feat: Add VideoSlice node with lazy operations on VideoInput

- Add VideoOp base class and SliceOp in _input/video_types.py
- Add sliced() method to VideoInput that returns a copy with operation appended
- Each subclass applies operations in get_components() and get_frame_count()
- After materialization, VideoFromFile delegates to internal VideoFromComponents
- Add VideoSlice node that uses video.sliced(start_frame, frame_count)
- Add tests for SliceOp, sliced() behavior, and materialization

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

@@ -1,3 +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.
+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

@@ -1,3 +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.
+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

@@ -1,3 +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.
+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/workflows/stable-release.yml

@@ -117,7 +117,7 @@ jobs:
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
 
-          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
           ./python.exe -s -m pip install -r requirements_comfyui.txt
           rm requirements_comfyui.txt
 

.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-launch.yml

@@ -13,7 +13,7 @@ jobs:
     - name: Checkout ComfyUI
       uses: actions/checkout@v4
       with:
-        repository: "comfyanonymous/ComfyUI"
+        repository: "Comfy-Org/ComfyUI"
         path: "ComfyUI"
     - uses: actions/setup-python@v4
       with:
@@ -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/update-ci-container.yml

@@ -0,0 +1,59 @@
+name: "CI: Update CI Container"
+
+on:
+  release:
+    types: [published]
+  workflow_dispatch:
+    inputs:
+      version:
+        description: 'ComfyUI version (e.g., v0.7.0)'
+        required: true
+        type: string
+
+jobs:
+  update-ci-container:
+    runs-on: ubuntu-latest
+    # Skip pre-releases unless manually triggered
+    if: github.event_name == 'workflow_dispatch' || !github.event.release.prerelease
+    steps:
+      - name: Get version
+        id: version
+        run: |
+          if [ "${{ github.event_name }}" = "release" ]; then
+            VERSION="${{ github.event.release.tag_name }}"
+          else
+            VERSION="${{ inputs.version }}"
+          fi
+          echo "version=$VERSION" >> $GITHUB_OUTPUT
+
+      - name: Checkout comfyui-ci-container
+        uses: actions/checkout@v4
+        with:
+          repository: comfy-org/comfyui-ci-container
+          token: ${{ secrets.CI_CONTAINER_PAT }}
+
+      - name: Check current version
+        id: current
+        run: |
+          CURRENT=$(grep -oP 'ARG COMFYUI_VERSION=\K.*' Dockerfile || echo "unknown")
+          echo "current_version=$CURRENT" >> $GITHUB_OUTPUT
+
+      - name: Update Dockerfile
+        run: |
+          VERSION="${{ steps.version.outputs.version }}"
+          sed -i "s/^ARG COMFYUI_VERSION=.*/ARG COMFYUI_VERSION=${VERSION}/" Dockerfile
+
+      - name: Create Pull Request
+        id: create-pr
+        uses: peter-evans/create-pull-request@v7
+        with:
+          token: ${{ secrets.CI_CONTAINER_PAT }}
+          branch: automation/comfyui-${{ steps.version.outputs.version }}
+          title: "chore: bump ComfyUI to ${{ steps.version.outputs.version }}"
+          body: |
+            Updates ComfyUI version from `${{ steps.current.outputs.current_version }}` to `${{ steps.version.outputs.version }}`
+
+            **Triggered by:** ${{ github.event_name == 'release' && format('[Release {0}]({1})', github.event.release.tag_name, github.event.release.html_url) || 'Manual workflow dispatch' }}
+
+          labels: automation
+          commit-message: "chore: bump ComfyUI to ${{ steps.version.outputs.version }}"

README.md

@@ -108,7 +108,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
 - Latent previews with [TAESD](#how-to-show-high-quality-previews)
 - Works fully offline: core will never download anything unless you want to.
-- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview).
+- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview) disable with: `--disable-api-nodes`
 - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
 
 Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
@@ -119,6 +119,9 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
    - 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
 
@@ -180,7 +183,7 @@ Simply download, extract with [7-Zip](https://7-zip.org) or with the windows exp
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
-Update your Nvidia drivers if it doesn't start.
+The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start.
 
 #### Alternative Downloads:
 
@@ -209,6 +212,8 @@ Python 3.14 works but you may encounter issues with the torch compile node. The
 
 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 and optimizations might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old.
+
 ### Instructions:
 
