_dim
         else:
             time_embed_dim_out = time_embed_dim
-        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(
+            time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device
+        )
 
         if post_act_fn is None:
             self.post_act = None
@@ -140,12 +175,22 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
     https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
     """
 
-    def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
+    def __init__(
+        self,
+        embedding_dim,
+        size_emb_dim,
+        use_additional_conditions: bool = False,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
         super().__init__()
 
         self.outdim = size_emb_dim
         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=embedding_dim, dtype=dtype, device=device, operations=operations)
+        self.timestep_embedder = TimestepEmbedding(
+            in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations
+        )
 
     def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
         timesteps_proj = self.time_proj(timestep)
@@ -164,15 +209,22 @@ class AdaLayerNormSingle(nn.Module):
         use_additional_conditions (`bool`): To use additional conditions for normalization or not.
     """
 
-    def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
+    def __init__(
+        self, embedding_dim: int, embedding_coefficient: int = 6, use_additional_conditions: bool = False, dtype=None, device=None, operations=None
+    ):
         super().__init__()
 
         self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
-            embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations
+            embedding_dim,
+            size_emb_dim=embedding_dim // 3,
+            use_additional_conditions=use_additional_conditions,
+            dtype=dtype,
+            device=device,
+            operations=operations,
         )
 
         self.silu = nn.SiLU()
-        self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device)
+        self.linear = operations.Linear(embedding_dim, embedding_coefficient * embedding_dim, bias=True, dtype=dtype, device=device)
 
     def forward(
         self,
@@ -186,6 +238,7 @@ class AdaLayerNormSingle(nn.Module):
         embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
         return self.linear(self.silu(embedded_timestep)), embedded_timestep
 
+
 class PixArtAlphaTextProjection(nn.Module):
     """
     Projects caption embeddings. Also handles dropout for classifier-free guidance.
@@ -193,18 +246,24 @@ class PixArtAlphaTextProjection(nn.Module):
     Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
     """
 
-    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None):
+    def __init__(
+        self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None
+    ):
         super().__init__()
         if out_features is None:
             out_features = hidden_size
-        self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device)
+        self.linear_1 = operations.Linear(
+            in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device
+        )
         if act_fn == "gelu_tanh":
             self.act_1 = nn.GELU(approximate="tanh")
         elif act_fn == "silu":
             self.act_1 = nn.SiLU()
         else:
             raise ValueError(f"Unknown activation function: {act_fn}")
-        self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(
+            in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device
+        )
 
     def forward(self, caption):
         hidden_states = self.linear_1(caption)
@@ -223,25 +282,28 @@ class GELU_approx(nn.Module):
 
 
 class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None):
+    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0.0, dtype=None, device=None, operations=None):
         super().__init__()
         inner_dim = int(dim * mult)
         project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations)
 
         self.net = nn.Sequential(
-            project_in,
-            nn.Dropout(dropout),
-            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
+            project_in, nn.Dropout(dropout), operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
         )
 
     def forward(self, x):
         return self.net(x)
 
+def apply_rotary_emb(input_tensor, freqs_cis):
+    cos_freqs, sin_freqs = freqs_cis[0], freqs_cis[1]
+    split_pe = freqs_cis[2] if len(freqs_cis) > 2 else False
+    return (
+        apply_split_rotary_emb(input_tensor, cos_freqs, sin_freqs)
+        if split_pe else
+        apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs)
+    )
 
-def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and pick the best/fastest one
-    cos_freqs = freqs_cis[0]
-    sin_freqs = freqs_cis[1]
-
+def apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs):  # TODO: remove duplicate funcs and pick the best/fastest one
     t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2)
     t1, t2 = t_dup.unbind(dim=-1)
     t_dup = torch.stack((-t2, t1), dim=-1)
@@ -251,9 +313,37 @@ def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and
 
     return out
 
+def apply_split_rotary_emb(input_tensor, cos, sin):
+    needs_reshape = False
+    if input_tensor.ndim != 4 and cos.ndim == 4:
+        B, H, T, _ = cos.shape
+        input_tensor = input_tensor.reshape(B, T, H, -1).swapaxes(1, 2)
+        needs_reshape = True
+    split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2)
+    first_half_input = split_input[..., :1, :]
+    second_half_input = split_input[..., 1:, :]
+    output = split_input * cos.unsqueeze(-2)
+    first_half_output = output[..., :1, :]
+    second_half_output = output[..., 1:, :]
+    first_half_output.addcmul_(-sin.unsqueeze(-2), second_half_input)
+    second_half_output.addcmul_(sin.unsqueeze(-2), first_half_input)
+    output = rearrange(output, "... d r -> ... (d r)")
+    return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output
+
 
 class CrossAttention(nn.Module):
-    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None):
+    def __init__(
+        self,
+        query_dim,
+        context_dim=None,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
         super().__init__()
         inner_dim = dim_head * heads
         context_dim = query_dim if context_dim is None else context_dim
@@ -269,9 +359,11 @@ class CrossAttention(nn.Module):
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
 
-        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
+        self.to_out = nn.Sequential(
+            operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout)
+        )
 
-    def forward(self, x, context=None, mask=None, pe=None):
+    def forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}):
         q = self.to_q(x)
         context = x if context is None else context
         k = self.to_k(context)
@@ -282,156 +374,505 @@ class CrossAttention(nn.Module):
 
         if pe is not None:
             q = apply_rotary_emb(q, pe)
-            k = apply_rotary_emb(k, pe)
+            k = apply_rotary_emb(k, pe if k_pe is None else k_pe)
 
         if mask is None:
-            out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
+            out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
         else:
-            out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
+            out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options)
         return self.to_out(out)
 
 
 class BasicTransformerBlock(nn.Module):
-    def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None):
+    def __init__(
+        self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None
+    ):
         super().__init__()
 
         self.attn_precision = attn_precision
-        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
+        self.attn1 = CrossAttention(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            context_dim=None,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
         self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations)
 
-        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
+        self.attn2 = CrossAttention(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
 
         self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype))
 
-    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None):
+    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}):
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2)
 
-        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe) * gate_msa
+        attn1_input = comfy.ldm.common_dit.rms_norm(x)
+        attn1_input = torch.addcmul(attn1_input, attn1_input, scale_msa).add_(shift_msa)
+        attn1_input = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options)
+        x.addcmul_(attn1_input, gate_msa)
+        del attn1_input
 
-        x += self.attn2(x, context=context, mask=attention_mask)
+        x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options)
 
-        y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
-        x += self.ff(y) * gate_mlp
+        y = comfy.ldm.common_dit.rms_norm(x)
+        y = torch.addcmul(y, y, scale_mlp).add_(shift_mlp)
+        x.addcmul_(self.ff(y), gate_mlp)
 
         return x
 
 def get_fractional_positions(indices_grid, max_pos):
+    n_pos_dims = indices_grid.shape[1]
+    assert n_pos_dims == len(max_pos), f'Number of position dimensions ({n_pos_dims}) must match max_pos length ({len(max_pos)})'
     fractional_positions = torch.stack(
-        [
-            indices_grid[:, i] / max_pos[i]
-            for i in range(3)
-        ],
-        dim=-1,
+        [indices_grid[:, i] / max_pos[i] for i in range(n_pos_dims)],
+        axis=-1,
     )
     return fractional_positions
 
 
-def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]):
-    dtype = torch.float32 #self.dtype
-
-    fractional_positions = get_fractional_positions(indices_grid, max_pos)
-
+@functools.lru_cache(maxsize=5)
+def generate_freq_grid_np(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, _ = None):
+    theta = positional_embedding_theta
     start = 1
     end = theta
-    device = fractional_positions.device
+
+    n_elem = 2 * positional_embedding_max_pos_count
+    pow_indices = np.power(
+        theta,
+        np.linspace(
+            _log_base(start, theta),
+            _log_base(end, theta),
+            inner_dim // n_elem,
+            dtype=np.float64,
+        ),
+    )
+    return torch.tensor(pow_indices * math.pi / 2, dtype=torch.float32)
+
+def generate_freq_grid_pytorch(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, device):
+    theta = positional_embedding_theta
+    start = 1
+    end = theta
+    n_elem = 2 * positional_embedding_max_pos_count
 
     indices = theta ** (
         torch.linspace(
             math.log(start, theta),
             math.log(end, theta),
-            dim // 6,
+            inner_dim // n_elem,
             device=device,
-            dtype=dtype,
+            dtype=torch.float32,
         )
     )
-    indices = indices.to(dtype=dtype)
+    indices = indices.to(dtype=torch.float32)
 
     indices = indices * math.pi / 2
 
+    return indices
+
+def generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid):
+    if use_middle_indices_grid:
+        assert(len(indices_grid.shape) == 4 and indices_grid.shape[-1] ==2)
+        indices_grid_start, indices_grid_end = indices_grid[..., 0], indices_grid[..., 1]
+        indices_grid = (indices_grid_start + indices_grid_end) / 2.0
+    elif len(indices_grid.shape) == 4:
+        indices_grid = indices_grid[..., 0]
+
+    # Get fractional positions and compute frequency indices
+    fractional_positions = get_fractional_positions(indices_grid, max_pos)
+    indices = indices.to(device=fractional_positions.device)
+
     freqs = (
         (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
         .transpose(-1, -2)
         .flatten(2)
     )
+    return freqs
 
+def interleaved_freqs_cis(freqs, pad_size):
     cos_freq = freqs.cos().repeat_interleave(2, dim=-1)
     sin_freq = freqs.sin().repeat_interleave(2, dim=-1)
-    if dim % 6 != 0:
-        cos_padding = torch.ones_like(cos_freq[:, :, : dim % 6])
-        sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
+    if pad_size != 0:
+        cos_padding = torch.ones_like(cos_freq[:, :, : pad_size])
+        sin_padding = torch.zeros_like(cos_freq[:, :, : pad_size])
         cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
         sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
-    return cos_freq.to(out_dtype), sin_freq.to(out_dtype)
+    return cos_freq, sin_freq
 
+def split_freqs_cis(freqs, pad_size, num_attention_heads):
+    cos_freq = freqs.cos()
+    sin_freq = freqs.sin()
 
-class LTXVModel(torch.nn.Module):
-    def __init__(self,
-                 in_channels=128,
-                 cross_attention_dim=2048,
-                 attention_head_dim=64,
-                 num_attention_heads=32,
+    if pad_size != 0:
+        cos_padding = torch.ones_like(cos_freq[:, :, :pad_size])
+        sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size])
 
-                 caption_channels=4096,
-                 num_layers=28,
+        cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1)
+        sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1)
 
+    # Reshape freqs to be compatible with multi-head attention
+    B , T, half_HD = cos_freq.shape
 
-                 positional_embedding_theta=10000.0,
-                 positional_embedding_max_pos=[20, 2048, 2048],
-                 causal_temporal_positioning=False,
-                 vae_scale_factors=(8, 32, 32),
-                 dtype=None, device=None, operations=None, **kwargs):
+    cos_freq = cos_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads)
+    sin_freq = sin_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads)
+
+    cos_freq = torch.swapaxes(cos_freq, 1, 2)  # (B,H,T,D//2)
+    sin_freq = torch.swapaxes(sin_freq, 1, 2)  # (B,H,T,D//2)
+    return cos_freq, sin_freq
+
+class LTXBaseModel(torch.nn.Module, ABC):
+    """
+    Abstract base class for LTX models (Lightricks Transformer models).
+
+    This class defines the common interface and shared functionality for all LTX models,
+    including LTXV (video) and LTXAV (audio-video) variants.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        cross_attention_dim: int,
+        attention_head_dim: int,
+        num_attention_heads: int,
+        caption_channels: int,
+        num_layers: int,
+        positional_embedding_theta: float = 10000.0,
+        positional_embedding_max_pos: list = [20, 2048, 2048],
+        causal_temporal_positioning: bool = False,
+        vae_scale_factors: tuple = (8, 32, 32),
+        use_middle_indices_grid=False,
+        timestep_scale_multiplier = 1000.0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
         super().__init__()
         self.generator = None
         self.vae_scale_factors = vae_scale_factors
+        self.use_middle_indices_grid = use_middle_indices_grid
         self.dtype = dtype
-        self.out_channels = in_channels
-        self.inner_dim = num_attention_heads * attention_head_dim
+        self.in_channels = in_channels
+        self.cross_attention_dim = cross_attention_dim
+        self.attention_head_dim = attention_head_dim
+        self.num_attention_heads = num_attention_heads
+        self.caption_channels = caption_channels
+        self.num_layers = num_layers
+        self.positional_embedding_theta = positional_embedding_theta
+        self.positional_embedding_max_pos = positional_embedding_max_pos
+        self.split_positional_embedding = LTXRopeType.from_dict(kwargs)
+        self.freq_grid_generator = (
+            generate_freq_grid_np if LTXFrequenciesPrecision.from_dict(kwargs) == LTXFrequenciesPrecision.FLOAT64
+            else generate_freq_grid_pytorch
+        )
         self.causal_temporal_positioning = causal_temporal_positioning
+        self.operations = operations
+        self.timestep_scale_multiplier = timestep_scale_multiplier
 
-        self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device)
+        # Common dimensions
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.out_channels = in_channels
+
+        # Initialize common components
+        self._init_common_components(device, dtype)
+
+        # Initialize model-specific components
+        self._init_model_components(device, dtype, **kwargs)
+
+        # Initialize transformer blocks
+        self._init_transformer_blocks(device, dtype, **kwargs)
+
+        # Initialize output components
+        self._init_output_components(device, dtype)
+
+    def _init_common_components(self, device, dtype):
+        """Initialize components common to all LTX models
+        - patchify_proj: Linear projection for patchifying input
+        - adaln_single: AdaLN layer for timestep embedding
+        - caption_projection: Linear projection for caption embedding
+        """
+        self.patchify_proj = self.operations.Linear(
+            self.in_channels, self.inner_dim, bias=True, dtype=dtype, device=device
+        )
 
         self.adaln_single = AdaLayerNormSingle(
-            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations
+            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=self.operations
         )
 
-        # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device)
-
         self.caption_projection = PixArtAlphaTextProjection(
-            in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations
+            in_features=self.caption_channels,
+            hidden_size=self.inner_dim,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
         )
 
+    @abstractmethod
+    def _init_model_components(self, device, dtype, **kwargs):
+        """Initialize model-specific components. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _init_transformer_blocks(self, device, dtype, **kwargs):
+        """Initialize transformer blocks. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _init_output_components(self, device, dtype):
+        """Initialize output components. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
+        """Process input data. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs):
+        """Process transformer blocks. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
+        """Process output data. Must be implemented by subclasses."""
+        pass
+
+    def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs):
+        """Prepare timestep embeddings."""
+        grid_mask = kwargs.get("grid_mask", None)
+        if grid_mask is not None:
+            timestep = timestep[:, grid_mask]
+
+        timestep = timestep * self.timestep_scale_multiplier
+        timestep, embedded_timestep = self.adaln_single(
+            timestep.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=hidden_dtype,
+        )
+
+        # Second dimension is 1 or number of tokens (if timestep_per_token)
+        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
+        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1])
+
+        return timestep, embedded_timestep
+
+    def _prepare_context(self, context, batch_size, x, attention_mask=None):
+        """Prepare context for transformer blocks."""
+        if self.caption_projection is not None:
+            context = self.caption_projection(context)
+            context = context.view(batch_size, -1, x.shape[-1])
+
+        return context, attention_mask
+
+    def _precompute_freqs_cis(
+        self,
+        indices_grid,
+        dim,
+        out_dtype,
+        theta=10000.0,
+        max_pos=[20, 2048, 2048],
+        use_middle_indices_grid=False,
+        num_attention_heads=32,
+    ):
+        split_mode = self.split_positional_embedding == LTXRopeType.SPLIT
+        indices = self.freq_grid_generator(theta, indices_grid.shape[1], dim, indices_grid.device)
+        freqs = generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid)
+
+        if split_mode:
+            expected_freqs = dim // 2
+            current_freqs = freqs.shape[-1]
+            pad_size = expected_freqs - current_freqs
+            cos_freq, sin_freq = split_freqs_cis(freqs, pad_size, num_attention_heads)
+        else:
+            # 2 because of cos and sin by 3 for (t, x, y), 1 for temporal only
+            n_elem = 2 * indices_grid.shape[1]
+            cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem)
+        return cos_freq.to(out_dtype), sin_freq.to(out_dtype), split_mode
+
+    def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype):
+        """Prepare positional embeddings."""
+        fractional_coords = pixel_coords.to(torch.float32)
+        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
+        pe = self._precompute_freqs_cis(
+            fractional_coords,
+            dim=self.inner_dim,
+            out_dtype=x_dtype,
+            max_pos=self.positional_embedding_max_pos,
+            use_middle_indices_grid=self.use_middle_indices_grid,
+            num_attention_heads=self.num_attention_heads,
+        )
+        return pe
+
+    def _prepare_attention_mask(self, attention_mask, x_dtype):
+        """Prepare attention mask."""
+        if attention_mask is not None and not torch.is_floating_point(attention_mask):
+            attention_mask = (attention_mask - 1).to(x_dtype).reshape(
+                (attention_mask.shape[0], 1, -1, attention_mask.shape[-1])
+            ) * torch.finfo(x_dtype).max
+        return attention_mask
+
+    def forward(
+        self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs
+    ):
+        """
+        Forward pass for LTX models.
+
+        Args:
+            x: Input tensor
+            timestep: Timestep tensor
+            context: Context tensor (e.g., text embeddings)
+            attention_mask: Attention mask tensor
+            frame_rate: Frame rate for temporal processing
+            transformer_options: Additional options for transformer blocks
+            keyframe_idxs: Keyframe indices for temporal processing
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed output tensor
+        """
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(
+                comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options
+            ),
+        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, denoise_mask=denoise_mask, **kwargs)
+
+    def _forward(
+        self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs
+    ):
+        """
+        Internal forward pass for LTX models.
+
+        Args:
+            x: Input tensor
+            timestep: Timestep tensor
+            context: Context tensor (e.g., text embeddings)
+            attention_mask: Attention mask tensor
+            frame_rate: Frame rate for temporal processing
+            transformer_options: Additional options for transformer blocks
+            keyframe_idxs: Keyframe indices for temporal processing
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed output tensor
+        """
+        if isinstance(x, list):
+            input_dtype = x[0].dtype
+            batch_size = x[0].shape[0]
+        else:
+            input_dtype = x.dtype
+            batch_size = x.shape[0]
+        # Process input
+        merged_args = {**transformer_options, **kwargs}
+        x, pixel_coords, additional_args = self._process_input(x, keyframe_idxs, denoise_mask, **merged_args)
+        merged_args.update(additional_args)
+
+        # Prepare timestep and context
+        timestep, embedded_timestep = self._prepare_timestep(timestep, batch_size, input_dtype, **merged_args)
+        context, attention_mask = self._prepare_context(context, batch_size, x, attention_mask)
+
+        # Prepare attention mask and positional embeddings
+        attention_mask = self._prepare_attention_mask(attention_mask, input_dtype)
+        pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype)
+
+        # Process transformer blocks
+        x = self._process_transformer_blocks(
+            x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args
+        )
+
+        # Process output
+        x = self._process_output(x, embedded_timestep, keyframe_idxs, **merged_args)
+        return x
+
+
+class LTXVModel(LTXBaseModel):
+    """LTXV model for video generation."""
+
+    def __init__(
+        self,
+        in_channels=128,
+        cross_attention_dim=2048,
+        attention_head_dim=64,
+        num_attention_heads=32,
+        caption_channels=4096,
+        num_layers=28,
+        positional_embedding_theta=10000.0,
+        positional_embedding_max_pos=[20, 2048, 2048],
+        causal_temporal_positioning=False,
+        vae_scale_factors=(8, 32, 32),
+        use_middle_indices_grid=False,
+        timestep_scale_multiplier = 1000.0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+            caption_channels=caption_channels,
+            num_layers=num_layers,
+            positional_embedding_theta=positional_embedding_theta,
+            positional_embedding_max_pos=positional_embedding_max_pos,
+            causal_temporal_positioning=causal_temporal_positioning,
+            vae_scale_factors=vae_scale_factors,
+            use_middle_indices_grid=use_middle_indices_grid,
+            timestep_scale_multiplier=timestep_scale_multiplier,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+            **kwargs,
+        )
+
+    def _init_model_components(self, device, dtype, **kwargs):
+        """Initialize LTXV-specific components."""
+        # No additional components needed for LTXV beyond base class
+        pass
+
+    def _init_transformer_blocks(self, device, dtype, **kwargs):
+        """Initialize transformer blocks for LTXV."""
         self.transformer_blocks = nn.ModuleList(
             [
                 BasicTransformerBlock(
                     self.inner_dim,
-                    num_attention_heads,
-                    attention_head_dim,
-                    context_dim=cross_attention_dim,
-                    # attn_precision=attn_precision,
-                    dtype=dtype, device=device, operations=operations
+                    self.num_attention_heads,
+                    self.attention_head_dim,
+                    context_dim=self.cross_attention_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=self.operations,
                 )
-                for d in range(num_layers)
+                for _ in range(self.num_layers)
             ]
         )
 
