Merge branch 'master' into toolkit/wire-essentials-categorization

Frontend will own blueprint categorization separately.

.github/scripts/check-ai-co-authors.sh

@@ -1,103 +0,0 @@
-#!/usr/bin/env bash
-# Checks pull request commits for AI agent Co-authored-by trailers.
-# Exits non-zero when any are found and prints fix instructions.
-set -euo pipefail
-
-base_sha="${1:?usage: check-ai-co-authors.sh <base_sha> <head_sha>}"
-head_sha="${2:?usage: check-ai-co-authors.sh <base_sha> <head_sha>}"
-
-# Known AI coding-agent trailer patterns (case-insensitive).
-# Each entry is an extended-regex fragment matched against Co-authored-by lines.
-AGENT_PATTERNS=(
-    # Anthropic — Claude Code / Amp
-    'noreply@anthropic\.com'
-    # Cursor
-    'cursoragent@cursor\.com'
-    # GitHub Copilot
-    'copilot-swe-agent\[bot\]'
-    'copilot@github\.com'
-    # OpenAI Codex
-    'noreply@openai\.com'
-    'codex@openai\.com'
-    # Aider
-    'aider@aider\.chat'
-    # Google — Gemini / Jules
-    'gemini@google\.com'
-    'jules@google\.com'
-    # Windsurf / Codeium
-    '@codeium\.com'
-    # Devin
-    'devin-ai-integration\[bot\]'
-    'devin@cognition\.ai'
-    'devin@cognition-labs\.com'
-    # Amazon Q Developer
-    'amazon-q-developer'
-    '@amazon\.com.*[Qq].[Dd]eveloper'
-    # Cline
-    'cline-bot'
-    'cline@cline\.ai'
-    # Continue
-    'continue-agent'
-    'continue@continue\.dev'
-    # Sourcegraph
-    'noreply@sourcegraph\.com'
-    # Generic catch-alls for common agent name patterns
-    'Co-authored-by:.*\b[Cc]laude\b'
-    'Co-authored-by:.*\b[Cc]opilot\b'
-    'Co-authored-by:.*\b[Cc]ursor\b'
-    'Co-authored-by:.*\b[Cc]odex\b'
-    'Co-authored-by:.*\b[Gg]emini\b'
-    'Co-authored-by:.*\b[Aa]ider\b'
-    'Co-authored-by:.*\b[Dd]evin\b'
-    'Co-authored-by:.*\b[Ww]indsurf\b'
-    'Co-authored-by:.*\b[Cc]line\b'
-    'Co-authored-by:.*\b[Aa]mazon Q\b'
-    'Co-authored-by:.*\b[Jj]ules\b'
-    'Co-authored-by:.*\bOpenCode\b'
-)
-
-# Build a single alternation regex from all patterns.
-regex=""
-for pattern in "${AGENT_PATTERNS[@]}"; do
-    if [[ -n "$regex" ]]; then
-        regex="${regex}|${pattern}"
-    else
-        regex="$pattern"
-    fi
-done
-
-# Collect Co-authored-by lines from every commit in the PR range.
-violations=""
-while IFS= read -r sha; do
-    message="$(git log -1 --format='%B' "$sha")"
-    matched_lines="$(echo "$message" | grep -iE "^Co-authored-by:" || true)"
-    if [[ -z "$matched_lines" ]]; then
-        continue
-    fi
-
-    while IFS= read -r line; do
-        if echo "$line" | grep -iqE "$regex"; then
-            short="$(git log -1 --format='%h' "$sha")"
-            violations="${violations}  ${short}: ${line}"$'\n'
-        fi
-    done <<< "$matched_lines"
-done < <(git rev-list "${base_sha}..${head_sha}")
-
-if [[ -n "$violations" ]]; then
-    echo "::error::AI agent Co-authored-by trailers detected in PR commits."
-    echo ""
-    echo "The following commits contain Co-authored-by trailers from AI coding agents:"
-    echo ""
-    echo "$violations"
-    echo "These trailers should be removed before merging."
-    echo ""
-    echo "To fix, rewrite the commit messages with:"
-    echo "  git rebase -i ${base_sha}"
-    echo ""
-    echo "and remove the Co-authored-by lines, then force-push your branch."
-    echo ""
-    echo "If you believe this is a false positive, please open an issue."
-    exit 1
-fi
-
-echo "No AI agent Co-authored-by trailers found."