 Git clone this repo.
@@ -224,7 +229,7 @@ AMD users can install rocm and pytorch with pip if you don't have it already ins
 
 ```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4```
 
-This is the command to install the nightly with ROCm 7.0 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.1 which might have some performance improvements:
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1```
 
@@ -235,7 +240,7 @@ These have less hardware support than the builds above but they work on windows.
 
 RDNA 3 (RX 7000 series):
 
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/```
+```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/```
 
 RDNA 3.5 (Strix halo/Ryzen AI Max+ 365):
 

alembic_db/versions/0001_assets.py

@@ -0,0 +1,174 @@
+"""
+Initial assets schema
+Revision ID: 0001_assets
+Revises: None
+Create Date: 2025-12-10 00:00:00
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+revision = "0001_assets"
+down_revision = None
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # ASSETS: content identity
+    op.create_table(
+        "assets",
+        sa.Column("id", sa.String(length=36), primary_key=True),
+        sa.Column("hash", sa.String(length=256), nullable=True),
+        sa.Column("size_bytes", sa.BigInteger(), nullable=False, server_default="0"),
+        sa.Column("mime_type", sa.String(length=255), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=False), nullable=False),
+        sa.CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"),
+    )
+    op.create_index("uq_assets_hash", "assets", ["hash"], unique=True)
+    op.create_index("ix_assets_mime_type", "assets", ["mime_type"])
+
+    # ASSETS_INFO: user-visible references
+    op.create_table(
+        "assets_info",
+        sa.Column("id", sa.String(length=36), primary_key=True),
+        sa.Column("owner_id", sa.String(length=128), nullable=False, server_default=""),
+        sa.Column("name", sa.String(length=512), nullable=False),
+        sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False),
+        sa.Column("preview_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="SET NULL"), nullable=True),
+        sa.Column("user_metadata", sa.JSON(), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=False), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=False), nullable=False),
+        sa.Column("last_access_time", sa.DateTime(timezone=False), nullable=False),
+        sa.UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
+    )
+    op.create_index("ix_assets_info_owner_id", "assets_info", ["owner_id"])
+    op.create_index("ix_assets_info_asset_id", "assets_info", ["asset_id"])
+    op.create_index("ix_assets_info_name", "assets_info", ["name"])
+    op.create_index("ix_assets_info_created_at", "assets_info", ["created_at"])
+    op.create_index("ix_assets_info_last_access_time", "assets_info", ["last_access_time"])
+    op.create_index("ix_assets_info_owner_name", "assets_info", ["owner_id", "name"])
+
+    # TAGS: normalized tag vocabulary
+    op.create_table(
+        "tags",
+        sa.Column("name", sa.String(length=512), primary_key=True),
+        sa.Column("tag_type", sa.String(length=32), nullable=False, server_default="user"),
+        sa.CheckConstraint("name = lower(name)", name="ck_tags_lowercase"),
+    )
+    op.create_index("ix_tags_tag_type", "tags", ["tag_type"])
+
+    # ASSET_INFO_TAGS: many-to-many for tags on AssetInfo
+    op.create_table(
+        "asset_info_tags",
+        sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("tag_name", sa.String(length=512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), nullable=False),
+        sa.Column("origin", sa.String(length=32), nullable=False, server_default="manual"),
+        sa.Column("added_at", sa.DateTime(timezone=False), nullable=False),
+        sa.PrimaryKeyConstraint("asset_info_id", "tag_name", name="pk_asset_info_tags"),
+    )
+    op.create_index("ix_asset_info_tags_tag_name", "asset_info_tags", ["tag_name"])
+    op.create_index("ix_asset_info_tags_asset_info_id", "asset_info_tags", ["asset_info_id"])
+
+    # ASSET_CACHE_STATE: N:1 local cache rows per Asset
+    op.create_table(
+        "asset_cache_state",
+        sa.Column("id", sa.Integer(), primary_key=True, autoincrement=True),
+        sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("file_path", sa.Text(), nullable=False),  # absolute local path to cached file
+        sa.Column("mtime_ns", sa.BigInteger(), nullable=True),
+        sa.Column("needs_verify", sa.Boolean(), nullable=False, server_default=sa.text("false")),
+        sa.CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
+        sa.UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
+    )
+    op.create_index("ix_asset_cache_state_file_path", "asset_cache_state", ["file_path"])
+    op.create_index("ix_asset_cache_state_asset_id", "asset_cache_state", ["asset_id"])
+
+    # ASSET_INFO_META: typed KV projection of user_metadata for filtering/sorting
+    op.create_table(
+        "asset_info_meta",
+        sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("key", sa.String(length=256), nullable=False),
+        sa.Column("ordinal", sa.Integer(), nullable=False, server_default="0"),
+        sa.Column("val_str", sa.String(length=2048), nullable=True),
+        sa.Column("val_num", sa.Numeric(38, 10), nullable=True),
+        sa.Column("val_bool", sa.Boolean(), nullable=True),
+        sa.Column("val_json", sa.JSON(), nullable=True),
+        sa.PrimaryKeyConstraint("asset_info_id", "key", "ordinal", name="pk_asset_info_meta"),
+    )
+    op.create_index("ix_asset_info_meta_key", "asset_info_meta", ["key"])
+    op.create_index("ix_asset_info_meta_key_val_str", "asset_info_meta", ["key", "val_str"])
+    op.create_index("ix_asset_info_meta_key_val_num", "asset_info_meta", ["key", "val_num"])
+    op.create_index("ix_asset_info_meta_key_val_bool", "asset_info_meta", ["key", "val_bool"])
+
+    # Tags vocabulary
+    tags_table = sa.table(
+        "tags",
+        sa.column("name", sa.String(length=512)),
+        sa.column("tag_type", sa.String()),
+    )
+    op.bulk_insert(
+        tags_table,
+        [
+            {"name": "models", "tag_type": "system"},
+            {"name": "input", "tag_type": "system"},
+            {"name": "output", "tag_type": "system"},
+
+            {"name": "configs", "tag_type": "system"},
+            {"name": "checkpoints", "tag_type": "system"},
+            {"name": "loras", "tag_type": "system"},
+            {"name": "vae", "tag_type": "system"},
+            {"name": "text_encoders", "tag_type": "system"},
+            {"name": "diffusion_models", "tag_type": "system"},
+            {"name": "clip_vision", "tag_type": "system"},
+            {"name": "style_models", "tag_type": "system"},
+            {"name": "embeddings", "tag_type": "system"},
+            {"name": "diffusers", "tag_type": "system"},
+            {"name": "vae_approx", "tag_type": "system"},
+            {"name": "controlnet", "tag_type": "system"},
+            {"name": "gligen", "tag_type": "system"},
+            {"name": "upscale_models", "tag_type": "system"},
+            {"name": "hypernetworks", "tag_type": "system"},
+            {"name": "photomaker", "tag_type": "system"},
+            {"name": "classifiers", "tag_type": "system"},
+
+            {"name": "encoder", "tag_type": "system"},
+            {"name": "decoder", "tag_type": "system"},
+
+            {"name": "missing", "tag_type": "system"},
+            {"name": "rescan", "tag_type": "system"},
+        ],
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_asset_info_meta_key_val_bool", table_name="asset_info_meta")
+    op.drop_index("ix_asset_info_meta_key_val_num", table_name="asset_info_meta")
+    op.drop_index("ix_asset_info_meta_key_val_str", table_name="asset_info_meta")
+    op.drop_index("ix_asset_info_meta_key", table_name="asset_info_meta")
+    op.drop_table("asset_info_meta")
+
+    op.drop_index("ix_asset_cache_state_asset_id", table_name="asset_cache_state")
+    op.drop_index("ix_asset_cache_state_file_path", table_name="asset_cache_state")
+    op.drop_constraint("uq_asset_cache_state_file_path", table_name="asset_cache_state")
+    op.drop_table("asset_cache_state")
+
+    op.drop_index("ix_asset_info_tags_asset_info_id", table_name="asset_info_tags")
+    op.drop_index("ix_asset_info_tags_tag_name", table_name="asset_info_tags")
+    op.drop_table("asset_info_tags")
+
+    op.drop_index("ix_tags_tag_type", table_name="tags")
+    op.drop_table("tags")
+
+    op.drop_constraint("uq_assets_info_asset_owner_name", table_name="assets_info")
+    op.drop_index("ix_assets_info_owner_name", table_name="assets_info")
+    op.drop_index("ix_assets_info_last_access_time", table_name="assets_info")
+    op.drop_index("ix_assets_info_created_at", table_name="assets_info")
+    op.drop_index("ix_assets_info_name", table_name="assets_info")
+    op.drop_index("ix_assets_info_asset_id", table_name="assets_info")
+    op.drop_index("ix_assets_info_owner_id", table_name="assets_info")
+    op.drop_table("assets_info")
+
+    op.drop_index("uq_assets_hash", table_name="assets")
+    op.drop_index("ix_assets_mime_type", table_name="assets")
+    op.drop_table("assets")

app/assets/api/routes.py

@@ -0,0 +1,102 @@
+import logging
+import uuid
+from aiohttp import web
+
+from pydantic import ValidationError
+
+import app.assets.manager as manager
+from app import user_manager
+from app.assets.api import schemas_in
+from app.assets.helpers import get_query_dict
+
+ROUTES = web.RouteTableDef()
+USER_MANAGER: user_manager.UserManager | None = None
+
+# UUID regex (canonical hyphenated form, case-insensitive)
+UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}"
+
+def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None:
+    global USER_MANAGER
+    USER_MANAGER = user_manager_instance
+    app.add_routes(ROUTES)
+
+def _error_response(status: int, code: str, message: str, details: dict | None = None) -> web.Response:
+    return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status)
+
+
+def _validation_error_response(code: str, ve: ValidationError) -> web.Response:
+    return _error_response(400, code, "Validation failed.", {"errors": ve.json()})
+
+
+@ROUTES.get("/api/assets")
+async def list_assets(request: web.Request) -> web.Response:
+    """
+    GET request to list assets.
+    """
+    query_dict = get_query_dict(request)
+    try:
+        q = schemas_in.ListAssetsQuery.model_validate(query_dict)
+    except ValidationError as ve:
+        return _validation_error_response("INVALID_QUERY", ve)
+
+    payload = manager.list_assets(
+        include_tags=q.include_tags,
+        exclude_tags=q.exclude_tags,
+        name_contains=q.name_contains,
+        metadata_filter=q.metadata_filter,
+        limit=q.limit,
+        offset=q.offset,
+        sort=q.sort,
+        order=q.order,
+        owner_id=USER_MANAGER.get_request_user_id(request),
+    )
+    return web.json_response(payload.model_dump(mode="json"))
+
+
+@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}")
+async def get_asset(request: web.Request) -> web.Response:
+    """
+    GET request to get an asset's info as JSON.
+    """
+    asset_info_id = str(uuid.UUID(request.match_info["id"]))
+    try:
+        result = manager.get_asset(
+            asset_info_id=asset_info_id,
+            owner_id=USER_MANAGER.get_request_user_id(request),
+        )
+    except ValueError as e:
+        return _error_response(404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id})
+    except Exception:
+        logging.exception(
+            "get_asset failed for asset_info_id=%s, owner_id=%s",
+            asset_info_id,
+            USER_MANAGER.get_request_user_id(request),
+        )
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+    return web.json_response(result.model_dump(mode="json"), status=200)
+
+
+@ROUTES.get("/api/tags")
+async def get_tags(request: web.Request) -> web.Response:
+    """
+    GET request to list all tags based on query parameters.
+    """
+    query_map = dict(request.rel_url.query)
+
+    try:
+        query = schemas_in.TagsListQuery.model_validate(query_map)
+    except ValidationError as e:
+        return web.json_response(
+            {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": e.errors()}},
+            status=400,
+        )
+
+    result = manager.list_tags(
+        prefix=query.prefix,
+        limit=query.limit,
+        offset=query.offset,
+        order=query.order,
+        include_zero=query.include_zero,
+        owner_id=USER_MANAGER.get_request_user_id(request),
+    )
+    return web.json_response(result.model_dump(mode="json"))

app/assets/api/schemas_in.py

@@ -0,0 +1,94 @@
+import json
+import uuid
+from typing import Any, Literal
+
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    conint,
+    field_validator,
+)
+
+
+class ListAssetsQuery(BaseModel):
+    include_tags: list[str] = Field(default_factory=list)
+    exclude_tags: list[str] = Field(default_factory=list)
+    name_contains: str | None = None
+
+    # Accept either a JSON string (query param) or a dict
+    metadata_filter: dict[str, Any] | None = None
+
+    limit: conint(ge=1, le=500) = 20
+    offset: conint(ge=0) = 0
+
+    sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at"
+    order: Literal["asc", "desc"] = "desc"
+
+    @field_validator("include_tags", "exclude_tags", mode="before")
+    @classmethod
+    def _split_csv_tags(cls, v):
+        # Accept "a,b,c" or ["a","b"] (we are liberal in what we accept)
+        if v is None:
+            return []
+        if isinstance(v, str):
+            return [t.strip() for t in v.split(",") if t.strip()]
+        if isinstance(v, list):
+            out: list[str] = []
+            for item in v:
+                if isinstance(item, str):
+                    out.extend([t.strip() for t in item.split(",") if t.strip()])
+            return out
+        return v
+
+    @field_validator("metadata_filter", mode="before")
+    @classmethod
+    def _parse_metadata_json(cls, v):
+        if v is None or isinstance(v, dict):
+            return v
+        if isinstance(v, str) and v.strip():
+            try:
+                parsed = json.loads(v)
+            except Exception as e:
+                raise ValueError(f"metadata_filter must be JSON: {e}") from e
+            if not isinstance(parsed, dict):
+                raise ValueError("metadata_filter must be a JSON object")
+            return parsed
+        return None
+
+
+class TagsListQuery(BaseModel):
+    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
+
+    prefix: str | None = Field(None, min_length=1, max_length=256)
+    limit: int = Field(100, ge=1, le=1000)
+    offset: int = Field(0, ge=0, le=10_000_000)
+    order: Literal["count_desc", "name_asc"] = "count_desc"
+    include_zero: bool = True
+
+    @field_validator("prefix")
+    @classmethod
+    def normalize_prefix(cls, v: str | None) -> str | None:
+        if v is None:
+            return v
+        v = v.strip()
+        return v.lower() or None
+
+
+class SetPreviewBody(BaseModel):
+    """Set or clear the preview for an AssetInfo. Provide an Asset.id or null."""
+    preview_id: str | None = None
+
+    @field_validator("preview_id", mode="before")
+    @classmethod
+    def _norm_uuid(cls, v):
+        if v is None:
+            return None
+        s = str(v).strip()
+        if not s:
+            return None
+        try:
+            uuid.UUID(s)
+        except Exception:
+            raise ValueError("preview_id must be a UUID")
+        return s

app/assets/api/schemas_out.py

@@ -0,0 +1,60 @@
+from datetime import datetime
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict, Field, field_serializer
+
+
+class AssetSummary(BaseModel):
+    id: str
+    name: str
+    asset_hash: str | None = None
+    size: int | None = None
+    mime_type: str | None = None
+    tags: list[str] = Field(default_factory=list)
+    preview_url: str | None = None
+    created_at: datetime | None = None
+    updated_at: datetime | None = None
+    last_access_time: datetime | None = None
+
+    model_config = ConfigDict(from_attributes=True)
+
+    @field_serializer("created_at", "updated_at", "last_access_time")
+    def _ser_dt(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+class AssetsList(BaseModel):
+    assets: list[AssetSummary]
+    total: int
+    has_more: bool
+
+
+class AssetDetail(BaseModel):
+    id: str
+    name: str
+    asset_hash: str | None = None
+    size: int | None = None
+    mime_type: str | None = None
+    tags: list[str] = Field(default_factory=list)
+    user_metadata: dict[str, Any] = Field(default_factory=dict)
+    preview_id: str | None = None
+    created_at: datetime | None = None
+    last_access_time: datetime | None = None
+
+    model_config = ConfigDict(from_attributes=True)
+
+    @field_serializer("created_at", "last_access_time")
+    def _ser_dt(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+class TagUsage(BaseModel):
+    name: str
+    count: int
+    type: str
+
+
+class TagsList(BaseModel):
+    tags: list[TagUsage] = Field(default_factory=list)
+    total: int
+    has_more: bool

app/assets/database/bulk_ops.py

@@ -0,0 +1,204 @@
+import os
+import uuid
+import sqlalchemy
+from typing import Iterable
+from sqlalchemy.orm import Session
+from sqlalchemy.dialects import sqlite
+
+from app.assets.helpers import utcnow
+from app.assets.database.models import Asset, AssetCacheState, AssetInfo, AssetInfoTag, AssetInfoMeta
+
+MAX_BIND_PARAMS = 800
+
+def _chunk_rows(rows: list[dict], cols_per_row: int, max_bind_params: int) -> Iterable[list[dict]]:
+    if not rows:
+        return []
+    rows_per_stmt = max(1, max_bind_params // max(1, cols_per_row))
+    for i in range(0, len(rows), rows_per_stmt):
+        yield rows[i:i + rows_per_stmt]
+
+def _iter_chunks(seq, n: int):
+    for i in range(0, len(seq), n):
+        yield seq[i:i + n]
+
+def _rows_per_stmt(cols: int) -> int:
+    return max(1, MAX_BIND_PARAMS // max(1, cols))
+
+
+def seed_from_paths_batch(
+    session: Session,
+    *,
+    specs: list[dict],
+    owner_id: str = "",
+) -> dict:
+    """Each spec is a dict with keys:
+      - abs_path: str
+      - size_bytes: int
+      - mtime_ns: int
+      - info_name: str
+      - tags: list[str]
+      - fname: Optional[str]
+    """
+    if not specs:
+        return {"inserted_infos": 0, "won_states": 0, "lost_states": 0}
+
+    now = utcnow()
+    asset_rows: list[dict] = []
+    state_rows: list[dict] = []
+    path_to_asset: dict[str, str] = {}
+    asset_to_info: dict[str, dict] = {}  # asset_id -> prepared info row
+    path_list: list[str] = []
+
+    for sp in specs:
+        ap = os.path.abspath(sp["abs_path"])
+        aid = str(uuid.uuid4())
+        iid = str(uuid.uuid4())
+        path_list.append(ap)
+        path_to_asset[ap] = aid
+
+        asset_rows.append(
+            {
+                "id": aid,
+                "hash": None,
+                "size_bytes": sp["size_bytes"],
+                "mime_type": None,
+                "created_at": now,
+            }
+        )
+        state_rows.append(
+            {
+                "asset_id": aid,
+                "file_path": ap,
+                "mtime_ns": sp["mtime_ns"],
+            }
+        )
+        asset_to_info[aid] = {
+            "id": iid,
+            "owner_id": owner_id,
+            "name": sp["info_name"],
+            "asset_id": aid,
+            "preview_id": None,
+            "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None,
+            "created_at": now,
+            "updated_at": now,
+            "last_access_time": now,
+            "_tags": sp["tags"],
+            "_filename": sp["fname"],
+        }
+
+    # insert all seed Assets (hash=NULL)
+    ins_asset = sqlite.insert(Asset)
+    for chunk in _iter_chunks(asset_rows, _rows_per_stmt(5)):
+        session.execute(ins_asset, chunk)
+
+    # try to claim AssetCacheState (file_path)
+    # Insert with ON CONFLICT DO NOTHING, then query to find which paths were actually inserted
+    ins_state = (
+        sqlite.insert(AssetCacheState)
+        .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
+    )
+    for chunk in _iter_chunks(state_rows, _rows_per_stmt(3)):
+        session.execute(ins_state, chunk)
+
+    # Query to find which of our paths won (were actually inserted)
+    winners_by_path: set[str] = set()
+    for chunk in _iter_chunks(path_list, MAX_BIND_PARAMS):
+        result = session.execute(
+            sqlalchemy.select(AssetCacheState.file_path)
+            .where(AssetCacheState.file_path.in_(chunk))
+            .where(AssetCacheState.asset_id.in_([path_to_asset[p] for p in chunk]))
+        )
+        winners_by_path.update(result.scalars().all())
+
+    all_paths_set = set(path_list)
+    losers_by_path = all_paths_set - winners_by_path
+    lost_assets = [path_to_asset[p] for p in losers_by_path]
+    if lost_assets:  # losers get their Asset removed
+        for id_chunk in _iter_chunks(lost_assets, MAX_BIND_PARAMS):
+            session.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(id_chunk)))
+
+    if not winners_by_path:
+        return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)}
+
+    # insert AssetInfo only for winners
+    # Insert with ON CONFLICT DO NOTHING, then query to find which were actually inserted
+    winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path]
+    ins_info = (
+        sqlite.insert(AssetInfo)
+        .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name])
+    )
+    for chunk in _iter_chunks(winner_info_rows, _rows_per_stmt(9)):
+        session.execute(ins_info, chunk)
+
+    # Query to find which info rows were actually inserted (by matching our generated IDs)
+    all_info_ids = [row["id"] for row in winner_info_rows]
+    inserted_info_ids: set[str] = set()
+    for chunk in _iter_chunks(all_info_ids, MAX_BIND_PARAMS):
+        result = session.execute(
+            sqlalchemy.select(AssetInfo.id).where(AssetInfo.id.in_(chunk))
+        )
+        inserted_info_ids.update(result.scalars().all())
+
+    # build and insert tag + meta rows for the AssetInfo
+    tag_rows: list[dict] = []
+    meta_rows: list[dict] = []
+    if inserted_info_ids:
+        for row in winner_info_rows:
+            iid = row["id"]
+            if iid not in inserted_info_ids:
+                continue
+            for t in row["_tags"]:
+                tag_rows.append({
+                    "asset_info_id": iid,
+                    "tag_name": t,
+                    "origin": "automatic",
+                    "added_at": now,
+                })
+            if row["_filename"]:
+                meta_rows.append(
+                    {
+                        "asset_info_id": iid,
+                        "key": "filename",
+                        "ordinal": 0,
+                        "val_str": row["_filename"],
+                        "val_num": None,
+                        "val_bool": None,
+                        "val_json": None,
+                    }
+                )
+
+    bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows, max_bind_params=MAX_BIND_PARAMS)
+    return {
+        "inserted_infos": len(inserted_info_ids),
+        "won_states": len(winners_by_path),
+        "lost_states": len(losers_by_path),
+    }
+
+
+def bulk_insert_tags_and_meta(
+    session: Session,
+    *,
+    tag_rows: list[dict],
+    meta_rows: list[dict],
+    max_bind_params: int,
+) -> None:
+    """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING.
+    - tag_rows keys: asset_info_id, tag_name, origin, added_at
+    - meta_rows keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json
+    """
+    if tag_rows:
+        ins_links = (
+            sqlite.insert(AssetInfoTag)
+            .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
+        )
+        for chunk in _chunk_rows(tag_rows, cols_per_row=4, max_bind_params=max_bind_params):
+            session.execute(ins_links, chunk)
+    if meta_rows:
+        ins_meta = (
+            sqlite.insert(AssetInfoMeta)
+            .on_conflict_do_nothing(
+                index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal]
+            )
+        )
+        for chunk in _chunk_rows(meta_rows, cols_per_row=7, max_bind_params=max_bind_params):
+            session.execute(ins_meta, chunk)

app/assets/database/models.py

@@ -0,0 +1,233 @@
+from __future__ import annotations
+
+import uuid
+from datetime import datetime
+
+from typing import Any
+from sqlalchemy import (
+    JSON,
+    BigInteger,
+    Boolean,
+    CheckConstraint,
+    DateTime,
+    ForeignKey,
+    Index,
+    Integer,
+    Numeric,
+    String,
+    Text,
+    UniqueConstraint,
+)
+from sqlalchemy.orm import Mapped, foreign, mapped_column, relationship
+
+from app.assets.helpers import utcnow
+from app.database.models import to_dict, Base
+
+
+class Asset(Base):
+    __tablename__ = "assets"
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    hash: Mapped[str | None] = mapped_column(String(256), nullable=True)
+    size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0)
+    mime_type: Mapped[str | None] = mapped_column(String(255))
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=utcnow
+    )
+
+    infos: Mapped[list[AssetInfo]] = relationship(
+        "AssetInfo",
+        back_populates="asset",
+        primaryjoin=lambda: Asset.id == foreign(AssetInfo.asset_id),
+        foreign_keys=lambda: [AssetInfo.asset_id],
+        cascade="all,delete-orphan",
+        passive_deletes=True,
+    )
+
+    preview_of: Mapped[list[AssetInfo]] = relationship(
+        "AssetInfo",
+        back_populates="preview_asset",
+        primaryjoin=lambda: Asset.id == foreign(AssetInfo.preview_id),
+        foreign_keys=lambda: [AssetInfo.preview_id],
+        viewonly=True,
+    )
+
+    cache_states: Mapped[list[AssetCacheState]] = relationship(
+        back_populates="asset",
+        cascade="all, delete-orphan",
+        passive_deletes=True,
+    )
+
+    __table_args__ = (
+        Index("uq_assets_hash", "hash", unique=True),
+        Index("ix_assets_mime_type", "mime_type"),
+        CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"),
+    )
+
+    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
+        return to_dict(self, include_none=include_none)
+
+    def __repr__(self) -> str:
+        return f"<Asset id={self.id} hash={(self.hash or '')[:12]}>"
+
+
+class AssetCacheState(Base):
+    __tablename__ = "asset_cache_state"
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
+    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False)
+    file_path: Mapped[str] = mapped_column(Text, nullable=False)
+    mtime_ns: Mapped[int | None] = mapped_column(BigInteger, nullable=True)
+    needs_verify: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+
+    asset: Mapped[Asset] = relationship(back_populates="cache_states")
+
+    __table_args__ = (
+        Index("ix_asset_cache_state_file_path", "file_path"),
+        Index("ix_asset_cache_state_asset_id", "asset_id"),
+        CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
+        UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
+    )
+
+    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
+        return to_dict(self, include_none=include_none)
+
+    def __repr__(self) -> str:
+        return f"<AssetCacheState id={self.id} asset_id={self.asset_id} path={self.file_path!r}>"
+
+
+class AssetInfo(Base):
+    __tablename__ = "assets_info"
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="")
+    name: Mapped[str] = mapped_column(String(512), nullable=False)
+    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False)
+    preview_id: Mapped[str | None] = mapped_column(String(36), ForeignKey("assets.id", ondelete="SET NULL"))
+    user_metadata: Mapped[dict[str, Any] | None] = mapped_column(JSON(none_as_null=True))
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+    updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+    last_access_time: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+
+    asset: Mapped[Asset] = relationship(
+        "Asset",
+        back_populates="infos",
+        foreign_keys=[asset_id],
+        lazy="selectin",
+    )
+    preview_asset: Mapped[Asset | None] = relationship(
+        "Asset",
+        back_populates="preview_of",
+        foreign_keys=[preview_id],
+    )
+
+    metadata_entries: Mapped[list[AssetInfoMeta]] = relationship(
+        back_populates="asset_info",
+        cascade="all,delete-orphan",
+        passive_deletes=True,
+    )
+
+    tag_links: Mapped[list[AssetInfoTag]] = relationship(
+        back_populates="asset_info",
+        cascade="all,delete-orphan",
+        passive_deletes=True,
+        overlaps="tags,asset_infos",
+    )
+
+    tags: Mapped[list[Tag]] = relationship(
+        secondary="asset_info_tags",
+        back_populates="asset_infos",
+        lazy="selectin",
+        viewonly=True,
+        overlaps="tag_links,asset_info_links,asset_infos,tag",
+    )
+
+    __table_args__ = (
+        UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
+        Index("ix_assets_info_owner_name", "owner_id", "name"),
+        Index("ix_assets_info_owner_id", "owner_id"),
+        Index("ix_assets_info_asset_id", "asset_id"),
+        Index("ix_assets_info_name", "name"),
+        Index("ix_assets_info_created_at", "created_at"),
+        Index("ix_assets_info_last_access_time", "last_access_time"),
+    )
+
+    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
+        data = to_dict(self, include_none=include_none)
+        data["tags"] = [t.name for t in self.tags]
+        return data
+
+    def __repr__(self) -> str:
+        return f"<AssetInfo id={self.id} name={self.name!r} asset_id={self.asset_id}>"
+
+
+class AssetInfoMeta(Base):
+    __tablename__ = "asset_info_meta"
+
+    asset_info_id: Mapped[str] = mapped_column(
+        String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True
+    )
+    key: Mapped[str] = mapped_column(String(256), primary_key=True)
+    ordinal: Mapped[int] = mapped_column(Integer, primary_key=True, default=0)
+
+    val_str: Mapped[str | None] = mapped_column(String(2048), nullable=True)
+    val_num: Mapped[float | None] = mapped_column(Numeric(38, 10), nullable=True)
+    val_bool: Mapped[bool | None] = mapped_column(Boolean, nullable=True)
+    val_json: Mapped[Any | None] = mapped_column(JSON(none_as_null=True), nullable=True)
+
+    asset_info: Mapped[AssetInfo] = relationship(back_populates="metadata_entries")
+
+    __table_args__ = (
+        Index("ix_asset_info_meta_key", "key"),
+        Index("ix_asset_info_meta_key_val_str", "key", "val_str"),
+        Index("ix_asset_info_meta_key_val_num", "key", "val_num"),
+        Index("ix_asset_info_meta_key_val_bool", "key", "val_bool"),
+    )
+
+
+class AssetInfoTag(Base):
+    __tablename__ = "asset_info_tags"
+
+    asset_info_id: Mapped[str] = mapped_column(
+        String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True
+    )
+    tag_name: Mapped[str] = mapped_column(
+        String(512), ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True
+    )
+    origin: Mapped[str] = mapped_column(String(32), nullable=False, default="manual")
+    added_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=utcnow
+    )
+
+    asset_info: Mapped[AssetInfo] = relationship(back_populates="tag_links")
+    tag: Mapped[Tag] = relationship(back_populates="asset_info_links")
+
+    __table_args__ = (
+        Index("ix_asset_info_tags_tag_name", "tag_name"),
+        Index("ix_asset_info_tags_asset_info_id", "asset_info_id"),
+    )
+
+
+class Tag(Base):
+    __tablename__ = "tags"
+
+    name: Mapped[str] = mapped_column(String(512), primary_key=True)
+    tag_type: Mapped[str] = mapped_column(String(32), nullable=False, default="user")
+
+    asset_info_links: Mapped[list[AssetInfoTag]] = relationship(
+        back_populates="tag",
+        overlaps="asset_infos,tags",
+    )
+    asset_infos: Mapped[list[AssetInfo]] = relationship(
+        secondary="asset_info_tags",
+        back_populates="tags",
+        viewonly=True,
+        overlaps="asset_info_links,tag_links,tags,asset_info",
+    )
+
+    __table_args__ = (
+        Index("ix_tags_tag_type", "tag_type"),
+    )
+
+    def __repr__(self) -> str:
+        return f"<Tag {self.name}>"

app/assets/database/queries.py

@@ -0,0 +1,267 @@
+import sqlalchemy as sa
+from collections import defaultdict
+from sqlalchemy import select, exists, func
+from sqlalchemy.orm import Session, contains_eager, noload
+from app.assets.database.models import Asset, AssetInfo, AssetInfoMeta, AssetInfoTag, Tag
+from app.assets.helpers import escape_like_prefix, normalize_tags
+from typing import Sequence
+
+
+def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement:
+    """Build owner visibility predicate for reads. Owner-less rows are visible to everyone."""
+    owner_id = (owner_id or "").strip()
+    if owner_id == "":
+        return AssetInfo.owner_id == ""
+    return AssetInfo.owner_id.in_(["", owner_id])
+
+
+def apply_tag_filters(
+    stmt: sa.sql.Select,
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+) -> sa.sql.Select:
+    """include_tags: every tag must be present; exclude_tags: none may be present."""
+    include_tags = normalize_tags(include_tags)
+    exclude_tags = normalize_tags(exclude_tags)
+
+    if include_tags:
+        for tag_name in include_tags:
+            stmt = stmt.where(
+                exists().where(
+                    (AssetInfoTag.asset_info_id == AssetInfo.id)
+                    & (AssetInfoTag.tag_name == tag_name)
+                )
+            )
+
+    if exclude_tags:
+        stmt = stmt.where(
+            ~exists().where(
+                (AssetInfoTag.asset_info_id == AssetInfo.id)
+                & (AssetInfoTag.tag_name.in_(exclude_tags))
+            )
+        )
+    return stmt
+
+def apply_metadata_filter(
+    stmt: sa.sql.Select,
+    metadata_filter: dict | None = None,
+) -> sa.sql.Select:
+    """Apply filters using asset_info_meta projection table."""
+    if not metadata_filter:
+        return stmt
+
+    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
+        return sa.exists().where(
+            AssetInfoMeta.asset_info_id == AssetInfo.id,
+            AssetInfoMeta.key == key,
+            *preds,
+        )
+
+    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
+        if value is None:
+            no_row_for_key = sa.not_(
+                sa.exists().where(
+                    AssetInfoMeta.asset_info_id == AssetInfo.id,
+                    AssetInfoMeta.key == key,
+                )
+            )
+            null_row = _exists_for_pred(
+                key,
+                AssetInfoMeta.val_json.is_(None),
+                AssetInfoMeta.val_str.is_(None),
+                AssetInfoMeta.val_num.is_(None),
+                AssetInfoMeta.val_bool.is_(None),
+            )
+            return sa.or_(no_row_for_key, null_row)
+
+        if isinstance(value, bool):
+            return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value))
+        if isinstance(value, (int, float)):
+            from decimal import Decimal
+            num = value if isinstance(value, Decimal) else Decimal(str(value))
+            return _exists_for_pred(key, AssetInfoMeta.val_num == num)
+        if isinstance(value, str):
+            return _exists_for_pred(key, AssetInfoMeta.val_str == value)
+        return _exists_for_pred(key, AssetInfoMeta.val_json == value)
+
+    for k, v in metadata_filter.items():
+        if isinstance(v, list):
+            ors = [_exists_clause_for_value(k, elem) for elem in v]
+            if ors:
+                stmt = stmt.where(sa.or_(*ors))
+        else:
+            stmt = stmt.where(_exists_clause_for_value(k, v))
+    return stmt
+
+
+def asset_exists_by_hash(session: Session, asset_hash: str) -> bool:
+    """
+    Check if an asset with a given hash exists in database.
+    """
+    row = (
+        session.execute(
+            select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1)
+        )
+    ).first()
+    return row is not None
+
+def get_asset_info_by_id(session: Session, asset_info_id: str) -> AssetInfo | None:
+    return session.get(AssetInfo, asset_info_id)
+
+def list_asset_infos_page(
+    session: Session,
+    owner_id: str = "",
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+    name_contains: str | None = None,
+    metadata_filter: dict | None = None,
+    limit: int = 20,
+    offset: int = 0,
+    sort: str = "created_at",
+    order: str = "desc",
+) -> tuple[list[AssetInfo], dict[str, list[str]], int]:
+    base = (
+        select(AssetInfo)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags))
+        .where(visible_owner_clause(owner_id))
+    )
+
+    if name_contains:
+        escaped, esc = escape_like_prefix(name_contains)
+        base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
+
+    base = apply_tag_filters(base, include_tags, exclude_tags)
+    base = apply_metadata_filter(base, metadata_filter)
+
+    sort = (sort or "created_at").lower()
+    order = (order or "desc").lower()
+    sort_map = {
+        "name": AssetInfo.name,
+        "created_at": AssetInfo.created_at,
+        "updated_at": AssetInfo.updated_at,
+        "last_access_time": AssetInfo.last_access_time,
+        "size": Asset.size_bytes,
+    }
+    sort_col = sort_map.get(sort, AssetInfo.created_at)
+    sort_exp = sort_col.desc() if order == "desc" else sort_col.asc()
+
+    base = base.order_by(sort_exp).limit(limit).offset(offset)
+
+    count_stmt = (
+        select(sa.func.count())
+        .select_from(AssetInfo)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .where(visible_owner_clause(owner_id))
+    )
+    if name_contains:
+        escaped, esc = escape_like_prefix(name_contains)
+        count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
+    count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags)
+    count_stmt = apply_metadata_filter(count_stmt, metadata_filter)
+
+    total = int((session.execute(count_stmt)).scalar_one() or 0)
+
+    infos = (session.execute(base)).unique().scalars().all()
+
+    id_list: list[str] = [i.id for i in infos]
+    tag_map: dict[str, list[str]] = defaultdict(list)
+    if id_list:
+        rows = session.execute(
+            select(AssetInfoTag.asset_info_id, Tag.name)
+            .join(Tag, Tag.name == AssetInfoTag.tag_name)
+            .where(AssetInfoTag.asset_info_id.in_(id_list))
+        )
+        for aid, tag_name in rows.all():
+            tag_map[aid].append(tag_name)
+
+    return infos, tag_map, total
+
+def fetch_asset_info_asset_and_tags(
+    session: Session,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> tuple[AssetInfo, Asset, list[str]] | None:
+    stmt = (
+        select(AssetInfo, Asset, Tag.name)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True)
+        .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True)
+        .where(
+            AssetInfo.id == asset_info_id,
+            visible_owner_clause(owner_id),
+        )
+        .options(noload(AssetInfo.tags))
+        .order_by(Tag.name.asc())
+    )
+
+    rows = (session.execute(stmt)).all()
+    if not rows:
+        return None
+
+    first_info, first_asset, _ = rows[0]
+    tags: list[str] = []
+    seen: set[str] = set()
+    for _info, _asset, tag_name in rows:
+        if tag_name and tag_name not in seen:
+            seen.add(tag_name)
+            tags.append(tag_name)
+    return first_info, first_asset, tags
+
+def list_tags_with_usage(
+    session: Session,
+    prefix: str | None = None,
+    limit: int = 100,
+    offset: int = 0,
+    include_zero: bool = True,
+    order: str = "count_desc",
+    owner_id: str = "",
+) -> tuple[list[tuple[str, str, int]], int]:
+    counts_sq = (
+        select(
+            AssetInfoTag.tag_name.label("tag_name"),
+            func.count(AssetInfoTag.asset_info_id).label("cnt"),
+        )
+        .select_from(AssetInfoTag)
+        .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id)
+        .where(visible_owner_clause(owner_id))
+        .group_by(AssetInfoTag.tag_name)
+        .subquery()
+    )
+
+    q = (
+        select(
+            Tag.name,
+            Tag.tag_type,
+            func.coalesce(counts_sq.c.cnt, 0).label("count"),
+        )
+        .select_from(Tag)
+        .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True)
+    )
+
+    if prefix:
+        escaped, esc = escape_like_prefix(prefix.strip().lower())
+        q = q.where(Tag.name.like(escaped + "%", escape=esc))
+
+    if not include_zero:
+        q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0)
+
+    if order == "name_asc":
+        q = q.order_by(Tag.name.asc())
+    else:
+        q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc())
+
+    total_q = select(func.count()).select_from(Tag)
+    if prefix:
+        escaped, esc = escape_like_prefix(prefix.strip().lower())
+        total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc))
+    if not include_zero:
+        total_q = total_q.where(
+            Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name))
+        )
+
+    rows = (session.execute(q.limit(limit).offset(offset))).all()
+    total = (session.execute(total_q)).scalar_one()
+
+    rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows]
+    return rows_norm, int(total or 0)

app/assets/database/tags.py

@@ -0,0 +1,62 @@
+from typing import Iterable
+
+import sqlalchemy
+from sqlalchemy.orm import Session
+from sqlalchemy.dialects import sqlite
+
+from app.assets.helpers import normalize_tags, utcnow
+from app.assets.database.models import Tag, AssetInfoTag, AssetInfo
+
+
+def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None:
+    wanted = normalize_tags(list(names))
+    if not wanted:
+        return
+    rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))]
+    ins = (
+            sqlite.insert(Tag)
+            .values(rows)
+            .on_conflict_do_nothing(index_elements=[Tag.name])
+        )
+    return session.execute(ins)
+
+def add_missing_tag_for_asset_id(
+    session: Session,
+    *,
+    asset_id: str,
+    origin: str = "automatic",
+) -> None:
+    select_rows = (
+        sqlalchemy.select(
+            AssetInfo.id.label("asset_info_id"),
+            sqlalchemy.literal("missing").label("tag_name"),
+            sqlalchemy.literal(origin).label("origin"),
+            sqlalchemy.literal(utcnow()).label("added_at"),
+        )
+        .where(AssetInfo.asset_id == asset_id)
+        .where(
+            sqlalchemy.not_(
+                sqlalchemy.exists().where((AssetInfoTag.asset_info_id == AssetInfo.id) & (AssetInfoTag.tag_name == "missing"))
+            )
+        )
+    )
+    session.execute(
+        sqlite.insert(AssetInfoTag)
+        .from_select(
+            ["asset_info_id", "tag_name", "origin", "added_at"],
+            select_rows,
+        )
+        .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
+    )
+
+def remove_missing_tag_for_asset_id(
+    session: Session,
+    *,
+    asset_id: str,
+) -> None:
+    session.execute(
+        sqlalchemy.delete(AssetInfoTag).where(
+            AssetInfoTag.asset_info_id.in_(sqlalchemy.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)),
+            AssetInfoTag.tag_name == "missing",
+        )
+    )

app/assets/hashing.py

@@ -0,0 +1,75 @@
+from blake3 import blake3
+from typing import IO
+import os
+import asyncio
+
+
+DEFAULT_CHUNK = 8 * 1024 *1024 # 8MB
+
+# NOTE: this allows hashing different representations of a file-like object
+def blake3_hash(
+    fp: str | IO[bytes],
+    chunk_size: int = DEFAULT_CHUNK,
+) -> str:
+    """
+    Returns a BLAKE3 hex digest for ``fp``, which may be:
+      - a filename (str/bytes) or PathLike
+      - an open binary file object
+    If ``fp`` is a file object, it must be opened in **binary** mode and support
+    ``read``, ``seek``, and ``tell``. The function will seek to the start before
+    reading and will attempt to restore the original position afterward.
+    """
+    # duck typing to check if input is a file-like object
+    if hasattr(fp, "read"):
+        return _hash_file_obj(fp, chunk_size)
+
+    with open(os.fspath(fp), "rb") as f:
+        return _hash_file_obj(f, chunk_size)
+
+
+async def blake3_hash_async(
+    fp: str | IO[bytes],
+    chunk_size: int = DEFAULT_CHUNK,
+) -> str:
+    """Async wrapper for ``blake3_hash_sync``.
+    Uses a worker thread so the event loop remains responsive.
+    """
+    # If it is a path, open inside the worker thread to keep I/O off the loop.
+    if hasattr(fp, "read"):
+        return await asyncio.to_thread(blake3_hash, fp, chunk_size)
+
+    def _worker() -> str:
+        with open(os.fspath(fp), "rb") as f:
+            return _hash_file_obj(f, chunk_size)
+
+    return await asyncio.to_thread(_worker)
+
+
+def _hash_file_obj(file_obj: IO, chunk_size: int = DEFAULT_CHUNK) -> str:
+    """
+    Hash an already-open binary file object by streaming in chunks.
+    - Seeks to the beginning before reading (if supported).
+    - Restores the original position afterward (if tell/seek are supported).
+    """
+    if chunk_size <= 0:
+        chunk_size = DEFAULT_CHUNK
+
+    # in case file object is already open and not at the beginning, track so can be restored after hashing
+    orig_pos = file_obj.tell()
+
+    try:
+        # seek to the beginning before reading
+        if orig_pos != 0:
+            file_obj.seek(0)
+
+        h = blake3()
+        while True:
+            chunk = file_obj.read(chunk_size)
+            if not chunk:
+                break
+            h.update(chunk)
+        return h.hexdigest()
+    finally:
+        # restore original position in file object, if needed
+        if orig_pos != 0:
+            file_obj.seek(orig_pos)

app/assets/helpers.py

@@ -0,0 +1,217 @@
+import contextlib
+import os
+from aiohttp import web
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Literal, Any
+
+import folder_paths
+
+
+RootType = Literal["models", "input", "output"]
+ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output")
+
+def get_query_dict(request: web.Request) -> dict[str, Any]:
+    """
+    Gets a dictionary of query parameters from the request.
+
+    'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic.
+    """
+    query_dict = {
+        key: request.query.getall(key) if len(request.query.getall(key)) > 1 else request.query.get(key)
+        for key in request.query.keys()
+    }
+    return query_dict
+
+def list_tree(base_dir: str) -> list[str]:
+    out: list[str] = []
+    base_abs = os.path.abspath(base_dir)
+    if not os.path.isdir(base_abs):
+        return out
+    for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False):
+        for name in filenames:
+            out.append(os.path.abspath(os.path.join(dirpath, name)))
+    return out
+
+def prefixes_for_root(root: RootType) -> list[str]:
+    if root == "models":
+        bases: list[str] = []
+        for _bucket, paths in get_comfy_models_folders():
+            bases.extend(paths)
+        return [os.path.abspath(p) for p in bases]
+    if root == "input":
+        return [os.path.abspath(folder_paths.get_input_directory())]
+    if root == "output":
+        return [os.path.abspath(folder_paths.get_output_directory())]
+    return []
+
+def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]:
+    """Escapes %, _ and the escape char itself in a LIKE prefix.
+    Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like().
+    """
+    s = s.replace(escape, escape + escape)  # escape the escape char first
+    s = s.replace("%", escape + "%").replace("_", escape + "_")  # escape LIKE wildcards
+    return s, escape
+
+def fast_asset_file_check(
+    *,
+    mtime_db: int | None,
+    size_db: int | None,
+    stat_result: os.stat_result,
+) -> bool:
+    if mtime_db is None:
+        return False
+    actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000))
+    if int(mtime_db) != int(actual_mtime_ns):
+        return False
+    sz = int(size_db or 0)
+    if sz > 0:
+        return int(stat_result.st_size) == sz
+    return True
+
+def utcnow() -> datetime:
+    """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC."""
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+def get_comfy_models_folders() -> list[tuple[str, list[str]]]:
+    """Build a list of (folder_name, base_paths[]) categories that are configured for model locations.
+
+    We trust `folder_paths.folder_names_and_paths` and include a category if
+    *any* of its base paths lies under the Comfy `models_dir`.
+    """
+    targets: list[tuple[str, list[str]]] = []
+    models_root = os.path.abspath(folder_paths.models_dir)
+    for name, values in folder_paths.folder_names_and_paths.items():
+        paths, _exts = values[0], values[1]  # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI
+        if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths):
+            targets.append((name, paths))
+    return targets
+
+def compute_relative_filename(file_path: str) -> str | None:
+    """
+    Return the model's path relative to the last well-known folder (the model category),
+    using forward slashes, eg:
+      /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors"
+      /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors"