+    def _init_output_components(self, device, dtype):
+        """Initialize output components for LTXV."""
         self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device))
-        self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
-
-        self.patchifier = SymmetricPatchifier(1)
-
-    def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
-        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
-            self._forward,
-            self,
-            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs)
-
-    def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
-        patches_replace = transformer_options.get("patches_replace", {})
-
-        orig_shape = list(x.shape)
+        self.norm_out = self.operations.LayerNorm(
+            self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
+        )
+        self.proj_out = self.operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
+        self.patchifier = SymmetricPatchifier(1, start_end=True)
 
+    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
+        """Process input for LTXV."""
+        additional_args = {"orig_shape": list(x.shape)}
         x, latent_coords = self.patchifier.patchify(x)
         pixel_coords = latent_to_pixel_coords(
             latent_coords=latent_coords,
@@ -439,50 +880,36 @@ class LTXVModel(torch.nn.Module):
             causal_fix=self.causal_temporal_positioning,
         )
 
+        grid_mask = None
         if keyframe_idxs is not None:
-            pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs
+            additional_args.update({ "orig_patchified_shape": list(x.shape)})
+            denoise_mask = self.patchifier.patchify(denoise_mask)[0]
+            grid_mask = ~torch.any(denoise_mask < 0, dim=-1)[0]
+            additional_args.update({"grid_mask": grid_mask})
+            x = x[:, grid_mask, :]
+            pixel_coords = pixel_coords[:, :, grid_mask, ...]
 
-        fractional_coords = pixel_coords.to(torch.float32)
-        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
+            kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:]
+            keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :]
+            pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs
 
         x = self.patchify_proj(x)
-        timestep = timestep * 1000.0
-
-        if attention_mask is not None and not torch.is_floating_point(attention_mask):
-            attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
-
-        pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype)
-
-        batch_size = x.shape[0]
-        timestep, embedded_timestep = self.adaln_single(
-            timestep.flatten(),
-            {"resolution": None, "aspect_ratio": None},
-            batch_size=batch_size,
-            hidden_dtype=x.dtype,
-        )
-        # Second dimension is 1 or number of tokens (if timestep_per_token)
-        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
-        embedded_timestep = embedded_timestep.view(
-            batch_size, -1, embedded_timestep.shape[-1]
-        )
-
-        # 2. Blocks
-        if self.caption_projection is not None:
-            batch_size = x.shape[0]
-            context = self.caption_projection(context)
-            context = context.view(
-                batch_size, -1, x.shape[-1]
-            )
+        return x, pixel_coords, additional_args
 
+    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs):
+        """Process transformer blocks for LTXV."""
+        patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+
         for i, block in enumerate(self.transformer_blocks):
             if ("double_block", i) in blocks_replace:
+
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"])
+                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
                 x = block(
@@ -490,19 +917,32 @@ class LTXVModel(torch.nn.Module):
                     context=context,
                     attention_mask=attention_mask,
                     timestep=timestep,
-                    pe=pe
+                    pe=pe,
+                    transformer_options=transformer_options,
                 )
 
-        # 3. Output
+        return x
+
+    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
+        """Process output for LTXV."""
+        # Apply scale-shift modulation
         scale_shift_values = (
             self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None]
         )
         shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
+
         x = self.norm_out(x)
-        # Modulation
         x = x * (1 + scale) + shift
         x = self.proj_out(x)
 
+        if keyframe_idxs is not None:
+            grid_mask = kwargs["grid_mask"]
+            orig_patchified_shape = kwargs["orig_patchified_shape"]
+            full_x = torch.zeros(orig_patchified_shape, dtype=x.dtype, device=x.device)
+            full_x[:, grid_mask, :] = x
+            x = full_x
+        # Unpatchify to restore original dimensions
+        orig_shape = kwargs["orig_shape"]
         x = self.patchifier.unpatchify(
             latents=x,
             output_height=orig_shape[3],

comfy/ldm/lightricks/symmetric_patchifier.py

@@ -21,20 +21,23 @@ def latent_to_pixel_coords(
     Returns:
         Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates.
     """
+    shape = [1] * latent_coords.ndim
+    shape[1] = -1
     pixel_coords = (
         latent_coords
-        * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None]
+        * torch.tensor(scale_factors, device=latent_coords.device).view(*shape)
     )
     if causal_fix:
         # Fix temporal scale for first frame to 1 due to causality
-        pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0)
+        pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0)
     return pixel_coords
 
 
 class Patchifier(ABC):
-    def __init__(self, patch_size: int):
+    def __init__(self, patch_size: int, start_end: bool=False):
         super().__init__()
         self._patch_size = (1, patch_size, patch_size)
+        self.start_end = start_end
 
     @abstractmethod
     def patchify(
@@ -71,11 +74,23 @@ class Patchifier(ABC):
             torch.arange(0, latent_width, self._patch_size[2], device=device),
             indexing="ij",
         )
-        latent_sample_coords = torch.stack(latent_sample_coords, dim=0)
-        latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
-        latent_coords = rearrange(
-            latent_coords, "b c f h w -> b c (f h w)", b=batch_size
+        latent_sample_coords_start = torch.stack(latent_sample_coords, dim=0)
+        delta = torch.tensor(self._patch_size, device=latent_sample_coords_start.device, dtype=latent_sample_coords_start.dtype)[:, None, None, None]
+        latent_sample_coords_end = latent_sample_coords_start + delta
+
+        latent_sample_coords_start = latent_sample_coords_start.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+        latent_sample_coords_start = rearrange(
+            latent_sample_coords_start, "b c f h w -> b c (f h w)", b=batch_size
         )
+        if self.start_end:
+            latent_sample_coords_end = latent_sample_coords_end.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+            latent_sample_coords_end = rearrange(
+                latent_sample_coords_end, "b c f h w -> b c (f h w)", b=batch_size
+            )
+
+            latent_coords = torch.stack((latent_sample_coords_start, latent_sample_coords_end), dim=-1)
+        else:
+            latent_coords = latent_sample_coords_start
         return latent_coords
 
 
@@ -115,3 +130,61 @@ class SymmetricPatchifier(Patchifier):
             q=self._patch_size[2],
         )
         return latents
+
+
+class AudioPatchifier(Patchifier):
+    def __init__(self, patch_size: int,
+        sample_rate=16000,
+        hop_length=160,
+        audio_latent_downsample_factor=4,
+        is_causal=True,
+        start_end=False,
+        shift = 0
+    ):
+        super().__init__(patch_size, start_end=start_end)
+        self.hop_length = hop_length
+        self.sample_rate = sample_rate
+        self.audio_latent_downsample_factor = audio_latent_downsample_factor
+        self.is_causal = is_causal
+        self.shift = shift
+
+    def copy_with_shift(self, shift):
+        return AudioPatchifier(
+            self.patch_size, self.sample_rate, self.hop_length, self.audio_latent_downsample_factor,
+            self.is_causal, self.start_end, shift
+        )
+
+    def _get_audio_latent_time_in_sec(self, start_latent, end_latent: int, dtype: torch.dtype, device=torch.device):
+        audio_latent_frame = torch.arange(start_latent, end_latent, dtype=dtype, device=device)
+        audio_mel_frame = audio_latent_frame * self.audio_latent_downsample_factor
+        if self.is_causal:
+            audio_mel_frame = (audio_mel_frame + 1 - self.audio_latent_downsample_factor).clip(min=0)
+        return audio_mel_frame * self.hop_length / self.sample_rate
+
+
+    def patchify(self, audio_latents: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        # audio_latents: (batch, channels, time, freq)
+        b, _, t, _ = audio_latents.shape
+        audio_latents = rearrange(
+            audio_latents,
+            "b c t f -> b t (c f)",
+        )
+
+        audio_latents_start_timings = self._get_audio_latent_time_in_sec(self.shift, t + self.shift, torch.float32, audio_latents.device)
+        audio_latents_start_timings = audio_latents_start_timings.unsqueeze(0).expand(b, -1).unsqueeze(1)
+
+        if self.start_end:
+            audio_latents_end_timings = self._get_audio_latent_time_in_sec(self.shift + 1, t + self.shift + 1, torch.float32, audio_latents.device)
+            audio_latents_end_timings = audio_latents_end_timings.unsqueeze(0).expand(b, -1).unsqueeze(1)
+
+            audio_latents_timings = torch.stack([audio_latents_start_timings, audio_latents_end_timings], dim=-1)
+        else:
+            audio_latents_timings = audio_latents_start_timings
+        return audio_latents, audio_latents_timings
+
+    def unpatchify(self, audio_latents: torch.Tensor, channels: int, freq: int) -> torch.Tensor:
+        # audio_latents: (batch, time, freq * channels)
+        audio_latents = rearrange(
+            audio_latents, "b t (c f) -> b c t f", c=channels, f=freq
+        )
+        return audio_latents

comfy/ldm/lightricks/vae/audio_vae.py

@@ -0,0 +1,286 @@
+import json
+from dataclasses import dataclass
+import math
+import torch
+import torchaudio
+
+import comfy.model_management
+import comfy.model_patcher
+import comfy.utils as utils
+from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution
+from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
+from comfy.ldm.lightricks.vae.causal_audio_autoencoder import (
+    CausalityAxis,
+    CausalAudioAutoencoder,
+)
+from comfy.ldm.lightricks.vocoders.vocoder import Vocoder
+
+LATENT_DOWNSAMPLE_FACTOR = 4
+
+
+@dataclass(frozen=True)
+class AudioVAEComponentConfig:
+    """Container for model component configuration extracted from metadata."""
+
+    autoencoder: dict
+    vocoder: dict
+
+    @classmethod
+    def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig":
+        assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE"
+
+        raw_config = metadata["config"]
+        if isinstance(raw_config, str):
+            parsed_config = json.loads(raw_config)
+        else:
+            parsed_config = raw_config
+
+        audio_config = parsed_config.get("audio_vae")
+        vocoder_config = parsed_config.get("vocoder")
+
+        assert audio_config is not None, "Audio VAE config is required for audio VAE"
+        assert vocoder_config is not None, "Vocoder config is required for audio VAE"
+
+        return cls(autoencoder=audio_config, vocoder=vocoder_config)
+
+
+class ModelDeviceManager:
+    """Manages device placement and GPU residency for the composed model."""
+
+    def __init__(self, module: torch.nn.Module):
+        load_device = comfy.model_management.get_torch_device()
+        offload_device = comfy.model_management.vae_offload_device()
+        self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device)
+
+    def ensure_model_loaded(self) -> None:
+        comfy.model_management.free_memory(
+            self.patcher.model_size(),
+            self.patcher.load_device,
+        )
+        comfy.model_management.load_model_gpu(self.patcher)
+
+    def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor:
+        return tensor.to(self.patcher.load_device)
+
+    @property
+    def load_device(self):
+        return self.patcher.load_device
+
+
+class AudioLatentNormalizer:
+    """Applies per-channel statistics in patch space and restores original layout."""
+
+    def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module):
+        self.patchifier = patchfier
+        self.statistics = statistics_processor
+
+    def normalize(self, latents: torch.Tensor) -> torch.Tensor:
+        channels = latents.shape[1]
+        freq = latents.shape[3]
+        patched, _ = self.patchifier.patchify(latents)
+        normalized = self.statistics.normalize(patched)
+        return self.patchifier.unpatchify(normalized, channels=channels, freq=freq)
+
+    def denormalize(self, latents: torch.Tensor) -> torch.Tensor:
+        channels = latents.shape[1]
+        freq = latents.shape[3]
+        patched, _ = self.patchifier.patchify(latents)
+        denormalized = self.statistics.un_normalize(patched)
+        return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq)
+
+
+class AudioPreprocessor:
+    """Prepares raw waveforms for the autoencoder by matching training conditions."""
+
+    def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int):
+        self.target_sample_rate = target_sample_rate
+        self.mel_bins = mel_bins
+        self.mel_hop_length = mel_hop_length
+        self.n_fft = n_fft
+
+    def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor:
+        if source_rate == self.target_sample_rate:
+            return waveform
+        return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate)
+
+    @staticmethod
+    def normalize_amplitude(
+        waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5
+    ) -> torch.Tensor:
+        waveform = waveform - waveform.mean(dim=2, keepdim=True)
+        peak = torch.max(torch.abs(waveform)) + eps
+        scale = peak.clamp(max=max_amplitude) / peak
+        return waveform * scale
+
+    def waveform_to_mel(
+        self, waveform: torch.Tensor, waveform_sample_rate: int, device
+    ) -> torch.Tensor:
+        waveform = self.resample(waveform, waveform_sample_rate)
+        waveform = self.normalize_amplitude(waveform)
+
+        mel_transform = torchaudio.transforms.MelSpectrogram(
+            sample_rate=self.target_sample_rate,
+            n_fft=self.n_fft,
+            win_length=self.n_fft,
+            hop_length=self.mel_hop_length,
+            f_min=0.0,
+            f_max=self.target_sample_rate / 2.0,
+            n_mels=self.mel_bins,
+            window_fn=torch.hann_window,
+            center=True,
+            pad_mode="reflect",
+            power=1.0,
+            mel_scale="slaney",
+            norm="slaney",
+        ).to(device)
+
+        mel = mel_transform(waveform)
+        mel = torch.log(torch.clamp(mel, min=1e-5))
+        return mel.permute(0, 1, 3, 2).contiguous()
+
+
+class AudioVAE(torch.nn.Module):
+    """High-level Audio VAE wrapper exposing encode and decode entry points."""
+
+    def __init__(self, state_dict: dict, metadata: dict):
+        super().__init__()
+
+        component_config = AudioVAEComponentConfig.from_metadata(metadata)
+
+        vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True)
+        vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True)
+
+        self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder)
+        self.vocoder = Vocoder(config=component_config.vocoder)
+
+        self.autoencoder.load_state_dict(vae_sd, strict=False)
+        self.vocoder.load_state_dict(vocoder_sd, strict=False)
+
+        autoencoder_config = self.autoencoder.get_config()
+        self.normalizer = AudioLatentNormalizer(
+            AudioPatchifier(
+                patch_size=1,
+                audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR,
+                sample_rate=autoencoder_config["sampling_rate"],
+                hop_length=autoencoder_config["mel_hop_length"],
+                is_causal=autoencoder_config["is_causal"],
+            ),
+            self.autoencoder.per_channel_statistics,
+        )
+
+        self.preprocessor = AudioPreprocessor(
+            target_sample_rate=autoencoder_config["sampling_rate"],
+            mel_bins=autoencoder_config["mel_bins"],
+            mel_hop_length=autoencoder_config["mel_hop_length"],
+            n_fft=autoencoder_config["n_fft"],
+        )
+
+        self.device_manager = ModelDeviceManager(self)
+
+    def encode(self, audio: dict) -> torch.Tensor:
+        """Encode a waveform dictionary into normalized latent tensors."""
+
+        waveform = audio["waveform"]
+        waveform_sample_rate = audio["sample_rate"]
+        input_device = waveform.device
+        # Ensure that Audio VAE is loaded on the correct device.
+        self.device_manager.ensure_model_loaded()
+
+        waveform = self.device_manager.move_to_load_device(waveform)
+        expected_channels = self.autoencoder.encoder.in_channels
+        if waveform.shape[1] != expected_channels:
+            raise ValueError(
+                f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
+            )
+
+        mel_spec = self.preprocessor.waveform_to_mel(
+            waveform, waveform_sample_rate, device=self.device_manager.load_device
+        )
+
+        latents = self.autoencoder.encode(mel_spec)
+        posterior = DiagonalGaussianDistribution(latents)
+        latent_mode = posterior.mode()
+
+        normalized = self.normalizer.normalize(latent_mode)
+        return normalized.to(input_device)
+
+    def decode(self, latents: torch.Tensor) -> torch.Tensor:
+        """Decode normalized latent tensors into an audio waveform."""
+        original_shape = latents.shape
+
+        # Ensure that Audio VAE is loaded on the correct device.
+        self.device_manager.ensure_model_loaded()
+
+        latents = self.device_manager.move_to_load_device(latents)
+        latents = self.normalizer.denormalize(latents)
+
+        target_shape = self.target_shape_from_latents(original_shape)
+        mel_spec = self.autoencoder.decode(latents, target_shape=target_shape)
+
+        waveform = self.run_vocoder(mel_spec)
+        return self.device_manager.move_to_load_device(waveform)
+
+    def target_shape_from_latents(self, latents_shape):
+        batch, _, time, _ = latents_shape
+        target_length = time * LATENT_DOWNSAMPLE_FACTOR
+        if self.autoencoder.causality_axis != CausalityAxis.NONE:
+            target_length -= LATENT_DOWNSAMPLE_FACTOR - 1
+        return (
+            batch,
+            self.autoencoder.decoder.out_ch,
+            target_length,
+            self.autoencoder.mel_bins,
+        )
+
+    def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int:
+        return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second)
+
+    def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor:
+        audio_channels = self.autoencoder.decoder.out_ch
+        vocoder_input = mel_spec.transpose(2, 3)
+
+        if audio_channels == 1:
+            vocoder_input = vocoder_input.squeeze(1)
+        elif audio_channels != 2:
+            raise ValueError(f"Unsupported audio_channels: {audio_channels}")
+
+        return self.vocoder(vocoder_input)
+
+    @property
+    def sample_rate(self) -> int:
+        return int(self.autoencoder.sampling_rate)
+
+    @property
+    def mel_hop_length(self) -> int:
+        return int(self.autoencoder.mel_hop_length)
+
+    @property
+    def mel_bins(self) -> int:
+        return int(self.autoencoder.mel_bins)
+
+    @property
+    def latent_channels(self) -> int:
+        return int(self.autoencoder.decoder.z_channels)
+
+    @property
+    def latent_frequency_bins(self) -> int:
+        return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR)
+
+    @property
+    def latents_per_second(self) -> float:
+        return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR
+
+    @property
+    def output_sample_rate(self) -> int:
+        output_rate = getattr(self.vocoder, "output_sample_rate", None)
+        if output_rate is not None:
+            return int(output_rate)
+        upsample_factor = getattr(self.vocoder, "upsample_factor", None)
+        if upsample_factor is None:
+            raise AttributeError(
+                "Vocoder is missing upsample_factor; cannot infer output sample rate"
+            )
+        return int(self.sample_rate * upsample_factor / self.mel_hop_length)
+
+    def memory_required(self, input_shape):
+        return self.device_manager.patcher.model_size()