.github/workflows/check-ai-co-authors.yml

@@ -1,19 +0,0 @@
-name: Check AI Co-Authors
-
-on:
-  pull_request:
-    branches: ['*']
-
-jobs:
-  check-ai-co-authors:
-    name: Check for AI agent co-author trailers
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Check commits for AI co-author trailers
-        run: bash .github/scripts/check-ai-co-authors.sh "${{ github.event.pull_request.base.sha }}" "${{ github.event.pull_request.head.sha }}"

alembic_db/env.py

@@ -8,7 +8,7 @@ from alembic import context
 config = context.config
 
 
-from app.database.models import Base, NAMING_CONVENTION
+from app.database.models import Base
 target_metadata = Base.metadata
 
 # other values from the config, defined by the needs of env.py,
@@ -51,10 +51,7 @@ def run_migrations_online() -> None:
 
     with connectable.connect() as connection:
         context.configure(
-            connection=connection,
-            target_metadata=target_metadata,
-            render_as_batch=True,
-            naming_convention=NAMING_CONVENTION,
+            connection=connection, target_metadata=target_metadata
         )
 
         with context.begin_transaction():

alembic_db/versions/0003_add_metadata_job_id.py

@@ -1,98 +0,0 @@
-"""
-Add system_metadata and job_id columns to asset_references.
-Change preview_id FK from assets.id to asset_references.id.
-
-Revision ID: 0003_add_metadata_job_id
-Revises: 0002_merge_to_asset_references
-Create Date: 2026-03-09
-"""
-
-from alembic import op
-import sqlalchemy as sa
-
-from app.database.models import NAMING_CONVENTION
-
-revision = "0003_add_metadata_job_id"
-down_revision = "0002_merge_to_asset_references"
-branch_labels = None
-depends_on = None
-
-
-def upgrade() -> None:
-    with op.batch_alter_table("asset_references") as batch_op:
-        batch_op.add_column(
-            sa.Column("system_metadata", sa.JSON(), nullable=True)
-        )
-        batch_op.add_column(
-            sa.Column("job_id", sa.String(length=36), nullable=True)
-        )
-
-    # Change preview_id FK from assets.id to asset_references.id (self-ref).
-    # Existing values are asset-content IDs that won't match reference IDs,
-    # so null them out first.
-    op.execute("UPDATE asset_references SET preview_id = NULL WHERE preview_id IS NOT NULL")
-    with op.batch_alter_table(
-        "asset_references", naming_convention=NAMING_CONVENTION
-    ) as batch_op:
-        batch_op.drop_constraint(
-            "fk_asset_references_preview_id_assets", type_="foreignkey"
-        )
-        batch_op.create_foreign_key(
-            "fk_asset_references_preview_id_asset_references",
-            "asset_references",
-            ["preview_id"],
-            ["id"],
-            ondelete="SET NULL",
-        )
-        batch_op.create_index(
-            "ix_asset_references_preview_id", ["preview_id"]
-        )
-
-    # Purge any all-null meta rows before adding the constraint
-    op.execute(
-        "DELETE FROM asset_reference_meta"
-        " WHERE val_str IS NULL AND val_num IS NULL AND val_bool IS NULL AND val_json IS NULL"
-    )
-    with op.batch_alter_table("asset_reference_meta") as batch_op:
-        batch_op.create_check_constraint(
-            "ck_asset_reference_meta_has_value",
-            "val_str IS NOT NULL OR val_num IS NOT NULL OR val_bool IS NOT NULL OR val_json IS NOT NULL",
-        )
-
-
-def downgrade() -> None:
-    # SQLite doesn't reflect CHECK constraints, so we must declare it
-    # explicitly via table_args for the batch recreate to find it.
-    # Use the fully-rendered constraint name to avoid the naming convention
-    # doubling the prefix.
-    with op.batch_alter_table(
-        "asset_reference_meta",
-        table_args=[
-            sa.CheckConstraint(
-                "val_str IS NOT NULL OR val_num IS NOT NULL OR val_bool IS NOT NULL OR val_json IS NOT NULL",
-                name="ck_asset_reference_meta_has_value",
-            ),
-        ],
-    ) as batch_op:
-        batch_op.drop_constraint(
-            "ck_asset_reference_meta_has_value", type_="check"
-        )
-
-    with op.batch_alter_table(
-        "asset_references", naming_convention=NAMING_CONVENTION
-    ) as batch_op:
-        batch_op.drop_index("ix_asset_references_preview_id")
-        batch_op.drop_constraint(
-            "fk_asset_references_preview_id_asset_references", type_="foreignkey"
-        )
-        batch_op.create_foreign_key(
-            "fk_asset_references_preview_id_assets",
-            "assets",
-            ["preview_id"],
-            ["id"],
-            ondelete="SET NULL",
-        )
-
-    with op.batch_alter_table("asset_references") as batch_op:
-        batch_op.drop_column("job_id")
-        batch_op.drop_column("system_metadata")

app/assets/api/routes.py

@@ -13,7 +13,6 @@ from pydantic import ValidationError
 import folder_paths
 from app import user_manager
 from app.assets.api import schemas_in, schemas_out
-from app.assets.services import schemas
 from app.assets.api.schemas_in import (
     AssetValidationError,
     UploadError,
@@ -39,7 +38,6 @@ from app.assets.services import (
     update_asset_metadata,
     upload_from_temp_path,
 )
-from app.assets.services.tagging import list_tag_histogram
 