+
+    For non-model paths, returns None.
+    NOTE: this is a temporary helper, used only for initializing metadata["filename"] field.
+    """
+    try:
+        root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path)
+    except ValueError:
+        return None
+
+    p = Path(rel_path)
+    parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)]
+    if not parts:
+        return None
+
+    if root_category == "models":
+        # parts[0] is the category ("checkpoints", "vae", etc) – drop it
+        inside = parts[1:] if len(parts) > 1 else [parts[0]]
+        return "/".join(inside)
+    return "/".join(parts)  # input/output: keep all parts
+
+
+def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]:
+    """Given an absolute or relative file path, determine which root category the path belongs to:
+      - 'input' if the file resides under `folder_paths.get_input_directory()`
+      - 'output' if the file resides under `folder_paths.get_output_directory()`
+      - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()`
+
+    Returns:
+        (root_category, relative_path_inside_that_root)
+        For 'models', the relative path is prefixed with the category name:
+            e.g. ('models', 'vae/test/sub/ae.safetensors')
+
+    Raises:
+        ValueError: if the path does not belong to input, output, or configured model bases.
+    """
+    fp_abs = os.path.abspath(file_path)
+
+    def _is_within(child: str, parent: str) -> bool:
+        try:
+            return os.path.commonpath([child, parent]) == parent
+        except Exception:
+            return False
+
+    def _rel(child: str, parent: str) -> str:
+        return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep)
+
+    # 1) input
+    input_base = os.path.abspath(folder_paths.get_input_directory())
+    if _is_within(fp_abs, input_base):
+        return "input", _rel(fp_abs, input_base)
+
+    # 2) output
+    output_base = os.path.abspath(folder_paths.get_output_directory())
+    if _is_within(fp_abs, output_base):
+        return "output", _rel(fp_abs, output_base)
+
+    # 3) models (check deepest matching base to avoid ambiguity)
+    best: tuple[int, str, str] | None = None  # (base_len, bucket, rel_inside_bucket)
+    for bucket, bases in get_comfy_models_folders():
+        for b in bases:
+            base_abs = os.path.abspath(b)
+            if not _is_within(fp_abs, base_abs):
+                continue
+            cand = (len(base_abs), bucket, _rel(fp_abs, base_abs))
+            if best is None or cand[0] > best[0]:
+                best = cand
+
+    if best is not None:
+        _, bucket, rel_inside = best
+        combined = os.path.join(bucket, rel_inside)
+        return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep)
+
+    raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}")
+
+def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:
+    """Return a tuple (name, tags) derived from a filesystem path.
+
+    Semantics:
+      - Root category is determined by `get_relative_to_root_category_path_of_asset`.
+      - The returned `name` is the base filename with extension from the relative path.
+      - The returned `tags` are:
+            [root_category] + parent folders of the relative path (in order)
+        For 'models', this means:
+            file '/.../ModelsDir/vae/test_tag/ae.safetensors'
+            -> root_category='models', some_path='vae/test_tag/ae.safetensors'
+            -> name='ae.safetensors', tags=['models', 'vae', 'test_tag']
+
+    Raises:
+        ValueError: if the path does not belong to input, output, or configured model bases.
+    """
+    root_category, some_path = get_relative_to_root_category_path_of_asset(file_path)
+    p = Path(some_path)
+    parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)]
+    return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts])))
+
+def normalize_tags(tags: list[str] | None) -> list[str]:
+    """
+    Normalize a list of tags by:
+      - Stripping whitespace and converting to lowercase.
+      - Removing duplicates.
+    """
+    return [t.strip().lower() for t in (tags or []) if (t or "").strip()]
+
+def collect_models_files() -> list[str]:
+    out: list[str] = []
+    for folder_name, bases in get_comfy_models_folders():
+        rel_files = folder_paths.get_filename_list(folder_name) or []
+        for rel_path in rel_files:
+            abs_path = folder_paths.get_full_path(folder_name, rel_path)
+            if not abs_path:
+                continue
+            abs_path = os.path.abspath(abs_path)
+            allowed = False
+            for b in bases:
+                base_abs = os.path.abspath(b)
+                with contextlib.suppress(Exception):
+                    if os.path.commonpath([abs_path, base_abs]) == base_abs:
+                        allowed = True
+                        break
+            if allowed:
+                out.append(abs_path)
+    return out

app/assets/manager.py

@@ -0,0 +1,123 @@
+from typing import Sequence
+
+from app.database.db import create_session
+from app.assets.api import schemas_out
+from app.assets.database.queries import (
+    asset_exists_by_hash,
+    fetch_asset_info_asset_and_tags,
+    list_asset_infos_page,
+    list_tags_with_usage,
+)
+
+
+def _safe_sort_field(requested: str | None) -> str:
+    if not requested:
+        return "created_at"
+    v = requested.lower()
+    if v in {"name", "created_at", "updated_at", "size", "last_access_time"}:
+        return v
+    return "created_at"
+
+
+def asset_exists(asset_hash: str) -> bool:
+    with create_session() as session:
+        return asset_exists_by_hash(session, asset_hash=asset_hash)
+
+def list_assets(
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+    name_contains: str | None = None,
+    metadata_filter: dict | None = None,
+    limit: int = 20,
+    offset: int = 0,
+    sort: str = "created_at",
+    order: str = "desc",
+    owner_id: str = "",
+) -> schemas_out.AssetsList:
+    sort = _safe_sort_field(sort)
+    order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower()
+
+    with create_session() as session:
+        infos, tag_map, total = list_asset_infos_page(
+            session,
+            owner_id=owner_id,
+            include_tags=include_tags,
+            exclude_tags=exclude_tags,
+            name_contains=name_contains,
+            metadata_filter=metadata_filter,
+            limit=limit,
+            offset=offset,
+            sort=sort,
+            order=order,
+        )
+
+    summaries: list[schemas_out.AssetSummary] = []
+    for info in infos:
+        asset = info.asset
+        tags = tag_map.get(info.id, [])
+        summaries.append(
+            schemas_out.AssetSummary(
+                id=info.id,
+                name=info.name,
+                asset_hash=asset.hash if asset else None,
+                size=int(asset.size_bytes) if asset else None,
+                mime_type=asset.mime_type if asset else None,
+                tags=tags,
+                preview_url=f"/api/assets/{info.id}/content",
+                created_at=info.created_at,
+                updated_at=info.updated_at,
+                last_access_time=info.last_access_time,
+            )
+        )
+
+    return schemas_out.AssetsList(
+        assets=summaries,
+        total=total,
+        has_more=(offset + len(summaries)) < total,
+    )
+
+def get_asset(asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail:
+    with create_session() as session:
+        res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
+        if not res:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        info, asset, tag_names = res
+        preview_id = info.preview_id
+
+    return schemas_out.AssetDetail(
+        id=info.id,
+        name=info.name,
+        asset_hash=asset.hash if asset else None,
+        size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
+        mime_type=asset.mime_type if asset else None,
+        tags=tag_names,
+        user_metadata=info.user_metadata or {},
+        preview_id=preview_id,
+        created_at=info.created_at,
+        last_access_time=info.last_access_time,
+    )
+
+def list_tags(
+    prefix: str | None = None,
+    limit: int = 100,
+    offset: int = 0,
+    order: str = "count_desc",
+    include_zero: bool = True,
+    owner_id: str = "",
+) -> schemas_out.TagsList:
+    limit = max(1, min(1000, limit))
+    offset = max(0, offset)
+
+    with create_session() as session:
+        rows, total = list_tags_with_usage(
+            session,
+            prefix=prefix,
+            limit=limit,
+            offset=offset,
+            include_zero=include_zero,
+            order=order,
+            owner_id=owner_id,
+        )
+
+    tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows]
+    return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total)

app/assets/scanner.py

@@ -0,0 +1,229 @@
+import contextlib
+import time
+import logging
+import os
+import sqlalchemy
+
+import folder_paths
+from app.database.db import create_session, dependencies_available
+from app.assets.helpers import (
+    collect_models_files, compute_relative_filename, fast_asset_file_check, get_name_and_tags_from_asset_path,
+    list_tree,prefixes_for_root, escape_like_prefix,
+    RootType
+)
+from app.assets.database.tags import add_missing_tag_for_asset_id, ensure_tags_exist, remove_missing_tag_for_asset_id
+from app.assets.database.bulk_ops import seed_from_paths_batch
+from app.assets.database.models import Asset, AssetCacheState, AssetInfo
+
+
+def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> None:
+    """
+    Scan the given roots and seed the assets into the database.
+    """
+    if not dependencies_available():
+        if enable_logging:
+            logging.warning("Database dependencies not available, skipping assets scan")
+        return
+    t_start = time.perf_counter()
+    created = 0
+    skipped_existing = 0
+    paths: list[str] = []
+    try:
+        existing_paths: set[str] = set()
+        for r in roots:
+            try:
+                survivors: set[str] = _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True)
+                if survivors:
+                    existing_paths.update(survivors)
+            except Exception as e:
+                logging.exception("fast DB scan failed for %s: %s", r, e)
+
+        if "models" in roots:
+            paths.extend(collect_models_files())
+        if "input" in roots:
+            paths.extend(list_tree(folder_paths.get_input_directory()))
+        if "output" in roots:
+            paths.extend(list_tree(folder_paths.get_output_directory()))
+
+        specs: list[dict] = []
+        tag_pool: set[str] = set()
+        for p in paths:
+            abs_p = os.path.abspath(p)
+            if abs_p in existing_paths:
+                skipped_existing += 1
+                continue
+            try:
+                stat_p = os.stat(abs_p, follow_symlinks=False)
+            except OSError:
+                continue
+            # skip empty files
+            if not stat_p.st_size:
+                continue
+            name, tags = get_name_and_tags_from_asset_path(abs_p)
+            specs.append(
+                {
+                    "abs_path": abs_p,
+                    "size_bytes": stat_p.st_size,
+                    "mtime_ns": getattr(stat_p, "st_mtime_ns", int(stat_p.st_mtime * 1_000_000_000)),
+                    "info_name": name,
+                    "tags": tags,
+                    "fname": compute_relative_filename(abs_p),
+                }
+            )
+            for t in tags:
+                tag_pool.add(t)
+        # if no file specs, nothing to do
+        if not specs:
+            return
+        with create_session() as sess:
+            if tag_pool:
+                ensure_tags_exist(sess, tag_pool, tag_type="user")
+
+            result = seed_from_paths_batch(sess, specs=specs, owner_id="")
+            created += result["inserted_infos"]
+            sess.commit()
+    finally:
+        if enable_logging:
+            logging.info(
+                "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, total_seen=%d)",
+                roots,
+                time.perf_counter() - t_start,
+                created,
+                skipped_existing,
+                len(paths),
+            )
+
+
+def _fast_db_consistency_pass(
+    root: RootType,
+    *,
+    collect_existing_paths: bool = False,
+    update_missing_tags: bool = False,
+) -> set[str] | None:
+    """Fast DB+FS pass for a root:
+      - Toggle needs_verify per state using fast check
+      - For hashed assets with at least one fast-ok state in this root: delete stale missing states
+      - For seed assets with all states missing: delete Asset and its AssetInfos
+      - Optionally add/remove 'missing' tags based on fast-ok in this root
+      - Optionally return surviving absolute paths
+    """
+    prefixes = prefixes_for_root(root)
+    if not prefixes:
+        return set() if collect_existing_paths else None
+
+    conds = []
+    for p in prefixes:
+        base = os.path.abspath(p)
+        if not base.endswith(os.sep):
+            base += os.sep
+        escaped, esc = escape_like_prefix(base)
+        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
+
+    with create_session() as sess:
+        rows = (
+            sess.execute(
+                sqlalchemy.select(
+                    AssetCacheState.id,
+                    AssetCacheState.file_path,
+                    AssetCacheState.mtime_ns,
+                    AssetCacheState.needs_verify,
+                    AssetCacheState.asset_id,
+                    Asset.hash,
+                    Asset.size_bytes,
+                )
+                .join(Asset, Asset.id == AssetCacheState.asset_id)
+                .where(sqlalchemy.or_(*conds))
+                .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
+            )
+        ).all()
+
+        by_asset: dict[str, dict] = {}
+        for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows:
+            acc = by_asset.get(aid)
+            if acc is None:
+                acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []}
+                by_asset[aid] = acc
+
+            fast_ok = False
+            try:
+                exists = True
+                fast_ok = fast_asset_file_check(
+                    mtime_db=mtime_db,
+                    size_db=acc["size_db"],
+                    stat_result=os.stat(fp, follow_symlinks=True),
+                )
+            except FileNotFoundError:
+                exists = False
+            except OSError:
+                exists = False
+
+            acc["states"].append({
+                "sid": sid,
+                "fp": fp,
+                "exists": exists,
+                "fast_ok": fast_ok,
+                "needs_verify": bool(needs_verify),
+            })
+
+        to_set_verify: list[int] = []
+        to_clear_verify: list[int] = []
+        stale_state_ids: list[int] = []
+        survivors: set[str] = set()
+
+        for aid, acc in by_asset.items():
+            a_hash = acc["hash"]
+            states = acc["states"]
+            any_fast_ok = any(s["fast_ok"] for s in states)
+            all_missing = all(not s["exists"] for s in states)
+
+            for s in states:
+                if not s["exists"]:
+                    continue
+                if s["fast_ok"] and s["needs_verify"]:
+                    to_clear_verify.append(s["sid"])
+                if not s["fast_ok"] and not s["needs_verify"]:
+                    to_set_verify.append(s["sid"])
+
+            if a_hash is None:
+                if states and all_missing:  # remove seed Asset completely, if no valid AssetCache exists
+                    sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id == aid))
+                    asset = sess.get(Asset, aid)
+                    if asset:
+                        sess.delete(asset)
+                else:
+                    for s in states:
+                        if s["exists"]:
+                            survivors.add(os.path.abspath(s["fp"]))
+                continue
+
+            if any_fast_ok:  # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records
+                for s in states:
+                    if not s["exists"]:
+                        stale_state_ids.append(s["sid"])
+                if update_missing_tags:
+                    with contextlib.suppress(Exception):
+                        remove_missing_tag_for_asset_id(sess, asset_id=aid)
+            elif update_missing_tags:
+                with contextlib.suppress(Exception):
+                    add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic")
+
+            for s in states:
+                if s["exists"]:
+                    survivors.add(os.path.abspath(s["fp"]))
+
+        if stale_state_ids:
+            sess.execute(sqlalchemy.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids)))
+        if to_set_verify:
+            sess.execute(
+                sqlalchemy.update(AssetCacheState)
+                .where(AssetCacheState.id.in_(to_set_verify))
+                .values(needs_verify=True)
+            )
+        if to_clear_verify:
+            sess.execute(
+                sqlalchemy.update(AssetCacheState)
+                .where(AssetCacheState.id.in_(to_clear_verify))
+                .values(needs_verify=False)
+            )
+        sess.commit()
+        return survivors if collect_existing_paths else None

app/database/models.py

@@ -1,14 +1,21 @@
-from sqlalchemy.orm import declarative_base
+from typing import Any
+from datetime import datetime
+from sqlalchemy.orm import DeclarativeBase
 
-Base = declarative_base()
+class Base(DeclarativeBase):
+    pass
 
-
-def to_dict(obj):
+def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]:
     fields = obj.__table__.columns.keys()
-    return {
-        field: (val.to_dict() if hasattr(val, "to_dict") else val)
-        for field in fields
-        if (val := getattr(obj, field))
-    }
+    out: dict[str, Any] = {}
+    for field in fields:
+        val = getattr(obj, field)
+        if val is None and not include_none:
+            continue
+        if isinstance(val, datetime):
+            out[field] = val.isoformat()
+        else:
+            out[field] = val
+    return out
 
 # TODO: Define models here

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

@@ -10,6 +10,7 @@ import hashlib
 
 class Source:
     custom_node = "custom_node"
+    templates = "templates"
 
 class SubgraphEntry(TypedDict):
     source: str
@@ -38,6 +39,18 @@ class CustomNodeSubgraphEntryInfo(TypedDict):
 class SubgraphManager:
     def __init__(self):
         self.cached_custom_node_subgraphs: dict[SubgraphEntry] | None = None
+        self.cached_blueprint_subgraphs: dict[SubgraphEntry] | None = None
+
+    def _create_entry(self, file: str, source: str, node_pack: str) -> tuple[str, SubgraphEntry]:
+        """Create a subgraph entry from a file path. Expects normalized path (forward slashes)."""
+        entry_id = hashlib.sha256(f"{source}{file}".encode()).hexdigest()
+        entry: SubgraphEntry = {
+            "source": source,
+            "name": os.path.splitext(os.path.basename(file))[0],
+            "path": file,
+            "info": {"node_pack": node_pack},
+        }
+        return entry_id, entry
 
     async def load_entry_data(self, entry: SubgraphEntry):
         with open(entry['path'], 'r') as f:
@@ -60,53 +73,60 @@ class SubgraphManager:
         return entries
 
     async def get_custom_node_subgraphs(self, loadedModules, force_reload=False):
-        # if not forced to reload and cached, return cache
+        """Load subgraphs from custom nodes."""
         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] = {}
 
+        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
+            pattern = os.path.join(folder, "*/subgraphs/*.json")
+            for file in glob.glob(pattern):
                 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
+                node_pack = "custom_nodes." + file.split('/')[-3]
+                entry_id, entry = self._create_entry(file, Source.custom_node, node_pack)
+                subgraphs_dict[entry_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:
+    async def get_blueprint_subgraphs(self, force_reload=False):
+        """Load subgraphs from the blueprints directory."""
+        if not force_reload and self.cached_blueprint_subgraphs is not None:
+            return self.cached_blueprint_subgraphs
+
+        subgraphs_dict: dict[SubgraphEntry] = {}
+        blueprints_dir = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'blueprints')
+
+        if os.path.exists(blueprints_dir):
+            for file in glob.glob(os.path.join(blueprints_dir, "*.json")):
+                file = file.replace('\\', '/')
+                entry_id, entry = self._create_entry(file, Source.templates, "comfyui")
+                subgraphs_dict[entry_id] = entry
+
+        self.cached_blueprint_subgraphs = subgraphs_dict
+        return subgraphs_dict
+
+    async def get_all_subgraphs(self, loadedModules, force_reload=False):
+        """Get all subgraphs from all sources (custom nodes and blueprints)."""
+        custom_node_subgraphs = await self.get_custom_node_subgraphs(loadedModules, force_reload)
+        blueprint_subgraphs = await self.get_blueprint_subgraphs(force_reload)
+        return {**custom_node_subgraphs, **blueprint_subgraphs}
+
+    async def get_subgraph(self, id: str, loadedModules):
+        """Get a specific subgraph by ID from any source."""
+        entry = (await self.get_all_subgraphs(loadedModules)).get(id)
+        if entry is not None and entry.get('data') 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
+            subgraphs_dict = await self.get_all_subgraphs(loadedModules)
             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)
+            subgraph = await self.get_subgraph(id, loadedModules)
             return web.json_response(await self.sanitize_entry(subgraph))

comfy/cli_args.py

@@ -231,6 +231,7 @@ database_default_path = os.path.abspath(
     os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
 )
 parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
+parser.add_argument("--disable-assets-autoscan", action="store_true", help="Disable asset scanning on startup for database synchronization.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()

comfy/clip_model.py

@@ -1,6 +1,59 @@
 import torch
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.ops
+import math
+
+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])
+
+def siglip2_flex_calc_resolution(oh, ow, patch_size, max_num_patches, eps=1e-5):
+    def scale_dim(size, scale):
+        scaled = math.ceil(size * scale / patch_size) * patch_size
+        return max(patch_size, int(scaled))
+
+    # Binary search for optimal scale
+    lo, hi = eps / 10, 100.0
+    while hi - lo >= eps:
+        mid = (lo + hi) / 2
+        h, w = scale_dim(oh, mid), scale_dim(ow, mid)
+        if (h // patch_size) * (w // patch_size) <= max_num_patches:
+            lo = mid
+        else:
+            hi = mid
+
+    return scale_dim(oh, lo), scale_dim(ow, lo)
+
+def siglip2_preprocess(image, size, patch_size, num_patches, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True):
+    if size > 0:
+        return clip_preprocess(image, size=size, mean=mean, std=std, crop=crop)
+
+    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)
+
+    b, c, h, w = image.shape
+    h, w = siglip2_flex_calc_resolution(h, w, patch_size, num_patches)
+
+    image = torch.nn.functional.interpolate(image, size=(h, w), mode="bilinear", antialias=True)
+    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):
@@ -156,6 +209,27 @@ class CLIPTextModel(torch.nn.Module):
         out = self.text_projection(x[2])
         return (x[0], x[1], out, x[2])
 
+def siglip2_pos_embed(embed_weight, embeds, orig_shape):
+    embed_weight_len = round(embed_weight.shape[0] ** 0.5)
+    embed_weight = comfy.ops.cast_to_input(embed_weight, embeds).movedim(1, 0).reshape(1, -1, embed_weight_len, embed_weight_len)
+    embed_weight = torch.nn.functional.interpolate(embed_weight, size=orig_shape, mode="bilinear", align_corners=False, antialias=True)
+    embed_weight = embed_weight.reshape(-1, embed_weight.shape[-2] * embed_weight.shape[-1]).movedim(0, 1)
+    return embeds + embed_weight
+
+class Siglip2Embeddings(torch.nn.Module):
+    def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", num_patches=None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.patch_embedding = operations.Linear(num_channels * patch_size * patch_size, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = operations.Embedding(num_patches, embed_dim, dtype=dtype, device=device)
+        self.patch_size = patch_size
+
+    def forward(self, pixel_values):
+        b, c, h, w = pixel_values.shape
+        img = pixel_values.movedim(1, -1).reshape(b, h // self.patch_size, self.patch_size, w // self.patch_size, self.patch_size, c)
+        img = img.permute(0, 1, 3, 2, 4, 5)
+        img = img.reshape(b, img.shape[1] * img.shape[2], -1)
+        img = self.patch_embedding(img)
+        return siglip2_pos_embed(self.position_embedding.weight, img, (h // self.patch_size, w // self.patch_size))
 
 class CLIPVisionEmbeddings(torch.nn.Module):
     def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", dtype=None, device=None, operations=None):
@@ -199,8 +273,11 @@ class CLIPVision(torch.nn.Module):
         intermediate_activation = config_dict["hidden_act"]
         model_type = config_dict["model_type"]
 
-        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations)
-        if model_type == "siglip_vision_model":
+        if model_type in ["siglip2_vision_model"]:
+            self.embeddings = Siglip2Embeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, num_patches=config_dict.get("num_patches", None), dtype=dtype, device=device, operations=operations)
+        else:
+            self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations)
+        if model_type in ["siglip_vision_model", "siglip2_vision_model"]:
             self.pre_layrnorm = lambda a: a
             self.output_layernorm = True
         else:

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
 
@@ -17,28 +16,12 @@ 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,
     "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "siglip2_vision_model": comfy.clip_model.CLIPVisionModelProjection,
     "dinov2": comfy.image_encoders.dino2.Dinov2Model,
 }
 
@@ -50,9 +33,10 @@ 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_type = config.get("model_type", "clip_vision_model")
-        model_class = IMAGE_ENCODERS.get(model_type)
-        if model_type == "siglip_vision_model":
+        self.model_type = config.get("model_type", "clip_vision_model")
+        self.config = config.copy()
+        model_class = IMAGE_ENCODERS.get(self.model_type)
+        if self.model_type == "siglip_vision_model":
             self.return_all_hidden_states = True
         else:
             self.return_all_hidden_states = False
@@ -73,12 +57,16 @@ 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()
+        if self.model_type == "siglip2_vision_model":
+            pixel_values = comfy.clip_model.siglip2_preprocess(image.to(self.load_device), size=self.image_size, patch_size=self.config.get("patch_size", 16), num_patches=self.config.get("num_patches", 256), mean=self.image_mean, std=self.image_std, crop=crop).float()
+        else:
+            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["image_sizes"] = [pixel_values.shape[1:]] * pixel_values.shape[0]
         if self.return_all_hidden_states:
             all_hs = out[1].to(comfy.model_management.intermediate_device())
             outputs["penultimate_hidden_states"] = all_hs[:, -2]
@@ -125,10 +113,14 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
     elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
         embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
         if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
-            if embed_shape == 729:
-                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
-            elif embed_shape == 1024:
-                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
+            patch_embedding_shape = sd["vision_model.embeddings.patch_embedding.weight"].shape
+            if len(patch_embedding_shape) == 2:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip2_base_naflex.json")
+            else:
+                if embed_shape == 729:
+                    json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
+                elif embed_shape == 1024:
+                    json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
         elif embed_shape == 577:
             if "multi_modal_projector.linear_1.bias" in sd:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")

comfy/clip_vision_siglip2_base_naflex.json

@@ -0,0 +1,14 @@
+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": -1,
+  "intermediate_size": 4304,
+  "model_type": "siglip2_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 16,
+  "num_patches": 256,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/context_windows.py

@@ -143,7 +143,7 @@ class IndexListContextHandler(ContextHandlerABC):
         # 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[0]}-{window[-1]} with center ratio {window.center_ratio:.3f}")
+            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:
@@ -188,6 +188,12 @@ class IndexListContextHandler(ContextHandlerABC):
                                 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  (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \

comfy/float.py

@@ -65,3 +65,147 @@ def stochastic_rounding(value, dtype, seed=0):
         return output
 
     return value.to(dtype=dtype)
+
+
+# TODO: improve this?
+def stochastic_float_to_fp4_e2m1(x, generator):
+    orig_shape = x.shape
+    sign = torch.signbit(x).to(torch.uint8)
+
+    exp = torch.floor(torch.log2(x.abs()) + 1.0).clamp(0, 3)
+    x += (torch.rand(x.size(), dtype=x.dtype, layout=x.layout, device=x.device, generator=generator) - 0.5) * (2 ** (exp - 2.0)) * 1.25
+
+    x = x.abs()
+    exp = torch.floor(torch.log2(x) + 1.1925).clamp(0, 3)
+
+    mantissa = torch.where(
+        exp > 0,
+        (x / (2.0 ** (exp - 1)) - 1.0) * 2.0,
+        (x * 2.0),
+        out=x
+    ).round().to(torch.uint8)
+    del x
+
+    exp = exp.to(torch.uint8)
+
+    fp4 = (sign << 3) | (exp << 1) | mantissa
+    del sign, exp, mantissa
+
+    fp4_flat = fp4.view(-1)
+    packed = (fp4_flat[0::2] << 4) | fp4_flat[1::2]
+    return packed.reshape(list(orig_shape)[:-1] + [-1])
+
+
+def to_blocked(input_matrix, flatten: bool = True) -> torch.Tensor:
+    """
+    Rearrange a large matrix by breaking it into blocks and applying the rearrangement pattern.
+    See:
+        https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout
+
+    Args:
+        input_matrix: Input tensor of shape (H, W)
+    Returns:
+        Rearranged tensor of shape (32*ceil_div(H,128), 16*ceil_div(W,4))
+    """
+
+    def ceil_div(a, b):
+        return (a + b - 1) // b
+
+    rows, cols = input_matrix.shape
+    n_row_blocks = ceil_div(rows, 128)
+    n_col_blocks = ceil_div(cols, 4)
+
+    # Calculate the padded shape
+    padded_rows = n_row_blocks * 128
+    padded_cols = n_col_blocks * 4
+
+    padded = input_matrix
+    if (rows, cols) != (padded_rows, padded_cols):
+        padded = torch.zeros(
+            (padded_rows, padded_cols),
+            device=input_matrix.device,
+            dtype=input_matrix.dtype,
+        )
+        padded[:rows, :cols] = input_matrix
+
+    # Rearrange the blocks
+    blocks = padded.view(n_row_blocks, 128, n_col_blocks, 4).permute(0, 2, 1, 3)
+    rearranged = blocks.reshape(-1, 4, 32, 4).transpose(1, 2).reshape(-1, 32, 16)
+    if flatten:
+        return rearranged.flatten()
+
+    return rearranged.reshape(padded_rows, padded_cols)
+
+
+def stochastic_round_quantize_nvfp4_block(x, per_tensor_scale, generator):
+    F4_E2M1_MAX = 6.0
+    F8_E4M3_MAX = 448.0
+
+    orig_shape = x.shape
+
+    block_size = 16
+
+    x = x.reshape(orig_shape[0], -1, block_size)
+    scaled_block_scales_fp8 = torch.clamp(((torch.amax(torch.abs(x), dim=-1)) / F4_E2M1_MAX) / per_tensor_scale.to(x.dtype), max=F8_E4M3_MAX).to(torch.float8_e4m3fn)
+    x = x / (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1)
+
+    x = x.view(orig_shape).nan_to_num()
+    data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator)
+    return data_lp, scaled_block_scales_fp8
+
+
+def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0):
+    def roundup(x: int, multiple: int) -> int:
+        """Round up x to the nearest multiple."""
+        return ((x + multiple - 1) // multiple) * multiple
+
+    generator = torch.Generator(device=x.device)
+    generator.manual_seed(seed)
+
+    # Handle padding
+    if pad_16x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 16)
+        padded_cols = roundup(cols, 16)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+
+    x, blocked_scaled = stochastic_round_quantize_nvfp4_block(x, per_tensor_scale, generator)
+    return x, to_blocked(blocked_scaled, flatten=False)
+
+
+def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=0, block_size=4096 * 4096):
+    def roundup(x: int, multiple: int) -> int:
+        """Round up x to the nearest multiple."""
+        return ((x + multiple - 1) // multiple) * multiple
+
+    orig_shape = x.shape
+
+    # Handle padding
+    if pad_16x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 16)
+        padded_cols = roundup(cols, 16)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+            # Note: We update orig_shape because the output tensor logic below assumes x.shape matches
+            # what we want to produce. If we pad here, we want the padded output.
+            orig_shape = x.shape
+
+    orig_shape = list(orig_shape)
+
+    output_fp4 = torch.empty(orig_shape[:-1] + [orig_shape[-1] // 2], dtype=torch.uint8, device=x.device)
+    output_block = torch.empty(orig_shape[:-1] + [orig_shape[-1] // 16], dtype=torch.float8_e4m3fn, device=x.device)
+
+    generator = torch.Generator(device=x.device)
+    generator.manual_seed(seed)
+
+    num_slices = max(1, (x.numel() / block_size))
+    slice_size = max(1, (round(x.shape[0] / num_slices)))
+
+    for i in range(0, x.shape[0], slice_size):
+        fp4, block = stochastic_round_quantize_nvfp4_block(x[i: i + slice_size], per_tensor_scale, generator=generator)
+        output_fp4[i:i + slice_size].copy_(fp4)
+        output_block[i:i + slice_size].copy_(block)
+
+    return output_fp4, to_blocked(output_block, flatten=False)

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/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:
@@ -1776,7 +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]
@@ -1797,7 +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

@@ -407,6 +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

comfy/ldm/anima/model.py

@@ -0,0 +1,202 @@
+from comfy.ldm.cosmos.predict2 import MiniTrainDIT
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+
+def rotate_half(x):
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(x, cos, sin, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    x_embed = (x * cos) + (rotate_half(x) * sin)
+    return x_embed
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, head_dim):
+        super().__init__()
+        self.rope_theta = 10000
+        inv_freq = 1.0 / (self.rope_theta ** (torch.arange(0, head_dim, 2, dtype=torch.int64).to(dtype=torch.float) / head_dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Attention(nn.Module):
+    def __init__(self, query_dim, context_dim, n_heads, head_dim, device=None, dtype=None, operations=None):
+        super().__init__()
+
+        inner_dim = head_dim * n_heads
+        self.n_heads = n_heads
+        self.head_dim = head_dim
+        self.query_dim = query_dim
+        self.context_dim = context_dim
+
+        self.q_proj = operations.Linear(query_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.k_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.v_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+
+        self.o_proj = operations.Linear(inner_dim, query_dim, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None, context=None, position_embeddings=None, position_embeddings_context=None):
+        context = x if context is None else context
+        input_shape = x.shape[:-1]
+        q_shape = (*input_shape, self.n_heads, self.head_dim)
+        context_shape = context.shape[:-1]
+        kv_shape = (*context_shape, self.n_heads, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(x).view(q_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(context).view(kv_shape)).transpose(1, 2)
+        value_states = self.v_proj(context).view(kv_shape).transpose(1, 2)
+
+        if position_embeddings is not None:
+            assert position_embeddings_context is not None
+            cos, sin = position_embeddings
+            query_states = apply_rotary_pos_emb(query_states, cos, sin)
+            cos, sin = position_embeddings_context
+            key_states = apply_rotary_pos_emb(key_states, cos, sin)
+
+        attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attn_mask=mask)
+
+        attn_output = attn_output.transpose(1, 2).reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output
+
+    def init_weights(self):
+        torch.nn.init.zeros_(self.o_proj.weight)
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, source_dim, model_dim, num_heads=16, mlp_ratio=4.0, use_self_attn=False, layer_norm=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.use_self_attn = use_self_attn
+
+        if self.use_self_attn:
+            self.norm_self_attn = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype)
+            self.self_attn = Attention(
+                query_dim=model_dim,
+                context_dim=model_dim,
+                n_heads=num_heads,
+                head_dim=model_dim//num_heads,
+                device=device,
+                dtype=dtype,
+                operations=operations,
+            )
+
+        self.norm_cross_attn = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype)
+        self.cross_attn = Attention(
+            query_dim=model_dim,
+            context_dim=source_dim,
+            n_heads=num_heads,
+            head_dim=model_dim//num_heads,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+        self.norm_mlp = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype)
+        self.mlp = nn.Sequential(
+            operations.Linear(model_dim, int(model_dim * mlp_ratio), device=device, dtype=dtype),
+            nn.GELU(),
+            operations.Linear(int(model_dim * mlp_ratio), model_dim, device=device, dtype=dtype)
+        )
+
+    def forward(self, x, context, target_attention_mask=None, source_attention_mask=None, position_embeddings=None, position_embeddings_context=None):
+        if self.use_self_attn:
+            normed = self.norm_self_attn(x)
+            attn_out = self.self_attn(normed, mask=target_attention_mask, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings)
+            x = x + attn_out
+
+        normed = self.norm_cross_attn(x)
+        attn_out = self.cross_attn(normed, mask=source_attention_mask, context=context, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings_context)
+        x = x + attn_out
+
+        x = x + self.mlp(self.norm_mlp(x))
+        return x
+
+    def init_weights(self):
+        torch.nn.init.zeros_(self.mlp[2].weight)
+        self.cross_attn.init_weights()
+
+
+class LLMAdapter(nn.Module):
+    def __init__(
+            self,
+            source_dim=1024,
+            target_dim=1024,
+            model_dim=1024,
+            num_layers=6,
+            num_heads=16,
+            use_self_attn=True,
+            layer_norm=False,
+            device=None,
+            dtype=None,
+            operations=None,
+        ):
+        super().__init__()
+
+        self.embed = operations.Embedding(32128, target_dim, device=device, dtype=dtype)
+        if model_dim != target_dim:
+            self.in_proj = operations.Linear(target_dim, model_dim, device=device, dtype=dtype)
+        else:
+            self.in_proj = nn.Identity()
+        self.rotary_emb = RotaryEmbedding(model_dim//num_heads)
+        self.blocks = nn.ModuleList([
+            TransformerBlock(source_dim, model_dim, num_heads=num_heads, use_self_attn=use_self_attn, layer_norm=layer_norm, device=device, dtype=dtype, operations=operations) for _ in range(num_layers)
+        ])
+        self.out_proj = operations.Linear(model_dim, target_dim, device=device, dtype=dtype)
+        self.norm = operations.RMSNorm(target_dim, eps=1e-6, device=device, dtype=dtype)
+
+    def forward(self, source_hidden_states, target_input_ids, target_attention_mask=None, source_attention_mask=None):
+        if target_attention_mask is not None:
+            target_attention_mask = target_attention_mask.to(torch.bool)
+            if target_attention_mask.ndim == 2:
+                target_attention_mask = target_attention_mask.unsqueeze(1).unsqueeze(1)
+
+        if source_attention_mask is not None:
+            source_attention_mask = source_attention_mask.to(torch.bool)
+            if source_attention_mask.ndim == 2:
+                source_attention_mask = source_attention_mask.unsqueeze(1).unsqueeze(1)
+
+        x = self.in_proj(self.embed(target_input_ids))
+        context = source_hidden_states
+        position_ids = torch.arange(x.shape[1], device=x.device).unsqueeze(0)
+        position_ids_context = torch.arange(context.shape[1], device=x.device).unsqueeze(0)
+        position_embeddings = self.rotary_emb(x, position_ids)
+        position_embeddings_context = self.rotary_emb(x, position_ids_context)
+        for block in self.blocks:
+            x = block(x, context, target_attention_mask=target_attention_mask, source_attention_mask=source_attention_mask, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings_context)
+        return self.norm(self.out_proj(x))
+
+
+class Anima(MiniTrainDIT):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations"))
+
+    def preprocess_text_embeds(self, text_embeds, text_ids):
+        if text_ids is not None:
+            return self.llm_adapter(text_embeds, text_ids)
+        else:
+            return text_embeds

comfy/ldm/chroma_radiance/model.py

@@ -270,7 +270,7 @@ class ChromaRadiance(Chroma):
         bad_keys = tuple(
             k
             for k, v in overrides.items()
-            if type(v) != type(getattr(params, k)) and (v is not None or k not in nullable_keys)
+            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)}"

comfy/ldm/flux/math.py

@@ -4,6 +4,7 @@ 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, transformer_options={}) -> Tensor:
@@ -13,7 +14,6 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme
     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():
@@ -28,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
     return out.to(dtype=torch.float32, device=pos.device)
 
-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])
+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)
 
-    return x_out.reshape(*x.shape).type_as(x)
+        x_out = freqs_cis[..., 0] * x_[..., 0]
+        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
 
-def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+        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/hunyuan_video/upsampler.py

@@ -3,7 +3,8 @@ 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, model_patcher
+import comfy.model_management
+import comfy.model_patcher
 
 class SRResidualCausalBlock3D(nn.Module):
     def __init__(self, channels: int):
@@ -102,13 +103,13 @@ UPSAMPLERS = {
 
 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.load_device = comfy.model_management.vae_device()
+        offload_device = comfy.model_management.vae_offload_device()
+        self.dtype = comfy.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)
+        self.patcher = comfy.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)
@@ -117,5 +118,5 @@ class HunyuanVideo15SRModel():
         return self.model.state_dict()
 
     def resample_latent(self, latent):
-        model_management.load_model_gpu(self.patcher)
+        comfy.model_management.load_model_gpu(self.patcher)
         return self.model(latent.to(self.load_device))

comfy/ldm/lightricks/av_model.py

@@ -0,0 +1,913 @@
+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 CompressedTimestep:
+    """Store video timestep embeddings in compressed form using per-frame indexing."""
+    __slots__ = ('data', 'batch_size', 'num_frames', 'patches_per_frame', 'feature_dim')
+
+    def __init__(self, tensor: torch.Tensor, patches_per_frame: int):
+        """
+        tensor: [batch_size, num_tokens, feature_dim] tensor where num_tokens = num_frames * patches_per_frame
+        patches_per_frame: Number of spatial patches per frame (height * width in latent space)
+        """
+        self.batch_size, num_tokens, self.feature_dim = tensor.shape
+
+        # Check if compression is valid (num_tokens must be divisible by patches_per_frame)
+        if num_tokens % patches_per_frame == 0 and num_tokens >= patches_per_frame:
+            self.patches_per_frame = patches_per_frame
+            self.num_frames = num_tokens // patches_per_frame
+
+            # Reshape to [batch, frames, patches_per_frame, feature_dim] and store one value per frame
+            # All patches in a frame are identical, so we only keep the first one
+            reshaped = tensor.view(self.batch_size, self.num_frames, patches_per_frame, self.feature_dim)
+            self.data = reshaped[:, :, 0, :].contiguous()  # [batch, frames, feature_dim]
+        else:
+            # Not divisible or too small - store directly without compression
+            self.patches_per_frame = 1
+            self.num_frames = num_tokens
+            self.data = tensor
+
+    def expand(self):
+        """Expand back to original tensor."""
+        if self.patches_per_frame == 1:
+            return self.data
+
+        # [batch, frames, feature_dim] -> [batch, frames, patches_per_frame, feature_dim] -> [batch, tokens, feature_dim]
+        expanded = self.data.unsqueeze(2).expand(self.batch_size, self.num_frames, self.patches_per_frame, self.feature_dim)
+        return expanded.reshape(self.batch_size, -1, self.feature_dim)
+
+    def expand_for_computation(self, scale_shift_table: torch.Tensor, batch_size: int, indices: slice = slice(None, None)):
+        """Compute ada values on compressed per-frame data, then expand spatially."""
+        num_ada_params = scale_shift_table.shape[0]
+
+        # No compression - compute directly
+        if self.patches_per_frame == 1:
+            num_tokens = self.data.shape[1]
+            dim_per_param = self.feature_dim // num_ada_params
+            reshaped = self.data.reshape(batch_size, num_tokens, num_ada_params, dim_per_param)[:, :, indices, :]
+            table_values = scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=self.data.device, dtype=self.data.dtype)
+            ada_values = (table_values + reshaped).unbind(dim=2)
+            return ada_values
+
+        # Compressed: compute on per-frame data then expand spatially
+        # Reshape: [batch, frames, feature_dim] -> [batch, frames, num_ada_params, dim_per_param]
+        frame_reshaped = self.data.reshape(batch_size, self.num_frames, num_ada_params, -1)[:, :, indices, :]
+        table_values = scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(
+            device=self.data.device, dtype=self.data.dtype
+        )
+        frame_ada = (table_values + frame_reshaped).unbind(dim=2)
+
+        # Expand each ada parameter spatially: [batch, frames, dim] -> [batch, frames, patches, dim] -> [batch, tokens, dim]
+        return tuple(
+            frame_val.unsqueeze(2).expand(batch_size, self.num_frames, self.patches_per_frame, -1)
+            .reshape(batch_size, -1, frame_val.shape[-1])
+            for frame_val in frame_ada
+        )
+
+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)
+    ):
+        if isinstance(timestep, CompressedTimestep):
+            return timestep.expand_for_computation(scale_shift_table, batch_size, indices)
+
+        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,
+        )
+
+        return (*scale_shift_ada_values, *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_scaled = timestep * self.timestep_scale_multiplier
+
+        v_timestep, v_embedded_timestep = self.adaln_single(
+            timestep_scaled.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=hidden_dtype,
+        )
+
+        # Calculate patches_per_frame from orig_shape: [batch, channels, frames, height, width]
+        # Video tokens are arranged as (frames * height * width), so patches_per_frame = height * width
+        orig_shape = kwargs.get("orig_shape")
+        v_patches_per_frame = None
+        if orig_shape is not None and len(orig_shape) == 5:
+            # orig_shape[3] = height, orig_shape[4] = width (in latent space)
+            v_patches_per_frame = orig_shape[3] * orig_shape[4]
+