comfy/ldm/lightricks/vae/causal_audio_autoencoder.py

@@ -0,0 +1,909 @@
+from __future__ import annotations
+import torch
+from torch import nn
+from torch.nn import functional as F
+from typing import Optional
+from enum import Enum
+from .pixel_norm import PixelNorm
+import comfy.ops
+import logging
+
+ops = comfy.ops.disable_weight_init
+
+
+class StringConvertibleEnum(Enum):
+    """
+    Base enum class that provides string-to-enum conversion functionality.
+
+    This mixin adds a str_to_enum() class method that handles conversion from
+    strings, None, or existing enum instances with case-insensitive matching.
+    """
+
+    @classmethod
+    def str_to_enum(cls, value):
+        """
+        Convert a string, enum instance, or None to the appropriate enum member.
+
+        Args:
+            value: Can be an enum instance of this class, a string, or None
+
+        Returns:
+            Enum member of this class
+
+        Raises:
+            ValueError: If the value cannot be converted to a valid enum member
+        """
+        # Already an enum instance of this class
+        if isinstance(value, cls):
+            return value
+
+        # None maps to NONE member if it exists
+        if value is None:
+            if hasattr(cls, "NONE"):
+                return cls.NONE
+            raise ValueError(f"{cls.__name__} does not have a NONE member to map None to")
+
+        # String conversion (case-insensitive)
+        if isinstance(value, str):
+            value_lower = value.lower()
+
+            # Try to match against enum values
+            for member in cls:
+                # Handle members with None values
+                if member.value is None:
+                    if value_lower == "none":
+                        return member
+                # Handle members with string values
+                elif isinstance(member.value, str) and member.value.lower() == value_lower:
+                    return member
+
+            # Build helpful error message with valid values
+            valid_values = []
+            for member in cls:
+                if member.value is None:
+                    valid_values.append("none")
+                elif isinstance(member.value, str):
+                    valid_values.append(member.value)
+
+            raise ValueError(f"Invalid {cls.__name__} string: '{value}'. " f"Valid values are: {valid_values}")
+
+        raise ValueError(
+            f"Cannot convert type {type(value).__name__} to {cls.__name__} enum. "
+            f"Expected string, None, or {cls.__name__} instance."
+        )
+
+
+class AttentionType(StringConvertibleEnum):
+    """Enum for specifying the attention mechanism type."""
+
+    VANILLA = "vanilla"
+    LINEAR = "linear"
+    NONE = "none"
+
+
+class CausalityAxis(StringConvertibleEnum):
+    """Enum for specifying the causality axis in causal convolutions."""
+
+    NONE = None
+    WIDTH = "width"
+    HEIGHT = "height"
+    WIDTH_COMPATIBILITY = "width-compatibility"
+
+
+def Normalize(in_channels, *, num_groups=32, normtype="group"):
+    if normtype == "group":
+        return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+    elif normtype == "pixel":
+        return PixelNorm(dim=1, eps=1e-6)
+    else:
+        raise ValueError(f"Invalid normalization type: {normtype}")
+
+
+class CausalConv2d(nn.Module):
+    """
+    A causal 2D convolution.
+
+    This layer ensures that the output at time `t` only depends on inputs
+    at time `t` and earlier. It achieves this by applying asymmetric padding
+    to the time dimension (width) before the convolution.
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        groups=1,
+        bias=True,
+        causality_axis: CausalityAxis = CausalityAxis.HEIGHT,
+    ):
+        super().__init__()
+
+        self.causality_axis = causality_axis
+
+        # Ensure kernel_size and dilation are tuples
+        kernel_size = nn.modules.utils._pair(kernel_size)
+        dilation = nn.modules.utils._pair(dilation)
+
+        # Calculate padding dimensions
+        pad_h = (kernel_size[0] - 1) * dilation[0]
+        pad_w = (kernel_size[1] - 1) * dilation[1]
+
+        # The padding tuple for F.pad is (pad_left, pad_right, pad_top, pad_bottom)
+        match self.causality_axis:
+            case CausalityAxis.NONE:
+                self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2)
+            case CausalityAxis.WIDTH | CausalityAxis.WIDTH_COMPATIBILITY:
+                self.padding = (pad_w, 0, pad_h // 2, pad_h - pad_h // 2)
+            case CausalityAxis.HEIGHT:
+                self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h, 0)
+            case _:
+                raise ValueError(f"Invalid causality_axis: {causality_axis}")
+
+        # The internal convolution layer uses no padding, as we handle it manually
+        self.conv = ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=stride,
+            padding=0,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+        )
+
+    def forward(self, x):
+        # Apply causal padding before convolution
+        x = F.pad(x, self.padding)
+        return self.conv(x)
+
+
+def make_conv2d(
+    in_channels,
+    out_channels,
+    kernel_size,
+    stride=1,
+    padding=None,
+    dilation=1,
+    groups=1,
+    bias=True,
+    causality_axis: Optional[CausalityAxis] = None,
+):
+    """
+    Create a 2D convolution layer that can be either causal or non-causal.
+
+    Args:
+        in_channels: Number of input channels
+        out_channels: Number of output channels
+        kernel_size: Size of the convolution kernel
+        stride: Convolution stride
+        padding: Padding (if None, will be calculated based on causal flag)
+        dilation: Dilation rate
+        groups: Number of groups for grouped convolution
+        bias: Whether to use bias
+        causality_axis: Dimension along which to apply causality.
+
+    Returns:
+        Either a regular Conv2d or CausalConv2d layer
+    """
+    if causality_axis is not None:
+        # For causal convolution, padding is handled internally by CausalConv2d
+        return CausalConv2d(in_channels, out_channels, kernel_size, stride, dilation, groups, bias, causality_axis)
+    else:
+        # For non-causal convolution, use symmetric padding if not specified
+        if padding is None:
+            if isinstance(kernel_size, int):
+                padding = kernel_size // 2
+            else:
+                padding = tuple(k // 2 for k in kernel_size)
+        return ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups,
+            bias,
+        )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.HEIGHT):
+        super().__init__()
+        self.with_conv = with_conv
+        self.causality_axis = causality_axis
+        if self.with_conv:
+            self.conv = make_conv2d(in_channels, in_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+            # Drop FIRST element in the causal axis to undo encoder's padding, while keeping the length 1 + 2 * n.
+            # For example, if the input is [0, 1, 2], after interpolation, the output is [0, 0, 1, 1, 2, 2].
+            # The causal convolution will pad the first element as [-, -, 0, 0, 1, 1, 2, 2],
+            # So the output elements rely on the following windows:
+            # 0: [-,-,0]
+            # 1: [-,0,0]
+            # 2: [0,0,1]
+            # 3: [0,1,1]
+            # 4: [1,1,2]
+            # 5: [1,2,2]
+            # Notice that the first and second elements in the output rely only on the first element in the input,
+            # while all other elements rely on two elements in the input.
+            # So we can drop the first element to undo the padding (rather than the last element).
+            # This is a no-op for non-causal convolutions.
+            match self.causality_axis:
+                case CausalityAxis.NONE:
+                    pass  # x remains unchanged
+                case CausalityAxis.HEIGHT:
+                    x = x[:, :, 1:, :]
+                case CausalityAxis.WIDTH:
+                    x = x[:, :, :, 1:]
+                case CausalityAxis.WIDTH_COMPATIBILITY:
+                    pass  # x remains unchanged
+                case _:
+                    raise ValueError(f"Invalid causality_axis: {self.causality_axis}")
+
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer that can use either a strided convolution
+    or average pooling. Supports standard and causal padding for the
+    convolutional mode.
+    """
+
+    def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.WIDTH):
+        super().__init__()
+        self.with_conv = with_conv
+        self.causality_axis = causality_axis
+
+        if self.causality_axis != CausalityAxis.NONE and not self.with_conv:
+            raise ValueError("causality is only supported when `with_conv=True`.")
+
+        if self.with_conv:
+            # Do time downsampling here
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            # (pad_left, pad_right, pad_top, pad_bottom)
+            match self.causality_axis:
+                case CausalityAxis.NONE:
+                    pad = (0, 1, 0, 1)
+                case CausalityAxis.WIDTH:
+                    pad = (2, 0, 0, 1)
+                case CausalityAxis.HEIGHT:
+                    pad = (0, 1, 2, 0)
+                case CausalityAxis.WIDTH_COMPATIBILITY:
+                    pad = (1, 0, 0, 1)
+                case _:
+                    raise ValueError(f"Invalid causality_axis: {self.causality_axis}")
+
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            # This branch is only taken if with_conv=False, which implies causality_axis is NONE.
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+        norm_type="group",
+        causality_axis: CausalityAxis = CausalityAxis.HEIGHT,
+    ):
+        super().__init__()
+        self.causality_axis = causality_axis
+
+        if self.causality_axis != CausalityAxis.NONE and norm_type == "group":
+            raise ValueError("Causal ResnetBlock with GroupNorm is not supported.")
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels, normtype=norm_type)
+        self.non_linearity = nn.SiLU()
+        self.conv1 = make_conv2d(in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels, normtype=norm_type)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = make_conv2d(out_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = make_conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis
+                )
+            else:
+                self.nin_shortcut = make_conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, causality_axis=causality_axis
+                )
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = self.non_linearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(self.non_linearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = self.non_linearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels, norm_type="group"):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels, normtype=norm_type)
+        self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w).contiguous()
+        q = q.permute(0, 2, 1).contiguous()  # b,hw,c
+        k = k.reshape(b, c, h * w).contiguous()  # b,c,hw
+        w_ = torch.bmm(q, k).contiguous()  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w).contiguous()
+        w_ = w_.permute(0, 2, 1).contiguous()  # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v, w_).contiguous()  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b, c, h, w).contiguous()
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+def make_attn(in_channels, attn_type="vanilla", norm_type="group"):
+    # Convert string to enum if needed
+    attn_type = AttentionType.str_to_enum(attn_type)
+
+    if attn_type != AttentionType.NONE:
+        logging.info(f"making attention of type '{attn_type.value}' with {in_channels} in_channels")
+    else:
+        logging.info(f"making identity attention with {in_channels} in_channels")
+
+    match attn_type:
+        case AttentionType.VANILLA:
+            return AttnBlock(in_channels, norm_type=norm_type)
+        case AttentionType.NONE:
+            return nn.Identity(in_channels)
+        case AttentionType.LINEAR:
+            raise NotImplementedError(f"Attention type {attn_type.value} is not supported yet.")
+        case _:
+            raise ValueError(f"Unknown attention type: {attn_type}")
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        attn_type="vanilla",
+        mid_block_add_attention=True,
+        norm_type="group",
+        causality_axis=CausalityAxis.WIDTH.value,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.z_channels = z_channels
+        self.double_z = double_z
+        self.norm_type = norm_type
+        # Convert string to enum if needed (for config loading)
+        causality_axis = CausalityAxis.str_to_enum(causality_axis)
+        self.attn_type = AttentionType.str_to_enum(attn_type)
+
+        # downsampling
+        self.conv_in = make_conv2d(
+            in_channels,
+            self.ch,
+            kernel_size=3,
+            stride=1,
+            causality_axis=causality_axis,
+        )
+
+        self.non_linearity = nn.SiLU()
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+
+            for _ in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        norm_type=self.norm_type,
+                        causality_axis=causality_axis,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type))
+
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv, causality_axis=causality_axis)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+        if mid_block_add_attention:
+            self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)
+        else:
+            self.mid.attn_1 = nn.Identity()
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+
+        # end
+        self.norm_out = Normalize(block_in, normtype=self.norm_type)
+        self.conv_out = make_conv2d(
+            block_in,
+            2 * z_channels if double_z else z_channels,
+            kernel_size=3,
+            stride=1,
+            causality_axis=causality_axis,
+        )
+
+    def forward(self, x):
+        """
+        Forward pass through the encoder.
+
+        Args:
+            x: Input tensor of shape [batch, channels, time, n_mels]
+
+        Returns:
+            Encoded latent representation
+        """
+        feature_maps = [self.conv_in(x)]
+
+        # Process each resolution level (from high to low resolution)
+        for resolution_level in range(self.num_resolutions):
+            # Apply residual blocks at current resolution level
+            for block_idx in range(self.num_res_blocks):
+                # Apply ResNet block with optional timestep embedding
+                current_features = self.down[resolution_level].block[block_idx](feature_maps[-1], temb=None)
+
+                # Apply attention if configured for this resolution level
+                if len(self.down[resolution_level].attn) > 0:
+                    current_features = self.down[resolution_level].attn[block_idx](current_features)
+
+                # Store processed features
+                feature_maps.append(current_features)
+
+            # Downsample spatial dimensions (except at the final resolution level)
+            if resolution_level != self.num_resolutions - 1:
+                downsampled_features = self.down[resolution_level].downsample(feature_maps[-1])
+                feature_maps.append(downsampled_features)
+
+        # === MIDDLE PROCESSING PHASE ===
+        # Take the lowest resolution features for middle processing
+        bottleneck_features = feature_maps[-1]
+
+        # Apply first middle ResNet block
+        bottleneck_features = self.mid.block_1(bottleneck_features, temb=None)
+
+        # Apply middle attention block
+        bottleneck_features = self.mid.attn_1(bottleneck_features)
+
+        # Apply second middle ResNet block
+        bottleneck_features = self.mid.block_2(bottleneck_features, temb=None)
+
+        # === OUTPUT PHASE ===
+        # Normalize the bottleneck features
+        output_features = self.norm_out(bottleneck_features)
+
+        # Apply non-linearity (SiLU activation)
+        output_features = self.non_linearity(output_features)
+
+        # Final convolution to produce latent representation
+        # [batch, channels, time, n_mels] -> [batch, 2 * z_channels if double_z else z_channels, time, n_mels]
+        return self.conv_out(output_features)
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        tanh_out=False,
+        attn_type="vanilla",
+        mid_block_add_attention=True,
+        norm_type="group",
+        causality_axis=CausalityAxis.WIDTH.value,
+        **ignorekwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.out_ch = out_ch
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+        self.norm_type = norm_type
+        self.z_channels = z_channels
+        # Convert string to enum if needed (for config loading)
+        causality_axis = CausalityAxis.str_to_enum(causality_axis)
+        self.attn_type = AttentionType.str_to_enum(attn_type)
+
+        # compute block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = make_conv2d(z_channels, block_in, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+        self.non_linearity = nn.SiLU()
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+        if mid_block_add_attention:
+            self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)
+        else:
+            self.mid.attn_1 = nn.Identity()
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        norm_type=self.norm_type,
+                        causality_axis=causality_axis,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv, causality_axis=causality_axis)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in, normtype=self.norm_type)
+        self.conv_out = make_conv2d(block_in, out_ch, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+    def _adjust_output_shape(self, decoded_output, target_shape):
+        """
+        Adjust output shape to match target dimensions for variable-length audio.
+
+        This function handles the common case where decoded audio spectrograms need to be
+        resized to match a specific target shape.
+
+        Args:
+            decoded_output: Tensor of shape (batch, channels, time, frequency)
+            target_shape: Target shape tuple (batch, channels, time, frequency)
+
+        Returns:
+            Tensor adjusted to match target_shape exactly
+        """
+        # Current output shape: (batch, channels, time, frequency)
+        _, _, current_time, current_freq = decoded_output.shape
+        _, target_channels, target_time, target_freq = target_shape
+
+        # Step 1: Crop first to avoid exceeding target dimensions
+        decoded_output = decoded_output[
+            :, :target_channels, : min(current_time, target_time), : min(current_freq, target_freq)
+        ]
+
+        # Step 2: Calculate padding needed for time and frequency dimensions
+        time_padding_needed = target_time - decoded_output.shape[2]
+        freq_padding_needed = target_freq - decoded_output.shape[3]
+
+        # Step 3: Apply padding if needed
+        if time_padding_needed > 0 or freq_padding_needed > 0:
+            # PyTorch padding format: (pad_left, pad_right, pad_top, pad_bottom)
+            # For audio: pad_left/right = frequency, pad_top/bottom = time
+            padding = (
+                0,
+                max(freq_padding_needed, 0),  # frequency padding (left, right)
+                0,
+                max(time_padding_needed, 0),  # time padding (top, bottom)
+            )
+            decoded_output = F.pad(decoded_output, padding)
+
+        # Step 4: Final safety crop to ensure exact target shape
+        decoded_output = decoded_output[:, :target_channels, :target_time, :target_freq]
+
+        return decoded_output
+
+    def get_config(self):
+        return {
+            "ch": self.ch,
+            "out_ch": self.out_ch,
+            "ch_mult": self.ch_mult,
+            "num_res_blocks": self.num_res_blocks,
+            "in_channels": self.in_channels,
+            "resolution": self.resolution,
+            "z_channels": self.z_channels,
+        }
+
+    def forward(self, latent_features, target_shape=None):
+        """
+        Decode latent features back to audio spectrograms.
+
+        Args:
+            latent_features: Encoded latent representation of shape (batch, channels, height, width)
+            target_shape: Optional target output shape (batch, channels, time, frequency)
+                         If provided, output will be cropped/padded to match this shape
+
+        Returns:
+            Reconstructed audio spectrogram of shape (batch, channels, time, frequency)
+        """
+        assert target_shape is not None, "Target shape is required for CausalAudioAutoencoder Decoder"
+
+        # Transform latent features to decoder's internal feature dimension
+        hidden_features = self.conv_in(latent_features)
+
+        # Middle processing
+        hidden_features = self.mid.block_1(hidden_features, temb=None)
+        hidden_features = self.mid.attn_1(hidden_features)
+        hidden_features = self.mid.block_2(hidden_features, temb=None)
+
+        # Upsampling
+        # Progressively increase spatial resolution from lowest to highest
+        for resolution_level in reversed(range(self.num_resolutions)):
+            # Apply residual blocks at current resolution level
+            for block_index in range(self.num_res_blocks + 1):
+                hidden_features = self.up[resolution_level].block[block_index](hidden_features, temb=None)
+
+                if len(self.up[resolution_level].attn) > 0:
+                    hidden_features = self.up[resolution_level].attn[block_index](hidden_features)
+
+            if resolution_level != 0:
+                hidden_features = self.up[resolution_level].upsample(hidden_features)
+
+        # Output
+        if self.give_pre_end:
+            # Return intermediate features before final processing (for debugging/analysis)
+            decoded_output = hidden_features
+        else:
+            # Standard output path: normalize, activate, and convert to output channels
+            # Final normalization layer
+            hidden_features = self.norm_out(hidden_features)
+
+            # Apply SiLU (Swish) activation function
+            hidden_features = self.non_linearity(hidden_features)
+
+            # Final convolution to map to output channels (typically 2 for stereo audio)
+            decoded_output = self.conv_out(hidden_features)
+
+            # Optional tanh activation to bound output values to [-1, 1] range
+            if self.tanh_out:
+                decoded_output = torch.tanh(decoded_output)
+
+        # Adjust shape for audio data
+        if target_shape is not None:
+            decoded_output = self._adjust_output_shape(decoded_output, target_shape)
+
+        return decoded_output
+
+
+class processor(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.register_buffer("std-of-means", torch.empty(128))
+        self.register_buffer("mean-of-means", torch.empty(128))
+
+    def un_normalize(self, x):
+        return (x * self.get_buffer("std-of-means").to(x)) + self.get_buffer("mean-of-means").to(x)
+
+    def normalize(self, x):
+        return (x - self.get_buffer("mean-of-means").to(x)) / self.get_buffer("std-of-means").to(x)
+
+
+class CausalAudioAutoencoder(nn.Module):
+    def __init__(self, config=None):
+        super().__init__()
+
+        if config is None:
+            config = self._guess_config()
+
+        # Extract encoder and decoder configs from the new format
+        model_config = config.get("model", {}).get("params", {})
+        variables_config = config.get("variables", {})
+
+        self.sampling_rate = variables_config.get(
+            "sampling_rate",
+            model_config.get("sampling_rate", config.get("sampling_rate", 16000)),
+        )
+        encoder_config = model_config.get("encoder", model_config.get("ddconfig", {}))
+        decoder_config = model_config.get("decoder", encoder_config)
+
+        # Load mel spectrogram parameters
+        self.mel_bins = encoder_config.get("mel_bins", 64)
+        self.mel_hop_length = model_config.get("preprocessing", {}).get("stft", {}).get("hop_length", 160)
+        self.n_fft = model_config.get("preprocessing", {}).get("stft", {}).get("filter_length", 1024)
+
+        # Store causality configuration at VAE level (not just in encoder internals)
+        causality_axis_value = encoder_config.get("causality_axis", CausalityAxis.WIDTH.value)
+        self.causality_axis = CausalityAxis.str_to_enum(causality_axis_value)
+        self.is_causal = self.causality_axis == CausalityAxis.HEIGHT
+
+        self.encoder = Encoder(**encoder_config)
+        self.decoder = Decoder(**decoder_config)
+
+        self.per_channel_statistics = processor()
+
+    def _guess_config(self):
+        encoder_config = {
+            # Required parameters - based on ltx-video-av-1679000 model metadata
+            "ch": 128,
+            "out_ch": 8,
+            "ch_mult": [1, 2, 4],  # Based on metadata: [1, 2, 4] not [1, 2, 4, 8]
+            "num_res_blocks": 2,
+            "attn_resolutions": [],  # Based on metadata: empty list, no attention
+            "dropout": 0.0,
+            "resamp_with_conv": True,
+            "in_channels": 2,  # stereo
+            "resolution": 256,
+            "z_channels": 8,
+            "double_z": True,
+            "attn_type": "vanilla",
+            "mid_block_add_attention": False,  # Based on metadata: false
+            "norm_type": "pixel",
+            "causality_axis": "height",  # Based on metadata
+            "mel_bins": 64,  # Based on metadata: mel_bins = 64
+        }
+
+        decoder_config = {
+            # Inherits encoder config, can override specific params
+            **encoder_config,
+            "out_ch": 2,  # Stereo audio output (2 channels)
+            "give_pre_end": False,
+            "tanh_out": False,
+        }
+
+        config = {
+            "_class_name": "CausalAudioAutoencoder",
+            "sampling_rate": 16000,
+            "model": {
+                "params": {
+                    "encoder": encoder_config,
+                    "decoder": decoder_config,
+                }
+            },
+        }
+
+        return config
+
+    def get_config(self):
+        return {
+            "sampling_rate": self.sampling_rate,
+            "mel_bins": self.mel_bins,
+            "mel_hop_length": self.mel_hop_length,
+            "n_fft": self.n_fft,
+            "causality_axis": self.causality_axis.value,
+            "is_causal": self.is_causal,
+        }
+
+    def encode(self, x):
+        return self.encoder(x)
+
+    def decode(self, x, target_shape=None):
+        return self.decoder(x, target_shape=target_shape)