+        # Reshape to [batch_size, num_tokens, dim] and compress for storage
+        v_timestep = CompressedTimestep(v_timestep.view(batch_size, -1, v_timestep.shape[-1]), v_patches_per_frame)
+        v_embedded_timestep = CompressedTimestep(v_embedded_timestep.view(batch_size, -1, v_embedded_timestep.shape[-1]), v_patches_per_frame)
+
+        # Prepare audio timestep
+        a_timestep = kwargs.get("a_timestep")
+        if a_timestep is not None:
+            a_timestep_scaled = a_timestep * self.timestep_scale_multiplier
+            a_timestep_flat = a_timestep_scaled.flatten()
+            timestep_flat = timestep_scaled.flatten()
+            av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier
+
+            # Cross-attention timesteps - compress these too
+            av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single(
+                a_timestep_flat,
+                {"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_flat,
+                {"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_flat * 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_flat * av_ca_factor,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+
+            # Compress cross-attention timesteps (only video side, audio is too small to benefit)
+            cross_av_timestep_ss = [
+                av_ca_audio_scale_shift_timestep.view(batch_size, -1, av_ca_audio_scale_shift_timestep.shape[-1]),
+                CompressedTimestep(av_ca_video_scale_shift_timestep.view(batch_size, -1, av_ca_video_scale_shift_timestep.shape[-1]), v_patches_per_frame),  # video - compressed
+                CompressedTimestep(av_ca_a2v_gate_noise_timestep.view(batch_size, -1, av_ca_a2v_gate_noise_timestep.shape[-1]), v_patches_per_frame),  # video - compressed
+                av_ca_v2a_gate_noise_timestep.view(batch_size, -1, av_ca_v2a_gate_noise_timestep.shape[-1]),
+            ]
+
+            a_timestep, a_embedded_timestep = self.audio_adaln_single(
+                a_timestep_flat,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            # Audio timesteps
+            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])
+        else:
+            a_timestep = timestep_scaled
+            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]
+
+        # Expand compressed video timestep if needed
+        if isinstance(v_embedded_timestep, CompressedTimestep):
+            v_embedded_timestep = v_embedded_timestep.expand()
+
+        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

@@ -1,13 +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
-import math
-from typing import Dict, Optional, Tuple
 
 from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
-from comfy.ldm.flux.math import apply_rope1
+
+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,
@@ -39,9 +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)
@@ -73,7 +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__()
 
@@ -90,7 +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
@@ -139,12 +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)
@@ -163,15 +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,
@@ -185,6 +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.
@@ -192,18 +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)
@@ -222,23 +282,68 @@ 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_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)
+    input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)")
+
+    out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs
+
+    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
@@ -254,9 +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, transformer_options={}):
+    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)
@@ -266,8 +373,8 @@ class CrossAttention(nn.Module):
         k = self.k_norm(k)
 
         if pe is not None:
-            q = apply_rope1(q.unsqueeze(1), pe).squeeze(1)
-            k = apply_rope1(k.unsqueeze(1), pe).squeeze(1)
+            q = apply_rotary_emb(q, 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, transformer_options=transformer_options)
@@ -277,14 +384,34 @@ class CrossAttention(nn.Module):
 
 
 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))
 
@@ -306,116 +433,446 @@ class BasicTransformerBlock(nn.Module):
         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
-    device = indices_grid.device
+@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
+
+    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),
+            inner_dim // n_elem,
+            device=device,
+            dtype=torch.float32,
+        )
+    )
+    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 = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2
+    indices = indices.to(device=fractional_positions.device)
 
-    # Compute frequencies and apply cos/sin
-    freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2)
-    cos_vals = freqs.cos().repeat_interleave(2, dim=-1)
-    sin_vals = freqs.sin().repeat_interleave(2, dim=-1)
+    freqs = (
+        (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
+        .transpose(-1, -2)
+        .flatten(2)
+    )
+    return freqs
 
-    # Pad if dim is not divisible by 6
-    if dim % 6 != 0:
-        padding_size = dim % 6
-        cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1)
-        sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1)
+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 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, sin_freq
 
-    # Reshape and extract one value per pair (since repeat_interleave duplicates each value)
-    cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype)  # [B, N, dim//2]
-    sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype)  # [B, N, dim//2]
+def split_freqs_cis(freqs, pad_size, num_attention_heads):
+    cos_freq = freqs.cos()
+    sin_freq = freqs.sin()
 
-    # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension
-    freqs_cis = torch.stack([
-        torch.stack([cos_vals, -sin_vals], dim=-1),
-        torch.stack([sin_vals, cos_vals], dim=-1)
-    ], dim=-2).unsqueeze(1)  # [B, 1, N, dim//2, 2, 2]
+    if pad_size != 0:
+        cos_padding = torch.ones_like(cos_freq[:, :, :pad_size])
+        sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size])
 
-    return freqs_cis
+        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
 
-class LTXVModel(torch.nn.Module):
-    def __init__(self,
-                 in_channels=128,
-                 cross_attention_dim=2048,
-                 attention_head_dim=64,
-                 num_attention_heads=32,
+    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)
 
-                 caption_channels=4096,
-                 num_layers=28,
+    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).
 
-                 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):
+    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,
@@ -423,44 +880,30 @@ 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"], transformer_options=args["transformer_options"])
@@ -478,16 +921,28 @@ class LTXVModel(torch.nn.Module):
                     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 = torch.addcmul(x, x, scale).add_(shift)
+        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 +1,279 @@
+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)
+
+    def waveform_to_mel(
+        self, waveform: torch.Tensor, waveform_sample_rate: int, device
+    ) -> torch.Tensor:
+        waveform = self.resample(waveform, waveform_sample_rate)
+
+        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:
+            if waveform.shape[1] == 1:
+                waveform = waveform.expand(-1, expected_channels, *waveform.shape[2:])
+            else:
+                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/vae/causal_conv3d.py

@@ -1,11 +1,11 @@
 from typing import Tuple, Union
 
+import threading
 import torch
 import torch.nn as nn
 import comfy.ops
 ops = comfy.ops.disable_weight_init
 
-
 class CausalConv3d(nn.Module):
     def __init__(
         self,
@@ -42,23 +42,34 @@ class CausalConv3d(nn.Module):
             padding_mode=spatial_padding_mode,
             groups=groups,
         )
+        self.temporal_cache_state={}
 
     def forward(self, x, causal: bool = True):
-        if causal:
-            first_frame_pad = x[:, :, :1, :, :].repeat(
-                (1, 1, self.time_kernel_size - 1, 1, 1)
-            )
-            x = torch.concatenate((first_frame_pad, x), dim=2)
-        else:
-            first_frame_pad = x[:, :, :1, :, :].repeat(
-                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
-            )
-            last_frame_pad = x[:, :, -1:, :, :].repeat(
-                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
-            )
-            x = torch.concatenate((first_frame_pad, x, last_frame_pad), dim=2)
-        x = self.conv(x)
-        return x
+        tid = threading.get_ident()
+
+        cached, is_end = self.temporal_cache_state.get(tid, (None, False))
+        if cached is None:
+            padding_length = self.time_kernel_size - 1
+            if not causal:
+                padding_length = padding_length // 2
+            if x.shape[2] == 0:
+                return x
+            cached = x[:, :, :1, :, :].repeat((1, 1, padding_length, 1, 1))
+        pieces = [ cached, x ]
+        if is_end and not causal:
+            pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
+
+        needs_caching = not is_end
+        if needs_caching and x.shape[2] >= self.time_kernel_size - 1:
+            needs_caching = False
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+
+        x = torch.cat(pieces, dim=2)
+
+        if needs_caching:
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+
+        return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
 
     @property
     def weight(self):

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+import threading
 import torch
 from torch import nn
 from functools import partial
@@ -6,12 +7,35 @@ import math
 from einops import rearrange
 from typing import List, Optional, Tuple, Union
 from .conv_nd_factory import make_conv_nd, make_linear_nd
+from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
 
+def mark_conv3d_ended(module):
+    tid = threading.get_ident()
+    for _, m in module.named_modules():
+        if isinstance(m, CausalConv3d):
+            current = m.temporal_cache_state.get(tid, (None, False))
+            m.temporal_cache_state[tid] = (current[0], True)
+
+def split2(tensor, split_point, dim=2):
+    return torch.split(tensor, [split_point, tensor.shape[dim] - split_point], dim=dim)
+
+def add_exchange_cache(dest, cache_in, new_input, dim=2):
+    if dest is not None:
+        if cache_in is not None:
+            cache_to_dest = min(dest.shape[dim], cache_in.shape[dim])
+            lead_in_dest, dest = split2(dest, cache_to_dest, dim=dim)
+            lead_in_source, cache_in = split2(cache_in, cache_to_dest, dim=dim)
+            lead_in_dest.add_(lead_in_source)
+        body, new_input = split2(new_input, dest.shape[dim], dim)
+        dest.add_(body)
+    return torch_cat_if_needed([cache_in, new_input], dim=dim)
+
 class Encoder(nn.Module):
     r"""
     The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
@@ -205,7 +229,7 @@ class Encoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+    def forward_orig(self, sample: torch.FloatTensor) -> torch.FloatTensor:
         r"""The forward method of the `Encoder` class."""
 
         sample = patchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
@@ -254,6 +278,22 @@ class Encoder(nn.Module):
 
         return sample
 
+    def forward(self, *args, **kwargs):
+        #No encoder support so just flag the end so it doesnt use the cache.
+        mark_conv3d_ended(self)
+        try:
+            return self.forward_orig(*args, **kwargs)
+        finally:
+            tid = threading.get_ident()
+            for _, module in self.named_modules():
+                # ComfyUI doesn't thread this kind of stuff today, but just in case
+                # we key on the thread to make it thread safe.
+                tid = threading.get_ident()
+                if hasattr(module, "temporal_cache_state"):
+                    module.temporal_cache_state.pop(tid, None)
+
+
+MAX_CHUNK_SIZE=(128 * 1024 ** 2)
 
 class Decoder(nn.Module):
     r"""
@@ -341,18 +381,6 @@ class Decoder(nn.Module):
                     timestep_conditioning=timestep_conditioning,
                     spatial_padding_mode=spatial_padding_mode,
                 )
-            elif block_name == "attn_res_x":
-                block = UNetMidBlock3D(
-                    dims=dims,
-                    in_channels=input_channel,
-                    num_layers=block_params["num_layers"],
-                    resnet_groups=norm_num_groups,
-                    norm_layer=norm_layer,
-                    inject_noise=block_params.get("inject_noise", False),
-                    timestep_conditioning=timestep_conditioning,
-                    attention_head_dim=block_params["attention_head_dim"],
-                    spatial_padding_mode=spatial_padding_mode,
-                )
             elif block_name == "res_x_y":
                 output_channel = output_channel // block_params.get("multiplier", 2)
                 block = ResnetBlock3D(
@@ -428,8 +456,9 @@ class Decoder(nn.Module):
             )
             self.last_scale_shift_table = nn.Parameter(torch.empty(2, output_channel))
 
+
     # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
-    def forward(
+    def forward_orig(
         self,
         sample: torch.FloatTensor,
         timestep: Optional[torch.Tensor] = None,
@@ -437,6 +466,7 @@ class Decoder(nn.Module):
         r"""The forward method of the `Decoder` class."""
         batch_size = sample.shape[0]
 
+        mark_conv3d_ended(self.conv_in)
         sample = self.conv_in(sample, causal=self.causal)
 
         checkpoint_fn = (
@@ -445,24 +475,12 @@ class Decoder(nn.Module):
             else lambda x: x
         )
 
-        scaled_timestep = None
+        timestep_shift_scale = None
         if self.timestep_conditioning:
             assert (
                 timestep is not None
             ), "should pass timestep with timestep_conditioning=True"
             scaled_timestep = timestep * self.timestep_scale_multiplier.to(dtype=sample.dtype, device=sample.device)
-
-        for up_block in self.up_blocks:
-            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
-                sample = checkpoint_fn(up_block)(
-                    sample, causal=self.causal, timestep=scaled_timestep
-                )
-            else:
-                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
-
-        sample = self.conv_norm_out(sample)
-
-        if self.timestep_conditioning:
             embedded_timestep = self.last_time_embedder(
                 timestep=scaled_timestep.flatten(),
                 resolution=None,
@@ -483,16 +501,62 @@ class Decoder(nn.Module):
                 embedded_timestep.shape[-2],
                 embedded_timestep.shape[-1],
             )
-            shift, scale = ada_values.unbind(dim=1)
-            sample = sample * (1 + scale) + shift
+            timestep_shift_scale = ada_values.unbind(dim=1)
 
-        sample = self.conv_act(sample)
-        sample = self.conv_out(sample, causal=self.causal)
+        output = []
+
+        def run_up(idx, sample, ended):
+            if idx >= len(self.up_blocks):
+                sample = self.conv_norm_out(sample)
+                if timestep_shift_scale is not None:
+                    shift, scale = timestep_shift_scale
+                    sample = sample * (1 + scale) + shift
+                sample = self.conv_act(sample)
+                if ended:
+                    mark_conv3d_ended(self.conv_out)
+                sample = self.conv_out(sample, causal=self.causal)
+                if sample is not None and sample.shape[2] > 0:
+                    output.append(sample)
+                return
+
+            up_block = self.up_blocks[idx]
+            if (ended):
+                mark_conv3d_ended(up_block)
+            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
+                sample = checkpoint_fn(up_block)(
+                    sample, causal=self.causal, timestep=scaled_timestep
+                )
+            else:
+                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
+
+            if sample is None or sample.shape[2] == 0:
+                return
+
+            total_bytes = sample.numel() * sample.element_size()
+            num_chunks = (total_bytes + MAX_CHUNK_SIZE - 1) // MAX_CHUNK_SIZE
+            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
+
+            for chunk_idx, sample1 in enumerate(samples):
+                run_up(idx + 1, sample1, ended and chunk_idx == len(samples) - 1)
+
+        run_up(0, sample, True)
+        sample = torch.cat(output, dim=2)
 
         sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
 
         return sample
 
+    def forward(self, *args, **kwargs):
+        try:
+            return self.forward_orig(*args, **kwargs)
+        finally:
+            for _, module in self.named_modules():
+                #ComfyUI doesn't thread this kind of stuff today, but just incase
+                #we key on the thread to make it thread safe.
+                tid = threading.get_ident()
+                if hasattr(module, "temporal_cache_state"):
+                    module.temporal_cache_state.pop(tid, None)
+
 
 class UNetMidBlock3D(nn.Module):
     """
@@ -663,8 +727,22 @@ class DepthToSpaceUpsample(nn.Module):
         )
         self.residual = residual
         self.out_channels_reduction_factor = out_channels_reduction_factor
+        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True, timestep: Optional[torch.Tensor] = None):
+        tid = threading.get_ident()
+        cached, drop_first_conv, drop_first_res = self.temporal_cache_state.get(tid, (None, True, True))
+        y = self.conv(x, causal=causal)
+        y = rearrange(
+            y,
+            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+        if self.stride[0] == 2 and y.shape[2] > 0 and drop_first_conv:
+            y = y[:, :, 1:, :, :]
+            drop_first_conv = False
         if self.residual:
             # Reshape and duplicate the input to match the output shape
             x_in = rearrange(
@@ -676,21 +754,20 @@ class DepthToSpaceUpsample(nn.Module):
             )
             num_repeat = math.prod(self.stride) // self.out_channels_reduction_factor
             x_in = x_in.repeat(1, num_repeat, 1, 1, 1)
-            if self.stride[0] == 2:
+            if self.stride[0] == 2 and x_in.shape[2] > 0 and drop_first_res:
                 x_in = x_in[:, :, 1:, :, :]
-        x = self.conv(x, causal=causal)
-        x = rearrange(
-            x,
-            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
-            p1=self.stride[0],
-            p2=self.stride[1],
-            p3=self.stride[2],
-        )
-        if self.stride[0] == 2:
-            x = x[:, :, 1:, :, :]
-        if self.residual:
-            x = x + x_in
-        return x
+                drop_first_res = False
+
+            if y.shape[2] == 0:
+                y = None
+
+            cached = add_exchange_cache(y, cached, x_in, dim=2)
+            self.temporal_cache_state[tid] = (cached, drop_first_conv, drop_first_res)
+
+        else:
+            self.temporal_cache_state[tid] = (None, drop_first_conv, False)
+
+        return y
 
 class LayerNorm(nn.Module):
     def __init__(self, dim, eps, elementwise_affine=True) -> None:
@@ -807,6 +884,8 @@ class ResnetBlock3D(nn.Module):
                 torch.randn(4, in_channels) / in_channels**0.5
             )
 
+        self.temporal_cache_state={}
+
     def _feed_spatial_noise(
         self, hidden_states: torch.FloatTensor, per_channel_scale: torch.FloatTensor
     ) -> torch.FloatTensor:
@@ -880,9 +959,12 @@ class ResnetBlock3D(nn.Module):
 
         input_tensor = self.conv_shortcut(input_tensor)
 
-        output_tensor = input_tensor + hidden_states
+        tid = threading.get_ident()
+        cached = self.temporal_cache_state.get(tid, None)
+        cached = add_exchange_cache(hidden_states, cached, input_tensor, dim=2)
+        self.temporal_cache_state[tid] = cached
 
-        return output_tensor
+        return hidden_states
 
 
 def patchify(x, patch_size_hw, patch_size_t=1):

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/model.py

@@ -13,10 +13,53 @@ 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
+import comfy.utils
 
 
-def modulate(x, scale):
-    return x * (1 + scale.unsqueeze(1))
+def invert_slices(slices, length):
+    sorted_slices = sorted(slices)
+    result = []
+    current = 0
+
+    for start, end in sorted_slices:
+        if current < start:
+            result.append((current, start))
+        current = max(current, end)
+
+    if current < length:
+        result.append((current, length))
+
+    return result
+
+
+def modulate(x, scale, timestep_zero_index=None):
+    if timestep_zero_index is None:
+        return x * (1 + scale.unsqueeze(1))
+    else:
+        scale = (1 + scale.unsqueeze(1))
+        actual_batch = scale.size(0) // 2
+        slices = timestep_zero_index
+        invert = invert_slices(timestep_zero_index, x.shape[1])
+        for s in slices:
+            x[:, s[0]:s[1]] *= scale[actual_batch:]
+        for s in invert:
+            x[:, s[0]:s[1]] *= scale[:actual_batch]
+        return x
+
+
+def apply_gate(gate, x, timestep_zero_index=None):
+    if timestep_zero_index is None:
+        return gate * x
+    else:
+        actual_batch = gate.size(0) // 2
+
+        slices = timestep_zero_index
+        invert = invert_slices(timestep_zero_index, x.shape[1])
+        for s in slices:
+            x[:, s[0]:s[1]] *= gate[actual_batch:]
+        for s in invert:
+            x[:, s[0]:s[1]] *= gate[:actual_batch]
+        return x
 
 #############################################################################
 #                               Core NextDiT Model                              #
@@ -258,6 +301,7 @@ class JointTransformerBlock(nn.Module):
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor]=None,
+        timestep_zero_index=None,
         transformer_options={},
     ):
         """
@@ -276,18 +320,18 @@ class JointTransformerBlock(nn.Module):
             assert adaln_input is not None
             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(
+            x = x + apply_gate(gate_msa.unsqueeze(1).tanh(), self.attention_norm2(
                 clamp_fp16(self.attention(
-                    modulate(self.attention_norm1(x), scale_msa),
+                    modulate(self.attention_norm1(x), scale_msa, timestep_zero_index=timestep_zero_index),
                     x_mask,
                     freqs_cis,
                     transformer_options=transformer_options,
-                ))
+                ))), timestep_zero_index=timestep_zero_index
             )
-            x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
+            x = x + apply_gate(gate_mlp.unsqueeze(1).tanh(), self.ffn_norm2(
                 clamp_fp16(self.feed_forward(
-                    modulate(self.ffn_norm1(x), scale_mlp),
-                ))
+                    modulate(self.ffn_norm1(x), scale_mlp, timestep_zero_index=timestep_zero_index),
+                ))), timestep_zero_index=timestep_zero_index
             )
         else:
             assert adaln_input is None
@@ -345,13 +389,37 @@ class FinalLayer(nn.Module):
             ),
         )
 
-    def forward(self, x, c):
+    def forward(self, x, c, timestep_zero_index=None):
         scale = self.adaLN_modulation(c)
-        x = modulate(self.norm_final(x), scale)
+        x = modulate(self.norm_final(x), scale, timestep_zero_index=timestep_zero_index)
         x = self.linear(x)
         return x
 
 
+def pad_zimage(feats, pad_token, pad_tokens_multiple):
+    pad_extra = (-feats.shape[1]) % pad_tokens_multiple
+    return torch.cat((feats, pad_token.to(device=feats.device, dtype=feats.dtype, copy=True).unsqueeze(0).repeat(feats.shape[0], pad_extra, 1)), dim=1), pad_extra
+
+
+def pos_ids_x(start_t, H_tokens, W_tokens, batch_size, device, transformer_options={}):
+    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)
+
+        h_start = rope_options.get("shift_y", 0.0)
+        w_start = rope_options.get("shift_x", 0.0)
+    x_pos_ids = torch.zeros((batch_size, H_tokens * W_tokens, 3), dtype=torch.float32, device=device)
+    x_pos_ids[:, :, 0] = start_t
+    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()
+    return x_pos_ids
+
+
 class NextDiT(nn.Module):
     """
     Diffusion model with a Transformer backbone.
@@ -378,6 +446,7 @@ class NextDiT(nn.Module):
         time_scale=1.0,
         pad_tokens_multiple=None,
         clip_text_dim=None,
+        siglip_feat_dim=None,
         image_model=None,
         device=None,
         dtype=None,
@@ -491,7 +560,43 @@ class NextDiT(nn.Module):
                 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"))
+
+        if siglip_feat_dim is not None:
+            self.siglip_embedder = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(siglip_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+                operation_settings.get("operations").Linear(
+                    siglip_feat_dim,
+                    dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+            self.siglip_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=False,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+            self.siglip_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
+        else:
+            self.siglip_embedder = None
+            self.siglip_refiner = None
+            self.siglip_pad_token = None
+
+        # 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:
@@ -530,70 +635,168 @@ class NextDiT(nn.Module):
             imgs = torch.stack(imgs, dim=0)
         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, 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
-        orig_x = x
-
-        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)
+    def embed_cap(self, cap_feats=None, offset=0, bsz=1, device=None, dtype=None):
+        if cap_feats is not None:
+            cap_feats = self.cap_embedder(cap_feats)
+            cap_feats_len = cap_feats.shape[1]
+            if self.pad_tokens_multiple is not None:
+                cap_feats, _ = pad_zimage(cap_feats, self.cap_pad_token, self.pad_tokens_multiple)
+        else:
+            cap_feats_len = 0
+            cap_feats = self.cap_pad_token.to(device=device, dtype=dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1)
 
         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
+        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 + offset
+        embeds = (cap_feats,)
+        freqs_cis = (self.rope_embedder(cap_pos_ids).movedim(1, 2),)
+        return embeds, freqs_cis, cap_feats_len
+
+    def embed_all(self, x, cap_feats=None, siglip_feats=None, offset=0, omni=False, transformer_options={}):
+        bsz = 1
+        pH = pW = self.patch_size
+        device = x.device
+        embeds, freqs_cis, cap_feats_len = self.embed_cap(cap_feats, offset=offset, bsz=bsz, device=device, dtype=x.dtype)
+
+        if (not omni) or self.siglip_embedder is None:
+            cap_feats_len = embeds[0].shape[1] + offset
+            embeds += (None,)
+            freqs_cis += (None,)
+        else:
+            cap_feats_len += offset
+            if siglip_feats is not None:
+                b, h, w, c = siglip_feats.shape
+                siglip_feats = siglip_feats.permute(0, 3, 1, 2).reshape(b, h * w, c)
+                siglip_feats = self.siglip_embedder(siglip_feats)
+                siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device)
+                siglip_pos_ids[:, :, 0] = cap_feats_len + 2
+                siglip_pos_ids[:, :, 1] = (torch.linspace(0, h * 8 - 1, steps=h, dtype=torch.float32, device=device).floor()).view(-1, 1).repeat(1, w).flatten()
+                siglip_pos_ids[:, :, 2] = (torch.linspace(0, w * 8 - 1, steps=w, dtype=torch.float32, device=device).floor()).view(1, -1).repeat(h, 1).flatten()
+                if self.siglip_pad_token is not None:
+                    siglip_feats, pad_extra = pad_zimage(siglip_feats, self.siglip_pad_token, self.pad_tokens_multiple)  # TODO: double check
+                    siglip_pos_ids = torch.nn.functional.pad(siglip_pos_ids, (0, 0, 0, pad_extra))
+            else:
+                if self.siglip_pad_token is not None:
+                    siglip_feats = self.siglip_pad_token.to(device=device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1)
+                    siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device)
+
+            if siglip_feats is None:
+                embeds += (None,)
+                freqs_cis += (None,)
+            else:
+                embeds += (siglip_feats,)
+                freqs_cis += (self.rope_embedder(siglip_pos_ids).movedim(1, 2),)
 
         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))
-
-        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)
-
-            h_start = rope_options.get("shift_y", 0.0)
-            w_start = rope_options.get("shift_x", 0.0)
-
-        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()
-
+        x_pos_ids = pos_ids_x(cap_feats_len + 1, H // pH, W // pW, bsz, device, transformer_options=transformer_options)
         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, pad_extra = pad_zimage(x, self.x_pad_token, self.pad_tokens_multiple)
             x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra))
 
-        freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
+        embeds += (x,)
+        freqs_cis += (self.rope_embedder(x_pos_ids).movedim(1, 2),)
+        return embeds, freqs_cis, cap_feats_len + len(freqs_cis) - 1
+
+
+    def patchify_and_embed(
+        self, x: torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
+        bsz = x.shape[0]
+        cap_mask = None  # TODO?
+        main_siglip = None
+        orig_x = x
+
+        embeds = ([], [], [])
+        freqs_cis = ([], [], [])
+        leftover_cap = []
+
+        start_t = 0
+        omni = len(ref_latents) > 0
+        if omni:
+            for i, ref in enumerate(ref_latents):
+                if i < len(ref_contexts):
+                    ref_con = ref_contexts[i]
+                else:
+                    ref_con = None
+                if i < len(siglip_feats):
+                    sig_feat = siglip_feats[i]
+                else:
+                    sig_feat = None
+
+                out = self.embed_all(ref, ref_con, sig_feat, offset=start_t, omni=omni, transformer_options=transformer_options)
+                for i, e in enumerate(out[0]):
+                    if e is not None:
+                        embeds[i].append(comfy.utils.repeat_to_batch_size(e, bsz))
+                        freqs_cis[i].append(out[1][i])
+                start_t = out[2]
+            leftover_cap = ref_contexts[len(ref_latents):]
+
+        H, W = x.shape[-2], x.shape[-1]
+        img_sizes = [(H, W)] * bsz
+        out = self.embed_all(x, cap_feats, main_siglip, offset=start_t, omni=omni, transformer_options=transformer_options)
+        img_len = out[0][-1].shape[1]
+        cap_len = out[0][0].shape[1]
+        for i, e in enumerate(out[0]):
+            if e is not None:
+                e = comfy.utils.repeat_to_batch_size(e, bsz)
+                embeds[i].append(e)
+                freqs_cis[i].append(out[1][i])
+        start_t = out[2]
+
+        for cap in leftover_cap:
+            out = self.embed_cap(cap, offset=start_t, bsz=bsz, device=x.device, dtype=x.dtype)
+            cap_len += out[0][0].shape[1]
+            embeds[0].append(comfy.utils.repeat_to_batch_size(out[0][0], bsz))
+            freqs_cis[0].append(out[1][0])
+            start_t += out[2]
 
         patches = transformer_options.get("patches", {})
 
         # refine context
+        cap_feats = torch.cat(embeds[0], dim=1)
+        cap_freqs_cis = torch.cat(freqs_cis[0], dim=1)
         for layer in self.context_refiner:
-            cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
+            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options)
+
+        feats = (cap_feats,)
+        fc = (cap_freqs_cis,)
+
+        if omni and len(embeds[1]) > 0:
+            siglip_mask = None
+            siglip_feats_combined = torch.cat(embeds[1], dim=1)
+            siglip_feats_freqs_cis = torch.cat(freqs_cis[1], dim=1)
+            if self.siglip_refiner is not None:
+                for layer in self.siglip_refiner:
+                    siglip_feats_combined = layer(siglip_feats_combined, siglip_mask, siglip_feats_freqs_cis, transformer_options=transformer_options)
+            feats += (siglip_feats_combined,)
+            fc += (siglip_feats_freqs_cis,)
 
         padded_img_mask = None
+        x = torch.cat(embeds[-1], dim=1)
+        fc_x = torch.cat(freqs_cis[-1], dim=1)
+        if omni:
+            timestep_zero_index = [(x.shape[1] - img_len, x.shape[1])]
+        else:
+            timestep_zero_index = 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)
+            x = layer(x, padded_img_mask, fc_x, t, timestep_zero_index=timestep_zero_index, 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"})
+                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": fc_x, "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)
+        padded_full_embed = torch.cat(feats + (x,), dim=1)
+        if timestep_zero_index is not None:
+            ind = padded_full_embed.shape[1] - x.shape[1]
+            timestep_zero_index = [(ind + x.shape[1] - img_len, ind + x.shape[1])]
+            timestep_zero_index.append((feats[0].shape[1] - cap_len, feats[0].shape[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
+        l_effective_cap_len = [padded_full_embed.shape[1] - img_len] * bsz
+        return padded_full_embed, mask, img_sizes, l_effective_cap_len, torch.cat(fc + (fc_x,), dim=1), timestep_zero_index
 
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -603,7 +806,11 @@ 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, transformer_options={}, **kwargs):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}, **kwargs):
+        omni = len(ref_latents) > 0
+        if omni:
+            timesteps = torch.cat([timesteps * 0, timesteps], dim=0)
+
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
@@ -618,20 +825,18 @@ class NextDiT(nn.Module):
         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((1, self.clip_text_dim), device=x.device, dtype=x.dtype)
+                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)
-        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)
+        img, mask, img_size, cap_size, freqs_cis, timestep_zero_index = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, ref_latents=ref_latents, ref_contexts=ref_contexts, siglip_feats=siglip_feats, transformer_options=transformer_options)
         freqs_cis = freqs_cis.to(img.device)
 
         transformer_options["total_blocks"] = len(self.layers)
@@ -639,7 +844,7 @@ class NextDiT(nn.Module):
         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)
+            img = layer(img, mask, freqs_cis, adaln_input, timestep_zero_index=timestep_zero_index, 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})
@@ -648,8 +853,7 @@ class NextDiT(nn.Module):
                     if "txt" in out:
                         img[:, :cap_size[0]] = out["txt"]
 
-        img = self.final_layer(img, adaln_input)
+        img = self.final_layer(img, adaln_input, timestep_zero_index=timestep_zero_index)
         img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w]
-
         return -img
 

comfy/ldm/modules/attention.py

@@ -30,6 +30,13 @@ except ImportError as e:
             raise e
         exit(-1)
 
+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
@@ -563,6 +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=())
@@ -650,6 +744,8 @@ 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():

comfy/ldm/modules/diffusionmodules/model.py

@@ -14,10 +14,13 @@ if model_management.xformers_enabled_vae():
     import xformers.ops
 
 def torch_cat_if_needed(xl, dim):
+    xl = [x for x in xl if x is not None and x.shape[dim] > 0]
     if len(xl) > 1:
         return torch.cat(xl, dim)
-    else:
+    elif len(xl) == 1:
         return xl[0]
+    else:
+        return None
 
 def get_timestep_embedding(timesteps, embedding_dim):
     """
@@ -394,7 +397,8 @@ 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)
@@ -548,7 +552,8 @@ 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)

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/qwen_image/model.py

@@ -61,7 +61,7 @@ def apply_rotary_emb(x, freqs_cis):
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None):
+    def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None):
         super().__init__()
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
         self.timestep_embedder = TimestepEmbedding(
@@ -72,9 +72,19 @@ class QwenTimestepProjEmbeddings(nn.Module):
             operations=operations
         )
 
-    def forward(self, timestep, hidden_states):
+        self.use_additional_t_cond = use_additional_t_cond
+        if self.use_additional_t_cond:
+            self.addition_t_embedding = operations.Embedding(2, embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, timestep, hidden_states, addition_t_cond=None):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))
+
+        if self.use_additional_t_cond:
+            if addition_t_cond is None:
+                addition_t_cond = torch.zeros((timesteps_emb.shape[0]), device=timesteps_emb.device, dtype=torch.long)
+            timesteps_emb += self.addition_t_embedding(addition_t_cond, out_dtype=timesteps_emb.dtype)
+
         return timesteps_emb
 
 
@@ -160,8 +170,14 @@ class Attention(nn.Module):
         joint_query = apply_rope1(joint_query, image_rotary_emb)
         joint_key = apply_rope1(joint_key, image_rotary_emb)
 
+        if encoder_hidden_states_mask is not None:
+            attn_mask = torch.zeros((batch_size, 1, seq_txt + seq_img), dtype=hidden_states.dtype, device=hidden_states.device)
+            attn_mask[:, 0, :seq_txt] = encoder_hidden_states_mask
+        else:
+            attn_mask = None
+
         joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
-                                                         attention_mask, transformer_options=transformer_options,
+                                                         attn_mask, transformer_options=transformer_options,
                                                          skip_reshape=True)
 
         txt_attn_output = joint_hidden_states[:, :seq_txt, :]
@@ -320,11 +336,11 @@ class QwenImageTransformer2DModel(nn.Module):
         num_attention_heads: int = 24,
         joint_attention_dim: int = 3584,
         pooled_projection_dim: int = 768,
-        guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
         default_ref_method="index",
         image_model=None,
         final_layer=True,
+        use_additional_t_cond=False,
         dtype=None,
         device=None,
         operations=None,
@@ -342,6 +358,7 @@ class QwenImageTransformer2DModel(nn.Module):
         self.time_text_embed = QwenTimestepProjEmbeddings(
             embedding_dim=self.inner_dim,
             pooled_projection_dim=pooled_projection_dim,
+            use_additional_t_cond=use_additional_t_cond,
             dtype=dtype,
             device=device,
             operations=operations
@@ -375,27 +392,33 @@ class QwenImageTransformer2DModel(nn.Module):
         patch_size = self.patch_size
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
         orig_shape = hidden_states.shape
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
-        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
-        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-3], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6)
+        hidden_states = hidden_states.reshape(orig_shape[0], orig_shape[-3] * (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        t_len = t
         h_len = ((h + (patch_size // 2)) // patch_size)
         w_len = ((w + (patch_size // 2)) // patch_size)
 
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
-        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) - (h_len // 2)
-        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) - (w_len // 2)
-        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device)
 
-    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
+        if t_len > 1:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(1)
+        else:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + 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).unsqueeze(0) - (h_len // 2)
+        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).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "t h w c -> b (t h w) c", b=bs), orig_shape
+
+    def forward(self, x, timestep, context, attention_mask=None, ref_latents=None, additional_t_cond=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, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
+        ).execute(x, timestep, context, attention_mask, ref_latents, additional_t_cond, transformer_options, **kwargs)
 
     def _forward(
         self,
@@ -403,8 +426,8 @@ class QwenImageTransformer2DModel(nn.Module):
         timesteps,
         context,
         attention_mask=None,
-        guidance: torch.Tensor = None,
         ref_latents=None,
+        additional_t_cond=None,
         transformer_options={},
         control=None,
         **kwargs
@@ -413,6 +436,9 @@ class QwenImageTransformer2DModel(nn.Module):
         encoder_hidden_states = context
         encoder_hidden_states_mask = attention_mask
 
+        if encoder_hidden_states_mask is not None and not torch.is_floating_point(encoder_hidden_states_mask):
+            encoder_hidden_states_mask = (encoder_hidden_states_mask - 1).to(x.dtype) * torch.finfo(x.dtype).max
+
         hidden_states, img_ids, orig_shape = self.process_img(x)
         num_embeds = hidden_states.shape[1]
 
@@ -423,12 +449,17 @@ class QwenImageTransformer2DModel(nn.Module):
             index = 0
             ref_method = kwargs.get("ref_latents_method", self.default_ref_method)
             index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero")
+            negative_ref_method = ref_method == "negative_index"
             timestep_zero = ref_method == "index_timestep_zero"
             for ref in ref_latents:
                 if index_ref_method:
                     index += 1
                     h_offset = 0
                     w_offset = 0
+                elif negative_ref_method:
+                    index -= 1
+                    h_offset = 0
+                    w_offset = 0
                 else:
                     index = 1
                     h_offset = 0
@@ -458,14 +489,7 @@ class QwenImageTransformer2DModel(nn.Module):
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
         encoder_hidden_states = self.txt_in(encoder_hidden_states)
 
-        if guidance is not None:
-            guidance = guidance * 1000
-
-        temb = (
-            self.time_text_embed(timestep, hidden_states)
-            if guidance is None
-            else self.time_text_embed(timestep, guidance, hidden_states)
-        )
+        temb = self.time_text_embed(timestep, hidden_states, additional_t_cond)
 
         patches_replace = transformer_options.get("patches_replace", {})
         patches = transformer_options.get("patches", {})
@@ -513,6 +537,6 @@ class QwenImageTransformer2DModel(nn.Module):
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
-        hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-3], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states.permute(0, 4, 1, 2, 5, 3, 6)
         return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/util.py

@@ -71,7 +71,7 @@ def count_params(model, verbose=False):
 
 
 def instantiate_from_config(config):
-    if not "target" in config:
+    if "target" not in config:
         if config == '__is_first_stage__':
             return None
         elif config == "__is_unconditional__":

comfy/ldm/wan/model.py

@@ -62,6 +62,8 @@ class WanSelfAttention(nn.Module):
             x(Tensor): Shape [B, L, num_heads, C / num_heads]
             freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
         """
+        patches = transformer_options.get("patches", {})
+
         b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
 
         def qkv_fn_q(x):
@@ -86,6 +88,10 @@ class WanSelfAttention(nn.Module):
             transformer_options=transformer_options,
         )
 
+        if "attn1_patch" in patches:
+            for p in patches["attn1_patch"]:
+                x = p({"x": x, "q": q, "k": k, "transformer_options": transformer_options})
+
         x = self.o(x)
         return x
 
@@ -225,6 +231,8 @@ class WanAttentionBlock(nn.Module):
         """
         # assert e.dtype == torch.float32
 
+        patches = transformer_options.get("patches", {})
+
         if e.ndim < 4:
             e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
         else:
@@ -242,6 +250,11 @@ class WanAttentionBlock(nn.Module):
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
+
+        if "attn2_patch" in patches:
+            for p in patches["attn2_patch"]:
+                x = p({"x": x, "transformer_options": transformer_options})
+
         y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
         x = torch.addcmul(x, y, repeat_e(e[5], x))
         return x
@@ -488,7 +501,7 @@ class WanModel(torch.nn.Module):
         self.blocks = nn.ModuleList([
             wan_attn_block_class(cross_attn_type, dim, ffn_dim, num_heads,
                                  window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
-            for _ in range(num_layers)
+            for i in range(num_layers)
         ])
 
         # head
@@ -541,6 +554,7 @@ class WanModel(torch.nn.Module):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
         x = x.flatten(2).transpose(1, 2)
 
         # time embeddings
@@ -738,6 +752,7 @@ class VaceWanModel(WanModel):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
         x = x.flatten(2).transpose(1, 2)
 
         # time embeddings

comfy/ldm/wan/model_multitalk.py

@@ -0,0 +1,500 @@
+import torch
+from einops import rearrange, repeat
+import comfy
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def calculate_x_ref_attn_map(visual_q, ref_k, ref_target_masks, split_num=8):
+    scale = 1.0 / visual_q.shape[-1] ** 0.5
+    visual_q = visual_q.transpose(1, 2) * scale
+
+    B, H, x_seqlens, K = visual_q.shape
+
+    x_ref_attn_maps = []
+    for class_idx, ref_target_mask in enumerate(ref_target_masks):
+        ref_target_mask = ref_target_mask.view(1, 1, 1, -1)
+
+        x_ref_attnmap = torch.zeros(B, H, x_seqlens, device=visual_q.device, dtype=visual_q.dtype)
+        chunk_size = min(max(x_seqlens // split_num, 1), x_seqlens)
+
+        for i in range(0, x_seqlens, chunk_size):
+            end_i = min(i + chunk_size, x_seqlens)
+
+            attn_chunk = visual_q[:, :, i:end_i] @ ref_k.permute(0, 2, 3, 1)  # B, H, chunk, ref_seqlens
+
+            # Apply softmax
+            attn_max = attn_chunk.max(dim=-1, keepdim=True).values
+            attn_chunk = (attn_chunk - attn_max).exp()
+            attn_sum = attn_chunk.sum(dim=-1, keepdim=True)
+            attn_chunk = attn_chunk / (attn_sum + 1e-8)
+
+            # Apply mask and sum
+            masked_attn = attn_chunk * ref_target_mask
+            x_ref_attnmap[:, :, i:end_i] = masked_attn.sum(-1) / (ref_target_mask.sum() + 1e-8)
+
+            del attn_chunk, masked_attn
+
+        # Average across heads
+        x_ref_attnmap = x_ref_attnmap.mean(dim=1)  # B, x_seqlens
+        x_ref_attn_maps.append(x_ref_attnmap)
+
+    del visual_q, ref_k
+
+    return torch.cat(x_ref_attn_maps, dim=0)
+
+def get_attn_map_with_target(visual_q, ref_k, shape, ref_target_masks=None, split_num=2):
+    """Args:
+        query (torch.tensor): B M H K
+        key (torch.tensor): B M H K
+        shape (tuple): (N_t, N_h, N_w)
+        ref_target_masks: [B, N_h * N_w]
+    """
+
+    N_t, N_h, N_w = shape
+
+    x_seqlens = N_h * N_w
+    ref_k     = ref_k[:, :x_seqlens]
+    _, seq_lens, heads, _ = visual_q.shape
+    class_num, _ = ref_target_masks.shape
+    x_ref_attn_maps = torch.zeros(class_num, seq_lens).to(visual_q)
+
+    split_chunk = heads // split_num
+
+    for i in range(split_num):
+        x_ref_attn_maps_perhead = calculate_x_ref_attn_map(
+            visual_q[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_k[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_target_masks
+            )
+        x_ref_attn_maps += x_ref_attn_maps_perhead
+
+    return x_ref_attn_maps / split_num
+
+
+def normalize_and_scale(column, source_range, target_range, epsilon=1e-8):
+    source_min, source_max = source_range
+    new_min, new_max = target_range
+    normalized = (column - source_min) / (source_max - source_min + epsilon)