comfy/ldm/lightricks/vocoders/vocoder.py

@@ -0,0 +1,213 @@
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+import comfy.ops
+import numpy as np
+
+ops = comfy.ops.disable_weight_init
+
+LRELU_SLOPE = 0.1
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[0],
+                    padding=get_padding(kernel_size, dilation[0]),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[1],
+                    padding=get_padding(kernel_size, dilation[1]),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[2],
+                    padding=get_padding(kernel_size, dilation[2]),
+                ),
+            ]
+        )
+
+        self.convs2 = nn.ModuleList(
+            [
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=1,
+                    padding=get_padding(kernel_size, 1),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=1,
+                    padding=get_padding(kernel_size, 1),
+                ),
+                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.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[0],
+                    padding=get_padding(kernel_size, dilation[0]),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[1],
+                    padding=get_padding(kernel_size, dilation[1]),
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+
+class Vocoder(torch.nn.Module):
+    """
+    Vocoder model for synthesizing audio from spectrograms, based on: https://github.com/jik876/hifi-gan.
+
+    """
+
+    def __init__(self, config=None):
+        super(Vocoder, self).__init__()
+
+        if config is None:
+            config = self.get_default_config()
+
+        resblock_kernel_sizes = config.get("resblock_kernel_sizes", [3, 7, 11])
+        upsample_rates = config.get("upsample_rates", [6, 5, 2, 2, 2])
+        upsample_kernel_sizes = config.get("upsample_kernel_sizes", [16, 15, 8, 4, 4])
+        resblock_dilation_sizes = config.get("resblock_dilation_sizes", [[1, 3, 5], [1, 3, 5], [1, 3, 5]])
+        upsample_initial_channel = config.get("upsample_initial_channel", 1024)
+        stereo = config.get("stereo", True)
+        resblock = config.get("resblock", "1")
+
+        self.output_sample_rate = config.get("output_sample_rate")
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        in_channels = 128 if stereo else 64
+        self.conv_pre = ops.Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3)
+        resblock_class = ResBlock1 if resblock == "1" else ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                ops.ConvTranspose1d(
+                    upsample_initial_channel // (2**i),
+                    upsample_initial_channel // (2 ** (i + 1)),
+                    k,
+                    u,
+                    padding=(k - u) // 2,
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.resblocks.append(resblock_class(ch, k, d))
+
+        out_channels = 2 if stereo else 1
+        self.conv_post = ops.Conv1d(ch, out_channels, 7, 1, padding=3)
+
+        self.upsample_factor = np.prod([self.ups[i].stride[0] for i in range(len(self.ups))])
+
+    def get_default_config(self):
+        """Generate default configuration for the vocoder."""
+
+        config = {
+            "resblock_kernel_sizes": [3, 7, 11],
+            "upsample_rates": [6, 5, 2, 2, 2],
+            "upsample_kernel_sizes": [16, 15, 8, 4, 4],
+            "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            "upsample_initial_channel": 1024,
+            "stereo": True,
+            "resblock": "1",
+        }
+
+        return config
+
+    def forward(self, x):
+        """
+        Forward pass of the vocoder.
+
+        Args:
+            x: Input spectrogram tensor. Can be:
+               - 3D: (batch_size, channels, time_steps) for mono
+               - 4D: (batch_size, 2, channels, time_steps) for stereo
+
+        Returns:
+            Audio tensor of shape (batch_size, out_channels, audio_length)
+        """
+        if x.dim() == 4:  # stereo
+            assert x.shape[1] == 2, "Input must have 2 channels for stereo"
+            x = torch.cat((x[:, 0, :, :], x[:, 1, :, :]), dim=1)
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            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 = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x

comfy/ldm/lumina/controlnet.py

@@ -0,0 +1,160 @@
+import torch
+from torch import nn
+
+from .model import JointTransformerBlock
+
+class ZImageControlTransformerBlock(JointTransformerBlock):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+        block_id=0,
+        operation_settings=None,
+    ):
+        super().__init__(layer_id, dim, n_heads, n_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, qk_norm, modulation, z_image_modulation=True, 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 ZImage_Control(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int = 3840,
+        n_heads: int = 30,
+        n_kv_heads: int = 30,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: float = (8.0 / 3.0),
+        norm_eps: float = 1e-5,
+        qk_norm: bool = True,
+        n_control_layers=6,
+        control_in_dim=16,
+        additional_in_dim=0,
+        broken=False,
+        refiner_control=False,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__()
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.broken = broken
+        self.additional_in_dim = additional_in_dim
+        self.control_in_dim = control_in_dim
+        n_refiner_layers = 2
+        self.n_control_layers = n_control_layers
+        self.control_layers = nn.ModuleList(
+            [
+                ZImageControlTransformerBlock(
+                    i,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    block_id=i,
+                    operation_settings=operation_settings,
+                )
+                for i in range(self.n_control_layers)
+            ]
+        )
+
+        all_x_embedder = {}
+        patch_size = 2
+        f_patch_size = 1
+        x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * (self.control_in_dim + self.additional_in_dim), dim, bias=True, device=device, dtype=dtype)
+        all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder
+
+        self.refiner_control = refiner_control
+
+        self.control_all_x_embedder = nn.ModuleDict(all_x_embedder)
+        if self.refiner_control:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    ZImageControlTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        block_id=layer_id,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+        else:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=True,
+                        z_image_modulation=True,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+
+    def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input):
+        patch_size = 2
+        f_patch_size = 1
+        pH = pW = patch_size
+        B, C, H, W = control_context.shape
+        control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
+
+        x_attn_mask = None
+        if not self.refiner_control:
+            for layer in self.control_noise_refiner:
+                control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input)
+
+        return control_context
+
+    def forward_noise_refiner_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
+        if self.refiner_control:
+            if self.broken:
+                if layer_id == 0:
+                    return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                if layer_id > 0:
+                    out = None
+                    for i in range(1, len(self.control_layers)):
+                        o, control_context = self.control_layers[i](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                        if out is None:
+                            out = o
+
+                    return (out, control_context)
+            else:
+                return self.control_noise_refiner[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+        else:
+            return (None, control_context)
+
+    def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
+        return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)

comfy/ldm/lumina/model.py

@@ -11,6 +11,7 @@ import comfy.ldm.common_dit
 from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
 import comfy.patcher_extension
 
 
@@ -21,6 +22,10 @@ def modulate(x, scale):
 #                               Core NextDiT Model                              #
 #############################################################################
 
+def clamp_fp16(x):
+    if x.dtype == torch.float16:
+        return torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+    return x
 
 class JointAttention(nn.Module):
     """Multi-head attention module."""
@@ -31,6 +36,7 @@ class JointAttention(nn.Module):
         n_heads: int,
         n_kv_heads: Optional[int],
         qk_norm: bool,
+        out_bias: bool = False,
         operation_settings={},
     ):
         """
@@ -59,7 +65,7 @@ class JointAttention(nn.Module):
         self.out = operation_settings.get("operations").Linear(
             n_heads * self.head_dim,
             dim,
-            bias=False,
+            bias=out_bias,
             device=operation_settings.get("device"),
             dtype=operation_settings.get("dtype"),
         )
@@ -70,40 +76,12 @@ class JointAttention(nn.Module):
         else:
             self.q_norm = self.k_norm = nn.Identity()
 
-    @staticmethod
-    def apply_rotary_emb(
-        x_in: torch.Tensor,
-        freqs_cis: torch.Tensor,
-    ) -> torch.Tensor:
-        """
-        Apply rotary embeddings to input tensors using the given frequency
-        tensor.
-
-        This function applies rotary embeddings to the given query 'xq' and
-        key 'xk' 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_in (torch.Tensor): Query or Key tensor to apply rotary embeddings.
-            freqs_cis (torch.Tensor): Precomputed frequency tensor for complex
-                exponentials.
-
-        Returns:
-            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor
-                and key tensor with rotary embeddings.
-        """
-
-        t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2)
-        t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
-        return t_out.reshape(*x_in.shape)
-
     def forward(
         self,
         x: torch.Tensor,
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
+        transformer_options={},
     ) -> torch.Tensor:
         """
 
@@ -133,14 +111,13 @@ class JointAttention(nn.Module):
         xq = self.q_norm(xq)
         xk = self.k_norm(xk)
 
-        xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis)
-        xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis)
+        xq, xk = apply_rope(xq, xk, freqs_cis)
 
         n_rep = self.n_local_heads // self.n_local_kv_heads
         if n_rep >= 1:
             xk = xk.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
             xv = xv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
-        output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True)
+        output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True, transformer_options=transformer_options)
 
         return self.out(output)
 
@@ -196,7 +173,7 @@ class FeedForward(nn.Module):
 
     # @torch.compile
     def _forward_silu_gating(self, x1, x3):
-        return F.silu(x1) * x3
+        return clamp_fp16(F.silu(x1) * x3)
 
     def forward(self, x):
         return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
@@ -214,6 +191,8 @@ class JointTransformerBlock(nn.Module):
         norm_eps: float,
         qk_norm: bool,
         modulation=True,
+        z_image_modulation=False,
+        attn_out_bias=False,
         operation_settings={},
     ) -> None:
         """
@@ -234,10 +213,10 @@ class JointTransformerBlock(nn.Module):
         super().__init__()
         self.dim = dim
         self.head_dim = dim // n_heads
-        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings)
+        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, out_bias=attn_out_bias, operation_settings=operation_settings)
         self.feed_forward = FeedForward(
             dim=dim,
-            hidden_dim=4 * dim,
+            hidden_dim=dim,
             multiple_of=multiple_of,
             ffn_dim_multiplier=ffn_dim_multiplier,
             operation_settings=operation_settings,
@@ -251,16 +230,27 @@ class JointTransformerBlock(nn.Module):
 
         self.modulation = modulation
         if modulation:
-            self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                operation_settings.get("operations").Linear(
-                    min(dim, 1024),
-                    4 * dim,
-                    bias=True,
-                    device=operation_settings.get("device"),
-                    dtype=operation_settings.get("dtype"),
-                ),
-            )
+            if z_image_modulation:
+                self.adaLN_modulation = nn.Sequential(
+                    operation_settings.get("operations").Linear(
+                        min(dim, 256),
+                        4 * dim,
+                        bias=True,
+                        device=operation_settings.get("device"),
+                        dtype=operation_settings.get("dtype"),
+                    ),
+                )
+            else:
+                self.adaLN_modulation = nn.Sequential(
+                    nn.SiLU(),
+                    operation_settings.get("operations").Linear(
+                        min(dim, 1024),
+                        4 * dim,
+                        bias=True,
+                        device=operation_settings.get("device"),
+                        dtype=operation_settings.get("dtype"),
+                    ),
+                )
 
     def forward(
         self,
@@ -268,6 +258,7 @@ class JointTransformerBlock(nn.Module):
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor]=None,
+        transformer_options={},
     ):
         """
         Perform a forward pass through the TransformerBlock.
@@ -286,25 +277,27 @@ class JointTransformerBlock(nn.Module):
             scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
 
             x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
-                self.attention(
+                clamp_fp16(self.attention(
                     modulate(self.attention_norm1(x), scale_msa),
                     x_mask,
                     freqs_cis,
-                )
+                    transformer_options=transformer_options,
+                ))
             )
             x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
-                self.feed_forward(
+                clamp_fp16(self.feed_forward(
                     modulate(self.ffn_norm1(x), scale_mlp),
-                )
+                ))
             )
         else:
             assert adaln_input is None
             x = x + self.attention_norm2(
-                self.attention(
+                clamp_fp16(self.attention(
                     self.attention_norm1(x),
                     x_mask,
                     freqs_cis,
-                )
+                    transformer_options=transformer_options,
+                ))
             )
             x = x + self.ffn_norm2(
                 self.feed_forward(
@@ -319,7 +312,7 @@ class FinalLayer(nn.Module):
     The final layer of NextDiT.
     """
 
-    def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}):
+    def __init__(self, hidden_size, patch_size, out_channels, z_image_modulation=False, operation_settings={}):
         super().__init__()
         self.norm_final = operation_settings.get("operations").LayerNorm(
             hidden_size,
@@ -336,10 +329,15 @@ class FinalLayer(nn.Module):
             dtype=operation_settings.get("dtype"),
         )
 
+        if z_image_modulation:
+            min_mod = 256
+        else:
+            min_mod = 1024
+
         self.adaLN_modulation = nn.Sequential(
             nn.SiLU(),
             operation_settings.get("operations").Linear(
-                min(hidden_size, 1024),
+                min(hidden_size, min_mod),
                 hidden_size,
                 bias=True,
                 device=operation_settings.get("device"),
@@ -369,12 +367,17 @@ class NextDiT(nn.Module):
         n_heads: int = 32,
         n_kv_heads: Optional[int] = None,
         multiple_of: int = 256,
-        ffn_dim_multiplier: Optional[float] = None,
+        ffn_dim_multiplier: float = 4.0,
         norm_eps: float = 1e-5,
         qk_norm: bool = False,
         cap_feat_dim: int = 5120,
         axes_dims: List[int] = (16, 56, 56),
         axes_lens: List[int] = (1, 512, 512),
+        rope_theta=10000.0,
+        z_image_modulation=False,
+        time_scale=1.0,
+        pad_tokens_multiple=None,
+        clip_text_dim=None,
         image_model=None,
         device=None,
         dtype=None,
@@ -386,6 +389,8 @@ class NextDiT(nn.Module):
         self.in_channels = in_channels
         self.out_channels = in_channels
         self.patch_size = patch_size
+        self.time_scale = time_scale
+        self.pad_tokens_multiple = pad_tokens_multiple
 
         self.x_embedder = operation_settings.get("operations").Linear(
             in_features=patch_size * patch_size * in_channels,
@@ -407,6 +412,7 @@ class NextDiT(nn.Module):
                     norm_eps,
                     qk_norm,
                     modulation=True,
+                    z_image_modulation=z_image_modulation,
                     operation_settings=operation_settings,
                 )
                 for layer_id in range(n_refiner_layers)
@@ -430,7 +436,7 @@ class NextDiT(nn.Module):
             ]
         )
 
-        self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings)
+        self.t_embedder = TimestepEmbedder(min(dim, 1024), output_size=256 if z_image_modulation else None, **operation_settings)
         self.cap_embedder = nn.Sequential(
             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(
@@ -442,6 +448,31 @@ class NextDiT(nn.Module):
             ),
         )
 
+        self.clip_text_pooled_proj = None
+
+        if clip_text_dim is not None:
+            self.clip_text_dim = clip_text_dim
+            self.clip_text_pooled_proj = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(clip_text_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+                operation_settings.get("operations").Linear(
+                    clip_text_dim,
+                    clip_text_dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+            self.time_text_embed = nn.Sequential(
+                nn.SiLU(),
+                operation_settings.get("operations").Linear(
+                    min(dim, 1024) + clip_text_dim,
+                    min(dim, 1024),
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+
         self.layers = nn.ModuleList(
             [
                 JointTransformerBlock(
@@ -453,18 +484,25 @@ class NextDiT(nn.Module):
                     ffn_dim_multiplier,
                     norm_eps,
                     qk_norm,
+                    z_image_modulation=z_image_modulation,
+                    attn_out_bias=False,
                     operation_settings=operation_settings,
                 )
                 for layer_id in range(n_layers)
             ]
         )
-        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)
+        # This norm final is in the lumina 2.0 code but isn't actually used for anything.
+        # 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, z_image_modulation=z_image_modulation, operation_settings=operation_settings)
+
+        if self.pad_tokens_multiple is not None:
+            self.x_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
+            self.cap_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
 
         assert (dim // n_heads) == sum(axes_dims)
         self.axes_dims = axes_dims
         self.axes_lens = axes_lens
-        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims)
+        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=rope_theta, axes_dim=axes_dims)
         self.dim = dim
         self.n_heads = n_heads
 
@@ -494,101 +532,68 @@ class NextDiT(nn.Module):
         return imgs
 
     def patchify_and_embed(
-        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens
+        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, transformer_options={}
     ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
         bsz = len(x)
         pH = pW = self.patch_size
         device = x[0].device
-        dtype = x[0].dtype
+        orig_x = x
 
-        if cap_mask is not None:
-            l_effective_cap_len = cap_mask.sum(dim=1).tolist()
-        else:
-            l_effective_cap_len = [num_tokens] * bsz
+        if self.pad_tokens_multiple is not None:
+            pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
+            cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1)
 
-        if cap_mask is not None and not torch.is_floating_point(cap_mask):
-            cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max
+        cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device)
+        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0
 
-        img_sizes = [(img.size(1), img.size(2)) for img in x]
-        l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes]
+        B, C, H, W = x.shape
+        x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
 
-        max_seq_len = max(
-            (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len))
-        )
-        max_cap_len = max(l_effective_cap_len)
-        max_img_len = max(l_effective_img_len)
+        rope_options = transformer_options.get("rope_options", None)
+        h_scale = 1.0
+        w_scale = 1.0
+        h_start = 0
+        w_start = 0
+        if rope_options is not None:
+            h_scale = rope_options.get("scale_y", 1.0)
+            w_scale = rope_options.get("scale_x", 1.0)
 
-        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device)
+            h_start = rope_options.get("shift_y", 0.0)
+            w_start = rope_options.get("shift_x", 0.0)
 
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-            H, W = img_sizes[i]
-            H_tokens, W_tokens = H // pH, W // pW
-            assert H_tokens * W_tokens == img_len
+        H_tokens, W_tokens = H // pH, W // pW
+        x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device)
+        x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1
+        x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+        x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
 
-            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device)
-            position_ids[i, cap_len:cap_len+img_len, 0] = cap_len
-            row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
-            col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
-            position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
-            position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
+        if self.pad_tokens_multiple is not None:
+            pad_extra = (-x.shape[1]) % self.pad_tokens_multiple
+            x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1)
+            x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra))
 
-        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype)
+        freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
 
-        # build freqs_cis for cap and image individually
-        cap_freqs_cis_shape = list(freqs_cis.shape)
-        # cap_freqs_cis_shape[1] = max_cap_len
-        cap_freqs_cis_shape[1] = cap_feats.shape[1]
-        cap_freqs_cis = torch.zeros(*cap_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 in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-            cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len]
-            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len]
+        patches = transformer_options.get("patches", {})
 
         # refine context
         for layer in self.context_refiner:
-            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis)
+            cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
 
-        # refine image
-        flat_x = []
-        for i in range(bsz):
-            img = x[i]
-            C, H, W = img.size()
-            img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1)
-            flat_x.append(img)
-        x = flat_x
-        padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype)
-        padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device)
-        for i in range(bsz):
-            padded_img_embed[i, :l_effective_img_len[i]] = x[i]
-            padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max
-
-        padded_img_embed = self.x_embedder(padded_img_embed)
-        padded_img_mask = padded_img_mask.unsqueeze(1)
-        for layer in self.noise_refiner:
-            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t)
-
-        if cap_mask is not None:
-            mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device)
-            mask[:, :max_cap_len] = cap_mask[:, :max_cap_len]
-        else:
-            mask = None
-
-        padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype)
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-
-            padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len]
-            padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len]
+        padded_img_mask = None
+        x_input = x
+        for i, layer in enumerate(self.noise_refiner):
+            x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options)
+            if "noise_refiner" in patches:
+                for p in patches["noise_refiner"]:
+                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
+                    if "img" in out:
+                        x = out["img"]
 
+        padded_full_embed = torch.cat((cap_feats, x), dim=1)
+        mask = None
+        img_sizes = [(H, W)] * bsz
+        l_effective_cap_len = [cap_feats.shape[1]] * bsz
         return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
 
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
@@ -599,7 +604,7 @@ class NextDiT(nn.Module):
         ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
 
     # def forward(self, x, t, cap_feats, cap_mask):
-    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, transformer_options={}, **kwargs):
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
@@ -611,20 +616,41 @@ class NextDiT(nn.Module):
         y: (N,) tensor of text tokens/features
         """
 
-        t = self.t_embedder(t, dtype=x.dtype)  # (N, D)
+        t = self.t_embedder(t * self.time_scale, dtype=x.dtype)  # (N, D)
         adaln_input = t
 
         cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
 
+        if self.clip_text_pooled_proj is not None:
+            pooled = kwargs.get("clip_text_pooled", None)
+            if pooled is not None:
+                pooled = self.clip_text_pooled_proj(pooled)
+            else:
+                pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype)
+
+            adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1))
+
+        patches = transformer_options.get("patches", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens)
-        freqs_cis = freqs_cis.to(x.device)
+        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options)
+        freqs_cis = freqs_cis.to(img.device)
 
-        for layer in self.layers:
-            x = layer(x, mask, freqs_cis, adaln_input)
+        transformer_options["total_blocks"] = len(self.layers)
+        transformer_options["block_type"] = "double"
+        img_input = img
+        for i, layer in enumerate(self.layers):
+            transformer_options["block_index"] = i
+            img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
+            if "double_block" in patches:
+                for p in patches["double_block"]:
+                    out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
+                    if "img" in out:
+                        img[:, cap_size[0]:] = out["img"]
+                    if "txt" in out:
+                        img[:, :cap_size[0]] = out["txt"]
 
-        x = self.final_layer(x, adaln_input)
-        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]
+        img = self.final_layer(img, adaln_input)
+        img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w]
 
-        return -x
+        return -img
 

comfy/ldm/mmaudio/vae/activations.py

@@ -0,0 +1,120 @@
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+import comfy.model_management
+
+class Snake(nn.Module):
+    '''
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:
+            self.alpha = Parameter(torch.empty(in_features))
+        else:
+            self.alpha = Parameter(torch.empty(in_features))
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        '''
+        alpha = comfy.model_management.cast_to(self.alpha, dtype=x.dtype, device=x.device).unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x
+
+
+class SnakeBeta(nn.Module):
+    '''
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha - trainable parameter that controls frequency
+            - beta - trainable parameter that controls magnitude
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(SnakeBeta, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:
+            self.alpha = Parameter(torch.empty(in_features))
+            self.beta = Parameter(torch.empty(in_features))
+        else:
+            self.alpha = Parameter(torch.empty(in_features))
+            self.beta = Parameter(torch.empty(in_features))
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        '''
+        alpha = comfy.model_management.cast_to(self.alpha, dtype=x.dtype, device=x.device).unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        beta = comfy.model_management.cast_to(self.beta, dtype=x.dtype, device=x.device).unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

comfy/ldm/mmaudio/vae/alias_free_torch.py

@@ -0,0 +1,157 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+import comfy.model_management
+
+if 'sinc' in dir(torch):
+    sinc = torch.sinc
+else:
+    # This code is adopted from adefossez's julius.core.sinc under the MIT License
+    # https://adefossez.github.io/julius/julius/core.html
+    #   LICENSE is in incl_licenses directory.
+    def sinc(x: torch.Tensor):
+        """
+        Implementation of sinc, i.e. sin(pi * x) / (pi * x)
+        __Warning__: Different to julius.sinc, the input is multiplied by `pi`!
+        """
+        return torch.where(x == 0,
+                           torch.tensor(1., device=x.device, dtype=x.dtype),
+                           torch.sin(math.pi * x) / math.pi / x)
+
+
+# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License
+# https://adefossez.github.io/julius/julius/lowpass.html
+#   LICENSE is in incl_licenses directory.
+def kaiser_sinc_filter1d(cutoff, half_width, kernel_size): # return filter [1,1,kernel_size]
+    even = (kernel_size % 2 == 0)
+    half_size = kernel_size // 2
+
+    #For kaiser window
+    delta_f = 4 * half_width
+    A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95
+    if A > 50.:
+        beta = 0.1102 * (A - 8.7)
+    elif A >= 21.:
+        beta = 0.5842 * (A - 21)**0.4 + 0.07886 * (A - 21.)
+    else:
+        beta = 0.
+    window = torch.kaiser_window(kernel_size, beta=beta, periodic=False)
+
+    # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio
+    if even:
+        time = (torch.arange(-half_size, half_size) + 0.5)
+    else:
+        time = torch.arange(kernel_size) - half_size
+    if cutoff == 0:
+        filter_ = torch.zeros_like(time)
+    else:
+        filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
+        # Normalize filter to have sum = 1, otherwise we will have a small leakage
+        # of the constant component in the input signal.
+        filter_ /= filter_.sum()
+        filter = filter_.view(1, 1, kernel_size)
+
+    return filter
+
+
+class LowPassFilter1d(nn.Module):
+    def __init__(self,
+                 cutoff=0.5,
+                 half_width=0.6,
+                 stride: int = 1,
+                 padding: bool = True,
+                 padding_mode: str = 'replicate',
+                 kernel_size: int = 12):
+        # kernel_size should be even number for stylegan3 setup,
+        # in this implementation, odd number is also possible.
+        super().__init__()
+        if cutoff < -0.:
+            raise ValueError("Minimum cutoff must be larger than zero.")
+        if cutoff > 0.5:
+            raise ValueError("A cutoff above 0.5 does not make sense.")
+        self.kernel_size = kernel_size
+        self.even = (kernel_size % 2 == 0)
+        self.pad_left = kernel_size // 2 - int(self.even)
+        self.pad_right = kernel_size // 2
+        self.stride = stride
+        self.padding = padding
+        self.padding_mode = padding_mode
+        filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size)
+        self.register_buffer("filter", filter)
+
+    #input [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        if self.padding:
+            x = F.pad(x, (self.pad_left, self.pad_right),
+                      mode=self.padding_mode)
+        out = F.conv1d(x, comfy.model_management.cast_to(self.filter.expand(C, -1, -1), dtype=x.dtype, device=x.device),
+                       stride=self.stride, groups=C)
+
+        return out
+
+
+class UpSample1d(nn.Module):
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.stride = ratio
+        self.pad = self.kernel_size // ratio - 1
+        self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
+        self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2
+        filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio,
+                                      half_width=0.6 / ratio,
+                                      kernel_size=self.kernel_size)
+        self.register_buffer("filter", filter)
+
+    # x: [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        x = F.pad(x, (self.pad, self.pad), mode='replicate')
+        x = self.ratio * F.conv_transpose1d(
+            x, comfy.model_management.cast_to(self.filter.expand(C, -1, -1), dtype=x.dtype, device=x.device), stride=self.stride, groups=C)
+        x = x[..., self.pad_left:-self.pad_right]
+
+        return x
+
+
+class DownSample1d(nn.Module):
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.lowpass = LowPassFilter1d(cutoff=0.5 / ratio,
+                                       half_width=0.6 / ratio,
+                                       stride=ratio,
+                                       kernel_size=self.kernel_size)
+
+    def forward(self, x):
+        xx = self.lowpass(x)
+
+        return xx
+
+class Activation1d(nn.Module):
+    def __init__(self,
+                 activation,
+                 up_ratio: int = 2,
+                 down_ratio: int = 2,
+                 up_kernel_size: int = 12,
+                 down_kernel_size: int = 12):
+        super().__init__()
+        self.up_ratio = up_ratio
+        self.down_ratio = down_ratio
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+    # x: [B,C,T]
+    def forward(self, x):
+        x = self.upsample(x)
+        x = self.act(x)
+        x = self.downsample(x)
+
+        return x