+    scaled = normalized * (new_max - new_min) + new_min
+    return scaled
+
+
+def rotate_half(x):
+    x = rearrange(x, "... (d r) -> ... d r", r=2)
+    x1, x2 = x.unbind(dim=-1)
+    x = torch.stack((-x2, x1), dim=-1)
+    return rearrange(x, "... d r -> ... (d r)")
+
+
+def get_audio_embeds(encoded_audio, audio_start, audio_end):
+    audio_embs = []
+    human_num = len(encoded_audio)
+    audio_frames = encoded_audio[0].shape[0]
+
+    indices = (torch.arange(4 + 1) - 2) * 1
+
+    for human_idx in range(human_num):
+        if audio_end > audio_frames: # in case of not enough audio for current window, pad with first audio frame as that's most likely silence
+            pad_len = audio_end - audio_frames
+            pad_shape = list(encoded_audio[human_idx].shape)
+            pad_shape[0] = pad_len
+            pad_tensor = encoded_audio[human_idx][:1].repeat(pad_len, *([1] * (encoded_audio[human_idx].dim() - 1)))
+            encoded_audio_in = torch.cat([encoded_audio[human_idx], pad_tensor], dim=0)
+        else:
+            encoded_audio_in = encoded_audio[human_idx]
+        center_indices = torch.arange(audio_start, audio_end, 1).unsqueeze(1) + indices.unsqueeze(0)
+        center_indices = torch.clamp(center_indices, min=0, max=encoded_audio_in.shape[0] - 1)
+        audio_emb = encoded_audio_in[center_indices].unsqueeze(0)
+        audio_embs.append(audio_emb)
+
+    return torch.cat(audio_embs, dim=0)
+
+
+def project_audio_features(audio_proj, encoded_audio, audio_start, audio_end):
+    audio_embs = get_audio_embeds(encoded_audio, audio_start, audio_end)
+
+    first_frame_audio_emb_s = audio_embs[:, :1, ...]
+    latter_frame_audio_emb = audio_embs[:, 1:, ...]
+    latter_frame_audio_emb = rearrange(latter_frame_audio_emb, "b (n_t n) w s c -> b n_t n w s c", n=4)
+
+    middle_index = audio_proj.seq_len // 2
+
+    latter_first_frame_audio_emb = latter_frame_audio_emb[:, :, :1, :middle_index+1, ...]
+    latter_first_frame_audio_emb = rearrange(latter_first_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_last_frame_audio_emb = latter_frame_audio_emb[:, :, -1:, middle_index:, ...]
+    latter_last_frame_audio_emb = rearrange(latter_last_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_middle_frame_audio_emb = latter_frame_audio_emb[:, :, 1:-1, middle_index:middle_index+1, ...]
+    latter_middle_frame_audio_emb = rearrange(latter_middle_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_frame_audio_emb_s = torch.cat([latter_first_frame_audio_emb, latter_middle_frame_audio_emb, latter_last_frame_audio_emb], dim=2)
+
+    audio_emb = audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s)
+    audio_emb = torch.cat(audio_emb.split(1), dim=2)
+
+    return audio_emb
+
+
+class RotaryPositionalEmbedding1D(torch.nn.Module):
+    def __init__(self,
+                 head_dim,
+                 ):
+        super().__init__()
+        self.head_dim = head_dim
+        self.base = 10000
+
+    def precompute_freqs_cis_1d(self, pos_indices):
+        freqs = 1.0 / (self.base ** (torch.arange(0, self.head_dim, 2)[: (self.head_dim // 2)].float() / self.head_dim))
+        freqs = freqs.to(pos_indices.device)
+        freqs = torch.einsum("..., f -> ... f", pos_indices.float(), freqs)
+        freqs = repeat(freqs, "... n -> ... (n r)", r=2)
+        return freqs
+
+    def forward(self, x, pos_indices):
+        freqs_cis = self.precompute_freqs_cis_1d(pos_indices)
+
+        x_ = x.float()
+
+        freqs_cis = freqs_cis.float().to(x.device)
+        cos, sin = freqs_cis.cos(), freqs_cis.sin()
+        cos, sin = rearrange(cos, 'n d -> 1 1 n d'), rearrange(sin, 'n d -> 1 1 n d')
+        x_ = (x_ * cos) + (rotate_half(x_) * sin)
+
+        return x_.type_as(x)
+
+class SingleStreamAttention(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.encoder_hidden_states_dim = encoder_hidden_states_dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q_linear = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.kv_linear = operations.Linear(encoder_hidden_states_dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, encoder_hidden_states: torch.Tensor, shape=None) -> torch.Tensor:
+        N_t, N_h, N_w = shape
+
+        expected_tokens = N_t * N_h * N_w
+        actual_tokens = x.shape[1]
+        x_extra = None
+
+        if actual_tokens != expected_tokens:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+
+        B = x.shape[0]
+        S = N_h * N_w
+        x = x.view(B * N_t, S, self.dim)
+
+        # get q for hidden_state
+        q = self.q_linear(x).view(B * N_t, S, self.num_heads, self.head_dim)
+
+        # get kv from encoder_hidden_states # shape: (B, N, num_heads, head_dim)
+        kv = self.kv_linear(encoder_hidden_states)
+        encoder_k, encoder_v = kv.view(B * N_t, encoder_hidden_states.shape[1], 2, self.num_heads, self.head_dim).unbind(2)
+
+        #print("q.shape", q.shape) #torch.Size([21, 1024, 40, 128])
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # linear transform
+        x = self.proj(x.reshape(B * N_t, S, self.dim))
+        x = x.view(B, N_t * S, self.dim)
+
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+class SingleStreamMultiAttention(SingleStreamAttention):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        class_range: int = 24,
+        class_interval: int = 4,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__(
+            dim=dim,
+            encoder_hidden_states_dim=encoder_hidden_states_dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            device=device,
+            dtype=dtype,
+            operations=operations
+        )
+
+        # Rotary-embedding layout parameters
+        self.class_interval = class_interval
+        self.class_range = class_range
+        self.max_humans = self.class_range // self.class_interval
+
+        # Constant bucket used for background tokens
+        self.rope_bak = int(self.class_range // 2)
+
+        self.rope_1d = RotaryPositionalEmbedding1D(self.head_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        shape=None,
+        x_ref_attn_map=None
+    ) -> torch.Tensor:
+        encoder_hidden_states = encoder_hidden_states.squeeze(0).to(x.device)
+        human_num = x_ref_attn_map.shape[0] if x_ref_attn_map is not None else 1
+        # Single-speaker fall-through
+        if human_num <= 1:
+            return super().forward(x, encoder_hidden_states, shape)
+
+        N_t, N_h, N_w = shape
+
+        x_extra = None
+        if x.shape[0] * N_t != encoder_hidden_states.shape[0]:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+        x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
+
+        # Query projection
+        B, N, C = x.shape
+        q = self.q_linear(x)
+        q = q.view(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+
+        # Use `class_range` logic for 2 speakers
+        rope_h1 = (0, self.class_interval)
+        rope_h2 = (self.class_range - self.class_interval, self.class_range)
+        rope_bak = int(self.class_range // 2)
+
+        # Normalize and scale attention maps for each speaker
+        max_values = x_ref_attn_map.max(1).values[:, None, None]
+        min_values = x_ref_attn_map.min(1).values[:, None, None]
+        max_min_values = torch.cat([max_values, min_values], dim=2)
+
+        human1_max_value, human1_min_value = max_min_values[0, :, 0].max(), max_min_values[0, :, 1].min()
+        human2_max_value, human2_min_value = max_min_values[1, :, 0].max(), max_min_values[1, :, 1].min()
+
+        human1 = normalize_and_scale(x_ref_attn_map[0], (human1_min_value, human1_max_value), rope_h1)
+        human2 = normalize_and_scale(x_ref_attn_map[1], (human2_min_value, human2_max_value), rope_h2)
+        back = torch.full((x_ref_attn_map.size(1),), rope_bak, dtype=human1.dtype, device=human1.device)
+
+        # Token-wise speaker dominance
+        max_indices = x_ref_attn_map.argmax(dim=0)
+        normalized_map = torch.stack([human1, human2, back], dim=1)
+        normalized_pos = normalized_map[torch.arange(x_ref_attn_map.size(1)), max_indices]
+
+        # Apply rotary to Q
+        q = rearrange(q, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        q = self.rope_1d(q, normalized_pos)
+        q = rearrange(q, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Keys / Values
+        _, N_a, _ = encoder_hidden_states.shape
+        encoder_kv = self.kv_linear(encoder_hidden_states)
+        encoder_kv = encoder_kv.view(B, N_a, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+        encoder_k, encoder_v = encoder_kv.unbind(0)
+
+        # Rotary for keys – assign centre of each speaker bucket to its context tokens
+        per_frame = torch.zeros(N_a, dtype=encoder_k.dtype, device=encoder_k.device)
+        per_frame[: per_frame.size(0) // 2] = (rope_h1[0] + rope_h1[1]) / 2
+        per_frame[per_frame.size(0) // 2 :] = (rope_h2[0] + rope_h2[1]) / 2
+        encoder_pos = torch.cat([per_frame] * N_t, dim=0)
+
+        encoder_k = rearrange(encoder_k, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        encoder_k = self.rope_1d(encoder_k, encoder_pos)
+        encoder_k = rearrange(encoder_k, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Final attention
+        q = rearrange(q, "B H M K -> B M H K")
+        encoder_k = rearrange(encoder_k, "B H M K -> B M H K")
+        encoder_v = rearrange(encoder_v, "B H M K -> B M H K")
+
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # Linear projection
+        x = x.reshape(B, N, C)
+        x = self.proj(x)
+
+        # Restore original layout
+        x = rearrange(x, "(B N_t) S C -> B (N_t S) C", N_t=N_t)
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+
+class MultiTalkAudioProjModel(torch.nn.Module):
+    def __init__(
+        self,
+        seq_len: int = 5,
+        seq_len_vf: int = 12,
+        blocks: int = 12,
+        channels: int = 768,
+        intermediate_dim: int = 512,
+        out_dim: int = 768,
+        context_tokens: int = 32,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = seq_len * blocks * channels
+        self.input_dim_vf = seq_len_vf * blocks * channels
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.out_dim = out_dim
+
+        # define multiple linear layers
+        self.proj1 = operations.Linear(self.input_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj1_vf = operations.Linear(self.input_dim_vf, intermediate_dim, device=device, dtype=dtype)
+        self.proj2 = operations.Linear(intermediate_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj3 = operations.Linear(intermediate_dim, context_tokens * out_dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(out_dim, device=device, dtype=dtype)
+
+    def forward(self, audio_embeds, audio_embeds_vf):
+        video_length = audio_embeds.shape[1] + audio_embeds_vf.shape[1]
+        B, _, _, S, C = audio_embeds.shape
+
+        # process audio of first frame
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+
+        # process audio of latter frame
+        audio_embeds_vf = rearrange(audio_embeds_vf, "bz f w b c -> (bz f) w b c")
+        batch_size_vf, window_size_vf, blocks_vf, channels_vf = audio_embeds_vf.shape
+        audio_embeds_vf = audio_embeds_vf.view(batch_size_vf, window_size_vf * blocks_vf * channels_vf)
+
+        # first projection
+        audio_embeds = torch.relu(self.proj1(audio_embeds))
+        audio_embeds_vf = torch.relu(self.proj1_vf(audio_embeds_vf))
+        audio_embeds = rearrange(audio_embeds, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_vf = rearrange(audio_embeds_vf, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_c = torch.concat([audio_embeds, audio_embeds_vf], dim=1)
+        batch_size_c, N_t, C_a = audio_embeds_c.shape
+        audio_embeds_c = audio_embeds_c.view(batch_size_c*N_t, C_a)
+
+        # second projection
+        audio_embeds_c = torch.relu(self.proj2(audio_embeds_c))
+
+        context_tokens = self.proj3(audio_embeds_c).reshape(batch_size_c*N_t, self.context_tokens, self.out_dim)
+
+        # normalization and reshape
+        context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length)
+
+        return context_tokens
+
+
+class WanMultiTalkAttentionBlock(torch.nn.Module):
+    def __init__(self, in_dim=5120, out_dim=768, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.audio_cross_attn = SingleStreamMultiAttention(in_dim, out_dim, num_heads=40, qkv_bias=True, device=device, dtype=dtype, operations=operations)
+        self.norm_x = operations.LayerNorm(in_dim, device=device, dtype=dtype, elementwise_affine=True)
+
+
+class MultiTalkGetAttnMapPatch:
+    def __init__(self, ref_target_masks=None):
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        x = kwargs["x"]
+
+        if self.ref_target_masks is not None:
+            x_ref_attn_map = get_attn_map_with_target(kwargs["q"], kwargs["k"], transformer_options["grid_sizes"], ref_target_masks=self.ref_target_masks.to(x.device))
+            transformer_options["x_ref_attn_map"] = x_ref_attn_map
+        return x
+
+
+class MultiTalkCrossAttnPatch:
+    def __init__(self, model_patch, audio_scale=1.0, ref_target_masks=None):
+        self.model_patch = model_patch
+        self.audio_scale = audio_scale
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        block_idx = transformer_options.get("block_index", None)
+        x = kwargs["x"]
+        if block_idx is None:
+            return torch.zeros_like(x)
+
+        audio_embeds = transformer_options.get("audio_embeds")
+        x_ref_attn_map = transformer_options.pop("x_ref_attn_map", None)
+
+        norm_x = self.model_patch.model.blocks[block_idx].norm_x(x)
+        x_audio = self.model_patch.model.blocks[block_idx].audio_cross_attn(
+            norm_x, audio_embeds.to(x.dtype),
+            shape=transformer_options["grid_sizes"],
+            x_ref_attn_map=x_ref_attn_map
+        )
+        x = x + x_audio * self.audio_scale
+        return x
+
+    def models(self):
+        return [self.model_patch]
+
+class MultiTalkApplyModelWrapper:
+    def __init__(self, init_latents):
+        self.init_latents = init_latents
+
+    def __call__(self, executor, x, *args, **kwargs):
+        x[:, :, :self.init_latents.shape[2]] = self.init_latents.to(x)
+        samples = executor(x, *args, **kwargs)
+        return samples
+
+
+class InfiniteTalkOuterSampleWrapper:
+    def __init__(self, motion_frames_latent, model_patch, is_extend=False):
+        self.motion_frames_latent = motion_frames_latent
+        self.model_patch = model_patch
+        self.is_extend = is_extend
+
+    def __call__(self, executor, *args, **kwargs):
+        model_patcher = executor.class_obj.model_patcher
+        model_options = executor.class_obj.model_options
+        process_latent_in = model_patcher.model.process_latent_in
+
+        # for InfiniteTalk, model input first latent(s) need to always be replaced on every step
+        if self.motion_frames_latent is not None:
+            wrappers = model_options["transformer_options"]["wrappers"]
+            w = wrappers.setdefault(comfy.patcher_extension.WrappersMP.APPLY_MODEL, {})
+            w["MultiTalk_apply_model"] = [MultiTalkApplyModelWrapper(process_latent_in(self.motion_frames_latent))]
+
+        # run the sampling process
+        result = executor(*args, **kwargs)
+
+        # insert motion frames before decoding
+        if self.is_extend:
+            overlap = self.motion_frames_latent.shape[2]
+            result = torch.cat([self.motion_frames_latent.to(result), result[:, :, overlap:]], dim=2)
+
+        return result
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            if self.motion_frames_latent is not None:
+                self.motion_frames_latent = self.motion_frames_latent.to(device_or_dtype)
+        return self

comfy/ldm/wan/vae.py

@@ -227,6 +227,7 @@ class Encoder3d(nn.Module):
     def __init__(self,
                  dim=128,
                  z_dim=4,
+                 input_channels=3,
                  dim_mult=[1, 2, 4, 4],
                  num_res_blocks=2,
                  attn_scales=[],
@@ -245,7 +246,7 @@ class Encoder3d(nn.Module):
         scale = 1.0
 
         # init block
-        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+        self.conv1 = CausalConv3d(input_channels, dims[0], 3, padding=1)
 
         # downsample blocks
         downsamples = []
@@ -331,6 +332,7 @@ class Decoder3d(nn.Module):
     def __init__(self,
                  dim=128,
                  z_dim=4,
+                 output_channels=3,
                  dim_mult=[1, 2, 4, 4],
                  num_res_blocks=2,
                  attn_scales=[],
@@ -378,7 +380,7 @@ class Decoder3d(nn.Module):
         # output blocks
         self.head = nn.Sequential(
             RMS_norm(out_dim, images=False), nn.SiLU(),
-            CausalConv3d(out_dim, 3, 3, padding=1))
+            CausalConv3d(out_dim, output_channels, 3, padding=1))
 
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         ## conv1
@@ -449,6 +451,7 @@ class WanVAE(nn.Module):
                  num_res_blocks=2,
                  attn_scales=[],
                  temperal_downsample=[True, True, False],
+                 image_channels=3,
                  dropout=0.0):
         super().__init__()
         self.dim = dim
@@ -460,11 +463,11 @@ class WanVAE(nn.Module):
         self.temperal_upsample = temperal_downsample[::-1]
 
         # modules
-        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+        self.encoder = Encoder3d(dim, z_dim * 2, image_channels, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_downsample, dropout)
         self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
         self.conv2 = CausalConv3d(z_dim, z_dim, 1)
-        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+        self.decoder = Decoder3d(dim, z_dim, image_channels, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
     def encode(self, x):

comfy/lora.py

@@ -322,6 +322,7 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["diffusion_model.{}".format(key_lora)] = to
                 key_map["transformer.{}".format(key_lora)] = to
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to
+                key_map[key_lora] = to
 
     if isinstance(model, comfy.model_base.Kandinsky5):
         for k in sdk:

comfy/model_base.py

@@ -20,6 +20,7 @@ import comfy.ldm.hunyuan3dv2_1
 import comfy.ldm.hunyuan3dv2_1.hunyuandit
 import torch
 import logging
+import comfy.ldm.lightricks.av_model
 from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
 from comfy.ldm.cascade.stage_c import StageC
 from comfy.ldm.cascade.stage_b import StageB
@@ -48,6 +49,7 @@ import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
+import comfy.ldm.anima.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -946,7 +948,7 @@ class GenmoMochi(BaseModel):
 
 class LTXV(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -977,6 +979,60 @@ class LTXV(BaseModel):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return latent_image
 
+class LTXAV(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.av_model.LTXAVModel) #TODO
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25))
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+
+        audio_denoise_mask = None
+        if denoise_mask is not None and "latent_shapes" in kwargs:
+            denoise_mask = utils.unpack_latents(denoise_mask, kwargs["latent_shapes"])
+            if len(denoise_mask) > 1:
+                audio_denoise_mask = denoise_mask[1]
+            denoise_mask = denoise_mask[0]
+
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+
+        if audio_denoise_mask is not None:
+            out["audio_denoise_mask"] = comfy.conds.CONDRegular(audio_denoise_mask)
+
+        keyframe_idxs = kwargs.get("keyframe_idxs", None)
+        if keyframe_idxs is not None:
+            out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs)
+
+        latent_shapes = kwargs.get("latent_shapes", None)
+        if latent_shapes is not None:
+            out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes)
+
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs):
+        v_timestep = timestep
+        a_timestep = timestep
+
+        if denoise_mask is not None:
+            v_timestep = self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0]
+        if audio_denoise_mask is not None:
+            a_timestep = self.diffusion_model.a_patchifier.patchify(((audio_denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (audio_denoise_mask.ndim - 1)))[:, :1, :, :1])[0]
+
+        return v_timestep, a_timestep
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
 class HunyuanVideo(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo)
@@ -1092,9 +1148,31 @@ class CosmosPredict2(BaseModel):
         sigma = (sigma / (sigma + 1))
         return latent_image / (1.0 - sigma)
 
+class Anima(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.anima.model.Anima)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        t5xxl_ids = kwargs.get("t5xxl_ids", None)
+        t5xxl_weights = kwargs.get("t5xxl_weights", None)
+        device = kwargs["device"]
+        if cross_attn is not None:
+            if t5xxl_ids is not None:
+                cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.unsqueeze(0).to(device=device))
+                if t5xxl_weights is not None:
+                    cross_attn *= t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+
+                if cross_attn.shape[1] < 512:
+                    cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, 0, 512 - cross_attn.shape[1]))
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
 class Lumina2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1110,10 +1188,39 @@ class Lumina2(BaseModel):
             if 'num_tokens' not in out:
                 out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1])
 
-        clip_text_pooled = kwargs["pooled_output"]  # Newbie
+        clip_text_pooled = kwargs.get("pooled_output", None)  # NewBie
         if clip_text_pooled is not None:
             out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled)
 
+        clip_vision_outputs = kwargs.get("clip_vision_outputs", list(map(lambda a: a.get("clip_vision_output"), kwargs.get("unclip_conditioning", [{}]))))  # Z Image omni
+        if clip_vision_outputs is not None and len(clip_vision_outputs) > 0:
+            sigfeats = []
+            for clip_vision_output in clip_vision_outputs:
+                if clip_vision_output is not None:
+                    image_size = clip_vision_output.image_sizes[0]
+                    shape = clip_vision_output.last_hidden_state.shape
+                    sigfeats.append(clip_vision_output.last_hidden_state.reshape(shape[0], image_size[1] // 16, image_size[2] // 16, shape[-1]))
+            if len(sigfeats) > 0:
+                out['siglip_feats'] = comfy.conds.CONDList(sigfeats)
+
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+
+        ref_contexts = kwargs.get("reference_latents_text_embeds", None)
+        if ref_contexts is not None:
+            out['ref_contexts'] = comfy.conds.CONDList(ref_contexts)
+
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))])
         return out
 
 class WAN21(BaseModel):
@@ -1471,6 +1578,9 @@ class QwenImage(BaseModel):
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)

comfy/model_detection.py

@@ -237,6 +237,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         else:
             dit_config["vec_in_dim"] = None
 
+        dit_config["num_heads"] = dit_config["hidden_size"] // sum(dit_config["axes_dim"])
+
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
         if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
@@ -251,7 +253,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["image_model"] = "chroma_radiance"
                 dit_config["in_channels"] = 3
                 dit_config["out_channels"] = 3
-                dit_config["patch_size"] = 16
+                dit_config["patch_size"] = state_dict.get('{}img_in_patch.weight'.format(key_prefix)).size(dim=-1)
                 dit_config["nerf_hidden_size"] = 64
                 dit_config["nerf_mlp_ratio"] = 4
                 dit_config["nerf_depth"] = 4
@@ -305,7 +307,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
     if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv
         dit_config = {}
-        dit_config["image_model"] = "ltxv"
+        dit_config["image_model"] = "ltxav" if f'{key_prefix}audio_adaln_single.linear.weight' in state_dict_keys else "ltxv"
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape
         dit_config["attention_head_dim"] = shape[0] // 32
@@ -430,8 +432,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["rope_theta"] = 10000.0
             dit_config["ffn_dim_multiplier"] = 4.0
             ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None)
-            if ctd_weight is not None:
+            if ctd_weight is not None:  # NewBie
                 dit_config["clip_text_dim"] = ctd_weight.shape[0]
+                # NewBie also sets axes_lens = [1024, 512, 512] but it's not used in ComfyUI
         elif dit_config["dim"] == 3840:  # Z image
             dit_config["n_heads"] = 30
             dit_config["n_kv_heads"] = 30
@@ -443,6 +446,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["time_scale"] = 1000.0
             if '{}cap_pad_token'.format(key_prefix) in state_dict_keys:
                 dit_config["pad_tokens_multiple"] = 32
+            sig_weight = state_dict.get('{}siglip_embedder.0.weight'.format(key_prefix), None)
+            if sig_weight is not None:
+                dit_config["siglip_feat_dim"] = sig_weight.shape[0]
 
         return dit_config
 
@@ -544,6 +550,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     if '{}blocks.0.mlp.layer1.weight'.format(key_prefix) in state_dict_keys:  # Cosmos predict2
         dit_config = {}
         dit_config["image_model"] = "cosmos_predict2"
+        if "{}llm_adapter.blocks.0.cross_attn.q_proj.weight".format(key_prefix) in state_dict_keys:
+            dit_config["image_model"] = "anima"
         dit_config["max_img_h"] = 240
         dit_config["max_img_w"] = 240
         dit_config["max_frames"] = 128
@@ -620,6 +628,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys:  # 2511
             dit_config["default_ref_method"] = "index_timestep_zero"
+        if "{}time_text_embed.addition_t_embedding.weight".format(key_prefix) in state_dict_keys:  # Layered
+            dit_config["use_additional_t_cond"] = True
+            dit_config["default_ref_method"] = "negative_index"
         return dit_config
 
     if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5

comfy/model_management.py

@@ -22,10 +22,10 @@ from enum import Enum
 from comfy.cli_args import args, PerformanceFeature
 import torch
 import sys
-import importlib
 import platform
 import weakref
 import gc
+import os
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -333,28 +333,42 @@ except:
 SUPPORT_FP8_OPS = args.supports_fp8_compute
 
 AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
+AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
 
 try:
     if is_amd():
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
-            torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
-            logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
+            if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
+                torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
+                logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
 
         try:
             rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
         except:
             rocm_version = (6, -1)
 
+        def aotriton_supported(gpu_arch):
+            path = torch.__path__[0]
+            path = os.path.join(os.path.join(path, "lib"), "aotriton.images")
+            gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path))))
+            if gpu_arch in gfx:
+                return True
+            if "{}x".format(gpu_arch[:-1]) in gfx:
+                return True
+            if "{}xx".format(gpu_arch[:-2]) in gfx:
+                return True
+            return False
+
         logging.info("AMD arch: {}".format(arch))
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
-            if importlib.util.find_spec('triton') is not None:  # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not.
+            if aotriton_supported(arch):  # AMD efficient attention implementation depends on aotriton.
                 if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
                     if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True
                 if rocm_version >= (7, 0):
-                   if any((a in arch) for a in ["gfx1201"]):
+                   if any((a in arch) for a in ["gfx1200", "gfx1201"]):
                        ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]):  # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0
@@ -453,7 +467,7 @@ def module_size(module):
     sd = module.state_dict()
     for k in sd:
         t = sd[k]
-        module_mem += t.nelement() * t.element_size()
+        module_mem += t.nbytes
     return module_mem
 
 class LoadedModel:
@@ -1016,8 +1030,8 @@ NUM_STREAMS = 0
 if args.async_offload is not None:
     NUM_STREAMS = args.async_offload
 else:
-    #  Enable by default on Nvidia
-    if is_nvidia():
+    #  Enable by default on Nvidia and AMD
+    if is_nvidia() or is_amd():
         NUM_STREAMS = 2
 
 if args.disable_async_offload:
@@ -1123,6 +1137,16 @@ if not args.disable_pinned_memory:
 
 PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
 
+def discard_cuda_async_error():
+    try:
+        a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+        b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+        _ = a + b
+        torch.cuda.synchronize()
+    except torch.AcceleratorError:
+        #Dump it! We already know about it from the synchronous return
+        pass
+
 def pin_memory(tensor):
     global TOTAL_PINNED_MEMORY
     if MAX_PINNED_MEMORY <= 0:
@@ -1143,7 +1167,7 @@ def pin_memory(tensor):
     if not tensor.is_contiguous():
         return False
 
-    size = tensor.numel() * tensor.element_size()
+    size = tensor.nbytes
     if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY:
         return False
 
@@ -1155,6 +1179,9 @@ def pin_memory(tensor):
         PINNED_MEMORY[ptr] = size
         TOTAL_PINNED_MEMORY += size
         return True
+    else:
+        logging.warning("Pin error.")
+        discard_cuda_async_error()
 
     return False
 
@@ -1167,7 +1194,7 @@ def unpin_memory(tensor):
         return False
 
     ptr = tensor.data_ptr()
-    size = tensor.numel() * tensor.element_size()
+    size = tensor.nbytes
 
     size_stored = PINNED_MEMORY.get(ptr, None)
     if size_stored is None:
@@ -1183,6 +1210,9 @@ def unpin_memory(tensor):
         if len(PINNED_MEMORY) == 0:
             TOTAL_PINNED_MEMORY = 0
         return True
+    else:
+        logging.warning("Unpin error.")
+        discard_cuda_async_error()
 
     return False
 
@@ -1485,6 +1515,16 @@ def supports_fp8_compute(device=None):
 
     return True
 
+def supports_nvfp4_compute(device=None):
+    if not is_nvidia():
+        return False
+
+    props = torch.cuda.get_device_properties(device)
+    if props.major < 10:
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):
@@ -1523,6 +1563,10 @@ def soft_empty_cache(force=False):
 def unload_all_models():
     free_memory(1e30, get_torch_device())
 
+def debug_memory_summary():
+    if is_amd() or is_nvidia():
+        return torch.cuda.memory.memory_summary()
+    return ""
 
 #TODO: might be cleaner to put this somewhere else
 import threading

comfy/model_patcher.py

@@ -718,6 +718,7 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
+                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -790,11 +791,12 @@ class ModelPatcher:
                 for param in params:
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
+            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)

comfy/ops.py

@@ -79,7 +79,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     if input is not None:
         if dtype is None:
             if isinstance(input, QuantizedTensor):
-                dtype = input._layout_params["orig_dtype"]
+                dtype = input.params.orig_dtype
             else:
                 dtype = input.dtype
         if bias_dtype is None:
@@ -412,26 +412,34 @@ def fp8_linear(self, input):
         return None
 
     input_dtype = input.dtype
+    input_shape = input.shape
+    tensor_3d = input.ndim == 3
 
-    if input.ndim == 3 or input.ndim == 2:
-        w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
-        scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
+    if tensor_3d:
+        input = input.reshape(-1, input_shape[2])
 
-        scale_input = torch.ones((), device=input.device, dtype=torch.float32)
-        input = torch.clamp(input, min=-448, max=448, out=input)
-        layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
-        quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
+    if input.ndim != 2:
+        return None
+    w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
+    scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
 
-        # Wrap weight in QuantizedTensor - this enables unified dispatch
-        # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
-        layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype}
-        quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
-        o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
+    scale_input = torch.ones((), device=input.device, dtype=torch.float32)
+    input = torch.clamp(input, min=-448, max=448, out=input)
+    input_fp8 = input.to(dtype).contiguous()
+    layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape))
+    quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input)
 
-        uncast_bias_weight(self, w, bias, offload_stream)
-        return o
+    # Wrap weight in QuantizedTensor - this enables unified dispatch
+    # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
+    layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape))
+    quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
+    o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
 
-    return None
+    uncast_bias_weight(self, w, bias, offload_stream)
+    if tensor_3d:
+        o = o.reshape((input_shape[0], input_shape[1], w.shape[0]))
+
+    return o
 
 class fp8_ops(manual_cast):
     class Linear(manual_cast.Linear):
@@ -477,14 +485,20 @@ if CUBLAS_IS_AVAILABLE:
 # ==============================================================================
 # Mixed Precision Operations
 # ==============================================================================
-from .quant_ops import QuantizedTensor, QUANT_ALGOS
+from .quant_ops import (
+    QuantizedTensor,
+    QUANT_ALGOS,
+    TensorCoreFP8Layout,
+    get_layout_class,
+)
 
 
-def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False):
+def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]):
     class MixedPrecisionOps(manual_cast):
         _quant_config = quant_config
         _compute_dtype = compute_dtype
         _full_precision_mm = full_precision_mm
+        _disabled = disabled
 
         class Linear(torch.nn.Module, CastWeightBiasOp):
             def __init__(
@@ -497,21 +511,33 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
             ) -> None:
                 super().__init__()
 
-                if dtype is None:
-                    dtype = MixedPrecisionOps._compute_dtype
-
-                self.factory_kwargs = {"device": device, "dtype": dtype}
+                self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype}
+                # self.factory_kwargs = {"device": device, "dtype": dtype}
 
                 self.in_features = in_features
                 self.out_features = out_features
-                self._has_bias = bias
+                if bias:
+                    self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs))
+                else:
+                    self.register_parameter("bias", None)
 
                 self.tensor_class = None
                 self._full_precision_mm = MixedPrecisionOps._full_precision_mm
+                self._full_precision_mm_config = False
 
             def reset_parameters(self):
                 return None
 
+            def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None):
+                key = f"{prefix}{param_name}"
+                value = state_dict.pop(key, None)
+                if value is not None:
+                    value = value.to(device=device)
+                    if dtype is not None:
+                        value = value.view(dtype=dtype)
+                    manually_loaded_keys.append(key)
+                return value
+
             def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                     strict, missing_keys, unexpected_keys, error_msgs):
 
@@ -520,7 +546,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 weight_key = f"{prefix}weight"
                 weight = state_dict.pop(weight_key, None)
                 if weight is None:
-                    raise ValueError(f"Missing weight for layer {layer_name}")
+                    logging.warning(f"Missing weight for layer {layer_name}")
+                    return
 
                 manually_loaded_keys = [weight_key]
 
@@ -529,49 +556,61 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                     layer_conf = json.loads(layer_conf.numpy().tobytes())
 
                 if layer_conf is None:
-                    dtype = self.factory_kwargs["dtype"]
-                    self.weight = torch.nn.Parameter(weight.to(device=device, dtype=dtype), requires_grad=False)
-                    if dtype != MixedPrecisionOps._compute_dtype:
-                        self.comfy_cast_weights = True
-                    if self._has_bias:
-                        self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=dtype))
-                    else:
-                        self.register_parameter("bias", None)
+                    self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
                 else:
                     self.quant_format = layer_conf.get("format", None)
+                    self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False)
                     if not self._full_precision_mm:
-                        self._full_precision_mm = layer_conf.get("full_precision_matrix_mult", False)
+                        self._full_precision_mm = self._full_precision_mm_config
+
+                    if self.quant_format in MixedPrecisionOps._disabled:
+                        self._full_precision_mm = True
 
                     if self.quant_format is None:
                         raise ValueError(f"Unknown quantization format for layer {layer_name}")
 
                     qconfig = QUANT_ALGOS[self.quant_format]
                     self.layout_type = qconfig["comfy_tensor_layout"]
+                    layout_cls = get_layout_class(self.layout_type)
 
-                    weight_scale_key = f"{prefix}weight_scale"
-                    scale = state_dict.pop(weight_scale_key, None)
-                    if scale is not None:
-                        scale = scale.to(device)
-                    layout_params = {
-                        'scale': scale,
-                        'orig_dtype': MixedPrecisionOps._compute_dtype,
-                        'block_size': qconfig.get("group_size", None),
-                    }
+                    # Load format-specific parameters
+                    if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]:
+                        # FP8: single tensor scale
+                        scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys)
 
-                    if scale is not None:
-                        manually_loaded_keys.append(weight_scale_key)
+                        params = layout_cls.Params(
+                            scale=scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+
+                    elif self.quant_format == "nvfp4":
+                        # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale)
+                        tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys)
+                        block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys,
+                                                             dtype=torch.float8_e4m3fn)
+
+                        if tensor_scale is None or block_scale is None:
+                            raise ValueError(f"Missing NVFP4 scales for layer {layer_name}")
+
+                        params = layout_cls.Params(
+                            scale=tensor_scale,
+                            block_scale=block_scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+                    else:
+                        raise ValueError(f"Unsupported quantization format: {self.quant_format}")
 
                     self.weight = torch.nn.Parameter(
-                        QuantizedTensor(weight.to(device=device, dtype=qconfig.get("storage_t", None)), self.layout_type, layout_params),
+                        QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params),
                         requires_grad=False
                     )
 
-                    if self._has_bias:
-                        self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=MixedPrecisionOps._compute_dtype))
-                    else:
-                        self.register_parameter("bias", None)
-
                     for param_name in qconfig["parameters"]:
+                        if param_name in {"weight_scale", "weight_scale_2"}:
+                            continue  # Already handled above
+
                         param_key = f"{prefix}{param_name}"
                         _v = state_dict.pop(param_key, None)
                         if _v is None:
@@ -586,13 +625,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                         missing_keys.remove(key)
 
             def state_dict(self, *args, destination=None, prefix="", **kwargs):
-                sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs)
+                if destination is not None:
+                    sd = destination
+                else:
+                    sd = {}
+
+                if self.bias is not None:
+                    sd["{}bias".format(prefix)] = self.bias
+
                 if isinstance(self.weight, QuantizedTensor):
-                    sd["{}weight_scale".format(prefix)] = self.weight._layout_params['scale']
+                    sd_out = self.weight.state_dict("{}weight".format(prefix))
+                    for k in sd_out:
+                        sd[k] = sd_out[k]
+
                     quant_conf = {"format": self.quant_format}
-                    if self._full_precision_mm:
+                    if self._full_precision_mm_config:
                         quant_conf["full_precision_matrix_mult"] = True
                     sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
+
+                    input_scale = getattr(self, 'input_scale', None)
+                    if input_scale is not None:
+                        sd["{}input_scale".format(prefix)] = input_scale
+                else:
+                    sd["{}weight".format(prefix)] = self.weight
                 return sd
 
             def _forward(self, input, weight, bias):
@@ -607,12 +662,33 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
             def forward(self, input, *args, **kwargs):
                 run_every_op()
 
-                if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                    return self.forward_comfy_cast_weights(input, *args, **kwargs)
+                input_shape = input.shape
+                reshaped_3d = False
+
                 if (getattr(self, 'layout_type', None) is not None and
-                    not isinstance(input, QuantizedTensor)):
-                    input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None), dtype=self.weight.dtype)
-                return self._forward(input, self.weight, self.bias)
+                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
+                    not getattr(self, 'comfy_force_cast_weights', False) and
+                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
+
+                    # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
+                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
+
+                    # Fall back to non-quantized for non-2D tensors
+                    if input_reshaped.ndim == 2:
+                        reshaped_3d = input.ndim == 3
+                        # dtype is now implicit in the layout class
+                        scale = getattr(self, 'input_scale', None)
+                        if scale is not None:
+                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
+                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
+
+                output = self.forward_comfy_cast_weights(input)
+
+                # Reshape output back to 3D if input was 3D
+                if reshaped_3d:
+                    output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
+
+                return output
 
             def convert_weight(self, weight, inplace=False, **kwargs):
                 if isinstance(weight, QuantizedTensor):
@@ -622,7 +698,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
             def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 if getattr(self, 'layout_type', None) is not None:
-                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True)
+                    # dtype is now implicit in the layout class
+                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True).to(self.weight.dtype)
                 else:
                     weight = weight.to(self.weight.dtype)
                 if return_weight:
@@ -649,10 +726,17 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
+    nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
 
     if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
         logging.info("Using mixed precision operations")
-        return mixed_precision_ops(model_config.quant_config, compute_dtype, full_precision_mm=not fp8_compute)
+        disabled = set()
+        if not nvfp4_compute:
+            disabled.add("nvfp4")
+        if not fp8_compute:
+            disabled.add("float8_e4m3fn")
+            disabled.add("float8_e5m2")
+        return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled)
 
     if (
         fp8_compute and

comfy/quant_ops.py

@@ -1,580 +1,174 @@
 import torch
 import logging
-from typing import Tuple, Dict
+
+try:
+    import comfy_kitchen as ck
+    from comfy_kitchen.tensor import (
+        QuantizedTensor,
+        QuantizedLayout,
+        TensorCoreFP8Layout as _CKFp8Layout,
+        TensorCoreNVFP4Layout as _CKNvfp4Layout,
+        register_layout_op,
+        register_layout_class,
+        get_layout_class,
+    )
+    _CK_AVAILABLE = True
+    if torch.version.cuda is None:
+        ck.registry.disable("cuda")
+    else:
+        cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
+        if cuda_version < (13,):
+            ck.registry.disable("cuda")
+            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
+
+    ck.registry.disable("triton")
+    for k, v in ck.list_backends().items():
+        logging.info(f"Found comfy_kitchen backend {k}: {v}")
+except ImportError as e:
+    logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
+    _CK_AVAILABLE = False
+
+    class QuantizedTensor:
+        pass
+
+    class _CKFp8Layout:
+        pass
+
+    class _CKNvfp4Layout:
+        pass
+
+    def register_layout_class(name, cls):
+        pass
+
+    def get_layout_class(name):
+        return None
+
 import comfy.float
 
-_LAYOUT_REGISTRY = {}
-_GENERIC_UTILS = {}
-
-
-def register_layout_op(torch_op, layout_type):
-    """
-    Decorator to register a layout-specific operation handler.
-    Args:
-        torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default)
-        layout_type: Layout class (e.g., TensorCoreFP8Layout)
-    Example:
-        @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout)
-        def fp8_linear(func, args, kwargs):
-            # FP8-specific linear implementation
-            ...
-    """
-    def decorator(handler_func):
-        if torch_op not in _LAYOUT_REGISTRY:
-            _LAYOUT_REGISTRY[torch_op] = {}
-        _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func
-        return handler_func
-    return decorator
-
-
-def register_generic_util(torch_op):
-    """
-    Decorator to register a generic utility that works for all layouts.
-    Args:
-        torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default)
-
-    Example:
-        @register_generic_util(torch.ops.aten.detach.default)
-        def generic_detach(func, args, kwargs):
-            # Works for any layout
-            ...
-    """
-    def decorator(handler_func):
-        _GENERIC_UTILS[torch_op] = handler_func
-        return handler_func
-    return decorator
-
-
-def _get_layout_from_args(args):
-    for arg in args:
-        if isinstance(arg, QuantizedTensor):
-            return arg._layout_type
-        elif isinstance(arg, (list, tuple)):
-            for item in arg:
-                if isinstance(item, QuantizedTensor):
-                    return item._layout_type
-    return None
-
-
-def _move_layout_params_to_device(params, device):
-    new_params = {}
-    for k, v in params.items():
-        if isinstance(v, torch.Tensor):
-            new_params[k] = v.to(device=device)
-        else:
-            new_params[k] = v
-    return new_params
-
-
-def _copy_layout_params(params):
-    new_params = {}
-    for k, v in params.items():
-        if isinstance(v, torch.Tensor):
-            new_params[k] = v.clone()
-        else:
-            new_params[k] = v
-    return new_params
-
-def _copy_layout_params_inplace(src, dst, non_blocking=False):
-    for k, v in src.items():
-        if isinstance(v, torch.Tensor):
-            dst[k].copy_(v, non_blocking=non_blocking)
-        else:
-            dst[k] = v
-
-class QuantizedLayout:
-    """
-    Base class for quantization layouts.
-
-    A layout encapsulates the format-specific logic for quantization/dequantization
-    and provides a uniform interface for extracting raw tensors needed for computation.
-
-    New quantization formats should subclass this and implement the required methods.
-    """
-    @classmethod
-    def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]:
-        raise NotImplementedError(f"{cls.__name__} must implement quantize()")
-
-    @staticmethod
-    def dequantize(qdata, **layout_params) -> torch.Tensor:
-        raise NotImplementedError("TensorLayout must implement dequantize()")
-
-    @classmethod
-    def get_plain_tensors(cls, qtensor) -> torch.Tensor:
-        raise NotImplementedError(f"{cls.__name__} must implement get_plain_tensors()")
-
-
-class QuantizedTensor(torch.Tensor):
-    """
-    Universal quantized tensor that works with any layout.
-
-    This tensor subclass uses a pluggable layout system to support multiple
-    quantization formats (FP8, INT4, INT8, etc.) without code duplication.
-
-    The layout_type determines format-specific behavior, while common operations
-    (detach, clone, to) are handled generically.
-
-    Attributes:
-        _qdata: The quantized tensor data
-        _layout_type: Layout class (e.g., TensorCoreFP8Layout)
-        _layout_params: Dict with layout-specific params (scale, zero_point, etc.)
-    """
-
-    @staticmethod
-    def __new__(cls, qdata, layout_type, layout_params):
-        """
-        Create a quantized tensor.
-
-        Args:
-            qdata: The quantized data tensor
-            layout_type: Layout class (subclass of QuantizedLayout)
-            layout_params: Dict with layout-specific parameters
-        """
-        return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False)
-
-    def __init__(self, qdata, layout_type, layout_params):
-        self._qdata = qdata
-        self._layout_type = layout_type
-        self._layout_params = layout_params
-
-    def __repr__(self):
-        layout_name = self._layout_type
-        param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2])
-        return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})"
-
-    @property
-    def layout_type(self):
-        return self._layout_type
-
-    def __tensor_flatten__(self):
-        """
-        Tensor flattening protocol for proper device movement.
-        """
-        inner_tensors = ["_qdata"]
-        ctx = {
-            "layout_type": self._layout_type,
-        }
-
-        tensor_params = {}
-        non_tensor_params = {}
-        for k, v in self._layout_params.items():
-            if isinstance(v, torch.Tensor):
-                tensor_params[k] = v
-            else:
-                non_tensor_params[k] = v
-
-        ctx["tensor_param_keys"] = list(tensor_params.keys())
-        ctx["non_tensor_params"] = non_tensor_params
-
-        for k, v in tensor_params.items():
-            attr_name = f"_layout_param_{k}"
-            object.__setattr__(self, attr_name, v)
-            inner_tensors.append(attr_name)
-
-        return inner_tensors, ctx
-
-    @staticmethod
-    def __tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride):
-        """
-        Tensor unflattening protocol for proper device movement.
-        Reconstructs the QuantizedTensor after device movement.
-        """
-        layout_type = ctx["layout_type"]
-        layout_params = dict(ctx["non_tensor_params"])
-
-        for key in ctx["tensor_param_keys"]:
-            attr_name = f"_layout_param_{key}"
-            layout_params[key] = inner_tensors[attr_name]
-
-        return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params)
-
-    @classmethod
-    def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor':
-        qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs)
-        return cls(qdata, layout_type, layout_params)
-
-    def dequantize(self) -> torch.Tensor:
-        return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params)
-
-    @classmethod
-    def __torch_dispatch__(cls, func, types, args=(), kwargs=None):
-        kwargs = kwargs or {}
-
-        # Step 1: Check generic utilities first (detach, clone, to, etc.)
-        if func in _GENERIC_UTILS:
-            return _GENERIC_UTILS[func](func, args, kwargs)
-
-        # Step 2: Check layout-specific handlers (linear, matmul, etc.)
-        layout_type = _get_layout_from_args(args)
-        if layout_type and func in _LAYOUT_REGISTRY:
-            handler = _LAYOUT_REGISTRY[func].get(layout_type)
-            if handler:
-                return handler(func, args, kwargs)
-
-        # Step 3: Fallback to dequantization
-        if isinstance(args[0] if args else None, QuantizedTensor):
-            logging.info(f"QuantizedTensor: Unhandled operation {func}, falling back to dequantization. kwargs={kwargs}")
-        return cls._dequant_and_fallback(func, args, kwargs)
-
-    @classmethod
-    def _dequant_and_fallback(cls, func, args, kwargs):
-        def dequant_arg(arg):
-            if isinstance(arg, QuantizedTensor):
-                return arg.dequantize()
-            elif isinstance(arg, (list, tuple)):
-                return type(arg)(dequant_arg(a) for a in arg)
-            return arg
-
-        new_args = dequant_arg(args)
-        new_kwargs = dequant_arg(kwargs)
-        return func(*new_args, **new_kwargs)
-
-    def data_ptr(self):
-        return self._qdata.data_ptr()
-
-    def is_pinned(self):
-        return self._qdata.is_pinned()
-
-    def is_contiguous(self, *arg, **kwargs):
-        return self._qdata.is_contiguous(*arg, **kwargs)
-
-    def storage(self):
-        return self._qdata.storage()
-
 # ==============================================================================
-# Generic Utilities (Layout-Agnostic Operations)
+# FP8 Layouts with Comfy-Specific Extensions
 # ==============================================================================
 
-def _create_transformed_qtensor(qt, transform_fn):
-    new_data = transform_fn(qt._qdata)
-    new_params = _copy_layout_params(qt._layout_params)
-    return QuantizedTensor(new_data, qt._layout_type, new_params)
+class _TensorCoreFP8LayoutBase(_CKFp8Layout):
+    FP8_DTYPE = None  # Must be overridden in subclass
 
-
-def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"):
-    if target_layout is not None and target_layout != torch.strided:
-        logging.warning(
-            f"QuantizedTensor: layout change requested to {target_layout}, "
-            f"but not supported. Ignoring layout."
-        )
-
-    # Handle device transfer
-    current_device = qt._qdata.device
-    if target_device is not None:
-        # Normalize device for comparison
-        if isinstance(target_device, str):
-            target_device = torch.device(target_device)
-        if isinstance(current_device, str):
-            current_device = torch.device(current_device)
-
-        if target_device != current_device:
-            logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}")
-            new_q_data = qt._qdata.to(device=target_device)
-            new_params = _move_layout_params_to_device(qt._layout_params, target_device)
-            if target_dtype is not None:
-                new_params["orig_dtype"] = target_dtype
-            new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params)
-            logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}")
-            return new_qt
-
-    logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original")
-    return qt
-
-
-@register_generic_util(torch.ops.aten.detach.default)
-def generic_detach(func, args, kwargs):
-    """Detach operation - creates a detached copy of the quantized tensor."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _create_transformed_qtensor(qt, lambda x: x.detach())
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.clone.default)
-def generic_clone(func, args, kwargs):
-    """Clone operation - creates a deep copy of the quantized tensor."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _create_transformed_qtensor(qt, lambda x: x.clone())
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten._to_copy.default)
-def generic_to_copy(func, args, kwargs):
-    """Device/dtype transfer operation - handles .to(device) calls."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _handle_device_transfer(
-            qt,
-            target_device=kwargs.get('device', None),
-            target_dtype=kwargs.get('dtype', None),
-            op_name="_to_copy"
-        )
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.to.dtype_layout)
-def generic_to_dtype_layout(func, args, kwargs):
-    """Handle .to(device) calls using the dtype_layout variant."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _handle_device_transfer(
-            qt,
-            target_device=kwargs.get('device', None),
-            target_dtype=kwargs.get('dtype', None),
-            target_layout=kwargs.get('layout', None),
-            op_name="to"
-        )
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.copy_.default)
-def generic_copy_(func, args, kwargs):
-    qt_dest = args[0]
-    src = args[1]
-    non_blocking = args[2] if len(args) > 2 else False
-    if isinstance(qt_dest, QuantizedTensor):
-        if isinstance(src, QuantizedTensor):
-            # Copy from another quantized tensor
-            qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking)
-            qt_dest._layout_type = src._layout_type
-            orig_dtype = qt_dest._layout_params["orig_dtype"]
-            _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking)
-            qt_dest._layout_params["orig_dtype"] = orig_dtype
-        else:
-            # Copy from regular tensor - just copy raw data
-            qt_dest._qdata.copy_(src)
-        return qt_dest
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.to.dtype)
-def generic_to_dtype(func, args, kwargs):
-    """Handle .to(dtype) calls - dtype conversion only."""
-    src = args[0]
-    if isinstance(src, QuantizedTensor):
-        # For dtype-only conversion, just change the orig_dtype, no real cast is needed
-        target_dtype = args[1] if len(args) > 1 else kwargs.get('dtype')
-        src._layout_params["orig_dtype"] = target_dtype
-        return src
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default)
-def generic_has_compatible_shallow_copy_type(func, args, kwargs):
-    return True
-
-
-@register_generic_util(torch.ops.aten.empty_like.default)
-def generic_empty_like(func, args, kwargs):
-    """Empty_like operation - creates an empty tensor with the same quantized structure."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        # Create empty tensor with same shape and dtype as the quantized data
-        hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"])
-        new_qdata = torch.empty_like(qt._qdata, **kwargs)
-
-        # Handle device transfer for layout params
-        target_device = kwargs.get('device', new_qdata.device)
-        new_params = _move_layout_params_to_device(qt._layout_params, target_device)
-
-        # Update orig_dtype if dtype is specified
-        new_params['orig_dtype'] = hp_dtype
-
-        return QuantizedTensor(new_qdata, qt._layout_type, new_params)
-    return func(*args, **kwargs)
-
-# ==============================================================================
-# FP8 Layout + Operation Handlers
-# ==============================================================================
-class TensorCoreFP8Layout(QuantizedLayout):
-    """
-    Storage format:
-    - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2)
-    - scale: Scalar tensor (float32) for dequantization
-    - orig_dtype: Original dtype before quantization (for casting back)
-    """
     @classmethod
-    def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False):
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if cls.FP8_DTYPE is None:
+            raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE")
+
         orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
 
         if isinstance(scale, str) and scale == "recalculate":
-            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max
+            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max
             if tensor.dtype not in [torch.float32, torch.bfloat16]:  # Prevent scale from being too small
                 tensor_info = torch.finfo(tensor.dtype)
                 scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
 
-        if scale is not None:
-            if not isinstance(scale, torch.Tensor):
-                scale = torch.tensor(scale)
-            scale = scale.to(device=tensor.device, dtype=torch.float32)
+        if scale is None:
+            scale = torch.ones((), device=tensor.device, dtype=torch.float32)
+        if not isinstance(scale, torch.Tensor):
+            scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32)
 
+        if stochastic_rounding > 0:
             if inplace_ops:
                 tensor *= (1.0 / scale).to(tensor.dtype)
             else:
                 tensor = tensor * (1.0 / scale).to(tensor.dtype)
+            qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding)
         else:
-            scale = torch.ones((), device=tensor.device, dtype=torch.float32)
+            qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE)
+
+        params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape)
+        return qdata, params
+
+
+class TensorCoreNVFP4Layout(_CKNvfp4Layout):
+    @classmethod
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if tensor.dim() != 2:
+            raise ValueError(f"NVFP4 requires 2D tensor, got {tensor.dim()}D")
+
+        orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        if scale is None or (isinstance(scale, str) and scale == "recalculate"):
+            scale = torch.amax(tensor.abs()) / (ck.float_utils.F8_E4M3_MAX * ck.float_utils.F4_E2M1_MAX)
+
+        if not isinstance(scale, torch.Tensor):
+            scale = torch.tensor(scale)
+        scale = scale.to(device=tensor.device, dtype=torch.float32)
+
+        padded_shape = cls.get_padded_shape(orig_shape)
+        needs_padding = padded_shape != orig_shape
 
         if stochastic_rounding > 0:
-            tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding)
+            qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4_by_block(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding)
         else:
-            lp_amax = torch.finfo(dtype).max
-            torch.clamp(tensor, min=-lp_amax, max=lp_amax, out=tensor)
-            tensor = tensor.to(dtype, memory_format=torch.contiguous_format)
+            qdata, block_scale = ck.quantize_nvfp4(tensor, scale, pad_16x=needs_padding)
 
-        layout_params = {
-            'scale': scale,
-            'orig_dtype': orig_dtype
-        }
-        return tensor, layout_params
+        params = cls.Params(
+            scale=scale,
+            orig_dtype=orig_dtype,
+            orig_shape=orig_shape,
+            block_scale=block_scale,
+        )
+        return qdata, params
 
-    @staticmethod
-    def dequantize(qdata, scale, orig_dtype, **kwargs):
-        plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype)
-        plain_tensor.mul_(scale)
-        return plain_tensor
 
-    @classmethod
-    def get_plain_tensors(cls, qtensor):
-        return qtensor._qdata, qtensor._layout_params['scale']
+class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase):
+    FP8_DTYPE = torch.float8_e4m3fn
+
+
+class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase):
+    FP8_DTYPE = torch.float8_e5m2
+
+
+# Backward compatibility alias - default to E4M3
+TensorCoreFP8Layout = TensorCoreFP8E4M3Layout
+
+
+# ==============================================================================
+# Registry
+# ==============================================================================
+
+register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
+register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
+register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
+register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
 
 QUANT_ALGOS = {
     "float8_e4m3fn": {
         "storage_t": torch.float8_e4m3fn,
         "parameters": {"weight_scale", "input_scale"},
-        "comfy_tensor_layout": "TensorCoreFP8Layout",
+        "comfy_tensor_layout": "TensorCoreFP8E4M3Layout",
+    },
+    "float8_e5m2": {
+        "storage_t": torch.float8_e5m2,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreFP8E5M2Layout",
+    },
+    "nvfp4": {
+        "storage_t": torch.uint8,
+        "parameters": {"weight_scale", "weight_scale_2", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreNVFP4Layout",
+        "group_size": 16,
     },
 }
 
-LAYOUTS = {
-    "TensorCoreFP8Layout": TensorCoreFP8Layout,
-}
 
+# ==============================================================================
+# Re-exports for backward compatibility
+# ==============================================================================
 
-@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout")
-def fp8_linear(func, args, kwargs):
-    input_tensor = args[0]
-    weight = args[1]
-    bias = args[2] if len(args) > 2 else None
-
-    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
-        plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
-        plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight)
-
-        out_dtype = kwargs.get("out_dtype")
-        if out_dtype is None:
-            out_dtype = input_tensor._layout_params['orig_dtype']
-
-        weight_t = plain_weight.t()
-
-        tensor_2d = False
-        if len(plain_input.shape) == 2:
-            tensor_2d = True
-            plain_input = plain_input.unsqueeze(1)
-
-        input_shape = plain_input.shape
-        if len(input_shape) != 3:
-            return None
-
-        try:
-            output = torch._scaled_mm(
-                plain_input.reshape(-1, input_shape[2]).contiguous(),
-                weight_t,
-                bias=bias,
-                scale_a=scale_a,
-                scale_b=scale_b,
-                out_dtype=out_dtype,
-            )
-
-            if isinstance(output, tuple):  # TODO: remove when we drop support for torch 2.4
-                output = output[0]
-
-            if not tensor_2d:
-                output = output.reshape((-1, input_shape[1], weight.shape[0]))
-
-            if output.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
-                output_scale = scale_a * scale_b
-                output_params = {
-                    'scale': output_scale,
-                    'orig_dtype': input_tensor._layout_params['orig_dtype']
-                }
-                return QuantizedTensor(output, "TensorCoreFP8Layout", output_params)
-            else:
-                return output
-
-        except Exception as e:
-            raise RuntimeError(f"FP8 _scaled_mm failed, falling back to dequantization: {e}")
-
-    # Case 2: DQ Fallback
-    if isinstance(weight, QuantizedTensor):
-        weight = weight.dequantize()
-    if isinstance(input_tensor, QuantizedTensor):
-        input_tensor = input_tensor.dequantize()
-
-    return torch.nn.functional.linear(input_tensor, weight, bias)
-
-def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None):
-    if out_dtype is None:
-        out_dtype = input_tensor._layout_params['orig_dtype']
-
-    plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
-    plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight)
-
-    output = torch._scaled_mm(
-        plain_input.contiguous(),
-        plain_weight,
-        bias=bias,
-        scale_a=scale_a,
-        scale_b=scale_b,
-        out_dtype=out_dtype,
-    )
-
-    if isinstance(output, tuple):  # TODO: remove when we drop support for torch 2.4
-        output = output[0]
-    return output
-
-@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout")
-def fp8_addmm(func, args, kwargs):
-    input_tensor = args[1]
-    weight = args[2]
-    bias = args[0]
-
-    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
-        return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None))
-
-    a = list(args)
-    if isinstance(args[0], QuantizedTensor):
-        a[0] = args[0].dequantize()
-    if isinstance(args[1], QuantizedTensor):
-        a[1] = args[1].dequantize()
-    if isinstance(args[2], QuantizedTensor):
-        a[2] = args[2].dequantize()
-
-    return func(*a, **kwargs)
-
-@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout")
-def fp8_mm(func, args, kwargs):
-    input_tensor = args[0]
-    weight = args[1]
-
-    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
-        return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None))
-
-    a = list(args)
-    if isinstance(args[0], QuantizedTensor):
-        a[0] = args[0].dequantize()
-    if isinstance(args[1], QuantizedTensor):
-        a[1] = args[1].dequantize()
-    return func(*a, **kwargs)
-
-@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout")
-@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout")
-def fp8_func(func, args, kwargs):
-    input_tensor = args[0]
-    if isinstance(input_tensor, QuantizedTensor):
-        plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
-        ar = list(args)
-        ar[0] = plain_input
-        return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params)
-    return func(*args, **kwargs)
+__all__ = [
+    "QuantizedTensor",
+    "QuantizedLayout",
+    "TensorCoreFP8Layout",
+    "TensorCoreFP8E4M3Layout",
+    "TensorCoreFP8E5M2Layout",
+    "TensorCoreNVFP4Layout",
+    "QUANT_ALGOS",
+    "register_layout_op",
+]

comfy/sampler_helpers.py

@@ -122,20 +122,20 @@ def estimate_memory(model, noise_shape, conds):
     minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
     return memory_required, minimum_memory_required
 
-def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
+def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _prepare_sampling,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
     )
-    return executor.execute(model, noise_shape, conds, model_options=model_options)
+    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load)
 
-def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
     memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory)
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load)
     real_model = model.model
 
     return real_model, conds, models

comfy/samplers.py

@@ -984,9 +984,6 @@ class CFGGuider:
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
-        if denoise_mask is not None:
-            denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device)
-
         noise = noise.to(device)
         latent_image = latent_image.to(device)
         sigmas = sigmas.to(device)
@@ -1013,6 +1010,24 @@ class CFGGuider:
         else:
             latent_shapes = [latent_image.shape]
 
+        if denoise_mask is not None:
+            if denoise_mask.is_nested:
+                denoise_masks = denoise_mask.unbind()
+                denoise_masks = denoise_masks[:len(latent_shapes)]
+            else:
+                denoise_masks = [denoise_mask]
+
+            for i in range(len(denoise_masks), len(latent_shapes)):
+                denoise_masks.append(torch.ones(latent_shapes[i]))
+
+            for i in range(len(denoise_masks)):
+                denoise_masks[i] = comfy.sampler_helpers.prepare_mask(denoise_masks[i], latent_shapes[i], self.model_patcher.load_device)
+
+            if len(denoise_masks) > 1:
+                denoise_mask, _ = comfy.utils.pack_latents(denoise_masks)
+            else:
+                denoise_mask = denoise_masks[0]
+
         self.conds = {}
         for k in self.original_conds:
             self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k]))

comfy/sd.py

@@ -55,6 +55,9 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.z_image
 import comfy.text_encoders.ovis
 import comfy.text_encoders.kandinsky5
+import comfy.text_encoders.jina_clip_2
+import comfy.text_encoders.newbie
+import comfy.text_encoders.anima
 
 import comfy.model_patcher
 import comfy.lora
@@ -216,7 +219,7 @@ class CLIP:
             if unprojected:
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-            self.load_model()
+            self.load_model(tokens)
             self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
             all_hooks.reset()
             self.patcher.patch_hooks(None)
@@ -264,7 +267,7 @@ class CLIP:
         if return_pooled == "unprojected":
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-        self.load_model()
+        self.load_model(tokens)
         self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
         o = self.cond_stage_model.encode_token_weights(tokens)
         cond, pooled = o[:2]
@@ -297,8 +300,11 @@ class CLIP:
             sd_clip[k] = sd_tokenizer[k]
         return sd_clip
 
-    def load_model(self):
-        model_management.load_model_gpu(self.patcher)
+    def load_model(self, tokens={}):
+        memory_used = 0
+        if hasattr(self.cond_stage_model, "memory_estimation_function"):
+            memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
         return self.patcher
 
     def get_key_patches(self):
@@ -321,6 +327,7 @@ class VAE:
         self.latent_channels = 4
         self.latent_dim = 2
         self.output_channels = 3
+        self.pad_channel_value = None
         self.process_input = lambda image: image * 2.0 - 1.0
         self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)
         self.working_dtypes = [torch.bfloat16, torch.float32]
@@ -435,6 +442,7 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
+                self.pad_channel_value = "replicate"
                 self.upscale_ratio = 2048
                 self.downscale_ratio =  2048
                 self.latent_dim = 1
@@ -472,8 +480,8 @@ class VAE:
                 self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config)
                 self.latent_channels = 128
                 self.latent_dim = 3
-                self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
-                self.memory_used_encode = lambda shape, dtype: (70 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (1200 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (80 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 32, 32)
                 self.upscale_index_formula = (8, 32, 32)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 7) / 8)), 32, 32)
@@ -546,7 +554,9 @@ class VAE:
                     self.downscale_index_formula = (4, 8, 8)
                     self.latent_dim = 3
                     self.latent_channels = 16
-                    ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                    self.output_channels = sd["encoder.conv1.weight"].shape[1]
+                    self.pad_channel_value = 1.0
+                    ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "dropout": 0.0}
                     self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
                     self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
                     self.memory_used_encode = lambda shape, dtype: (1500 if shape[2]<=4 else 6000) * shape[3] * shape[4] * model_management.dtype_size(dtype)
@@ -582,6 +592,7 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype)
                 self.latent_channels = 8
                 self.output_channels = 2
+                self.pad_channel_value = "replicate"
                 self.upscale_ratio = 4096
                 self.downscale_ratio = 4096
                 self.latent_dim = 2
@@ -625,14 +636,13 @@ class VAE:
                 self.upscale_index_formula = (4, 16, 16)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16)
                 self.downscale_index_formula = (4, 16, 16)
-                if self.latent_channels == 48: # Wan 2.2
+                if self.latent_channels in [48, 128]: # Wan 2.2 and LTX2
                     self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=None) # taehv doesn't need scaling
-                    self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently"))
+                    self.process_input = self.process_output = lambda image: image
                     self.process_output = lambda image: image
                     self.memory_used_decode = lambda shape, dtype: (1800 * (max(1, (shape[-3] ** 0.7 * 0.1)) * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype))
                 elif self.latent_channels == 32 and sd["decoder.22.bias"].shape[0] == 12: # lighttae_hv15
                     self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=comfy.latent_formats.HunyuanVideo15)
-                    self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently"))
                     self.memory_used_decode = lambda shape, dtype: (1200 * (max(1, (shape[-3] ** 0.7 * 0.05)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype))
                 else:
                     if sd["decoder.1.weight"].dtype == torch.float16: # taehv currently only available in float16, so assume it's not lighttaew2_1 as otherwise state dicts are identical
@@ -690,17 +700,28 @@ class VAE:
             raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
 
     def vae_encode_crop_pixels(self, pixels):
-        if not self.crop_input:
-            return pixels
+        if self.crop_input:
+            downscale_ratio = self.spacial_compression_encode()
 
-        downscale_ratio = self.spacial_compression_encode()
+            dims = pixels.shape[1:-1]
+            for d in range(len(dims)):
+                x = (dims[d] // downscale_ratio) * downscale_ratio
+                x_offset = (dims[d] % downscale_ratio) // 2
+                if x != dims[d]:
+                    pixels = pixels.narrow(d + 1, x_offset, x)
 
-        dims = pixels.shape[1:-1]
-        for d in range(len(dims)):
-            x = (dims[d] // downscale_ratio) * downscale_ratio
-            x_offset = (dims[d] % downscale_ratio) // 2
-            if x != dims[d]:
-                pixels = pixels.narrow(d + 1, x_offset, x)
+        if pixels.shape[-1] > self.output_channels:
+            pixels = pixels[..., :self.output_channels]
+        elif pixels.shape[-1] < self.output_channels:
+            if self.pad_channel_value is not None:
+                if isinstance(self.pad_channel_value, str):
+                    mode = self.pad_channel_value
+                    value = None
+                else:
+                    mode = "constant"
+                    value = self.pad_channel_value
+
+                pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value)
         return pixels
 
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
@@ -992,6 +1013,8 @@ class CLIPType(Enum):
     OVIS = 21
     KANDINSKY5 = 22
     KANDINSKY5_IMAGE = 23
+    NEWBIE = 24
+    FLUX2 = 25
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -1022,6 +1045,10 @@ class TEModel(Enum):
     MISTRAL3_24B_PRUNED_FLUX2 = 15
     QWEN3_4B = 16
     QWEN3_2B = 17
+    GEMMA_3_12B = 18
+    JINA_CLIP_2 = 19
+    QWEN3_8B = 20
+    QWEN3_06B = 21
 
 
 def detect_te_model(sd):
@@ -1031,11 +1058,13 @@ def detect_te_model(sd):
         return TEModel.CLIP_H
     if "text_model.encoder.layers.0.mlp.fc1.weight" in sd:
         return TEModel.CLIP_L
+    if "model.encoder.layers.0.mixer.Wqkv.weight" in sd:
+        return TEModel.JINA_CLIP_2
     if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd:
         weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"]
-        if weight.shape[-1] == 4096:
+        if weight.shape[0] == 10240:
             return TEModel.T5_XXL
-        elif weight.shape[-1] == 2048:
+        elif weight.shape[0] == 5120:
             return TEModel.T5_XL
     if 'encoder.block.23.layer.1.DenseReluDense.wi.weight' in sd:
         return TEModel.T5_XXL_OLD
@@ -1045,6 +1074,8 @@ def detect_te_model(sd):
             return TEModel.BYT5_SMALL_GLYPH
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
+        if 'model.layers.47.self_attn.q_norm.weight' in sd:
+            return TEModel.GEMMA_3_12B
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
             return TEModel.GEMMA_3_4B
         return TEModel.GEMMA_2_2B
@@ -1061,6 +1092,10 @@ def detect_te_model(sd):
                 return TEModel.QWEN3_4B
             elif weight.shape[0] == 2048:
                 return TEModel.QWEN3_2B
+            elif weight.shape[0] == 4096:
+                return TEModel.QWEN3_8B
+            elif weight.shape[0] == 1024:
+                return TEModel.QWEN3_06B
         if weight.shape[0] == 5120:
             if "model.layers.39.post_attention_layernorm.weight" in sd:
                 return TEModel.MISTRAL3_24B
@@ -1186,11 +1221,24 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer
             tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None)
         elif te_model == TEModel.QWEN3_4B:
-            clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer
+            if clip_type == CLIPType.FLUX or clip_type == CLIPType.FLUX2:
+                clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_4b")
+                clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer
+            else:
+                clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer
         elif te_model == TEModel.QWEN3_2B:
             clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer
+        elif te_model == TEModel.QWEN3_8B:
+            clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_8b")
+            clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer8B
+        elif te_model == TEModel.JINA_CLIP_2:
+            clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper
+            clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper
+        elif te_model == TEModel.QWEN3_06B:
+            clip_target.clip = comfy.text_encoders.anima.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.anima.AnimaTokenizer
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
@@ -1246,6 +1294,21 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.KANDINSKY5_IMAGE:
             clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage
+        elif clip_type == CLIPType.LTXV:
+            clip_target.clip = comfy.text_encoders.lt.ltxav_te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.lt.LTXAVGemmaTokenizer
+            tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+        elif clip_type == CLIPType.NEWBIE:
+            clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer
+            if "model.layers.0.self_attn.q_norm.weight" in clip_data[0]:
+                clip_data_gemma = clip_data[0]
+                clip_data_jina = clip_data[1]
+            else:
+                clip_data_gemma = clip_data[1]
+                clip_data_jina = clip_data[0]
+            tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
+            tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer

comfy/sd1_clip.py

@@ -466,7 +466,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
     return embed_out
 
 class SDTokenizer:
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}):
         if tokenizer_path is None:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
         self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
@@ -513,6 +513,8 @@ class SDTokenizer:
         self.embedding_size = embedding_size
         self.embedding_key = embedding_key
 
+        self.disable_weights = disable_weights
+
     def _try_get_embedding(self, embedding_name:str):
         '''
         Takes a potential embedding name and tries to retrieve it.
@@ -547,7 +549,7 @@ class SDTokenizer:
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
-        if kwargs.get("disable_weights", False):
+        if kwargs.get("disable_weights", self.disable_weights):
             parsed_weights = [(text, 1.0)]
         else:
             parsed_weights = token_weights(text, 1.0)

comfy/supported_models.py

@@ -23,6 +23,7 @@ import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
+import comfy.text_encoders.anima
 
 from . import supported_models_base
 from . import latent_formats
@@ -763,17 +764,31 @@ class Flux2(Flux):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * 2.36
+        self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * (unet_config['hidden_size'] / 2604)
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.Flux2(self, device=device)
         return out
 
     def clip_target(self, state_dict={}):
-        return None # TODO
         pref = self.text_encoder_key_prefix[0]
-        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(**t5_detect))
+        detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))
+        if len(detect) > 0:
+            detect["model_type"] = "qwen3_4b"
+            return supported_models_base.ClipTarget(comfy.text_encoders.flux.KleinTokenizer, comfy.text_encoders.flux.klein_te(**detect))
+
+        detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_8b.transformer.".format(pref))
+        if len(detect) > 0:
+            detect["model_type"] = "qwen3_8b"
+            return supported_models_base.ClipTarget(comfy.text_encoders.flux.KleinTokenizer8B, comfy.text_encoders.flux.klein_te(**detect))
+
+        detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}mistral3_24b.transformer.".format(pref))
+        if len(detect) > 0:
+            if "{}mistral3_24b.transformer.model.layers.39.post_attention_layernorm.weight".format(pref) not in state_dict:
+                detect["pruned"] = True
+            return supported_models_base.ClipTarget(comfy.text_encoders.flux.Flux2Tokenizer, comfy.text_encoders.flux.flux2_te(**detect))
+
+        return None
 
 class GenmoMochi(supported_models_base.BASE):
     unet_config = {
@@ -836,6 +851,21 @@ class LTXV(supported_models_base.BASE):
         t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.lt.LTXVT5Tokenizer, comfy.text_encoders.lt.ltxv_te(**t5_detect))
 
+class LTXAV(LTXV):
+    unet_config = {
+        "image_model": "ltxav",
+    }
+
+    latent_format = latent_formats.LTXAV
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = 0.077  # TODO
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.LTXAV(self, device=device)
+        return out
+
 class HunyuanVideo(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan_video",
@@ -977,6 +1007,36 @@ class CosmosT2IPredict2(supported_models_base.BASE):
         t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.cosmos.CosmosT5Tokenizer, comfy.text_encoders.cosmos.te(**t5_detect))
 
+class Anima(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "anima",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    memory_usage_factor = 1.0
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Anima(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect))
+
 class CosmosI2VPredict2(CosmosT2IPredict2):
     unet_config = {
         "image_model": "cosmos_predict2",
@@ -1027,7 +1087,7 @@ class ZImage(Lumina2):
         "shift": 3.0,
     }
 
-    memory_usage_factor = 2.0
+    memory_usage_factor = 2.8
 
     supported_inference_dtypes = [torch.bfloat16, torch.float32]
 
@@ -1536,6 +1596,6 @@ class Kandinsky5Image(Kandinsky5):
         return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/taesd/taehv.py

@@ -112,7 +112,8 @@ def apply_model_with_memblocks(model, x, parallel, show_progress_bar):
 
 
 class TAEHV(nn.Module):
-    def __init__(self, latent_channels, parallel=False, decoder_time_upscale=(True, True), decoder_space_upscale=(True, True, True), latent_format=None, show_progress_bar=True):
+    def __init__(self, latent_channels, parallel=False, encoder_time_downscale=(True, True, False), decoder_time_upscale=(False, True, True), decoder_space_upscale=(True, True, True),
+                 latent_format=None, show_progress_bar=False):
         super().__init__()
         self.image_channels = 3
         self.patch_size = 1
@@ -124,6 +125,9 @@ class TAEHV(nn.Module):
         self.process_out = latent_format().process_out if latent_format is not None else (lambda x: x)
         if self.latent_channels in [48, 32]: # Wan 2.2 and HunyuanVideo1.5
             self.patch_size = 2
+        elif self.latent_channels == 128: # LTX2
+            self.patch_size, self.latent_channels, encoder_time_downscale, decoder_time_upscale = 4, 128, (True, True, True), (True, True, True)
+
         if self.latent_channels == 32: # HunyuanVideo1.5
             act_func = nn.LeakyReLU(0.2, inplace=True)
         else: # HunyuanVideo, Wan 2.1
@@ -131,41 +135,54 @@ class TAEHV(nn.Module):
 
         self.encoder = nn.Sequential(
             conv(self.image_channels*self.patch_size**2, 64), act_func,
-            TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func),
-            TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func),
-            TPool(64, 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func),
+            TPool(64, 2 if encoder_time_downscale[0] else 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func),
+            TPool(64, 2 if encoder_time_downscale[1] else 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func),
+            TPool(64, 2 if encoder_time_downscale[2] else 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func),
             conv(64, self.latent_channels),
         )
         n_f = [256, 128, 64, 64]
-        self.frames_to_trim = 2**sum(decoder_time_upscale) - 1
+
         self.decoder = nn.Sequential(
             Clamp(), conv(self.latent_channels, n_f[0]), act_func,
-            MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[0] else 1), TGrow(n_f[0], 1), conv(n_f[0], n_f[1], bias=False),
-            MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[1] else 1), TGrow(n_f[1], 2 if decoder_time_upscale[0] else 1), conv(n_f[1], n_f[2], bias=False),
-            MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[2] else 1), TGrow(n_f[2], 2 if decoder_time_upscale[1] else 1), conv(n_f[2], n_f[3], bias=False),
+            MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[0] else 1), TGrow(n_f[0], 2 if decoder_time_upscale[0] else 1), conv(n_f[0], n_f[1], bias=False),
+            MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[1] else 1), TGrow(n_f[1], 2 if decoder_time_upscale[1] else 1), conv(n_f[1], n_f[2], bias=False),
+            MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[2] else 1), TGrow(n_f[2], 2 if decoder_time_upscale[2] else 1), conv(n_f[2], n_f[3], bias=False),
             act_func, conv(n_f[3], self.image_channels*self.patch_size**2),
         )
-        @property
-        def show_progress_bar(self):
-            return self._show_progress_bar
 
-        @show_progress_bar.setter
-        def show_progress_bar(self, value):
-            self._show_progress_bar = value
+        self.t_downscale = 2**sum(t.stride == 2 for t in self.encoder if isinstance(t, TPool))
+        self.t_upscale = 2**sum(t.stride == 2 for t in self.decoder if isinstance(t, TGrow))
+        self.frames_to_trim = self.t_upscale - 1
+        self._show_progress_bar = show_progress_bar
+
+    @property
+    def show_progress_bar(self):
+        return self._show_progress_bar
+
+    @show_progress_bar.setter
+    def show_progress_bar(self, value):
+        self._show_progress_bar = value
 
     def encode(self, x, **kwargs):
-        if self.patch_size > 1: x = F.pixel_unshuffle(x, self.patch_size)
         x = x.movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
-        if x.shape[1] % 4 != 0:
-            # pad at end to multiple of 4
-            n_pad = 4 - x.shape[1] % 4
+        if self.patch_size > 1:
+            B, T, C, H, W = x.shape
+            x = x.reshape(B * T, C, H, W)
+            x = F.pixel_unshuffle(x, self.patch_size)
+            x = x.reshape(B, T, C * self.patch_size ** 2, H // self.patch_size, W // self.patch_size)
+        if x.shape[1] % self.t_downscale != 0:
+            # pad at end to multiple of t_downscale
+            n_pad = self.t_downscale - x.shape[1] % self.t_downscale
             padding = x[:, -1:].repeat_interleave(n_pad, dim=1)
             x = torch.cat([x, padding], 1)
         x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar).movedim(2, 1)
         return self.process_out(x)
 
     def decode(self, x, **kwargs):
+        x = x.unsqueeze(0) if x.ndim == 4 else x  # [T, C, H, W] -> [1, T, C, H, W]
+        x = x.movedim(1, 2) if x.shape[1] != self.latent_channels else x  # [B, T, C, H, W] or [B, C, T, H, W]
         x = self.process_in(x).movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
         x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar)
-        if self.patch_size > 1: x = F.pixel_shuffle(x, self.patch_size)
+        if self.patch_size > 1:
+            x = F.pixel_shuffle(x, self.patch_size)
         return x[:, self.frames_to_trim:].movedim(2, 1)

comfy/text_encoders/anima.py

@@ -0,0 +1,61 @@
+from transformers import Qwen2Tokenizer, T5TokenizerFast
+import comfy.text_encoders.llama
+from comfy import sd1_clip
+import os
+import torch
+
+
+class Qwen3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+
+class T5XXLTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data)
+
+class AnimaTokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.qwen3_06b = Qwen3Tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = {}
+        qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["qwen3_06b"] = [[(token, 1.0) for token, _ in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
+        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.t5xxl.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return {}
+
+
+class Qwen3_06BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class AnimaTEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="qwen3_06b", clip_model=Qwen3_06BModel, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        out = super().encode_token_weights(token_weight_pairs)
+        out[2]["t5xxl_ids"] = torch.tensor(list(map(lambda a: a[0], token_weight_pairs["t5xxl"][0])), dtype=torch.int)
+        out[2]["t5xxl_weights"] = torch.tensor(list(map(lambda a: a[1], token_weight_pairs["t5xxl"][0])))
+        return out
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class AnimaTEModel_(AnimaTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return AnimaTEModel_

comfy/text_encoders/cosmos.py

@@ -36,7 +36,7 @@ def te(dtype_t5=None, t5_quantization_metadata=None):
             if t5_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype is None:
+            if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return CosmosTEModel_

comfy/text_encoders/flux.py

@@ -3,7 +3,7 @@ import comfy.text_encoders.t5
 import comfy.text_encoders.sd3_clip
 import comfy.text_encoders.llama
 import comfy.model_management
-from transformers import T5TokenizerFast, LlamaTokenizerFast
+from transformers import T5TokenizerFast, LlamaTokenizerFast, Qwen2Tokenizer
 import torch
 import os
 import json
@@ -172,3 +172,60 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
                 model_options["num_layers"] = 30
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return Flux2TEModel_
+
+class Qwen3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+
+class Qwen3Tokenizer8B(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+
+class KleinTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"):
+        if name == "qwen3_4b":
+            tokenizer = Qwen3Tokenizer
+        elif name == "qwen3_8b":
+            tokenizer = Qwen3Tokenizer8B
+
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name=name, tokenizer=tokenizer)
+        self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        return tokens
+
+class KleinTokenizer8B(KleinTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_8b"):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name=name)
+
+class Qwen3_4BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer=[9, 18, 27], layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Qwen3_8BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer=[9, 18, 27], layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_8B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+def klein_te(dtype_llama=None, llama_quantization_metadata=None, model_type="qwen3_4b"):
+    if model_type == "qwen3_4b":
+        model = Qwen3_4BModel
+    elif model_type == "qwen3_8b":
+        model = Qwen3_8BModel
+
+    class Flux2TEModel_(Flux2TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model)
+    return Flux2TEModel_

comfy/text_encoders/genmo.py