comfy/ldm/mmaudio/vae/autoencoder.py

@@ -0,0 +1,156 @@
+from typing import Literal
+
+import torch
+import torch.nn as nn
+
+from .distributions import DiagonalGaussianDistribution
+from .vae import VAE_16k
+from .bigvgan import BigVGANVocoder
+import logging
+
+try:
+    import torchaudio
+except:
+    logging.warning("torchaudio missing, MMAudio VAE model will be broken")
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, *, norm_fn):
+    return norm_fn(torch.clamp(x, min=clip_val) * C)
+
+
+def spectral_normalize_torch(magnitudes, norm_fn):
+    output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn)
+    return output
+
+class MelConverter(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        sampling_rate: float,
+        n_fft: int,
+        num_mels: int,
+        hop_size: int,
+        win_size: int,
+        fmin: float,
+        fmax: float,
+        norm_fn,
+    ):
+        super().__init__()
+        self.sampling_rate = sampling_rate
+        self.n_fft = n_fft
+        self.num_mels = num_mels
+        self.hop_size = hop_size
+        self.win_size = win_size
+        self.fmin = fmin
+        self.fmax = fmax
+        self.norm_fn = norm_fn
+
+        # mel = librosa_mel_fn(sr=self.sampling_rate,
+        #                      n_fft=self.n_fft,
+        #                      n_mels=self.num_mels,
+        #                      fmin=self.fmin,
+        #                      fmax=self.fmax)
+        # mel_basis = torch.from_numpy(mel).float()
+        mel_basis = torch.empty((num_mels, 1 + n_fft // 2))
+        hann_window = torch.hann_window(self.win_size)
+
+        self.register_buffer('mel_basis', mel_basis)
+        self.register_buffer('hann_window', hann_window)
+
+    @property
+    def device(self):
+        return self.mel_basis.device
+
+    def forward(self, waveform: torch.Tensor, center: bool = False) -> torch.Tensor:
+        waveform = waveform.clamp(min=-1., max=1.).to(self.device)
+
+        waveform = torch.nn.functional.pad(
+            waveform.unsqueeze(1),
+            [int((self.n_fft - self.hop_size) / 2),
+             int((self.n_fft - self.hop_size) / 2)],
+            mode='reflect')
+        waveform = waveform.squeeze(1)
+
+        spec = torch.stft(waveform,
+                          self.n_fft,
+                          hop_length=self.hop_size,
+                          win_length=self.win_size,
+                          window=self.hann_window,
+                          center=center,
+                          pad_mode='reflect',
+                          normalized=False,
+                          onesided=True,
+                          return_complex=True)
+
+        spec = torch.view_as_real(spec)
+        spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+        spec = torch.matmul(self.mel_basis, spec)
+        spec = spectral_normalize_torch(spec, self.norm_fn)
+
+        return spec
+
+class AudioAutoencoder(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        # ckpt_path: str,
+        mode=Literal['16k', '44k'],
+        need_vae_encoder: bool = True,
+    ):
+        super().__init__()
+
+        assert mode == "16k", "Only 16k mode is supported currently."
+        self.mel_converter = MelConverter(sampling_rate=16_000,
+                            n_fft=1024,
+                            num_mels=80,
+                            hop_size=256,
+                            win_size=1024,
+                            fmin=0,
+                            fmax=8_000,
+                            norm_fn=torch.log10)
+
+        self.vae = VAE_16k().eval()
+
+        bigvgan_config = {
+            "resblock": "1",
+            "num_mels": 80,
+            "upsample_rates": [4, 4, 2, 2, 2, 2],
+            "upsample_kernel_sizes": [8, 8, 4, 4, 4, 4],
+            "upsample_initial_channel": 1536,
+            "resblock_kernel_sizes": [3, 7, 11],
+            "resblock_dilation_sizes": [
+                [1, 3, 5],
+                [1, 3, 5],
+                [1, 3, 5],
+            ],
+            "activation": "snakebeta",
+            "snake_logscale": True,
+        }
+
+        self.vocoder = BigVGANVocoder(
+            bigvgan_config
+        ).eval()
+
+    @torch.inference_mode()
+    def encode_audio(self, x) -> DiagonalGaussianDistribution:
+        # x: (B * L)
+        mel = self.mel_converter(x)
+        dist = self.vae.encode(mel)
+
+        return dist
+
+    @torch.no_grad()
+    def decode(self, z):
+        mel_decoded = self.vae.decode(z)
+        audio = self.vocoder(mel_decoded)
+
+        audio = torchaudio.functional.resample(audio, 16000, 44100)
+        return audio
+
+    @torch.no_grad()
+    def encode(self, audio):
+        audio = audio.mean(dim=1)
+        audio = torchaudio.functional.resample(audio, 44100, 16000)
+        dist = self.encode_audio(audio)
+        return dist.mean

comfy/ldm/mmaudio/vae/bigvgan.py

@@ -0,0 +1,219 @@
+# Copyright (c) 2022 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+import torch.nn as nn
+from types import SimpleNamespace
+from . import activations
+from .alias_free_torch import Activation1d
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+class AMPBlock1(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5), activation=None):
+        super(AMPBlock1, self).__init__()
+        self.h = h
+
+        self.convs1 = nn.ModuleList([
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[0],
+                       padding=get_padding(kernel_size, dilation[0])),
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[1],
+                       padding=get_padding(kernel_size, dilation[1])),
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[2],
+                       padding=get_padding(kernel_size, dilation[2]))
+        ])
+
+        self.convs2 = nn.ModuleList([
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1)),
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1)),
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1))
+        ])
+
+        self.num_layers = len(self.convs1) + len(self.convs2)  # total number of conv layers
+
+        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        acts1, acts2 = self.activations[::2], self.activations[1::2]
+        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2):
+            xt = a1(x)
+            xt = c1(xt)
+            xt = a2(xt)
+            xt = c2(xt)
+            x = xt + x
+
+        return x
+
+
+class AMPBlock2(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3), activation=None):
+        super(AMPBlock2, self).__init__()
+        self.h = h
+
+        self.convs = nn.ModuleList([
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[0],
+                       padding=get_padding(kernel_size, dilation[0])),
+                ops.Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[1],
+                       padding=get_padding(kernel_size, dilation[1]))
+        ])
+
+        self.num_layers = len(self.convs)  # total number of conv layers
+
+        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        for c, a in zip(self.convs, self.activations):
+            xt = a(x)
+            xt = c(xt)
+            x = xt + x
+
+        return x
+
+
+class BigVGANVocoder(torch.nn.Module):
+    # this is our main BigVGAN model. Applies anti-aliased periodic activation for resblocks.
+    def __init__(self, h):
+        super().__init__()
+        if isinstance(h, dict):
+            h = SimpleNamespace(**h)
+        self.h = h
+
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+
+        # pre conv
+        self.conv_pre = ops.Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
+
+        # define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
+        resblock = AMPBlock1 if h.resblock == '1' else AMPBlock2
+
+        # transposed conv-based upsamplers. does not apply anti-aliasing
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                nn.ModuleList([
+                        ops.ConvTranspose1d(h.upsample_initial_channel // (2**i),
+                                        h.upsample_initial_channel // (2**(i + 1)),
+                                        k,
+                                        u,
+                                        padding=(k - u) // 2)
+                ]))
+
+        # residual blocks using anti-aliased multi-periodicity composition modules (AMP)
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2**(i + 1))
+            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock(h, ch, k, d, activation=h.activation))
+
+        # post conv
+        if h.activation == "snake":  # periodic nonlinearity with snake function and anti-aliasing
+            activation_post = activations.Snake(ch, alpha_logscale=h.snake_logscale)
+            self.activation_post = Activation1d(activation=activation_post)
+        elif h.activation == "snakebeta":  # periodic nonlinearity with snakebeta function and anti-aliasing
+            activation_post = activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale)
+            self.activation_post = Activation1d(activation=activation_post)
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+        self.conv_post = ops.Conv1d(ch, 1, 7, 1, padding=3)
+
+
+    def forward(self, x):
+        # pre conv
+        x = self.conv_pre(x)
+
+        for i in range(self.num_upsamples):
+            # upsampling
+            for i_up in range(len(self.ups[i])):
+                x = self.ups[i][i_up](x)
+            # AMP blocks
+            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
+
+        # post conv
+        x = self.activation_post(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x

comfy/ldm/mmaudio/vae/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean, device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape, device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

comfy/ldm/mmaudio/vae/vae.py

@@ -0,0 +1,358 @@
+import logging
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from .vae_modules import (AttnBlock1D, Downsample1D, ResnetBlock1D,
+                                                 Upsample1D, nonlinearity)
+from .distributions import DiagonalGaussianDistribution
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+log = logging.getLogger()
+
+DATA_MEAN_80D = [
+    -1.6058, -1.3676, -1.2520, -1.2453, -1.2078, -1.2224, -1.2419, -1.2439, -1.2922, -1.2927,
+    -1.3170, -1.3543, -1.3401, -1.3836, -1.3907, -1.3912, -1.4313, -1.4152, -1.4527, -1.4728,
+    -1.4568, -1.5101, -1.5051, -1.5172, -1.5623, -1.5373, -1.5746, -1.5687, -1.6032, -1.6131,
+    -1.6081, -1.6331, -1.6489, -1.6489, -1.6700, -1.6738, -1.6953, -1.6969, -1.7048, -1.7280,
+    -1.7361, -1.7495, -1.7658, -1.7814, -1.7889, -1.8064, -1.8221, -1.8377, -1.8417, -1.8643,
+    -1.8857, -1.8929, -1.9173, -1.9379, -1.9531, -1.9673, -1.9824, -2.0042, -2.0215, -2.0436,
+    -2.0766, -2.1064, -2.1418, -2.1855, -2.2319, -2.2767, -2.3161, -2.3572, -2.3954, -2.4282,
+    -2.4659, -2.5072, -2.5552, -2.6074, -2.6584, -2.7107, -2.7634, -2.8266, -2.8981, -2.9673
+]
+
+DATA_STD_80D = [
+    1.0291, 1.0411, 1.0043, 0.9820, 0.9677, 0.9543, 0.9450, 0.9392, 0.9343, 0.9297, 0.9276, 0.9263,
+    0.9242, 0.9254, 0.9232, 0.9281, 0.9263, 0.9315, 0.9274, 0.9247, 0.9277, 0.9199, 0.9188, 0.9194,
+    0.9160, 0.9161, 0.9146, 0.9161, 0.9100, 0.9095, 0.9145, 0.9076, 0.9066, 0.9095, 0.9032, 0.9043,
+    0.9038, 0.9011, 0.9019, 0.9010, 0.8984, 0.8983, 0.8986, 0.8961, 0.8962, 0.8978, 0.8962, 0.8973,
+    0.8993, 0.8976, 0.8995, 0.9016, 0.8982, 0.8972, 0.8974, 0.8949, 0.8940, 0.8947, 0.8936, 0.8939,
+    0.8951, 0.8956, 0.9017, 0.9167, 0.9436, 0.9690, 1.0003, 1.0225, 1.0381, 1.0491, 1.0545, 1.0604,
+    1.0761, 1.0929, 1.1089, 1.1196, 1.1176, 1.1156, 1.1117, 1.1070
+]
+
+DATA_MEAN_128D = [
+    -3.3462, -2.6723, -2.4893, -2.3143, -2.2664, -2.3317, -2.1802, -2.4006, -2.2357, -2.4597,
+    -2.3717, -2.4690, -2.5142, -2.4919, -2.6610, -2.5047, -2.7483, -2.5926, -2.7462, -2.7033,
+    -2.7386, -2.8112, -2.7502, -2.9594, -2.7473, -3.0035, -2.8891, -2.9922, -2.9856, -3.0157,
+    -3.1191, -2.9893, -3.1718, -3.0745, -3.1879, -3.2310, -3.1424, -3.2296, -3.2791, -3.2782,
+    -3.2756, -3.3134, -3.3509, -3.3750, -3.3951, -3.3698, -3.4505, -3.4509, -3.5089, -3.4647,
+    -3.5536, -3.5788, -3.5867, -3.6036, -3.6400, -3.6747, -3.7072, -3.7279, -3.7283, -3.7795,
+    -3.8259, -3.8447, -3.8663, -3.9182, -3.9605, -3.9861, -4.0105, -4.0373, -4.0762, -4.1121,
+    -4.1488, -4.1874, -4.2461, -4.3170, -4.3639, -4.4452, -4.5282, -4.6297, -4.7019, -4.7960,
+    -4.8700, -4.9507, -5.0303, -5.0866, -5.1634, -5.2342, -5.3242, -5.4053, -5.4927, -5.5712,
+    -5.6464, -5.7052, -5.7619, -5.8410, -5.9188, -6.0103, -6.0955, -6.1673, -6.2362, -6.3120,
+    -6.3926, -6.4797, -6.5565, -6.6511, -6.8130, -6.9961, -7.1275, -7.2457, -7.3576, -7.4663,
+    -7.6136, -7.7469, -7.8815, -8.0132, -8.1515, -8.3071, -8.4722, -8.7418, -9.3975, -9.6628,
+    -9.7671, -9.8863, -9.9992, -10.0860, -10.1709, -10.5418, -11.2795, -11.3861
+]
+
+DATA_STD_128D = [
+    2.3804, 2.4368, 2.3772, 2.3145, 2.2803, 2.2510, 2.2316, 2.2083, 2.1996, 2.1835, 2.1769, 2.1659,
+    2.1631, 2.1618, 2.1540, 2.1606, 2.1571, 2.1567, 2.1612, 2.1579, 2.1679, 2.1683, 2.1634, 2.1557,
+    2.1668, 2.1518, 2.1415, 2.1449, 2.1406, 2.1350, 2.1313, 2.1415, 2.1281, 2.1352, 2.1219, 2.1182,
+    2.1327, 2.1195, 2.1137, 2.1080, 2.1179, 2.1036, 2.1087, 2.1036, 2.1015, 2.1068, 2.0975, 2.0991,
+    2.0902, 2.1015, 2.0857, 2.0920, 2.0893, 2.0897, 2.0910, 2.0881, 2.0925, 2.0873, 2.0960, 2.0900,
+    2.0957, 2.0958, 2.0978, 2.0936, 2.0886, 2.0905, 2.0845, 2.0855, 2.0796, 2.0840, 2.0813, 2.0817,
+    2.0838, 2.0840, 2.0917, 2.1061, 2.1431, 2.1976, 2.2482, 2.3055, 2.3700, 2.4088, 2.4372, 2.4609,
+    2.4731, 2.4847, 2.5072, 2.5451, 2.5772, 2.6147, 2.6529, 2.6596, 2.6645, 2.6726, 2.6803, 2.6812,
+    2.6899, 2.6916, 2.6931, 2.6998, 2.7062, 2.7262, 2.7222, 2.7158, 2.7041, 2.7485, 2.7491, 2.7451,
+    2.7485, 2.7233, 2.7297, 2.7233, 2.7145, 2.6958, 2.6788, 2.6439, 2.6007, 2.4786, 2.2469, 2.1877,
+    2.1392, 2.0717, 2.0107, 1.9676, 1.9140, 1.7102, 0.9101, 0.7164
+]
+
+
+class VAE(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        data_dim: int,
+        embed_dim: int,
+        hidden_dim: int,
+    ):
+        super().__init__()
+
+        if data_dim == 80:
+            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_80D, dtype=torch.float32))
+            self.data_std = nn.Buffer(torch.tensor(DATA_STD_80D, dtype=torch.float32))
+        elif data_dim == 128:
+            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_128D, dtype=torch.float32))
+            self.data_std = nn.Buffer(torch.tensor(DATA_STD_128D, dtype=torch.float32))
+
+        self.data_mean = self.data_mean.view(1, -1, 1)
+        self.data_std = self.data_std.view(1, -1, 1)
+
+        self.encoder = Encoder1D(
+            dim=hidden_dim,
+            ch_mult=(1, 2, 4),
+            num_res_blocks=2,
+            attn_layers=[3],
+            down_layers=[0],
+            in_dim=data_dim,
+            embed_dim=embed_dim,
+        )
+        self.decoder = Decoder1D(
+            dim=hidden_dim,
+            ch_mult=(1, 2, 4),
+            num_res_blocks=2,
+            attn_layers=[3],
+            down_layers=[0],
+            in_dim=data_dim,
+            out_dim=data_dim,
+            embed_dim=embed_dim,
+        )
+
+        self.embed_dim = embed_dim
+        # self.quant_conv = nn.Conv1d(2 * embed_dim, 2 * embed_dim, 1)
+        # self.post_quant_conv = nn.Conv1d(embed_dim, embed_dim, 1)
+
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        pass
+
+    def encode(self, x: torch.Tensor, normalize: bool = True) -> DiagonalGaussianDistribution:
+        if normalize:
+            x = self.normalize(x)
+        moments = self.encoder(x)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z: torch.Tensor, unnormalize: bool = True) -> torch.Tensor:
+        dec = self.decoder(z)
+        if unnormalize:
+            dec = self.unnormalize(dec)
+        return dec
+
+    def normalize(self, x: torch.Tensor) -> torch.Tensor:
+        return (x - comfy.model_management.cast_to(self.data_mean, dtype=x.dtype, device=x.device)) / comfy.model_management.cast_to(self.data_std, dtype=x.dtype, device=x.device)
+
+    def unnormalize(self, x: torch.Tensor) -> torch.Tensor:
+        return x * comfy.model_management.cast_to(self.data_std, dtype=x.dtype, device=x.device) + comfy.model_management.cast_to(self.data_mean, dtype=x.dtype, device=x.device)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        sample_posterior: bool = True,
+        rng: Optional[torch.Generator] = None,
+        normalize: bool = True,
+        unnormalize: bool = True,
+    ) -> tuple[torch.Tensor, DiagonalGaussianDistribution]:
+
+        posterior = self.encode(x, normalize=normalize)
+        if sample_posterior:
+            z = posterior.sample(rng)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, unnormalize=unnormalize)
+        return dec, posterior
+
+    def load_weights(self, src_dict) -> None:
+        self.load_state_dict(src_dict, strict=True)
+
+    @property
+    def device(self) -> torch.device:
+        return next(self.parameters()).device
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def remove_weight_norm(self):
+        return self
+
+
+class Encoder1D(nn.Module):
+
+    def __init__(self,
+                 *,
+                 dim: int,
+                 ch_mult: tuple[int] = (1, 2, 4, 8),
+                 num_res_blocks: int,
+                 attn_layers: list[int] = [],
+                 down_layers: list[int] = [],
+                 resamp_with_conv: bool = True,
+                 in_dim: int,
+                 embed_dim: int,
+                 double_z: bool = True,
+                 kernel_size: int = 3,
+                 clip_act: float = 256.0):
+        super().__init__()
+        self.dim = dim
+        self.num_layers = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.in_channels = in_dim
+        self.clip_act = clip_act
+        self.down_layers = down_layers
+        self.attn_layers = attn_layers
+        self.conv_in = ops.Conv1d(in_dim, self.dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+
+        in_ch_mult = (1, ) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        # downsampling
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_layers):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = dim * in_ch_mult[i_level]
+            block_out = dim * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock1D(in_dim=block_in,
+                                  out_dim=block_out,
+                                  kernel_size=kernel_size,
+                                  use_norm=True))
+                block_in = block_out
+                if i_level in attn_layers:
+                    attn.append(AttnBlock1D(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level in down_layers:
+                down.downsample = Downsample1D(block_in, resamp_with_conv)
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock1D(in_dim=block_in,
+                                         out_dim=block_in,
+                                         kernel_size=kernel_size,
+                                         use_norm=True)
+        self.mid.attn_1 = AttnBlock1D(block_in)
+        self.mid.block_2 = ResnetBlock1D(in_dim=block_in,
+                                         out_dim=block_in,
+                                         kernel_size=kernel_size,
+                                         use_norm=True)
+
+        # end
+        self.conv_out = ops.Conv1d(block_in,
+                                 2 * embed_dim if double_z else embed_dim,
+                                 kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+
+        self.learnable_gain = nn.Parameter(torch.zeros([]))
+
+    def forward(self, x):
+
+        # downsampling
+        h = self.conv_in(x)
+        for i_level in range(self.num_layers):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                h = h.clamp(-self.clip_act, self.clip_act)
+            if i_level in self.down_layers:
+                h = self.down[i_level].downsample(h)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        h = h.clamp(-self.clip_act, self.clip_act)
+
+        # end
+        h = nonlinearity(h)
+        h = self.conv_out(h) * (self.learnable_gain + 1)
+        return h
+
+
+class Decoder1D(nn.Module):
+
+    def __init__(self,
+                 *,
+                 dim: int,
+                 out_dim: int,
+                 ch_mult: tuple[int] = (1, 2, 4, 8),
+                 num_res_blocks: int,
+                 attn_layers: list[int] = [],
+                 down_layers: list[int] = [],
+                 kernel_size: int = 3,
+                 resamp_with_conv: bool = True,
+                 in_dim: int,
+                 embed_dim: int,
+                 clip_act: float = 256.0):
+        super().__init__()
+        self.ch = dim
+        self.num_layers = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.in_channels = in_dim
+        self.clip_act = clip_act
+        self.down_layers = [i + 1 for i in down_layers]  # each downlayer add one
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = dim * ch_mult[self.num_layers - 1]
+
+        # z to block_in
+        self.conv_in = ops.Conv1d(embed_dim, block_in, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
+        self.mid.attn_1 = AttnBlock1D(block_in)
+        self.mid.block_2 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_layers)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = dim * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock1D(in_dim=block_in, out_dim=block_out, use_norm=True))
+                block_in = block_out
+                if i_level in attn_layers:
+                    attn.append(AttnBlock1D(block_in))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level in self.down_layers:
+                up.upsample = Upsample1D(block_in, resamp_with_conv)
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.conv_out = ops.Conv1d(block_in, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+        self.learnable_gain = nn.Parameter(torch.zeros([]))
+
+    def forward(self, z):
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        h = h.clamp(-self.clip_act, self.clip_act)
+
+        # upsampling
+        for i_level in reversed(range(self.num_layers)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+                h = h.clamp(-self.clip_act, self.clip_act)
+            if i_level in self.down_layers:
+                h = self.up[i_level].upsample(h)
+
+        h = nonlinearity(h)
+        h = self.conv_out(h) * (self.learnable_gain + 1)
+        return h
+
+
+def VAE_16k(**kwargs) -> VAE:
+    return VAE(data_dim=80, embed_dim=20, hidden_dim=384, **kwargs)
+
+
+def VAE_44k(**kwargs) -> VAE:
+    return VAE(data_dim=128, embed_dim=40, hidden_dim=512, **kwargs)
+
+
+def get_my_vae(name: str, **kwargs) -> VAE:
+    if name == '16k':
+        return VAE_16k(**kwargs)
+    if name == '44k':
+        return VAE_44k(**kwargs)
+    raise ValueError(f'Unknown model: {name}')
+

comfy/ldm/mmaudio/vae/vae_modules.py

@@ -0,0 +1,121 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import vae_attention
+import math
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def nonlinearity(x):
+    # swish
+    return torch.nn.functional.silu(x) / 0.596
+
+def mp_sum(a, b, t=0.5):
+    return a.lerp(b, t) / math.sqrt((1 - t)**2 + t**2)
+
+def normalize(x, dim=None, eps=1e-4):
+    if dim is None:
+        dim = list(range(1, x.ndim))
+    norm = torch.linalg.vector_norm(x, dim=dim, keepdim=True, dtype=torch.float32)
+    norm = torch.add(eps, norm, alpha=math.sqrt(norm.numel() / x.numel()))
+    return x / norm.to(x.dtype)
+
+class ResnetBlock1D(nn.Module):
+
+    def __init__(self, *, in_dim, out_dim=None, conv_shortcut=False, kernel_size=3, use_norm=True):
+        super().__init__()
+        self.in_dim = in_dim
+        out_dim = in_dim if out_dim is None else out_dim
+        self.out_dim = out_dim
+        self.use_conv_shortcut = conv_shortcut
+        self.use_norm = use_norm
+
+        self.conv1 = ops.Conv1d(in_dim, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+        self.conv2 = ops.Conv1d(out_dim, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+        if self.in_dim != self.out_dim:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = ops.Conv1d(in_dim, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
+            else:
+                self.nin_shortcut = ops.Conv1d(in_dim, out_dim, kernel_size=1, padding=0, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+
+        # pixel norm
+        if self.use_norm:
+            x = normalize(x, dim=1)
+
+        h = x
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        h = nonlinearity(h)
+        h = self.conv2(h)
+
+        if self.in_dim != self.out_dim:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return mp_sum(x, h, t=0.3)
+
+
+class AttnBlock1D(nn.Module):
+
+    def __init__(self, in_channels, num_heads=1):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.num_heads = num_heads
+        self.qkv = ops.Conv1d(in_channels, in_channels * 3, kernel_size=1, padding=0, bias=False)
+        self.proj_out = ops.Conv1d(in_channels, in_channels, kernel_size=1, padding=0, bias=False)
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x):
+        h = x
+        y = self.qkv(h)
+        y = y.reshape(y.shape[0], -1, 3, y.shape[-1])
+        q, k, v = normalize(y, dim=1).unbind(2)
+
+        h = self.optimized_attention(q, k, v)
+        h = self.proj_out(h)
+
+        return mp_sum(x, h, t=0.3)
+
+
+class Upsample1D(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = ops.Conv1d(in_channels, in_channels, kernel_size=3, padding=1, bias=False)
+
+    def forward(self, x):
+        x = F.interpolate(x, scale_factor=2.0, mode='nearest-exact')  # support 3D tensor(B,C,T)
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample1D(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv1 = ops.Conv1d(in_channels, in_channels, kernel_size=1, padding=0, bias=False)
+            self.conv2 = ops.Conv1d(in_channels, in_channels, kernel_size=1, padding=0, bias=False)
+
+    def forward(self, x):
+
+        if self.with_conv:
+            x = self.conv1(x)
+
+        x = F.avg_pool1d(x, kernel_size=2, stride=2)
+
+        if self.with_conv:
+            x = self.conv2(x)
+
+        return x

comfy/ldm/models/autoencoder.py

@@ -9,6 +9,8 @@ from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistri
 from comfy.ldm.util import get_obj_from_str, instantiate_from_config
 from comfy.ldm.modules.ema import LitEma
 import comfy.ops
+from einops import rearrange
+import comfy.model_management
 
 class DiagonalGaussianRegularizer(torch.nn.Module):
     def __init__(self, sample: bool = False):
@@ -26,6 +28,12 @@ class DiagonalGaussianRegularizer(torch.nn.Module):
             z = posterior.mode()
         return z, None
 
+class EmptyRegularizer(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
+        return z, None
 
 class AbstractAutoencoder(torch.nn.Module):
     """
@@ -173,6 +181,21 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
         self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1)
         self.embed_dim = embed_dim
 
+        if ddconfig.get("batch_norm_latent", False):
+            self.bn_eps = 1e-4
+            self.bn_momentum = 0.1
+            self.ps = [2, 2]
+            self.bn = torch.nn.BatchNorm2d(math.prod(self.ps) * ddconfig["z_channels"],
+                                           eps=self.bn_eps,
+                                           momentum=self.bn_momentum,
+                                           affine=False,
+                                           track_running_stats=True,
+                                           )
+            self.bn.eval()
+        else:
+            self.bn = None
+
+
     def get_autoencoder_params(self) -> list:
         params = super().get_autoencoder_params()
         return params
@@ -195,11 +218,36 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
             z = torch.cat(z, 0)
 
         z, reg_log = self.regularization(z)
+
+        if self.bn is not None:
+            z = rearrange(z,
+                          "... c (i pi) (j pj)  -> ... (c pi pj) i j",
+                          pi=self.ps[0],
+                          pj=self.ps[1],
+                          )
+
+            z = torch.nn.functional.batch_norm(z,
+                                               comfy.model_management.cast_to(self.bn.running_mean, dtype=z.dtype, device=z.device),
+                                               comfy.model_management.cast_to(self.bn.running_var, dtype=z.dtype, device=z.device),
+                                               momentum=self.bn_momentum,
+                                               eps=self.bn_eps)
+
         if return_reg_log:
             return z, reg_log
         return z
 
     def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
+        if self.bn is not None:
+            s = torch.sqrt(comfy.model_management.cast_to(self.bn.running_var.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + self.bn_eps)
+            m = comfy.model_management.cast_to(self.bn.running_mean.view(1, -1, 1, 1), dtype=z.dtype, device=z.device)
+            z = z * s + m
+            z = rearrange(
+                z,
+                "... (c pi pj) i j -> ... c (i pi) (j pj)",
+                pi=self.ps[0],
+                pj=self.ps[1],
+            )
+
         if self.max_batch_size is None:
             dec = self.post_quant_conv(z)
             dec = self.decoder(dec, **decoder_kwargs)

comfy/ldm/modules/attention.py

@@ -5,8 +5,9 @@ import torch
 import torch.nn.functional as F
 from torch import nn, einsum
 from einops import rearrange, repeat
-from typing import Optional
+from typing import Optional, Any, Callable, Union
 import logging
+import functools
 
 from .diffusionmodules.util import AlphaBlender, timestep_embedding
 from .sub_quadratic_attention import efficient_dot_product_attention
@@ -17,23 +18,52 @@ if model_management.xformers_enabled():
     import xformers
     import xformers.ops
 
-if model_management.sage_attention_enabled():
-    try:
-        from sageattention import sageattn
-    except ModuleNotFoundError as e:
+SAGE_ATTENTION_IS_AVAILABLE = False
+try:
+    from sageattention import sageattn
+    SAGE_ATTENTION_IS_AVAILABLE = True
+except ImportError as e:
+    if model_management.sage_attention_enabled():
         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():
-    try:
-        from flash_attn import flash_attn_func
-    except ModuleNotFoundError:
+SAGE_ATTENTION3_IS_AVAILABLE = False
+try:
+    from sageattn3 import sageattn3_blackwell
+    SAGE_ATTENTION3_IS_AVAILABLE = True
+except ImportError:
+    pass
+
+FLASH_ATTENTION_IS_AVAILABLE = False
+try:
+    from flash_attn import flash_attn_func
+    FLASH_ATTENTION_IS_AVAILABLE = True
+except ImportError:
+    if model_management.flash_attention_enabled():
         logging.error(f"\n\nTo use the `--use-flash-attention` feature, the `flash-attn` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install flash-attn")
         exit(-1)
 
+REGISTERED_ATTENTION_FUNCTIONS = {}
+def register_attention_function(name: str, func: Callable):
+    # avoid replacing existing functions
+    if name not in REGISTERED_ATTENTION_FUNCTIONS:
+        REGISTERED_ATTENTION_FUNCTIONS[name] = func
+    else:
+        logging.warning(f"Attention function {name} already registered, skipping registration.")
+
+def get_attention_function(name: str, default: Any=...) -> Union[Callable, None]:
+    if name == "optimized":
+        return optimized_attention
+    elif name not in REGISTERED_ATTENTION_FUNCTIONS:
+        if default is ...:
+            raise KeyError(f"Attention function {name} not found.")
+        else:
+            return default
+    return REGISTERED_ATTENTION_FUNCTIONS[name]
+
 from comfy.cli_args import args
 import comfy.ops
 ops = comfy.ops.disable_weight_init
@@ -91,7 +121,27 @@ class FeedForward(nn.Module):
 def Normalize(in_channels, dtype=None, device=None):
     return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
 
-def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+
+def wrap_attn(func):
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs):
+        remove_attn_wrapper_key = False
+        try:
+            if "_inside_attn_wrapper" not in kwargs:
+                transformer_options = kwargs.get("transformer_options", None)
+                remove_attn_wrapper_key = True
+                kwargs["_inside_attn_wrapper"] = True
+                if transformer_options is not None:
+                    if "optimized_attention_override" in transformer_options:
+                        return transformer_options["optimized_attention_override"](func, *args, **kwargs)
+            return func(*args, **kwargs)
+        finally:
+            if remove_attn_wrapper_key:
+                del kwargs["_inside_attn_wrapper"]
+    return wrapper
+
+@wrap_attn
+def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
@@ -159,8 +209,8 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         )
     return out
 
-
-def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     attn_precision = get_attn_precision(attn_precision, query.dtype)
 
     if skip_reshape:
@@ -230,7 +280,8 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
         hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2)
     return hidden_states
 
-def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
@@ -359,7 +410,8 @@ try:
 except:
     pass
 
-def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     b = q.shape[0]
     dim_head = q.shape[-1]
     # check to make sure xformers isn't broken
@@ -374,7 +426,7 @@ def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_resh
             disabled_xformers = True
 
     if disabled_xformers:
-        return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape)
+        return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape, **kwargs)
 
     if skip_reshape:
         # b h k d -> b k h d
@@ -427,8 +479,8 @@ else:
     #TODO: other GPUs ?
     SDP_BATCH_LIMIT = 2**31
 
-
-def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     if skip_reshape:
         b, _, _, dim_head = q.shape
     else:
@@ -470,8 +522,9 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
     return out
 
-
-def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    exception_fallback = False
     if skip_reshape:
         b, _, _, dim_head = q.shape
         tensor_layout = "HND"
@@ -496,12 +549,14 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
     except Exception as e:
         logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+        exception_fallback = True
+    if exception_fallback:
         if tensor_layout == "NHD":
             q, k, v = map(
                 lambda t: t.transpose(1, 2),
                 (q, k, v),
             )
-        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape, **kwargs)
 
     if tensor_layout == "HND":
         if not skip_output_reshape:
@@ -515,6 +570,93 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
             out = out.reshape(b, -1, heads * dim_head)
     return out
 
+@wrap_attn
+def attention3_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    exception_fallback = False
+    if (q.device.type != "cuda" or
+        q.dtype not in (torch.float16, torch.bfloat16) or
+        mask is not None):
+        return attention_pytorch(
+            q, k, v, heads,
+            mask=mask,
+            attn_precision=attn_precision,
+            skip_reshape=skip_reshape,
+            skip_output_reshape=skip_output_reshape,
+            **kwargs
+        )
+
+    if skip_reshape:
+        B, H, L, D = q.shape
+        if H != heads:
+            return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=True,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+        q_s, k_s, v_s = q, k, v
+        N = q.shape[2]
+        dim_head = D
+    else:
+        B, N, inner_dim = q.shape
+        if inner_dim % heads != 0:
+            return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=False,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+        dim_head = inner_dim // heads
+
+    if dim_head >= 256 or N <= 1024:
+        return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=skip_reshape,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+
+    if not skip_reshape:
+        q_s, k_s, v_s = map(
+            lambda t: t.view(B, -1, heads, dim_head).permute(0, 2, 1, 3).contiguous(),
+            (q, k, v),
+        )
+        B, H, L, D = q_s.shape
+
+    try:
+        out = sageattn3_blackwell(q_s, k_s, v_s, is_causal=False)
+    except Exception as e:
+        exception_fallback = True
+        logging.error("Error running SageAttention3: %s, falling back to pytorch attention.", e)
+
+    if exception_fallback:
+        if not skip_reshape:
+            del q_s, k_s, v_s
+        return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=False,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+
+    if skip_reshape:
+        if not skip_output_reshape:
+            out = out.permute(0, 2, 1, 3).reshape(B, L, H * D)
+    else:
+        if skip_output_reshape:
+            pass
+        else:
+            out = out.permute(0, 2, 1, 3).reshape(B, L, H * D)
+
+    return out
 
 try:
     @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
@@ -534,8 +676,8 @@ except AttributeError as error:
                     dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
         assert False, f"Could not define flash_attn_wrapper: {FLASH_ATTN_ERROR}"
 
-
-def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     if skip_reshape:
         b, _, _, dim_head = q.shape
     else:
@@ -555,7 +697,8 @@ def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
             mask = mask.unsqueeze(1)
 
     try:
-        assert mask is None
+        if mask is not None:
+            raise RuntimeError("Mask must not be set for Flash attention")
         out = flash_attn_wrapper(
             q.transpose(1, 2),
             k.transpose(1, 2),
@@ -597,6 +740,21 @@ else:
 
 optimized_attention_masked = optimized_attention
 
+
+# register core-supported attention functions
+if SAGE_ATTENTION_IS_AVAILABLE:
+    register_attention_function("sage", attention_sage)
+if SAGE_ATTENTION3_IS_AVAILABLE:
+    register_attention_function("sage3", attention3_sage)
+if FLASH_ATTENTION_IS_AVAILABLE:
+    register_attention_function("flash", attention_flash)
+if model_management.xformers_enabled():
+    register_attention_function("xformers", attention_xformers)
+register_attention_function("pytorch", attention_pytorch)
+register_attention_function("sub_quad", attention_sub_quad)
+register_attention_function("split", attention_split)
+
+
 def optimized_attention_for_device(device, mask=False, small_input=False):
     if small_input:
         if model_management.pytorch_attention_enabled():
@@ -629,7 +787,7 @@ class CrossAttention(nn.Module):
 
         self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
 
-    def forward(self, x, context=None, value=None, mask=None):
+    def forward(self, x, context=None, value=None, mask=None, transformer_options={}):
         q = self.to_q(x)
         context = default(context, x)
         k = self.to_k(context)
@@ -640,9 +798,9 @@ class CrossAttention(nn.Module):
             v = self.to_v(context)
 
         if mask is None:
-            out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
+            out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
         else:
-            out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
+            out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options)
         return self.to_out(out)
 
 
@@ -746,7 +904,7 @@ class BasicTransformerBlock(nn.Module):
             n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options)
             n = self.attn1.to_out(n)
         else:
-            n = self.attn1(n, context=context_attn1, value=value_attn1)
+            n = self.attn1(n, context=context_attn1, value=value_attn1, transformer_options=transformer_options)
 
         if "attn1_output_patch" in transformer_patches:
             patch = transformer_patches["attn1_output_patch"]
@@ -786,7 +944,7 @@ class BasicTransformerBlock(nn.Module):
                 n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options)
                 n = self.attn2.to_out(n)
             else:
-                n = self.attn2(n, context=context_attn2, value=value_attn2)
+                n = self.attn2(n, context=context_attn2, value=value_attn2, transformer_options=transformer_options)
 
         if "attn2_output_patch" in transformer_patches:
             patch = transformer_patches["attn2_output_patch"]
@@ -1017,7 +1175,7 @@ class SpatialVideoTransformer(SpatialTransformer):
 