@@ -32,7 +32,7 @@ def mochi_te(dtype_t5=None, t5_quantization_metadata=None):
             if t5_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype is None:
+            if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return MochiTEModel_

comfy/text_encoders/hunyuan_video.py

@@ -10,9 +10,11 @@ import comfy.utils
 
 def llama_detect(state_dict, prefix=""):
     out = {}
-    t5_key = "{}model.norm.weight".format(prefix)
-    if t5_key in state_dict:
-        out["dtype_llama"] = state_dict[t5_key].dtype
+    norm_keys = ["{}model.norm.weight".format(prefix), "{}model.layers.0.input_layernorm.weight".format(prefix)]
+    for norm_key in norm_keys:
+        if norm_key in state_dict:
+            out["dtype_llama"] = state_dict[norm_key].dtype
+            break
 
     quant = comfy.utils.detect_layer_quantization(state_dict, prefix)
     if quant is not None:

comfy/text_encoders/jina_clip_2.py

@@ -0,0 +1,219 @@
+# Jina CLIP v2 and Jina Embeddings v3 both use their modified XLM-RoBERTa architecture. Reference implementation:
+# Jina CLIP v2 (both text and vision): https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/modeling_clip.py
+# Jina XLM-RoBERTa (text only): http://huggingface.co/jinaai/xlm-roberta-flash-implementation/blob/2b6bc3f30750b3a9648fe9b63448c09920efe9be/modeling_xlm_roberta.py
+
+from dataclasses import dataclass
+
+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+
+import comfy.model_management
+import comfy.ops
+from comfy import sd1_clip
+from .spiece_tokenizer import SPieceTokenizer
+
+class JinaClip2Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        # The official NewBie uses max_length=8000, but Jina Embeddings v3 actually supports 8192
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='jina_clip_2', tokenizer_class=SPieceTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=False, max_length=8192, min_length=1, pad_token=1, end_token=2, tokenizer_args={"add_bos": True, "add_eos": True}, tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+class JinaClip2TokenizerWrapper(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=JinaClip2Tokenizer, name="jina_clip_2")
+
+# https://huggingface.co/jinaai/jina-embeddings-v3/blob/343dbf534c76fe845f304fa5c2d1fd87e1e78918/config.json
+@dataclass
+class XLMRobertaConfig:
+    vocab_size: int = 250002
+    type_vocab_size: int = 1
+    hidden_size: int = 1024
+    num_hidden_layers: int = 24
+    num_attention_heads: int = 16
+    rotary_emb_base: float = 20000.0
+    intermediate_size: int = 4096
+    hidden_act: str = "gelu"
+    hidden_dropout_prob: float = 0.1
+    attention_probs_dropout_prob: float = 0.1
+    layer_norm_eps: float = 1e-05
+    bos_token_id: int = 0
+    eos_token_id: int = 2
+    pad_token_id: int = 1
+
+class XLMRobertaEmbeddings(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.word_embeddings = ops.Embedding(config.vocab_size, embed_dim, padding_idx=config.pad_token_id, device=device, dtype=dtype)
+        self.token_type_embeddings = ops.Embedding(config.type_vocab_size, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, input_ids=None, embeddings=None):
+        if input_ids is not None and embeddings is None:
+            embeddings = self.word_embeddings(input_ids)
+
+        if embeddings is not None:
+            token_type_ids = torch.zeros(embeddings.shape[1], device=embeddings.device, dtype=torch.int32)
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings = embeddings + token_type_embeddings
+        return embeddings
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, dim, base, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+
+    def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
+        if seqlen > self._seq_len_cached or self._cos_cached is None or self._cos_cached.device != device or self._cos_cached.dtype != dtype:
+            self._seq_len_cached = seqlen
+            t = torch.arange(seqlen, device=device, dtype=torch.float32)
+            freqs = torch.outer(t, self.inv_freq.to(device=t.device))
+            emb = torch.cat((freqs, freqs), dim=-1)
+            self._cos_cached = emb.cos().to(dtype)
+            self._sin_cached = emb.sin().to(dtype)
+
+    def forward(self, q, k):
+        batch, seqlen, heads, head_dim = q.shape
+        self._update_cos_sin_cache(seqlen, device=q.device, dtype=q.dtype)
+
+        cos = self._cos_cached[:seqlen].view(1, seqlen, 1, head_dim)
+        sin = self._sin_cached[:seqlen].view(1, seqlen, 1, head_dim)
+
+        def rotate_half(x):
+            size = x.shape[-1] // 2
+            x1, x2 = x[..., :size], x[..., size:]
+            return torch.cat((-x2, x1), dim=-1)
+
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+        k_embed = (k * cos) + (rotate_half(k) * sin)
+        return q_embed, k_embed
+
+class MHA(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = embed_dim // config.num_attention_heads
+
+        self.rotary_emb = RotaryEmbedding(self.head_dim, config.rotary_emb_base, device=device)
+        self.Wqkv = ops.Linear(embed_dim, 3 * embed_dim, device=device, dtype=dtype)
+        self.out_proj = ops.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        qkv = self.Wqkv(x)
+        batch_size, seq_len, _ = qkv.shape
+        qkv = qkv.view(batch_size, seq_len, 3, self.num_heads, self.head_dim)
+        q, k, v = qkv.unbind(2)
+
+        q, k = self.rotary_emb(q, k)
+
+        # NHD -> HND
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+
+        out = optimized_attention(q, k, v, heads=self.num_heads, mask=mask, skip_reshape=True)
+        return self.out_proj(out)
+
+class MLP(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.fc1 = ops.Linear(config.hidden_size, config.intermediate_size, device=device, dtype=dtype)
+        self.activation = F.gelu
+        self.fc2 = ops.Linear(config.intermediate_size, config.hidden_size, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.activation(x)
+        x = self.fc2(x)
+        return x
+
+class Block(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.mixer = MHA(config, device=device, dtype=dtype, ops=ops)
+        self.dropout1 = nn.Dropout(config.hidden_dropout_prob)
+        self.norm1 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
+        self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
+        self.dropout2 = nn.Dropout(config.hidden_dropout_prob)
+        self.norm2 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, mask=None, optimized_attention=None):
+        mixer_out = self.mixer(hidden_states, mask=mask, optimized_attention=optimized_attention)
+        hidden_states = self.norm1(self.dropout1(mixer_out) + hidden_states)
+        mlp_out = self.mlp(hidden_states)
+        hidden_states = self.norm2(self.dropout2(mlp_out) + hidden_states)
+        return hidden_states
+
+class XLMRobertaEncoder(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.layers = nn.ModuleList([Block(config, device=device, dtype=dtype, ops=ops) for _ in range(config.num_hidden_layers)])
+
+    def forward(self, hidden_states, attention_mask=None):
+        optimized_attention = comfy.ldm.modules.attention.optimized_attention_for_device(hidden_states.device, mask=attention_mask is not None, small_input=True)
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, mask=attention_mask, optimized_attention=optimized_attention)
+        return hidden_states
+
+class XLMRobertaModel_(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.embeddings = XLMRobertaEmbeddings(config, device=device, dtype=dtype, ops=ops)
+        self.emb_ln = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
+        self.emb_drop = nn.Dropout(config.hidden_dropout_prob)
+        self.encoder = XLMRobertaEncoder(config, device=device, dtype=dtype, ops=ops)
+
+    def forward(self, input_ids, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
+        x = self.embeddings(input_ids=input_ids, embeddings=embeds)
+        x = self.emb_ln(x)
+        x = self.emb_drop(x)
+
+        mask = None
+        if attention_mask is not None:
+            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, 1, attention_mask.shape[-1]))
+            mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
+
+        sequence_output = self.encoder(x, attention_mask=mask)
+
+        # Mean pool, see https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/hf_model.py
+        pooled_output = None
+        if attention_mask is None:
+            pooled_output = sequence_output.mean(dim=1)
+        else:
+            attention_mask = attention_mask.to(sequence_output.dtype)
+            pooled_output = (sequence_output * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(dim=-1, keepdim=True)
+
+        # Intermediate output is not yet implemented, use None for placeholder
+        return sequence_output, None, pooled_output
+
+class XLMRobertaModel(nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.config = XLMRobertaConfig(**config_dict)
+        self.model = XLMRobertaModel_(self.config, device=device, dtype=dtype, ops=operations)
+        self.num_layers = self.config.num_hidden_layers
+
+    def get_input_embeddings(self):
+        return self.model.embeddings.word_embeddings
+
+    def set_input_embeddings(self, embeddings):
+        self.model.embeddings.word_embeddings = embeddings
+
+    def forward(self, *args, **kwargs):
+        return self.model(*args, **kwargs)
+
+class JinaClip2TextModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, textmodel_json_config={}, model_class=XLMRobertaModel, special_tokens={"start": 0, "end": 2, "pad": 1}, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
+
+class JinaClip2TextModelWrapper(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, clip_model=JinaClip2TextModel, name="jina_clip_2", model_options=model_options)

comfy/text_encoders/llama.py

@@ -3,13 +3,12 @@ import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any
 import math
-import logging
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
 import comfy.ldm.common_dit
+import comfy.clip_model
 
-import comfy.model_management
 from . import qwen_vl
 
 @dataclass
@@ -78,6 +77,28 @@ class Qwen25_3BConfig:
     rope_scale = None
     final_norm: bool = True
 
+@dataclass
+class Qwen3_06BConfig:
+    vocab_size: int = 151936
+    hidden_size: int = 1024
+    intermediate_size: int = 3072
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 32768
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -100,6 +121,28 @@ class Qwen3_4BConfig:
     rope_scale = None
     final_norm: bool = True
 
+@dataclass
+class Qwen3_8BConfig:
+    vocab_size: int = 151936
+    hidden_size: int = 4096
+    intermediate_size: int = 12288
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+
 @dataclass
 class Ovis25_2BConfig:
     vocab_size: int = 151936
@@ -177,7 +220,7 @@ class Gemma3_4B_Config:
     num_key_value_heads: int = 4
     max_position_embeddings: int = 131072
     rms_norm_eps: float = 1e-6
-    rope_theta = [10000.0, 1000000.0]
+    rope_theta = [1000000.0, 10000.0]
     transformer_type: str = "gemma3"
     head_dim = 256
     rms_norm_add = True
@@ -186,10 +229,35 @@ class Gemma3_4B_Config:
     rope_dims = None
     q_norm = "gemma3"
     k_norm = "gemma3"
-    sliding_attention = [False, False, False, False, False, 1024]
-    rope_scale = [1.0, 8.0]
+    sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
+    rope_scale = [8.0, 1.0]
     final_norm: bool = True
 
+@dataclass
+class Gemma3_12B_Config:
+    vocab_size: int = 262208
+    hidden_size: int = 3840
+    intermediate_size: int = 15360
+    num_hidden_layers: int = 48
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 131072
+    rms_norm_eps: float = 1e-6
+    rope_theta = [1000000.0, 10000.0]
+    transformer_type: str = "gemma3"
+    head_dim = 256
+    rms_norm_add = True
+    mlp_activation = "gelu_pytorch_tanh"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
+    rope_scale = [8.0, 1.0]
+    final_norm: bool = True
+    vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
+    mm_tokens_per_image = 256
+
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
         super().__init__()
@@ -370,7 +438,7 @@ class TransformerBlockGemma2(nn.Module):
         self.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
 
-        if config.sliding_attention is not None:  # TODO: implement. (Not that necessary since models are trained on less than 1024 tokens)
+        if config.sliding_attention is not None:
             self.sliding_attention = config.sliding_attention[index % len(config.sliding_attention)]
         else:
             self.sliding_attention = False
@@ -387,7 +455,12 @@ class TransformerBlockGemma2(nn.Module):
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
                 if x.shape[1] > self.sliding_attention:
-                    logging.warning("Warning: sliding attention not implemented, results may be incorrect")
+                    sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype)
+                    sliding_mask.tril_(diagonal=-self.sliding_attention)
+                    if attention_mask is not None:
+                        attention_mask = attention_mask + sliding_mask
+                    else:
+                        attention_mask = sliding_mask
                 freqs_cis = freqs_cis[1]
             else:
                 freqs_cis = freqs_cis[0]
@@ -517,6 +590,41 @@ class Llama2_(nn.Module):
 
         return x, intermediate
 
+
+class Gemma3MultiModalProjector(torch.nn.Module):
+    def __init__(self, config, dtype, device, operations):
+        super().__init__()
+
+        self.mm_input_projection_weight = nn.Parameter(
+            torch.empty(config.vision_config["hidden_size"], config.hidden_size, device=device, dtype=dtype)
+        )
+
+        self.mm_soft_emb_norm = RMSNorm(config.vision_config["hidden_size"], eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
+
+        self.patches_per_image = int(config.vision_config["image_size"] // config.vision_config["patch_size"])
+        self.tokens_per_side = int(config.mm_tokens_per_image**0.5)
+        self.kernel_size = self.patches_per_image // self.tokens_per_side
+        self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size)
+
+    def forward(self, vision_outputs: torch.Tensor):
+        batch_size, _, seq_length = vision_outputs.shape
+
+        reshaped_vision_outputs = vision_outputs.transpose(1, 2)
+        reshaped_vision_outputs = reshaped_vision_outputs.reshape(
+            batch_size, seq_length, self.patches_per_image, self.patches_per_image
+        )
+        reshaped_vision_outputs = reshaped_vision_outputs.contiguous()
+
+        pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs)
+        pooled_vision_outputs = pooled_vision_outputs.flatten(2)
+        pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2)
+
+        normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs)
+
+        projected_vision_outputs = torch.matmul(normed_vision_outputs, comfy.model_management.cast_to_device(self.mm_input_projection_weight, device=normed_vision_outputs.device, dtype=normed_vision_outputs.dtype))
+        return projected_vision_outputs.type_as(vision_outputs)
+
+
 class BaseLlama:
     def get_input_embeddings(self):
         return self.model.embed_tokens
@@ -555,6 +663,15 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_06B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_06BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen3_4B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
@@ -564,6 +681,15 @@ class Qwen3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_8B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_8BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Ovis25_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
@@ -633,3 +759,21 @@ class Gemma3_4B(BaseLlama, torch.nn.Module):
 
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
+
+class Gemma3_12B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Gemma3_12B_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations)
+        self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations)
+        self.dtype = dtype
+        self.image_size = config.vision_config["image_size"]
+
+    def preprocess_embed(self, embed, device):
+        if embed["type"] == "image":
+            image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True)
+            return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None
+        return None, None

comfy/text_encoders/lt.py

@@ -1,7 +1,11 @@
 from comfy import sd1_clip
 import os
 from transformers import T5TokenizerFast
+from .spiece_tokenizer import SPieceTokenizer
 import comfy.text_encoders.genmo
+from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
+import torch
+import comfy.utils
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -16,3 +20,133 @@ class LTXVT5Tokenizer(sd1_clip.SD1Tokenizer):
 
 def ltxv_te(*args, **kwargs):
     return comfy.text_encoders.genmo.mochi_te(*args, **kwargs)
+
+
+class Gemma3_12BTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer)
+
+class Gemma3_12BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs):
+        text = llama_template.format(text)
+        text_tokens = super().tokenize_with_weights(text, return_word_ids)
+        embed_count = 0
+        for k in text_tokens:
+            tt = text_tokens[k]
+            for r in tt:
+                for i in range(len(r)):
+                    if r[i][0] == 262144:
+                        if image_embeds is not None and embed_count < image_embeds.shape[0]:
+                            r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:]
+                            embed_count += 1
+        return text_tokens
+
+class LTXAVTEModel(torch.nn.Module):
+    def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
+        super().__init__()
+        self.dtypes = set()
+        self.dtypes.add(dtype)
+
+        self.gemma3_12b = Gemma3_12BModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None)
+        self.dtypes.add(dtype_llama)
+
+        operations = self.gemma3_12b.operations # TODO
+        self.text_embedding_projection = operations.Linear(3840 * 49, 3840, bias=False, dtype=dtype, device=device)
+
+        self.audio_embeddings_connector = Embeddings1DConnector(
+            split_rope=True,
+            double_precision_rope=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.video_embeddings_connector = Embeddings1DConnector(
+            split_rope=True,
+            double_precision_rope=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def set_clip_options(self, options):
+        self.execution_device = options.get("execution_device", self.execution_device)
+        self.gemma3_12b.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.gemma3_12b.reset_clip_options()
+        self.execution_device = None
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs = token_weight_pairs["gemma3_12b"]
+
+        out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
+        out_device = out.device
+        if comfy.model_management.should_use_bf16(self.execution_device):
+            out = out.to(device=self.execution_device, dtype=torch.bfloat16)
+        out = out.movedim(1, -1).to(self.execution_device)
+        out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
+        out = out.reshape((out.shape[0], out.shape[1], -1))
+        out = self.text_embedding_projection(out)
+        out = out.float()
+        out_vid = self.video_embeddings_connector(out)[0]
+        out_audio = self.audio_embeddings_connector(out)[0]
+        out = torch.concat((out_vid, out_audio), dim=-1)
+
+        return out.to(out_device), pooled
+
+    def load_sd(self, sd):
+        if "model.layers.47.self_attn.q_norm.weight" in sd:
+            return self.gemma3_12b.load_sd(sd)
+        else:
+            sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True)
+            if len(sdo) == 0:
+                sdo = sd
+
+            missing_all = []
+            unexpected_all = []
+
+            for prefix, component in [("text_embedding_projection.", self.text_embedding_projection), ("video_embeddings_connector.", self.video_embeddings_connector), ("audio_embeddings_connector.", self.audio_embeddings_connector)]:
+                component_sd = {k.replace(prefix, ""): v for k, v in sdo.items() if k.startswith(prefix)}
+                if component_sd:
+                    missing, unexpected = component.load_state_dict(component_sd, strict=False)
+                    missing_all.extend([f"{prefix}{k}" for k in missing])
+                    unexpected_all.extend([f"{prefix}{k}" for k in unexpected])
+
+            return (missing_all, unexpected_all)
+
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        constant = 6.0
+        if comfy.model_management.should_use_bf16(device):
+            constant /= 2.0
+
+        token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        return num_tokens * constant * 1024 * 1024
+
+def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
+    class LTXAVTEModel_(LTXAVTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
+    return LTXAVTEModel_

comfy/text_encoders/lumina2.py

@@ -14,7 +14,7 @@ class Gemma2BTokenizer(sd1_clip.SDTokenizer):
 class Gemma3_4BTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, disable_weights=True, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
@@ -33,6 +33,11 @@ class Gemma2_2BModel(sd1_clip.SDClipModel):
 
 class Gemma3_4BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 class LuminaModel(sd1_clip.SD1ClipModel):

comfy/text_encoders/newbie.py

@@ -0,0 +1,62 @@
+import torch
+
+import comfy.model_management
+import comfy.text_encoders.jina_clip_2
+import comfy.text_encoders.lumina2
+
+class NewBieTokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.gemma = comfy.text_encoders.lumina2.Gemma3_4BTokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["gemma_spiece_model"]})
+        self.jina = comfy.text_encoders.jina_clip_2.JinaClip2Tokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["jina_spiece_model"]})
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = {}
+        out["gemma"] = self.gemma.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["jina"] = self.jina.tokenize_with_weights(text, return_word_ids, **kwargs)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        raise NotImplementedError
+
+    def state_dict(self):
+        return {}
+
+class NewBieTEModel(torch.nn.Module):
+    def __init__(self, dtype_gemma=None, device="cpu", dtype=None, model_options={}):
+        super().__init__()
+        dtype_gemma = comfy.model_management.pick_weight_dtype(dtype_gemma, dtype, device)
+        self.gemma = comfy.text_encoders.lumina2.Gemma3_4BModel(device=device, dtype=dtype_gemma, model_options=model_options)
+        self.jina = comfy.text_encoders.jina_clip_2.JinaClip2TextModel(device=device, dtype=dtype, model_options=model_options)
+        self.dtypes = {dtype, dtype_gemma}
+
+    def set_clip_options(self, options):
+        self.gemma.set_clip_options(options)
+        self.jina.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.gemma.reset_clip_options()
+        self.jina.reset_clip_options()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_gemma = token_weight_pairs["gemma"]
+        token_weight_pairs_jina = token_weight_pairs["jina"]
+
+        gemma_out, gemma_pooled, gemma_extra = self.gemma.encode_token_weights(token_weight_pairs_gemma)
+        jina_out, jina_pooled, jina_extra = self.jina.encode_token_weights(token_weight_pairs_jina)
+
+        return gemma_out, jina_pooled, gemma_extra
+
+    def load_sd(self, sd):
+        if "model.layers.0.self_attn.q_norm.weight" in sd:
+            return self.gemma.load_sd(sd)
+        else:
+            return self.jina.load_sd(sd)
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class NewBieTEModel_(NewBieTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            super().__init__(dtype_gemma=dtype_llama, device=device, dtype=dtype, model_options=model_options)
+    return NewBieTEModel_

comfy/text_encoders/ovis.py

@@ -61,6 +61,7 @@ def te(dtype_llama=None, llama_quantization_metadata=None):
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
                 model_options["quantization_metadata"] = llama_quantization_metadata
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return OvisTEModel_

comfy/text_encoders/pixart_t5.py

@@ -36,7 +36,7 @@ def pixart_te(dtype_t5=None, t5_quantization_metadata=None):
             if t5_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype is None:
+            if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return PixArtTEModel_

comfy/text_encoders/z_image.py

@@ -40,6 +40,7 @@ def te(dtype_llama=None, llama_quantization_metadata=None):
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
                 model_options["quantization_metadata"] = llama_quantization_metadata
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return ZImageTEModel_

comfy/utils.py

@@ -30,6 +30,7 @@ from torch.nn.functional import interpolate
 from einops import rearrange
 from comfy.cli_args import args
 import json
+import time
 
 MMAP_TORCH_FILES = args.mmap_torch_files
 DISABLE_MMAP = args.disable_mmap
@@ -610,6 +611,14 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
                         "ff_context.net.0.proj.bias": "txt_mlp.0.bias",
                         "ff_context.net.2.weight": "txt_mlp.2.weight",
                         "ff_context.net.2.bias": "txt_mlp.2.bias",
+                        "ff.linear_in.weight": "img_mlp.0.weight",  # LyCoris LoKr
+                        "ff.linear_in.bias": "img_mlp.0.bias",
+                        "ff.linear_out.weight": "img_mlp.2.weight",
+                        "ff.linear_out.bias": "img_mlp.2.bias",
+                        "ff_context.linear_in.weight": "txt_mlp.0.weight",
+                        "ff_context.linear_in.bias": "txt_mlp.0.bias",
+                        "ff_context.linear_out.weight": "txt_mlp.2.weight",
+                        "ff_context.linear_out.bias": "txt_mlp.2.bias",
                         "attn.norm_q.weight": "img_attn.norm.query_norm.scale",
                         "attn.norm_k.weight": "img_attn.norm.key_norm.scale",
                         "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale",
@@ -638,6 +647,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
                         "proj_out.bias": "linear2.bias",
                         "attn.norm_q.weight": "norm.query_norm.scale",
                         "attn.norm_k.weight": "norm.key_norm.scale",
+                        "attn.to_qkv_mlp_proj.weight": "linear1.weight", # Flux 2
+                        "attn.to_out.weight": "linear2.weight", # Flux 2
                     }
 
         for k in block_map:
@@ -928,7 +939,9 @@ def bislerp(samples, width, height):
     return result.to(orig_dtype)
 
 def lanczos(samples, width, height):
-    images = [Image.fromarray(np.clip(255. * image.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]
+    #the below API is strict and expects grayscale to be squeezed
+    samples = samples.squeeze(1) if samples.shape[1] == 1 else samples.movedim(1, -1)
+    images = [Image.fromarray(np.clip(255. * image.cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]
     images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images]
     images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images]
     result = torch.stack(images)
@@ -1097,6 +1110,10 @@ def set_progress_bar_global_hook(function):
     global PROGRESS_BAR_HOOK
     PROGRESS_BAR_HOOK = function
 
+# Throttle settings for progress bar updates to reduce WebSocket flooding
+PROGRESS_THROTTLE_MIN_INTERVAL = 0.1  # 100ms minimum between updates
+PROGRESS_THROTTLE_MIN_PERCENT = 0.5   # 0.5% minimum progress change
+
 class ProgressBar:
     def __init__(self, total, node_id=None):
         global PROGRESS_BAR_HOOK
@@ -1104,6 +1121,8 @@ class ProgressBar:
         self.current = 0
         self.hook = PROGRESS_BAR_HOOK
         self.node_id = node_id
+        self._last_update_time = 0.0
+        self._last_sent_value = -1
 
     def update_absolute(self, value, total=None, preview=None):
         if total is not None:
@@ -1112,7 +1131,29 @@ class ProgressBar:
             value = self.total
         self.current = value
         if self.hook is not None:
-            self.hook(self.current, self.total, preview, node_id=self.node_id)
+            current_time = time.perf_counter()
+            is_first = (self._last_sent_value < 0)
+            is_final = (value >= self.total)
+            has_preview = (preview is not None)
+
+            # Always send immediately for previews, first update, or final update
+            if has_preview or is_first or is_final:
+                self.hook(self.current, self.total, preview, node_id=self.node_id)
+                self._last_update_time = current_time
+                self._last_sent_value = value
+                return
+
+            # Apply throttling for regular progress updates
+            if self.total > 0:
+                percent_changed = ((value - max(0, self._last_sent_value)) / self.total) * 100
+            else:
+                percent_changed = 100
+            time_elapsed = current_time - self._last_update_time
+
+            if time_elapsed >= PROGRESS_THROTTLE_MIN_INTERVAL and percent_changed >= PROGRESS_THROTTLE_MIN_PERCENT:
+                self.hook(self.current, self.total, preview, node_id=self.node_id)
+                self._last_update_time = current_time
+                self._last_sent_value = value
 
     def update(self, value):
         self.update_absolute(self.current + value)
@@ -1198,7 +1239,7 @@ def unpack_latents(combined_latent, latent_shapes):
             combined_latent = combined_latent[:, :, cut:]
             output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:]))
     else:
-        output_tensors = combined_latent
+        output_tensors = [combined_latent]
     return output_tensors
 
 def detect_layer_quantization(state_dict, prefix):
@@ -1230,6 +1271,8 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}):
             out_sd = {}
             layers = {}
             for k in list(state_dict.keys()):
+                if k == scaled_fp8_key:
+                    continue
                 if not k.startswith(model_prefix):
                     out_sd[k] = state_dict[k]
                     continue

comfy_api/latest/__init__.py

@@ -10,7 +10,6 @@ from ._input_impl import VideoFromFile, VideoFromComponents
 from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL
 from . import _io_public as io
 from . import _ui_public as ui
-# from comfy_api.latest._resources import _RESOURCES as resources  #noqa: F401
 from comfy_execution.utils import get_executing_context
 from comfy_execution.progress import get_progress_state, PreviewImageTuple
 from PIL import Image

comfy_api/latest/_input/__init__.py

@@ -1,10 +1,12 @@
 from .basic_types import ImageInput, AudioInput, MaskInput, LatentInput
-from .video_types import VideoInput
+from .video_types import VideoInput, VideoOp, SliceOp
 
 __all__ = [
     "ImageInput",
     "AudioInput",
     "VideoInput",
+    "VideoOp",
+    "SliceOp",
     "MaskInput",
     "LatentInput",
 ]

comfy_api/latest/_input/video_types.py

@@ -1,11 +1,48 @@
 from __future__ import annotations
 from abc import ABC, abstractmethod
+from dataclasses import dataclass
 from fractions import Fraction
 from typing import Optional, Union, IO
+import copy
 import io
 import av
 from .._util import VideoContainer, VideoCodec, VideoComponents
 
+
+class VideoOp(ABC):
+    """Base class for lazy video operations."""
+
+    @abstractmethod
+    def apply(self, components: VideoComponents) -> VideoComponents:
+        pass
+
+    @abstractmethod
+    def compute_frame_count(self, input_frame_count: int) -> int:
+        pass
+
+
+@dataclass(frozen=True)
+class SliceOp(VideoOp):
+    """Extract a range of frames from the video."""
+    start_frame: int
+    frame_count: int
+
+    def apply(self, components: VideoComponents) -> VideoComponents:
+        total = components.images.shape[0]
+        start = max(0, min(self.start_frame, total))
+        end = min(start + self.frame_count, total)
+        return VideoComponents(
+            images=components.images[start:end],
+            audio=components.audio,
+            frame_rate=components.frame_rate,
+            metadata=getattr(components, 'metadata', None),
+        )
+
+    def compute_frame_count(self, input_frame_count: int) -> int:
+        start = max(0, min(self.start_frame, input_frame_count))
+        return min(self.frame_count, input_frame_count - start)
+
+
 class VideoInput(ABC):
     """
     Abstract base class for video input types.
@@ -21,6 +58,12 @@ class VideoInput(ABC):
         """
         pass
 
+    def sliced(self, start_frame: int, frame_count: int) -> "VideoInput":
+        """Return a copy of this video with a slice operation appended."""
+        new = copy.copy(self)
+        new._operations = getattr(self, '_operations', []) + [SliceOp(start_frame, frame_count)]
+        return new
+
     @abstractmethod
     def save_to(
         self,

comfy_api/latest/_input_impl/__init__.py

@@ -1,7 +1,8 @@
 from .video_types import VideoFromFile, VideoFromComponents
+from .._input import SliceOp
 
 __all__ = [
-    # Implementations
     "VideoFromFile",
     "VideoFromComponents",
+    "SliceOp",
 ]

comfy_api/latest/_input_impl/video_types.py

@@ -3,7 +3,7 @@ from av.container import InputContainer
 from av.subtitles.stream import SubtitleStream
 from fractions import Fraction
 from typing import Optional
-from .._input import AudioInput, VideoInput
+from .._input import AudioInput, VideoInput, VideoOp
 import av
 import io
 import json
@@ -63,6 +63,8 @@ class VideoFromFile(VideoInput):
         containing the file contents.
         """
         self.__file = file
+        self._operations: list[VideoOp] = []
+        self.__materialized: Optional[VideoFromComponents] = None
 
     def get_stream_source(self) -> str | io.BytesIO:
         """
@@ -161,6 +163,10 @@ class VideoFromFile(VideoInput):
 
             if frame_count == 0:
                 raise ValueError(f"Could not determine frame count for file '{self.__file}'")
+
+            # Apply operations to get final frame count
+            for op in self._operations:
+                frame_count = op.compute_frame_count(frame_count)
             return frame_count
 
     def get_frame_rate(self) -> Fraction:
@@ -239,10 +245,18 @@ class VideoFromFile(VideoInput):
         return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
 
     def get_components(self) -> VideoComponents:
+        if self.__materialized is not None:
+            return self.__materialized.get_components()
+
         if isinstance(self.__file, io.BytesIO):
             self.__file.seek(0)  # Reset the BytesIO object to the beginning
         with av.open(self.__file, mode='r') as container:
-            return self.get_components_internal(container)
+            components = self.get_components_internal(container)
+            for op in self._operations:
+                components = op.apply(components)
+            self.__materialized = VideoFromComponents(components)
+            self._operations = []
+            return components
         raise ValueError(f"No video stream found in file '{self.__file}'")
 
     def save_to(
@@ -317,14 +331,27 @@ class VideoFromComponents(VideoInput):
 
     def __init__(self, components: VideoComponents):
         self.__components = components
+        self._operations: list[VideoOp] = []
 
     def get_components(self) -> VideoComponents:
+        if self._operations:
+            components = self.__components
+            for op in self._operations:
+                components = op.apply(components)
+            self.__components = components
+            self._operations = []
         return VideoComponents(
             images=self.__components.images,
             audio=self.__components.audio,
             frame_rate=self.__components.frame_rate
         )
 
+    def get_frame_count(self) -> int:
+        count = int(self.__components.images.shape[0])
+        for op in self._operations:
+            count = op.compute_frame_count(count)
+        return count
+
     def save_to(
         self,
         path: str,
@@ -332,6 +359,9 @@ class VideoFromComponents(VideoInput):
         codec: VideoCodec = VideoCodec.AUTO,
         metadata: Optional[dict] = None
     ):
+        # Materialize ops before saving
+        components = self.get_components()
+
         if format != VideoContainer.AUTO and format != VideoContainer.MP4:
             raise ValueError("Only MP4 format is supported for now")
         if codec != VideoCodec.AUTO and codec != VideoCodec.H264:
@@ -345,22 +375,22 @@ class VideoFromComponents(VideoInput):
                 for key, value in metadata.items():
                     output.metadata[key] = json.dumps(value)
 
-            frame_rate = Fraction(round(self.__components.frame_rate * 1000), 1000)
+            frame_rate = Fraction(round(components.frame_rate * 1000), 1000)
             # Create a video stream
             video_stream = output.add_stream('h264', rate=frame_rate)
-            video_stream.width = self.__components.images.shape[2]
-            video_stream.height = self.__components.images.shape[1]
+            video_stream.width = components.images.shape[2]
+            video_stream.height = components.images.shape[1]
             video_stream.pix_fmt = 'yuv420p'
 
             # Create an audio stream
             audio_sample_rate = 1
             audio_stream: Optional[av.AudioStream] = None
-            if self.__components.audio:
-                audio_sample_rate = int(self.__components.audio['sample_rate'])
+            if components.audio:
+                audio_sample_rate = int(components.audio['sample_rate'])
                 audio_stream = output.add_stream('aac', rate=audio_sample_rate)
 
             # Encode video
-            for i, frame in enumerate(self.__components.images):
+            for i, frame in enumerate(components.images):
                 img = (frame * 255).clamp(0, 255).byte().cpu().numpy() # shape: (H, W, 3)
                 frame = av.VideoFrame.from_ndarray(img, format='rgb24')
                 frame = frame.reformat(format='yuv420p')  # Convert to YUV420P as required by h264
@@ -371,10 +401,10 @@ class VideoFromComponents(VideoInput):
             packet = video_stream.encode(None)
             output.mux(packet)
 
-            if audio_stream and self.__components.audio:
-                waveform = self.__components.audio['waveform']
-                waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
-                frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo')
+            if audio_stream and components.audio:
+                waveform = components.audio['waveform']
+                waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * components.images.shape[0])]
+                frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().cpu().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo')
                 frame.sample_rate = audio_sample_rate
                 frame.pts = 0
                 output.mux(audio_stream.encode(frame))

comfy_api/latest/_io.py

@@ -26,11 +26,9 @@ if TYPE_CHECKING:
     from comfy_api.input import VideoInput
 from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class,
     prune_dict, shallow_clone_class)
-from ._resources import Resources, ResourcesLocal
 from comfy_execution.graph_utils import ExecutionBlocker
-from ._util import MESH, VOXEL
+from ._util import MESH, VOXEL, SVG as _SVG
 
-# from comfy_extras.nodes_images import SVG as SVG_ # NOTE: needs to be moved before can be imported due to circular reference
 
 class FolderType(str, Enum):
     input = "input"
@@ -77,16 +75,6 @@ class NumberDisplay(str, Enum):
     slider = "slider"
 
 
-class _StringIOType(str):
-    def __ne__(self, value: object) -> bool:
-        if self == "*" or value == "*":
-            return False
-        if not isinstance(value, str):
-            return True
-        a = frozenset(self.split(","))
-        b = frozenset(value.split(","))
-        return not (b.issubset(a) or a.issubset(b))
-
 class _ComfyType(ABC):
     Type = Any
     io_type: str = None
@@ -126,8 +114,7 @@ def comfytype(io_type: str, **kwargs):
             new_cls.__module__ = cls.__module__
             new_cls.__doc__ = cls.__doc__
             # assign ComfyType attributes, if needed
-        # NOTE: use __ne__ trick for io_type (see node_typing.IO.__ne__ for details)
-        new_cls.io_type = _StringIOType(io_type)
+        new_cls.io_type = io_type
         if hasattr(new_cls, "Input") and new_cls.Input is not None:
             new_cls.Input.Parent = new_cls
         if hasattr(new_cls, "Output") and new_cls.Output is not None:
@@ -166,7 +153,7 @@ class Input(_IO_V3):
     '''
     Base class for a V3 Input.
     '''
-    def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None):
+    def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
         super().__init__()
         self.id = id
         self.display_name = display_name
@@ -174,6 +161,8 @@ class Input(_IO_V3):
         self.tooltip = tooltip
         self.lazy = lazy
         self.extra_dict = extra_dict if extra_dict is not None else {}
+        self.rawLink = raw_link
+        self.advanced = advanced
 
     def as_dict(self):
         return prune_dict({
@@ -181,10 +170,12 @@ class Input(_IO_V3):
             "optional": self.optional,
             "tooltip": self.tooltip,
             "lazy": self.lazy,
+            "rawLink": self.rawLink,
+            "advanced": self.advanced,
         }) | prune_dict(self.extra_dict)
 
     def get_io_type(self):
-        return _StringIOType(self.io_type)
+        return self.io_type
 
     def get_all(self) -> list[Input]:
         return [self]
@@ -195,8 +186,8 @@ class WidgetInput(Input):
     '''
     def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
                  default: Any=None,
-                 socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None):
-        super().__init__(id, display_name, optional, tooltip, lazy, extra_dict)
+                 socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+        super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link, advanced)
         self.default = default
         self.socketless = socketless
         self.widget_type = widget_type
@@ -218,13 +209,14 @@ class Output(_IO_V3):
     def __init__(self, id: str=None, display_name: str=None, tooltip: str=None,
                  is_output_list=False):
         self.id = id
-        self.display_name = display_name
+        self.display_name = display_name if display_name else id
         self.tooltip = tooltip
         self.is_output_list = is_output_list
 
     def as_dict(self):
+        display_name = self.display_name if self.display_name else self.id
         return prune_dict({
-            "display_name": self.display_name,
+            "display_name": display_name,
             "tooltip": self.tooltip,
             "is_output_list": self.is_output_list,
         })
@@ -252,8 +244,8 @@ class Boolean(ComfyTypeIO):
         '''Boolean input.'''
         def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
                     default: bool=None, label_on: str=None, label_off: str=None,
-                    socketless: bool=None, force_input: bool=None):
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input)
+                    socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced)
             self.label_on = label_on
             self.label_off = label_off
             self.default: bool
@@ -272,8 +264,8 @@ class Int(ComfyTypeIO):
         '''Integer input.'''
         def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
                     default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None,
-                    display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None):
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input)
+                    display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced)
             self.min = min
             self.max = max
             self.step = step
@@ -298,8 +290,8 @@ class Float(ComfyTypeIO):
         '''Float input.'''
         def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
                     default: float=None, min: float=None, max: float=None, step: float=None, round: float=None,
-                    display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None):
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input)
+                    display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced)
             self.min = min
             self.max = max
             self.step = step
@@ -324,8 +316,8 @@ class String(ComfyTypeIO):
         '''String input.'''
         def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
                     multiline=False, placeholder: str=None, default: str=None, dynamic_prompts: bool=None,
-                    socketless: bool=None, force_input: bool=None):
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input)
+                    socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced)
             self.multiline = multiline
             self.placeholder = placeholder
             self.dynamic_prompts = dynamic_prompts
@@ -358,12 +350,15 @@ class Combo(ComfyTypeIO):
             image_folder: FolderType=None,
             remote: RemoteOptions=None,
             socketless: bool=None,
+            extra_dict=None,
+            raw_link: bool=None,
+            advanced: bool=None,
         ):
             if isinstance(options, type) and issubclass(options, Enum):
                 options = [v.value for v in options]
             if isinstance(default, Enum):
                 default = default.value
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless)
+            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link, advanced)
             self.multiselect = False
             self.options = options
             self.control_after_generate = control_after_generate
@@ -387,10 +382,6 @@ class Combo(ComfyTypeIO):
             super().__init__(id, display_name, tooltip, is_output_list)
             self.options = options if options is not None else []
 
-        @property
-        def io_type(self):
-            return self.options
-
 @comfytype(io_type="COMBO")
 class MultiCombo(ComfyTypeI):
     '''Multiselect Combo input (dropdown for selecting potentially more than one value).'''
@@ -399,8 +390,8 @@ class MultiCombo(ComfyTypeI):
     class Input(Combo.Input):
         def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
                     default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None,
-                    socketless: bool=None):
-            super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless)
+                    socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+            super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link, advanced=advanced)
             self.multiselect = True
             self.placeholder = placeholder
             self.chip = chip
@@ -433,9 +424,9 @@ class Webcam(ComfyTypeIO):
         Type = str
         def __init__(
                 self, id: str, display_name: str=None, optional=False,
-                tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None
+                tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None
         ):
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless)
+            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link, advanced)
 
 
 @comfytype(io_type="MASK")
@@ -656,7 +647,7 @@ class Video(ComfyTypeIO):
 
 @comfytype(io_type="SVG")
 class SVG(ComfyTypeIO):
-    Type = Any # TODO: SVG class is defined in comfy_extras/nodes_images.py, causing circular reference; should be moved to somewhere else before referenced directly in v3
+    Type = _SVG
 
 @comfytype(io_type="LORA_MODEL")
 class LoraModel(ComfyTypeIO):
@@ -763,7 +754,7 @@ class AnyType(ComfyTypeIO):
     Type = Any
 
 @comfytype(io_type="MODEL_PATCH")
-class MODEL_PATCH(ComfyTypeIO):
+class ModelPatch(ComfyTypeIO):
     Type = Any
 
 @comfytype(io_type="AUDIO_ENCODER")
@@ -788,7 +779,7 @@ class MultiType:
         '''
         Input that permits more than one input type; if `id` is an instance of `ComfyType.Input`, then that input will be used to create a widget (if applicable) with overridden values.
         '''
-        def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None):
+        def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
             # if id is an Input, then use that Input with overridden values
             self.input_override = None
             if isinstance(id, Input):
@@ -801,7 +792,7 @@ class MultiType:
                 # if is a widget input, make sure widget_type is set appropriately
                 if isinstance(self.input_override, WidgetInput):
                     self.input_override.widget_type = self.input_override.get_io_type()
-            super().__init__(id, display_name, optional, tooltip, lazy, extra_dict)
+            super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link, advanced)
             self._io_types = types
 
         @property
@@ -855,8 +846,8 @@ class MatchType(ComfyTypeIO):
 
     class Input(Input):