             B, S, C = x_mix.shape
             x_mix = rearrange(x_mix, "(b t) s c -> (b s) t c", t=timesteps)
-            x_mix = mix_block(x_mix, context=time_context) #TODO: transformer_options
+            x_mix = mix_block(x_mix, context=time_context, transformer_options=transformer_options)
             x_mix = rearrange(
                 x_mix, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
             )

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -211,12 +211,14 @@ class TimestepEmbedder(nn.Module):
     Embeds scalar timesteps into vector representations.
     """
 
-    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None):
         super().__init__()
+        if output_size is None:
+            output_size = hidden_size
         self.mlp = nn.Sequential(
             operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device),
             nn.SiLU(),
-            operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device),
+            operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device),
         )
         self.frequency_embedding_size = frequency_embedding_size
 
@@ -606,7 +608,7 @@ def block_mixing(*args, use_checkpoint=True, **kwargs):
         return _block_mixing(*args, **kwargs)
 
 
-def _block_mixing(context, x, context_block, x_block, c):
+def _block_mixing(context, x, context_block, x_block, c, transformer_options={}):
     context_qkv, context_intermediates = context_block.pre_attention(context, c)
 
     if x_block.x_block_self_attn:
@@ -622,6 +624,7 @@ def _block_mixing(context, x, context_block, x_block, c):
     attn = optimized_attention(
         qkv[0], qkv[1], qkv[2],
         heads=x_block.attn.num_heads,
+        transformer_options=transformer_options,
     )
     context_attn, x_attn = (
         attn[:, : context_qkv[0].shape[1]],
@@ -637,6 +640,7 @@ def _block_mixing(context, x, context_block, x_block, c):
         attn2 = optimized_attention(
                 x_qkv2[0], x_qkv2[1], x_qkv2[2],
                 heads=x_block.attn2.num_heads,
+                transformer_options=transformer_options,
             )
         x = x_block.post_attention_x(x_attn, attn2, *x_intermediates)
     else:
@@ -958,10 +962,10 @@ class MMDiT(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["txt"], out["img"] = self.joint_blocks[i](args["txt"], args["img"], c=args["vec"])
+                    out["txt"], out["img"] = self.joint_blocks[i](args["txt"], args["img"], c=args["vec"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": c_mod}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": c_mod, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 context = out["txt"]
                 x = out["img"]
             else:
@@ -970,6 +974,7 @@ class MMDiT(nn.Module):
                     x,
                     c=c_mod,
                     use_checkpoint=self.use_checkpoint,
+                    transformer_options=transformer_options,
                 )
             if control is not None:
                 control_o = control.get("output")

comfy/ldm/modules/diffusionmodules/model.py

@@ -13,6 +13,12 @@ if model_management.xformers_enabled_vae():
     import xformers
     import xformers.ops
 
+def torch_cat_if_needed(xl, dim):
+    if len(xl) > 1:
+        return torch.cat(xl, dim)
+    else:
+        return xl[0]
+
 def get_timestep_embedding(timesteps, embedding_dim):
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models:
@@ -43,6 +49,37 @@ def Normalize(in_channels, num_groups=32):
     return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
 
 
+class CarriedConv3d(nn.Module):
+    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs):
+        super().__init__()
+        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None):
+
+    x = xl[0]
+    xl.clear()
+
+    if isinstance(op, CarriedConv3d):
+        if conv_carry_in is None:
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate')
+        else:
+            carry_len = conv_carry_in[0].shape[2]
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
+            x = torch.cat([conv_carry_in.pop(0), x], dim=2)
+
+        if conv_carry_out is not None:
+            to_push = x[:, :, -2:, :, :].clone()
+            conv_carry_out.append(to_push)
+
+    out = op(x)
+
+    return out
+
+
 class VideoConv3d(nn.Module):
     def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs):
         super().__init__()
@@ -89,29 +126,24 @@ class Upsample(nn.Module):
                                         stride=1,
                                         padding=1)
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         scale_factor = self.scale_factor
         if isinstance(scale_factor, (int, float)):
             scale_factor = (scale_factor,) * (x.ndim - 2)
 
         if x.ndim == 5 and scale_factor[0] > 1.0:
-            t = x.shape[2]
-            if t > 1:
-                a, b = x.split((1, t - 1), dim=2)
-                del x
-                b = interpolate_up(b, scale_factor)
-            else:
-                a = x
-
-            a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)
-            if t > 1:
-                x = torch.cat((a, b), dim=2)
-            else:
-                x = a
+            results = []
+            if conv_carry_in is None:
+                first = x[:, :, :1, :, :]
+                results.append(interpolate_up(first.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2))
+                x = x[:, :, 1:, :, :]
+            if x.shape[2] > 0:
+                results.append(interpolate_up(x, scale_factor))
+            x = torch_cat_if_needed(results, dim=2)
         else:
             x = interpolate_up(x, scale_factor)
         if self.with_conv:
-            x = self.conv(x)
+            x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
         return x
 
 
@@ -127,17 +159,20 @@ class Downsample(nn.Module):
                                         stride=stride,
                                         padding=0)
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         if self.with_conv:
-            if x.ndim == 4:
+            if isinstance(self.conv, CarriedConv3d):
+                x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+            elif x.ndim == 4:
                 pad = (0, 1, 0, 1)
                 mode = "constant"
                 x = torch.nn.functional.pad(x, pad, mode=mode, value=0)
+                x = self.conv(x)
             elif x.ndim == 5:
                 pad = (1, 1, 1, 1, 2, 0)
                 mode = "replicate"
                 x = torch.nn.functional.pad(x, pad, mode=mode)
-            x = self.conv(x)
+                x = self.conv(x)
         else:
             x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
         return x
@@ -145,7 +180,7 @@ class Downsample(nn.Module):
 
 class ResnetBlock(nn.Module):
     def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
-                 dropout, temb_channels=512, conv_op=ops.Conv2d):
+                 dropout=0.0, temb_channels=512, conv_op=ops.Conv2d, norm_op=Normalize):
         super().__init__()
         self.in_channels = in_channels
         out_channels = in_channels if out_channels is None else out_channels
@@ -153,7 +188,7 @@ class ResnetBlock(nn.Module):
         self.use_conv_shortcut = conv_shortcut
 
         self.swish = torch.nn.SiLU(inplace=True)
-        self.norm1 = Normalize(in_channels)
+        self.norm1 = norm_op(in_channels)
         self.conv1 = conv_op(in_channels,
                                      out_channels,
                                      kernel_size=3,
@@ -162,7 +197,7 @@ class ResnetBlock(nn.Module):
         if temb_channels > 0:
             self.temb_proj = ops.Linear(temb_channels,
                                              out_channels)
-        self.norm2 = Normalize(out_channels)
+        self.norm2 = norm_op(out_channels)
         self.dropout = torch.nn.Dropout(dropout, inplace=True)
         self.conv2 = conv_op(out_channels,
                                      out_channels,
@@ -183,23 +218,23 @@ class ResnetBlock(nn.Module):
                                                     stride=1,
                                                     padding=0)
 
-    def forward(self, x, temb):
+    def forward(self, x, temb=None, conv_carry_in=None, conv_carry_out=None):
         h = x
         h = self.norm1(h)
-        h = self.swish(h)
-        h = self.conv1(h)
+        h = [ self.swish(h) ]
+        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
 
         if temb is not None:
             h = h + self.temb_proj(self.swish(temb))[:,:,None,None]
 
         h = self.norm2(h)
         h = self.swish(h)
-        h = self.dropout(h)
-        h = self.conv2(h)
+        h = [ self.dropout(h) ]
+        h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
 
         if self.in_channels != self.out_channels:
             if self.use_conv_shortcut:
-                x = self.conv_shortcut(x)
+                x = conv_carry_causal_3d([x], self.conv_shortcut, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
             else:
                 x = self.nin_shortcut(x)
 
@@ -279,6 +314,7 @@ def pytorch_attention(q, k, v):
     orig_shape = q.shape
     B = orig_shape[0]
     C = orig_shape[1]
+    oom_fallback = False
     q, k, v = map(
         lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(),
         (q, k, v),
@@ -289,6 +325,8 @@ def pytorch_attention(q, k, v):
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")
+        oom_fallback = True
+    if oom_fallback:
         out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
     return out
 
@@ -305,11 +343,11 @@ def vae_attention():
         return normal_attention
 
 class AttnBlock(nn.Module):
-    def __init__(self, in_channels, conv_op=ops.Conv2d):
+    def __init__(self, in_channels, conv_op=ops.Conv2d, norm_op=Normalize):
         super().__init__()
         self.in_channels = in_channels
 
-        self.norm = Normalize(in_channels)
+        self.norm = norm_op(in_channels)
         self.q = conv_op(in_channels,
                                  in_channels,
                                  kernel_size=1,
@@ -356,7 +394,8 @@ class Model(nn.Module):
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
                  resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
         super().__init__()
-        if use_linear_attn: attn_type = "linear"
+        if use_linear_attn:
+            attn_type = "linear"
         self.ch = ch
         self.temb_ch = self.ch*4
         self.num_resolutions = len(ch_mult)
@@ -510,16 +549,22 @@ class Encoder(nn.Module):
                  conv3d=False, time_compress=None,
                  **ignore_kwargs):
         super().__init__()
-        if use_linear_attn: attn_type = "linear"
+        if use_linear_attn:
+            attn_type = "linear"
         self.ch = ch
         self.temb_ch = 0
         self.num_resolutions = len(ch_mult)
         self.num_res_blocks = num_res_blocks
         self.resolution = resolution
         self.in_channels = in_channels
+        self.carried = False
 
         if conv3d:
-            conv_op = VideoConv3d
+            if not attn_resolutions:
+                conv_op = CarriedConv3d
+                self.carried = True
+            else:
+                conv_op = VideoConv3d
             mid_attn_conv_op = ops.Conv3d
         else:
             conv_op = ops.Conv2d
@@ -532,6 +577,7 @@ class Encoder(nn.Module):
                                        stride=1,
                                        padding=1)
 
+        self.time_compress = 1
         curr_res = resolution
         in_ch_mult = (1,)+tuple(ch_mult)
         self.in_ch_mult = in_ch_mult
@@ -558,10 +604,15 @@ class Encoder(nn.Module):
                 if time_compress is not None:
                     if (self.num_resolutions - 1 - i_level) > math.log2(time_compress):
                         stride = (1, 2, 2)
+                else:
+                    self.time_compress *= 2
                 down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op)
                 curr_res = curr_res // 2
             self.down.append(down)
 
+        if time_compress is not None:
+            self.time_compress = time_compress
+
         # middle
         self.mid = nn.Module()
         self.mid.block_1 = ResnetBlock(in_channels=block_in,
@@ -587,15 +638,42 @@ class Encoder(nn.Module):
     def forward(self, x):
         # timestep embedding
         temb = None
-        # downsampling
-        h = self.conv_in(x)
-        for i_level in range(self.num_resolutions):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](h, temb)
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-            if i_level != self.num_resolutions-1:
-                h = self.down[i_level].downsample(h)
+
+        if self.carried:
+            xl = [x[:, :, :1, :, :]]
+            if x.shape[2] > self.time_compress:
+                tc = self.time_compress
+                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // tc) * tc, :, :], tc * 2, dim = 2)
+            x = xl
+        else:
+            x = [x]
+        out = []
+
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+
+            # downsampling
+            x1 = [ x1 ]
+            h1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
+
+            for i_level in range(self.num_resolutions):
+                for i_block in range(self.num_res_blocks):
+                    h1 = self.down[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out)
+                    if len(self.down[i_level].attn) > 0:
+                        assert i == 0 #carried should not happen if attn exists
+                        h1 = self.down[i_level].attn[i_block](h1)
+                if i_level != self.num_resolutions-1:
+                    h1 = self.down[i_level].downsample(h1, conv_carry_in, conv_carry_out)
+
+            out.append(h1)
+            conv_carry_in = conv_carry_out
+
+        h = torch_cat_if_needed(out, dim=2)
+        del out
 
         # middle
         h = self.mid.block_1(h, temb)
@@ -604,15 +682,15 @@ class Encoder(nn.Module):
 
         # end
         h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h)
+        h = [ nonlinearity(h) ]
+        h = conv_carry_causal_3d(h, self.conv_out)
         return h
 
 
 class Decoder(nn.Module):
     def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
-                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 resolution, z_channels, tanh_out=False, use_linear_attn=False,
                  conv_out_op=ops.Conv2d,
                  resnet_op=ResnetBlock,
                  attn_op=AttnBlock,
@@ -626,12 +704,18 @@ class Decoder(nn.Module):
         self.num_res_blocks = num_res_blocks
         self.resolution = resolution
         self.in_channels = in_channels
-        self.give_pre_end = give_pre_end
         self.tanh_out = tanh_out
+        self.carried = False
 
         if conv3d:
-            conv_op = VideoConv3d
-            conv_out_op = VideoConv3d
+            if not attn_resolutions and resnet_op == ResnetBlock:
+                conv_op = CarriedConv3d
+                conv_out_op = CarriedConv3d
+                self.carried = True
+            else:
+                conv_op = VideoConv3d
+                conv_out_op = VideoConv3d
+
             mid_attn_conv_op = ops.Conv3d
         else:
             conv_op = ops.Conv2d
@@ -706,29 +790,43 @@ class Decoder(nn.Module):
         temb = None
 
         # z to block_in
-        h = self.conv_in(z)
+        h = conv_carry_causal_3d([z], self.conv_in)
 
         # middle
         h = self.mid.block_1(h, temb, **kwargs)
         h = self.mid.attn_1(h, **kwargs)
         h = self.mid.block_2(h, temb, **kwargs)
 
+        if self.carried:
+            h = torch.split(h, 2, dim=2)
+        else:
+            h = [ h ]
+        out = []
+
+        conv_carry_in = None
+
         # upsampling
-        for i_level in reversed(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks+1):
-                h = self.up[i_level].block[i_block](h, temb, **kwargs)
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h, **kwargs)
-            if i_level != 0:
-                h = self.up[i_level].upsample(h)
+        for i, h1 in enumerate(h):
+            conv_carry_out = []
+            if i == len(h) - 1:
+                conv_carry_out = None
+            for i_level in reversed(range(self.num_resolutions)):
+                for i_block in range(self.num_res_blocks+1):
+                    h1 = self.up[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out, **kwargs)
+                    if len(self.up[i_level].attn) > 0:
+                        assert i == 0 #carried should not happen if attn exists
+                        h1 = self.up[i_level].attn[i_block](h1, **kwargs)
+                if i_level != 0:
+                    h1 = self.up[i_level].upsample(h1, conv_carry_in, conv_carry_out)
 
-        # end
-        if self.give_pre_end:
-            return h
+            h1 = self.norm_out(h1)
+            h1 = [ nonlinearity(h1) ]
+            h1 = conv_carry_causal_3d(h1, self.conv_out, conv_carry_in, conv_carry_out)
+            if self.tanh_out:
+                h1 = torch.tanh(h1)
+            out.append(h1)
+            conv_carry_in = conv_carry_out
 
-        h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h, **kwargs)
-        if self.tanh_out:
-            h = torch.tanh(h)
-        return h
+        out = torch_cat_if_needed(out, dim=2)
+
+        return out

comfy/ldm/modules/ema.py

@@ -45,7 +45,7 @@ class LitEma(nn.Module):
                     shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
                     shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
                 else:
-                    assert not key in self.m_name2s_name
+                    assert key not in self.m_name2s_name
 
     def copy_to(self, model):
         m_param = dict(model.named_parameters())
@@ -54,7 +54,7 @@ class LitEma(nn.Module):
             if m_param[key].requires_grad:
                 m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
             else:
-                assert not key in self.m_name2s_name
+                assert key not in self.m_name2s_name
 
     def store(self, parameters):
         """

comfy/ldm/omnigen/omnigen2.py

@@ -120,7 +120,7 @@ class Attention(nn.Module):
             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:
+    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, transformer_options={}) -> torch.Tensor:
         batch_size, sequence_length, _ = hidden_states.shape
 
         query = self.to_q(hidden_states)
@@ -146,7 +146,7 @@ class Attention(nn.Module):
             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 = optimized_attention_masked(query, key, value, self.heads, attention_mask, skip_reshape=True, transformer_options=transformer_options)
         hidden_states = self.to_out[0](hidden_states)
         return hidden_states
 
@@ -182,16 +182,16 @@ class OmniGen2TransformerBlock(nn.Module):
         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:
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, image_rotary_emb: torch.Tensor, temb: Optional[torch.Tensor] = None, transformer_options={}) -> 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)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb, transformer_options=transformer_options)
             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)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb, transformer_options=transformer_options)
             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)
@@ -390,7 +390,7 @@ class OmniGen2Transformer2DModel(nn.Module):
             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):
+    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, transformer_options={}):
         batch_size = len(hidden_states)
 
         hidden_states = self.x_embedder(hidden_states)
@@ -405,17 +405,17 @@ class OmniGen2Transformer2DModel(nn.Module):
                     shift += ref_img_len
 
         for layer in self.noise_refiner:
-            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
+            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb, transformer_options=transformer_options)
 
         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)
+                ref_image_hidden_states = layer(ref_image_hidden_states, padded_ref_img_mask, ref_img_rotary_emb, temb, transformer_options=transformer_options)
 
             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):
+    def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, transformer_options={}, **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
@@ -444,7 +444,7 @@ class OmniGen2Transformer2DModel(nn.Module):
         )
 
         for layer in self.context_refiner:
-            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)
+            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb, transformer_options=transformer_options)
 
         img_len = hidden_states.shape[1]
         combined_img_hidden_states = self.img_patch_embed_and_refine(
@@ -453,13 +453,14 @@ class OmniGen2Transformer2DModel(nn.Module):
             noise_rotary_emb, ref_img_rotary_emb,
             l_effective_ref_img_len, l_effective_img_len,
             temb,
+            transformer_options=transformer_options,
         )
 
         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 = layer(hidden_states, attention_mask, rotary_emb, temb, transformer_options=transformer_options)
 
         hidden_states = self.norm_out(hidden_states, temb)
 

comfy/ldm/qwen_image/controlnet.py

@@ -44,7 +44,7 @@ class QwenImageControlNetModel(QwenImageTransformer2DModel):
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
         ids = torch.cat((txt_ids, img_ids), dim=1)
-        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)