         def __init__(self, id: str, template: MatchType.Template,
-                    display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None):
-            super().__init__(id, display_name, optional, tooltip, lazy, extra_dict)
+                    display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
+            super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link, advanced)
             self.template = template
 
         def as_dict(self):
@@ -867,6 +858,8 @@ class MatchType(ComfyTypeIO):
     class Output(Output):
         def __init__(self, template: MatchType.Template, id: str=None, display_name: str=None, tooltip: str=None,
                      is_output_list=False):
+            if not id and not display_name:
+                display_name = "MATCHTYPE"
             super().__init__(id, display_name, tooltip, is_output_list)
             self.template = template
 
@@ -879,24 +872,30 @@ class DynamicInput(Input, ABC):
     '''
     Abstract class for dynamic input registration.
     '''
-    def get_dynamic(self) -> list[Input]:
-        return []
-
-    def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''):
-        pass
+    pass
 
 
 class DynamicOutput(Output, ABC):
     '''
     Abstract class for dynamic output registration.
     '''
-    def __init__(self, id: str=None, display_name: str=None, tooltip: str=None,
-                 is_output_list=False):
-        super().__init__(id, display_name, tooltip, is_output_list)
+    pass
 
-    def get_dynamic(self) -> list[Output]:
-        return []
 
+def handle_prefix(prefix_list: list[str] | None, id: str | None = None) -> list[str]:
+    if prefix_list is None:
+        prefix_list = []
+    if id is not None:
+        prefix_list = prefix_list + [id]
+    return prefix_list
+
+def finalize_prefix(prefix_list: list[str] | None, id: str | None = None) -> str:
+    assert not (prefix_list is None and id is None)
+    if prefix_list is None:
+        return id
+    elif id is not None:
+        prefix_list = prefix_list + [id]
+    return ".".join(prefix_list)
 
 @comfytype(io_type="COMFY_AUTOGROW_V3")
 class Autogrow(ComfyTypeI):
@@ -933,14 +932,6 @@ class Autogrow(ComfyTypeI):
         def validate(self):
             self.input.validate()
 
-        def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''):
-            real_inputs = []
-            for name, input in self.cached_inputs.items():
-                if name in live_inputs:
-                    real_inputs.append(input)
-            add_to_input_dict_v1(d, real_inputs, live_inputs, curr_prefix)
-            add_dynamic_id_mapping(d, real_inputs, curr_prefix)
-
     class TemplatePrefix(_AutogrowTemplate):
         def __init__(self, input: Input, prefix: str, min: int=1, max: int=10):
             super().__init__(input)
@@ -985,22 +976,63 @@ class Autogrow(ComfyTypeI):
                 "template": self.template.as_dict(),
             })
 
-        def get_dynamic(self) -> list[Input]:
-            return self.template.get_all()
-
         def get_all(self) -> list[Input]:
             return [self] + self.template.get_all()
 
         def validate(self):
             self.template.validate()
 
-        def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''):
-            curr_prefix = f"{curr_prefix}{self.id}."
-            # need to remove self from expected inputs dictionary; replaced by template inputs in frontend
-            for inner_dict in d.values():
-                if self.id in inner_dict:
-                    del inner_dict[self.id]
-            self.template.expand_schema_for_dynamic(d, live_inputs, curr_prefix)
+    @staticmethod
+    def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None):
+        # NOTE: purposely do not include self in out_dict; instead use only the template inputs
+        # need to figure out names based on template type
+        is_names = ("names" in value[1]["template"])
+        is_prefix = ("prefix" in value[1]["template"])
+        input = value[1]["template"]["input"]
+        if is_names:
+            min = value[1]["template"]["min"]
+            names = value[1]["template"]["names"]
+            max = len(names)
+        elif is_prefix:
+            prefix = value[1]["template"]["prefix"]
+            min = value[1]["template"]["min"]
+            max = value[1]["template"]["max"]
+            names = [f"{prefix}{i}" for i in range(max)]
+        # need to create a new input based on the contents of input
+        template_input = None
+        template_required = True
+        for _input_type, dict_input in input.items():
+            # for now, get just the first value from dict_input; if not required, min can be ignored
+            if len(dict_input) == 0:
+                continue
+            template_input = list(dict_input.values())[0]
+            template_required = _input_type == "required"
+            break
+        if template_input is None:
+            raise Exception("template_input could not be determined from required or optional; this should never happen.")
+        new_dict = {}
+        new_dict_added_to = False
+        # first, add possible inputs into out_dict
+        for i, name in enumerate(names):
+            expected_id = finalize_prefix(curr_prefix, name)
+            # required
+            if i < min and template_required:
+                out_dict["required"][expected_id] = template_input
+                type_dict = new_dict.setdefault("required", {})
+            # optional
+            else:
+                out_dict["optional"][expected_id] = template_input
+                type_dict = new_dict.setdefault("optional", {})
+            if expected_id in live_inputs:
+                # NOTE: prefix gets added in parse_class_inputs
+                type_dict[name] = template_input
+                new_dict_added_to = True
+        # account for the edge case that all inputs are optional and no values are received
+        if not new_dict_added_to:
+            finalized_prefix = finalize_prefix(curr_prefix)
+            out_dict["dynamic_paths"][finalized_prefix] = finalized_prefix
+            out_dict["dynamic_paths_default_value"][finalized_prefix] = DynamicPathsDefaultValue.EMPTY_DICT
+        parse_class_inputs(out_dict, live_inputs, new_dict, curr_prefix)
 
 @comfytype(io_type="COMFY_DYNAMICCOMBO_V3")
 class DynamicCombo(ComfyTypeI):
@@ -1023,23 +1055,6 @@ class DynamicCombo(ComfyTypeI):
             super().__init__(id, display_name, optional, tooltip, lazy, extra_dict)
             self.options = options
 
-        def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''):
-            # check if dynamic input's id is in live_inputs
-            if self.id in live_inputs:
-                curr_prefix = f"{curr_prefix}{self.id}."
-                key = live_inputs[self.id]
-                selected_option = None
-                for option in self.options:
-                    if option.key == key:
-                        selected_option = option
-                        break
-                if selected_option is not None:
-                    add_to_input_dict_v1(d, selected_option.inputs, live_inputs, curr_prefix)
-                    add_dynamic_id_mapping(d, selected_option.inputs, curr_prefix, self)
-
-        def get_dynamic(self) -> list[Input]:
-            return [input for option in self.options for input in option.inputs]
-
         def get_all(self) -> list[Input]:
             return [self] + [input for option in self.options for input in option.inputs]
 
@@ -1054,6 +1069,24 @@ class DynamicCombo(ComfyTypeI):
                 for input in option.inputs:
                     input.validate()
 
+    @staticmethod
+    def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None):
+        finalized_id = finalize_prefix(curr_prefix)
+        if finalized_id in live_inputs:
+            key = live_inputs[finalized_id]
+            selected_option = None
+            # get options from dict
+            options: list[dict[str, str | dict[str, Any]]] = value[1]["options"]
+            for option in options:
+                if option["key"] == key:
+                    selected_option = option
+                    break
+            if selected_option is not None:
+                parse_class_inputs(out_dict, live_inputs, selected_option["inputs"], curr_prefix)
+                # add self to inputs
+                out_dict[input_type][finalized_id] = value
+                out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1])
+
 @comfytype(io_type="COMFY_DYNAMICSLOT_V3")
 class DynamicSlot(ComfyTypeI):
     Type = dict[str, Any]
@@ -1076,17 +1109,8 @@ class DynamicSlot(ComfyTypeI):
                 self.force_input = True
                 self.slot.force_input = True
 
-        def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''):
-            if self.id in live_inputs:
-                curr_prefix = f"{curr_prefix}{self.id}."
-                add_to_input_dict_v1(d, self.inputs, live_inputs, curr_prefix)
-                add_dynamic_id_mapping(d, [self.slot] + self.inputs, curr_prefix)
-
-        def get_dynamic(self) -> list[Input]:
-            return [self.slot] + self.inputs
-
         def get_all(self) -> list[Input]:
-            return [self] + [self.slot] + self.inputs
+            return [self.slot] + self.inputs
 
         def as_dict(self):
             return super().as_dict() | prune_dict({
@@ -1100,17 +1124,55 @@ class DynamicSlot(ComfyTypeI):
             for input in self.inputs:
                 input.validate()
 
-def add_dynamic_id_mapping(d: dict[str, Any], inputs: list[Input], curr_prefix: str, self: DynamicInput=None):
-    dynamic = d.setdefault("dynamic_paths", {})
-    if self is not None:
-        dynamic[self.id] = f"{curr_prefix}{self.id}"
-    for i in inputs:
-        if not isinstance(i, DynamicInput):
-            dynamic[f"{i.id}"] = f"{curr_prefix}{i.id}"
+    @staticmethod
+    def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None):
+        finalized_id = finalize_prefix(curr_prefix)
+        if finalized_id in live_inputs:
+            inputs = value[1]["inputs"]
+            parse_class_inputs(out_dict, live_inputs, inputs, curr_prefix)
+            # add self to inputs
+            out_dict[input_type][finalized_id] = value
+            out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1])
+
+@comfytype(io_type="IMAGECOMPARE")
+class ImageCompare(ComfyTypeI):
+  Type = dict
+
+  class Input(WidgetInput):
+      def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None,
+                   socketless: bool=True, advanced: bool=None):
+          super().__init__(id, display_name, optional, tooltip, None, None, socketless, None, None, None, None, advanced)
+
+      def as_dict(self):
+          return super().as_dict()
+
+DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {}
+def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]):
+    DYNAMIC_INPUT_LOOKUP[io_type] = func
+
+def get_dynamic_input_func(io_type: str) -> Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]:
+    return DYNAMIC_INPUT_LOOKUP[io_type]
+
+def setup_dynamic_input_funcs():
+    # Autogrow.Input
+    register_dynamic_input_func(Autogrow.io_type, Autogrow._expand_schema_for_dynamic)
+    # DynamicCombo.Input
+    register_dynamic_input_func(DynamicCombo.io_type, DynamicCombo._expand_schema_for_dynamic)
+    # DynamicSlot.Input
+    register_dynamic_input_func(DynamicSlot.io_type, DynamicSlot._expand_schema_for_dynamic)
+
+if len(DYNAMIC_INPUT_LOOKUP) == 0:
+    setup_dynamic_input_funcs()
 
 class V3Data(TypedDict):
     hidden_inputs: dict[str, Any]
+    'Dictionary where the keys are the hidden input ids and the values are the values of the hidden inputs.'
     dynamic_paths: dict[str, Any]
+    'Dictionary where the keys are the input ids and the values dictate how to turn the inputs into a nested dictionary.'
+    dynamic_paths_default_value: dict[str, Any]
+    'Dictionary where the keys are the input ids and the values are a string from DynamicPathsDefaultValue for the inputs if value is None.'
+    create_dynamic_tuple: bool
+    'When True, the value of the dynamic input will be in the format (value, path_key).'
 
 class HiddenHolder:
     def __init__(self, unique_id: str, prompt: Any,
@@ -1146,6 +1208,10 @@ class HiddenHolder:
             api_key_comfy_org=d.get(Hidden.api_key_comfy_org, None),
         )
 
+    @classmethod
+    def from_v3_data(cls, v3_data: V3Data | None) -> HiddenHolder:
+        return cls.from_dict(v3_data["hidden_inputs"] if v3_data else None)
+
 class Hidden(str, Enum):
     '''
     Enumerator for requesting hidden variables in nodes.
@@ -1182,6 +1248,8 @@ class NodeInfoV1:
     deprecated: bool=None
     experimental: bool=None
     api_node: bool=None
+    price_badge: dict | None = None
+    search_aliases: list[str]=None
 
 @dataclass
 class NodeInfoV3:
@@ -1191,11 +1259,77 @@ class NodeInfoV3:
     name: str=None
     display_name: str=None
     description: str=None
+    python_module: Any = None
     category: str=None
     output_node: bool=None
     deprecated: bool=None
     experimental: bool=None
     api_node: bool=None
+    price_badge: dict | None = None
+
+
+@dataclass
+class PriceBadgeDepends:
+    widgets: list[str] = field(default_factory=list)
+    inputs: list[str] = field(default_factory=list)
+    input_groups: list[str] = field(default_factory=list)
+
+    def validate(self) -> None:
+        if not isinstance(self.widgets, list) or any(not isinstance(x, str) for x in self.widgets):
+            raise ValueError("PriceBadgeDepends.widgets must be a list[str].")
+        if not isinstance(self.inputs, list) or any(not isinstance(x, str) for x in self.inputs):
+            raise ValueError("PriceBadgeDepends.inputs must be a list[str].")
+        if not isinstance(self.input_groups, list) or any(not isinstance(x, str) for x in self.input_groups):
+            raise ValueError("PriceBadgeDepends.input_groups must be a list[str].")
+
+    def as_dict(self, schema_inputs: list["Input"]) -> dict[str, Any]:
+        # Build lookup: widget_id -> io_type
+        input_types: dict[str, str] = {}
+        for inp in schema_inputs:
+            all_inputs = inp.get_all()
+            input_types[inp.id] = inp.get_io_type()  # First input is always the parent itself
+            for nested_inp in all_inputs[1:]:
+                # For DynamicCombo/DynamicSlot, nested inputs are prefixed with parent ID
+                # to match frontend naming convention (e.g., "should_texture.enable_pbr")
+                prefixed_id = f"{inp.id}.{nested_inp.id}"
+                input_types[prefixed_id] = nested_inp.get_io_type()
+
+        # Enrich widgets with type information, raising error for unknown widgets
+        widgets_data: list[dict[str, str]] = []
+        for w in self.widgets:
+            if w not in input_types:
+                raise ValueError(
+                    f"PriceBadge depends_on.widgets references unknown widget '{w}'. "
+                    f"Available widgets: {list(input_types.keys())}"
+                )
+            widgets_data.append({"name": w, "type": input_types[w]})
+
+        return {
+            "widgets": widgets_data,
+            "inputs": self.inputs,
+            "input_groups": self.input_groups,
+        }
+
+
+@dataclass
+class PriceBadge:
+    expr: str
+    depends_on: PriceBadgeDepends = field(default_factory=PriceBadgeDepends)
+    engine: str = field(default="jsonata")
+
+    def validate(self) -> None:
+        if self.engine != "jsonata":
+            raise ValueError(f"Unsupported PriceBadge.engine '{self.engine}'. Only 'jsonata' is supported.")
+        if not isinstance(self.expr, str) or not self.expr.strip():
+            raise ValueError("PriceBadge.expr must be a non-empty string.")
+        self.depends_on.validate()
+
+    def as_dict(self, schema_inputs: list["Input"]) -> dict[str, Any]:
+        return {
+            "engine": self.engine,
+            "depends_on": self.depends_on.as_dict(schema_inputs),
+            "expr": self.expr,
+        }
 
 
 @dataclass
@@ -1213,6 +1347,8 @@ class Schema:
     hidden: list[Hidden] = field(default_factory=list)
     description: str=""
     """Node description, shown as a tooltip when hovering over the node."""
+    search_aliases: list[str] = field(default_factory=list)
+    """Alternative names for search. Useful for synonyms, abbreviations, or old names after renaming."""
     is_input_list: bool = False
     """A flag indicating if this node implements the additional code necessary to deal with OUTPUT_IS_LIST nodes.
 
@@ -1241,6 +1377,8 @@ class Schema:
     """Flags a node as experimental, informing users that it may change or not work as expected."""
     is_api_node: bool=False
     """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview."""
+    price_badge: PriceBadge | None = None
+    """Optional client-evaluated pricing badge declaration for this node."""
     not_idempotent: bool=False
     """Flags a node as not idempotent; when True, the node will run and not reuse the cached outputs when identical inputs are provided on a different node in the graph."""
     enable_expand: bool=False
@@ -1251,61 +1389,58 @@ class Schema:
         - verify ids on inputs and outputs are unique - both internally and in relation to each other
         '''
         nested_inputs: list[Input] = []
-        if self.inputs is not None:
-            for input in self.inputs:
+        for input in self.inputs:
+            if not isinstance(input, DynamicInput):
                 nested_inputs.extend(input.get_all())
-        input_ids = [i.id for i in nested_inputs] if nested_inputs is not None else []
-        output_ids = [o.id for o in self.outputs] if self.outputs is not None else []
+        input_ids = [i.id for i in nested_inputs]
+        output_ids = [o.id for o in self.outputs]
         input_set = set(input_ids)
         output_set = set(output_ids)
-        issues = []
+        issues: list[str] = []
         # verify ids are unique per list
         if len(input_set) != len(input_ids):
             issues.append(f"Input ids must be unique, but {[item for item, count in Counter(input_ids).items() if count > 1]} are not.")
         if len(output_set) != len(output_ids):
             issues.append(f"Output ids must be unique, but {[item for item, count in Counter(output_ids).items() if count > 1]} are not.")
-        # verify ids are unique between lists
-        intersection = input_set & output_set
-        if len(intersection) > 0:
-            issues.append(f"Ids must be unique between inputs and outputs, but {intersection} are not.")
         if len(issues) > 0:
             raise ValueError("\n".join(issues))
         # validate inputs and outputs
-        if self.inputs is not None:
-            for input in self.inputs:
-                input.validate()
-        if self.outputs is not None:
-            for output in self.outputs:
-                output.validate()
+        for input in self.inputs:
+            input.validate()
+        for output in self.outputs:
+            output.validate()
+        if self.price_badge is not None:
+            self.price_badge.validate()
 
     def finalize(self):
         """Add hidden based on selected schema options, and give outputs without ids default ids."""
+        # ensure inputs, outputs, and hidden are lists
+        if self.inputs is None:
+            self.inputs = []
+        if self.outputs is None:
+            self.outputs = []
+        if self.hidden is None:
+            self.hidden = []
         # if is an api_node, will need key-related hidden
         if self.is_api_node:
-            if self.hidden is None:
-                self.hidden = []
             if Hidden.auth_token_comfy_org not in self.hidden:
                 self.hidden.append(Hidden.auth_token_comfy_org)
             if Hidden.api_key_comfy_org not in self.hidden:
                 self.hidden.append(Hidden.api_key_comfy_org)
         # if is an output_node, will need prompt and extra_pnginfo
         if self.is_output_node:
-            if self.hidden is None:
-                self.hidden = []
             if Hidden.prompt not in self.hidden:
                 self.hidden.append(Hidden.prompt)
             if Hidden.extra_pnginfo not in self.hidden:
                 self.hidden.append(Hidden.extra_pnginfo)
         # give outputs without ids default ids
-        if self.outputs is not None:
-            for i, output in enumerate(self.outputs):
-                if output.id is None:
-                    output.id = f"_{i}_{output.io_type}_"
+        for i, output in enumerate(self.outputs):
+            if output.id is None:
+                output.id = f"_{i}_{output.io_type}_"
 
-    def get_v1_info(self, cls, live_inputs: dict[str, Any]=None) -> NodeInfoV1:
-        # NOTE: live_inputs will not be used anymore very soon and this will be done another way
+    def get_v1_info(self, cls) -> NodeInfoV1:
         # get V1 inputs
-        input = create_input_dict_v1(self.inputs, live_inputs)
+        input = create_input_dict_v1(self.inputs)
         if self.hidden:
             for hidden in self.hidden:
                 input.setdefault("hidden", {})[hidden.name] = (hidden.value,)
@@ -1349,7 +1484,9 @@ class Schema:
             deprecated=self.is_deprecated,
             experimental=self.is_experimental,
             api_node=self.is_api_node,
-            python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes")
+            python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"),
+            price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None,
+            search_aliases=self.search_aliases if self.search_aliases else None,
         )
         return info
 
@@ -1381,48 +1518,82 @@ class Schema:
             deprecated=self.is_deprecated,
             experimental=self.is_experimental,
             api_node=self.is_api_node,
-            python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes")
+            python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"),
+            price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None,
         )
         return info
 
+def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], include_hidden=False) -> tuple[dict[str, Any], V3Data]:
+    out_dict = {
+        "required": {},
+        "optional": {},
+        "dynamic_paths": {},
+        "dynamic_paths_default_value": {},
+    }
+    d = d.copy()
+    # ignore hidden for parsing
+    hidden = d.pop("hidden", None)
+    parse_class_inputs(out_dict, live_inputs, d)
+    if hidden is not None and include_hidden:
+        out_dict["hidden"] = hidden
+    v3_data = {}
+    dynamic_paths = out_dict.pop("dynamic_paths", None)
+    if dynamic_paths is not None and len(dynamic_paths) > 0:
+        v3_data["dynamic_paths"] = dynamic_paths
+    # this list is used for autogrow, in the case all inputs are optional and no values are passed
+    dynamic_paths_default_value = out_dict.pop("dynamic_paths_default_value", None)
+    if dynamic_paths_default_value is not None and len(dynamic_paths_default_value) > 0:
+        v3_data["dynamic_paths_default_value"] = dynamic_paths_default_value
+    return out_dict, hidden, v3_data
 
-def create_input_dict_v1(inputs: list[Input], live_inputs: dict[str, Any]=None) -> dict:
+def parse_class_inputs(out_dict: dict[str, Any], live_inputs: dict[str, Any], curr_dict: dict[str, Any], curr_prefix: list[str] | None=None) -> None:
+    for input_type, inner_d in curr_dict.items():
+        for id, value in inner_d.items():
+            io_type = value[0]
+            if io_type in DYNAMIC_INPUT_LOOKUP:
+                # dynamic inputs need to be handled with lookup functions
+                dynamic_input_func = get_dynamic_input_func(io_type)
+                new_prefix = handle_prefix(curr_prefix, id)
+                dynamic_input_func(out_dict, live_inputs, value, input_type, new_prefix)
+            else:
+                # non-dynamic inputs get directly transferred
+                finalized_id = finalize_prefix(curr_prefix, id)
+                out_dict[input_type][finalized_id] = value
+                if curr_prefix:
+                    out_dict["dynamic_paths"][finalized_id] = finalized_id
+
+def create_input_dict_v1(inputs: list[Input]) -> dict:
     input = {
         "required": {}
     }
-    add_to_input_dict_v1(input, inputs, live_inputs)
+    for i in inputs:
+        add_to_dict_v1(i, input)
     return input
 
-def add_to_input_dict_v1(d: dict[str, Any], inputs: list[Input], live_inputs: dict[str, Any]=None, curr_prefix=''):
-    for i in inputs:
-        if isinstance(i, DynamicInput):
-            add_to_dict_v1(i, d)
-            if live_inputs is not None:
-                i.expand_schema_for_dynamic(d, live_inputs, curr_prefix)
-        else:
-            add_to_dict_v1(i, d)
-
-def add_to_dict_v1(i: Input, d: dict, dynamic_dict: dict=None):
+def add_to_dict_v1(i: Input, d: dict):
     key = "optional" if i.optional else "required"
     as_dict = i.as_dict()
     # for v1, we don't want to include the optional key
     as_dict.pop("optional", None)
-    if dynamic_dict is None:
-        value = (i.get_io_type(), as_dict)
-    else:
-        value = (i.get_io_type(), as_dict, dynamic_dict)
-    d.setdefault(key, {})[i.id] = value
+    d.setdefault(key, {})[i.id] = (i.get_io_type(), as_dict)
 
 def add_to_dict_v3(io: Input | Output, d: dict):
     d[io.id] = (io.get_io_type(), io.as_dict())
 
+class DynamicPathsDefaultValue:
+    EMPTY_DICT = "empty_dict"
+
 def build_nested_inputs(values: dict[str, Any], v3_data: V3Data):
     paths = v3_data.get("dynamic_paths", None)
+    default_value_dict = v3_data.get("dynamic_paths_default_value", {})
     if paths is None:
         return values
     values = values.copy()
+
     result = {}
 
+    create_tuple = v3_data.get("create_dynamic_tuple", False)
+
     for key, path in paths.items():
         parts = path.split(".")
         current = result
@@ -1431,7 +1602,15 @@ def build_nested_inputs(values: dict[str, Any], v3_data: V3Data):
             is_last = (i == len(parts) - 1)
 
             if is_last:
-                current[p] = values.pop(key, None)
+                value = values.pop(key, None)
+                if value is None:
+                    # see if a default value was provided for this key
+                    default_option = default_value_dict.get(key, None)
+                    if default_option == DynamicPathsDefaultValue.EMPTY_DICT:
+                        value = {}
+                if create_tuple:
+                    value = (value, key)
+                current[p] = value
             else:
                 current = current.setdefault(p, {})
 
@@ -1446,7 +1625,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal):
     SCHEMA = None
 
     # filled in during execution
-    resources: Resources = None
     hidden: HiddenHolder = None
 
     @classmethod
@@ -1493,7 +1671,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal):
         return [name for name in kwargs if kwargs[name] is None]
 
     def __init__(self):
-        self.local_resources: ResourcesLocal = None
         self.__class__.VALIDATE_CLASS()
 
     @classmethod
@@ -1556,12 +1733,12 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal):
 
     @final
     @classmethod
-    def PREPARE_CLASS_CLONE(cls, v3_data: V3Data) -> type[ComfyNode]:
+    def PREPARE_CLASS_CLONE(cls, v3_data: V3Data | None) -> type[ComfyNode]:
         """Creates clone of real node class to prevent monkey-patching."""
         c_type: type[ComfyNode] = cls if is_class(cls) else type(cls)
         type_clone: type[ComfyNode] = shallow_clone_class(c_type)
         # set hidden
-        type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"])
+        type_clone.hidden = HiddenHolder.from_v3_data(v3_data)
         return type_clone
 
     @final
@@ -1678,19 +1855,10 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal):
 
     @final
     @classmethod
-    def INPUT_TYPES(cls, include_hidden=True, return_schema=False, live_inputs=None) -> dict[str, dict] | tuple[dict[str, dict], Schema, V3Data]:
+    def INPUT_TYPES(cls) -> dict[str, dict]:
         schema = cls.FINALIZE_SCHEMA()
-        info = schema.get_v1_info(cls, live_inputs)
-        input = info.input
-        if not include_hidden:
-            input.pop("hidden", None)
-        if return_schema:
-            v3_data: V3Data = {}
-            dynamic = input.pop("dynamic_paths", None)
-            if dynamic is not None:
-                v3_data["dynamic_paths"] = dynamic
-            return input, schema, v3_data
-        return input
+        info = schema.get_v1_info(cls)
+        return info.input
 
     @final
     @classmethod
@@ -1809,7 +1977,7 @@ class NodeOutput(_NodeOutputInternal):
         return self.args if len(self.args) > 0 else None
 
     @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> "NodeOutput":
+    def from_dict(cls, data: dict[str, Any]) -> NodeOutput:
         args = ()
         ui = None
         expand = None
@@ -1870,6 +2038,7 @@ __all__ = [
     "ControlNet",
     "Vae",
     "Model",
+    "ModelPatch",
     "ClipVision",
     "ClipVisionOutput",
     "AudioEncoder",
@@ -1904,8 +2073,8 @@ __all__ = [
     "Tracks",
     # Dynamic Types
     "MatchType",
-    # "DynamicCombo",
-    # "Autogrow",
+    "DynamicCombo",
+    "Autogrow",
     # Other classes
     "HiddenHolder",
     "Hidden",
@@ -1917,4 +2086,7 @@ __all__ = [
     "add_to_dict_v1",
     "add_to_dict_v3",
     "V3Data",
+    "ImageCompare",
+    "PriceBadgeDepends",
+    "PriceBadge",
 ]

comfy_api/latest/_resources.py

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

comfy_api/latest/_util/__init__.py

@@ -1,5 +1,6 @@
 from .video_types import VideoContainer, VideoCodec, VideoComponents
 from .geometry_types import VOXEL, MESH
+from .image_types import SVG
 
 __all__ = [
     # Utility Types
@@ -8,4 +9,5 @@ __all__ = [
     "VideoComponents",
     "VOXEL",
     "MESH",
+    "SVG",
 ]

comfy_api/latest/_util/image_types.py

@@ -0,0 +1,18 @@
+from io import BytesIO
+
+
+class SVG:
+    """Stores SVG representations via a list of BytesIO objects."""
+
+    def __init__(self, data: list[BytesIO]):
+        self.data = data
+
+    def combine(self, other: 'SVG') -> 'SVG':
+        return SVG(self.data + other.data)
+
+    @staticmethod
+    def combine_all(svgs: list['SVG']) -> 'SVG':
+        all_svgs_list: list[BytesIO] = []
+        for svg_item in svgs:
+            all_svgs_list.extend(svg_item.data)
+        return SVG(all_svgs_list)

comfy_api_nodes/README.md

@@ -1,65 +0,0 @@
-# ComfyUI API Nodes
-
-## Introduction 
-
-Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview).
-
-## Development
-
-While developing, you should be testing against the Staging environment. To test against staging:
-
-**Install ComfyUI_frontend**
-
-Follow the instructions [here](https://github.com/Comfy-Org/ComfyUI_frontend) to start the frontend server. By default, it will connect to Staging authentication. 
-
-> **Hint:** If you use --front-end-version argument for ComfyUI, it will use production authentication.
-
-```bash
-python run main.py --comfy-api-base https://stagingapi.comfy.org
-```
-
-To authenticate to staging, please login and then ask one of Comfy Org team to whitelist you for access to staging.
-
-API stubs are generated through automatic codegen tools from OpenAPI definitions. Since the Comfy Org OpenAPI definition contains many things from the Comfy Registry as well, we use redocly/cli to filter out only the paths relevant for API nodes.
-
-### Redocly Instructions 
-
-**Tip**
-When developing locally, use the `redocly-dev.yaml` file to generate pydantic models. This lets you use stubs for APIs that are not marked `Released` yet.
-
-Before your API node PR merges, make sure to add the `Released` tag to the `openapi.yaml` file and test in staging.
-
-```bash
-# Download the OpenAPI file from staging server.
-curl -o openapi.yaml https://stagingapi.comfy.org/openapi
-
-# Filter out unneeded API definitions.
-npm install -g @redocly/cli
-redocly bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly-dev.yaml --remove-unused-components
-
-# Generate the pydantic datamodels for validation.
-datamodel-codegen --use-subclass-enum --field-constraints --strict-types bytes --input filtered-openapi.yaml --output comfy_api_nodes/apis/__init__.py --output-model-type pydantic_v2.BaseModel
-
-```
-
-
-# Merging to Master
-
-Before merging to comfyanonymous/ComfyUI master, follow these steps:
-
-1. Add the "Released" tag to the ComfyUI OpenAPI yaml file for each endpoint you are using in the nodes. 
-1. Make sure the ComfyUI API is deployed to prod with your changes.
-1. Run the code generation again with `redocly.yaml` and the production OpenAPI yaml file.
-
-```bash
-# Download the OpenAPI file from prod server.
-curl -o openapi.yaml https://api.comfy.org/openapi
-
-# Filter out unneeded API definitions.
-npm install -g @redocly/cli
-redocly bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly.yaml --remove-unused-components
-
-# Generate the pydantic datamodels for validation.
-datamodel-codegen --use-subclass-enum --field-constraints --strict-types bytes --input filtered-openapi.yaml --output comfy_api_nodes/apis/__init__.py --output-model-type pydantic_v2.BaseModel
-
-```

comfy_api_nodes/apis/bria.py

@@ -0,0 +1,61 @@
+from typing import TypedDict
+
+from pydantic import BaseModel, Field
+
+
+class InputModerationSettings(TypedDict):
+    prompt_content_moderation: bool
+    visual_input_moderation: bool
+    visual_output_moderation: bool
+
+
+class BriaEditImageRequest(BaseModel):
+    instruction: str | None = Field(...)
+    structured_instruction: str | None = Field(
+        ...,
+        description="Use this instead of instruction for precise, programmatic control.",
+    )
+    images: list[str] = Field(
+        ...,
+        description="Required. Publicly available URL or Base64-encoded. Must contain exactly one item.",
+    )
+    mask: str | None = Field(
+        None,
+        description="Mask image (black and white). Black areas will be preserved, white areas will be edited. "
+        "If omitted, the edit applies to the entire image. "
+        "The input image and the the input mask must be of the same size.",
+    )
+    negative_prompt: str | None = Field(None)
+    guidance_scale: float = Field(...)
+    model_version: str = Field(...)
+    steps_num: int = Field(...)
+    seed: int = Field(...)
+    ip_signal: bool = Field(
+        False,
+        description="If true, returns a warning for potential IP content in the instruction.",
+    )
+    prompt_content_moderation: bool = Field(
+        False, description="If true, returns 422 on instruction moderation failure."
+    )
+    visual_input_content_moderation: bool = Field(
+        False, description="If true, returns 422 on images or mask moderation failure."
+    )
+    visual_output_content_moderation: bool = Field(
+        False, description="If true, returns 422 on visual output moderation failure."
+    )
+
+
+class BriaStatusResponse(BaseModel):
+    request_id: str = Field(...)
+    status_url: str = Field(...)
+    warning: str | None = Field(None)
+
+
+class BriaResult(BaseModel):
+    structured_prompt: str = Field(...)
+    image_url: str = Field(...)
+
+
+class BriaResponse(BaseModel):
+    status: str = Field(...)
+    result: BriaResult | None = Field(None)

comfy_api_nodes/apis/bytedance.py

@@ -10,7 +10,7 @@ class Text2ImageTaskCreationRequest(BaseModel):
     size: str | None = Field(None)
     seed: int | None = Field(0, ge=0, le=2147483647)
     guidance_scale: float | None = Field(..., ge=1.0, le=10.0)
-    watermark: bool | None = Field(True)
+    watermark: bool | None = Field(False)
 
 
 class Image2ImageTaskCreationRequest(BaseModel):
@@ -21,7 +21,7 @@ class Image2ImageTaskCreationRequest(BaseModel):
     size: str | None = Field("adaptive")
     seed: int | None = Field(..., ge=0, le=2147483647)
     guidance_scale: float | None = Field(..., ge=1.0, le=10.0)
-    watermark: bool | None = Field(True)
+    watermark: bool | None = Field(False)
 
 
 class Seedream4Options(BaseModel):
@@ -37,7 +37,7 @@ class Seedream4TaskCreationRequest(BaseModel):
     seed: int = Field(..., ge=0, le=2147483647)
     sequential_image_generation: str = Field("disabled")
     sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15))
-    watermark: bool = Field(True)
+    watermark: bool = Field(False)
 
 
 class ImageTaskCreationResponse(BaseModel):
@@ -65,11 +65,13 @@ class TaskImageContent(BaseModel):
 class Text2VideoTaskCreationRequest(BaseModel):
     model: str = Field(...)
     content: list[TaskTextContent] = Field(..., min_length=1)
+    generate_audio: bool | None = Field(...)
 
 
 class Image2VideoTaskCreationRequest(BaseModel):
     model: str = Field(...)
     content: list[TaskTextContent | TaskImageContent] = Field(..., min_length=2)
+    generate_audio: bool | None = Field(...)
 
 
 class TaskCreationResponse(BaseModel):
@@ -141,4 +143,9 @@ VIDEO_TASKS_EXECUTION_TIME = {
         "720p": 65,
         "1080p": 100,
     },
+    "seedance-1-5-pro-251215": {
+        "480p": 80,
+        "720p": 100,
+        "1080p": 150,
+    },
 }

comfy_api_nodes/apis/gemini.py

@@ -133,6 +133,7 @@ class GeminiImageGenerateContentRequest(BaseModel):
     systemInstruction: GeminiSystemInstructionContent | None = Field(None)
     tools: list[GeminiTool] | None = Field(None)
     videoMetadata: GeminiVideoMetadata | None = Field(None)
+    uploadImagesToStorage: bool = Field(True)
 
 
 class GeminiGenerateContentRequest(BaseModel):

comfy_api_nodes/apis/ideogram.py

@@ -0,0 +1,292 @@
+from enum import Enum
+from typing import Optional, List, Dict, Any, Union
+from datetime import datetime
+
+from pydantic import BaseModel, Field, RootModel, StrictBytes
+
+
+class IdeogramColorPalette1(BaseModel):
+    name: str = Field(..., description='Name of the preset color palette')
+
+
+class Member(BaseModel):
+    color: Optional[str] = Field(
+        None, description='Hexadecimal color code', pattern='^#[0-9A-Fa-f]{6}$'
+    )
+    weight: Optional[float] = Field(
+        None, description='Optional weight for the color (0-1)', ge=0.0, le=1.0
+    )
+
+
+class IdeogramColorPalette2(BaseModel):
+    members: List[Member] = Field(
+        ..., description='Array of color definitions with optional weights'
+    )
+
+
+class IdeogramColorPalette(
+    RootModel[Union[IdeogramColorPalette1, IdeogramColorPalette2]]
+):
+    root: Union[IdeogramColorPalette1, IdeogramColorPalette2] = Field(
+        ...,
+        description='A color palette specification that can either use a preset name or explicit color definitions with weights',
+    )
+
+
+class ImageRequest(BaseModel):
+    aspect_ratio: Optional[str] = Field(
+        None,
+        description="Optional. The aspect ratio (e.g., 'ASPECT_16_9', 'ASPECT_1_1'). Cannot be used with resolution. Defaults to 'ASPECT_1_1' if unspecified.",
+    )
+    color_palette: Optional[Dict[str, Any]] = Field(
+        None, description='Optional. Color palette object. Only for V_2, V_2_TURBO.'
+    )
+    magic_prompt_option: Optional[str] = Field(
+        None, description="Optional. MagicPrompt usage ('AUTO', 'ON', 'OFF')."
+    )
+    model: str = Field(..., description="The model used (e.g., 'V_2', 'V_2A_TURBO')")
+    negative_prompt: Optional[str] = Field(
+        None,
+        description='Optional. Description of what to exclude. Only for V_1, V_1_TURBO, V_2, V_2_TURBO.',
+    )
+    num_images: Optional[int] = Field(
+        1,
+        description='Optional. Number of images to generate (1-8). Defaults to 1.',
+        ge=1,
+        le=8,
+    )
+    prompt: str = Field(
+        ..., description='Required. The prompt to use to generate the image.'
+    )
+    resolution: Optional[str] = Field(
+        None,
+        description="Optional. Resolution (e.g., 'RESOLUTION_1024_1024'). Only for model V_2. Cannot be used with aspect_ratio.",
+    )
+    seed: Optional[int] = Field(
+        None,
+        description='Optional. A number between 0 and 2147483647.',
+        ge=0,
+        le=2147483647,
+    )
+    style_type: Optional[str] = Field(
+        None,
+        description="Optional. Style type ('AUTO', 'GENERAL', 'REALISTIC', 'DESIGN', 'RENDER_3D', 'ANIME'). Only for models V_2 and above.",
+    )
+
+
+class IdeogramGenerateRequest(BaseModel):
+    image_request: ImageRequest = Field(
+        ..., description='The image generation request parameters.'
+    )
+
+
+class Datum(BaseModel):
+    is_image_safe: Optional[bool] = Field(
+        None, description='Indicates whether the image is considered safe.'
+    )
+    prompt: Optional[str] = Field(
+        None, description='The prompt used to generate this image.'
+    )
+    resolution: Optional[str] = Field(
+        None, description="The resolution of the generated image (e.g., '1024x1024')."
+    )
+    seed: Optional[int] = Field(
+        None, description='The seed value used for this generation.'
+    )
+    style_type: Optional[str] = Field(
+        None,
+        description="The style type used for generation (e.g., 'REALISTIC', 'ANIME').",
+    )
+    url: Optional[str] = Field(None, description='URL to the generated image.')
+
+
+class IdeogramGenerateResponse(BaseModel):
+    created: Optional[datetime] = Field(
+        None, description='Timestamp when the generation was created.'
+    )
+    data: Optional[List[Datum]] = Field(
+        None, description='Array of generated image information.'
+    )
+
+
+class StyleCode(RootModel[str]):
+    root: str = Field(..., pattern='^[0-9A-Fa-f]{8}$')
+
+
+class Datum1(BaseModel):
+    is_image_safe: Optional[bool] = None
+    prompt: Optional[str] = None
+    resolution: Optional[str] = None
+    seed: Optional[int] = None
+    style_type: Optional[str] = None
+    url: Optional[str] = None
+
+
+class IdeogramV3IdeogramResponse(BaseModel):
+    created: Optional[datetime] = None
+    data: Optional[List[Datum1]] = None
+
+
+class RenderingSpeed1(str, Enum):
+    TURBO = 'TURBO'
+    DEFAULT = 'DEFAULT'
+    QUALITY = 'QUALITY'
+
+
+class IdeogramV3ReframeRequest(BaseModel):
+    color_palette: Optional[Dict[str, Any]] = None
+    image: Optional[StrictBytes] = None
+    num_images: Optional[int] = Field(None, ge=1, le=8)
+    rendering_speed: Optional[RenderingSpeed1] = None
+    resolution: str
+    seed: Optional[int] = Field(None, ge=0, le=2147483647)
+    style_codes: Optional[List[str]] = None
+    style_reference_images: Optional[List[StrictBytes]] = None
+
+
+class MagicPrompt(str, Enum):
+    AUTO = 'AUTO'
+    ON = 'ON'
+    OFF = 'OFF'
+
+
+class StyleType(str, Enum):
+    AUTO = 'AUTO'
+    GENERAL = 'GENERAL'
+    REALISTIC = 'REALISTIC'
+    DESIGN = 'DESIGN'
+
+
+class IdeogramV3RemixRequest(BaseModel):
+    aspect_ratio: Optional[str] = None
+    color_palette: Optional[Dict[str, Any]] = None
+    image: Optional[StrictBytes] = None
+    image_weight: Optional[int] = Field(50, ge=1, le=100)
+    magic_prompt: Optional[MagicPrompt] = None
+    negative_prompt: Optional[str] = None
+    num_images: Optional[int] = Field(None, ge=1, le=8)
+    prompt: str
+    rendering_speed: Optional[RenderingSpeed1] = None
+    resolution: Optional[str] = None
+    seed: Optional[int] = Field(None, ge=0, le=2147483647)
+    style_codes: Optional[List[str]] = None
+    style_reference_images: Optional[List[StrictBytes]] = None
+    style_type: Optional[StyleType] = None
+
+
+class IdeogramV3ReplaceBackgroundRequest(BaseModel):
+    color_palette: Optional[Dict[str, Any]] = None
+    image: Optional[StrictBytes] = None
+    magic_prompt: Optional[MagicPrompt] = None
+    num_images: Optional[int] = Field(None, ge=1, le=8)
+    prompt: str
+    rendering_speed: Optional[RenderingSpeed1] = None
+    seed: Optional[int] = Field(None, ge=0, le=2147483647)
+    style_codes: Optional[List[str]] = None
+    style_reference_images: Optional[List[StrictBytes]] = None
+
+
+class ColorPalette(BaseModel):
+    name: str = Field(..., description='Name of the color palette', examples=['PASTEL'])
+
+
+class MagicPrompt2(str, Enum):
+    ON = 'ON'
+    OFF = 'OFF'
+
+
+class StyleType1(str, Enum):
+    AUTO = 'AUTO'
+    GENERAL = 'GENERAL'
+    REALISTIC = 'REALISTIC'
+    DESIGN = 'DESIGN'
+    FICTION = 'FICTION'
+
+
+class RenderingSpeed(str, Enum):
+    DEFAULT = 'DEFAULT'
+    TURBO = 'TURBO'
+    QUALITY = 'QUALITY'
+
+
+class IdeogramV3EditRequest(BaseModel):
+    color_palette: Optional[IdeogramColorPalette] = None
+    image: Optional[StrictBytes] = Field(
+        None,
+        description='The image being edited (max size 10MB); only JPEG, WebP and PNG formats are supported at this time.',
+    )
+    magic_prompt: Optional[str] = Field(
+        None,
+        description='Determine if MagicPrompt should be used in generating the request or not.',
+    )
+    mask: Optional[StrictBytes] = Field(
+        None,
+        description='A black and white image of the same size as the image being edited (max size 10MB). Black regions in the mask should match up with the regions of the image that you would like to edit; only JPEG, WebP and PNG formats are supported at this time.',
+    )
+    num_images: Optional[int] = Field(
+        None, description='The number of images to generate.'
+    )
+    prompt: str = Field(
+        ..., description='The prompt used to describe the edited result.'
+    )
+    rendering_speed: RenderingSpeed
+    seed: Optional[int] = Field(
+        None, description='Random seed. Set for reproducible generation.'
+    )
+    style_codes: Optional[List[StyleCode]] = Field(
+        None,
+        description='A list of 8 character hexadecimal codes representing the style of the image. Cannot be used in conjunction with style_reference_images or style_type.',
+    )
+    style_reference_images: Optional[List[StrictBytes]] = Field(
+        None,
+        description='A set of images to use as style references (maximum total size 10MB across all style references). The images should be in JPEG, PNG or WebP format.',
+    )
+    character_reference_images: Optional[List[str]] = Field(
+        None,
+        description='Generations with character reference are subject to the character reference pricing. A set of images to use as character references (maximum total size 10MB across all character references), currently only supports 1 character reference image. The images should be in JPEG, PNG or WebP format.'
+    )
+    character_reference_images_mask: Optional[List[str]] = Field(
+        None,
+        description='Optional masks for character reference images. When provided, must match the number of character_reference_images. Each mask should be a grayscale image of the same dimensions as the corresponding character reference image. The images should be in JPEG, PNG or WebP format.'
+    )
+
+
+class IdeogramV3Request(BaseModel):
+    aspect_ratio: Optional[str] = Field(
+        None, description='Aspect ratio in format WxH', examples=['1x3']
+    )
+    color_palette: Optional[ColorPalette] = None
+    magic_prompt: Optional[MagicPrompt2] = Field(
+        None, description='Whether to enable magic prompt enhancement'
+    )
+    negative_prompt: Optional[str] = Field(
+        None, description='Text prompt specifying what to avoid in the generation'
+    )
+    num_images: Optional[int] = Field(
+        None, description='Number of images to generate', ge=1
+    )
+    prompt: str = Field(..., description='The text prompt for image generation')
+    rendering_speed: RenderingSpeed
+    resolution: Optional[str] = Field(
+        None, description='Image resolution in format WxH', examples=['1280x800']
+    )
+    seed: Optional[int] = Field(
+        None, description='Seed value for reproducible generation'
+    )
+    style_codes: Optional[List[StyleCode]] = Field(
+        None, description='Array of style codes in hexadecimal format'
+    )
+    style_reference_images: Optional[List[str]] = Field(
+        None, description='Array of reference image URLs or identifiers'
+    )
+    style_type: Optional[StyleType1] = Field(
+        None, description='The type of style to apply'
+    )
+    character_reference_images: Optional[List[str]] = Field(
+        None,
+        description='Generations with character reference are subject to the character reference pricing. A set of images to use as character references (maximum total size 10MB across all character references), currently only supports 1 character reference image. The images should be in JPEG, PNG or WebP format.'
+    )
+    character_reference_images_mask: Optional[List[str]] = Field(
+        None,
+        description='Optional masks for character reference images. When provided, must match the number of character_reference_images. Each mask should be a grayscale image of the same dimensions as the corresponding character reference image. The images should be in JPEG, PNG or WebP format.'
+    )

comfy_api_nodes/apis/kling.py

@@ -102,3 +102,12 @@ class ImageToVideoWithAudioRequest(BaseModel):
     prompt: str = Field(...)
     mode: str = Field("pro")
     sound: str = Field(..., description="'on' or 'off'")
+
+
+class MotionControlRequest(BaseModel):
+    prompt: str = Field(...)
+    image_url: str = Field(...)
+    video_url: str = Field(...)
+    keep_original_sound: str = Field(...)
+    character_orientation: str = Field(...)
+    mode: str = Field(..., description="'pro' or 'std'")