 ROUTES = web.RouteTableDef()
 USER_MANAGER: user_manager.UserManager | None = None
@@ -124,61 +122,6 @@ def _validate_sort_field(requested: str | None) -> str:
     return "created_at"
 
 
-def _build_preview_url_from_view(tags: list[str], user_metadata: dict[str, Any] | None) -> str | None:
-    """Build a /api/view preview URL from asset tags and user_metadata filename."""
-    if not user_metadata:
-        return None
-    filename = user_metadata.get("filename")
-    if not filename:
-        return None
-
-    if "input" in tags:
-        view_type = "input"
-    elif "output" in tags:
-        view_type = "output"
-    else:
-        return None
-
-    subfolder = ""
-    if "/" in filename:
-        subfolder, filename = filename.rsplit("/", 1)
-
-    encoded_filename = urllib.parse.quote(filename, safe="")
-    url = f"/api/view?type={view_type}&filename={encoded_filename}"
-    if subfolder:
-        url += f"&subfolder={urllib.parse.quote(subfolder, safe='')}"
-    return url
-
-
-def _build_asset_response(result: schemas.AssetDetailResult | schemas.UploadResult) -> schemas_out.Asset:
-    """Build an Asset response from a service result."""
-    if result.ref.preview_id:
-        preview_detail = get_asset_detail(result.ref.preview_id)
-        if preview_detail:
-            preview_url = _build_preview_url_from_view(preview_detail.tags, preview_detail.ref.user_metadata)
-        else:
-            preview_url = None
-    else:
-        preview_url = _build_preview_url_from_view(result.tags, result.ref.user_metadata)
-    return schemas_out.Asset(
-        id=result.ref.id,
-        name=result.ref.name,
-        asset_hash=result.asset.hash if result.asset else None,
-        size=int(result.asset.size_bytes) if result.asset else None,
-        mime_type=result.asset.mime_type if result.asset else None,
-        tags=result.tags,
-        preview_url=preview_url,
-        preview_id=result.ref.preview_id,
-        user_metadata=result.ref.user_metadata or {},
-        metadata=result.ref.system_metadata,
-        job_id=result.ref.job_id,
-        prompt_id=result.ref.job_id,  # deprecated: mirrors job_id for cloud compat
-        created_at=result.ref.created_at,
-        updated_at=result.ref.updated_at,
-        last_access_time=result.ref.last_access_time,
-    )
-
-
 @ROUTES.head("/api/assets/hash/{hash}")
 @_require_assets_feature_enabled
 async def head_asset_by_hash(request: web.Request) -> web.Response:
@@ -221,7 +164,20 @@ async def list_assets_route(request: web.Request) -> web.Response:
         order=order,
     )
 
-    summaries = [_build_asset_response(item) for item in result.items]
+    summaries = [
+        schemas_out.AssetSummary(
+            id=item.ref.id,
+            name=item.ref.name,
+            asset_hash=item.asset.hash if item.asset else None,
+            size=int(item.asset.size_bytes) if item.asset else None,
+            mime_type=item.asset.mime_type if item.asset else None,
+            tags=item.tags,
+            created_at=item.ref.created_at,
+            updated_at=item.ref.updated_at,
+            last_access_time=item.ref.last_access_time,
+        )
+        for item in result.items
+    ]
 
     payload = schemas_out.AssetsList(
         assets=summaries,
@@ -251,7 +207,18 @@ async def get_asset_route(request: web.Request) -> web.Response:
                 {"id": reference_id},
             )
 
-        payload = _build_asset_response(result)
+        payload = schemas_out.AssetDetail(
+            id=result.ref.id,
+            name=result.ref.name,
+            asset_hash=result.asset.hash if result.asset else None,
+            size=int(result.asset.size_bytes) if result.asset else None,
+            mime_type=result.asset.mime_type if result.asset else None,
+            tags=result.tags,
+            user_metadata=result.ref.user_metadata or {},
+            preview_id=result.ref.preview_id,
+            created_at=result.ref.created_at,
+            last_access_time=result.ref.last_access_time,
+        )
     except ValueError as e:
         return _build_error_response(
             404, "ASSET_NOT_FOUND", str(e), {"id": reference_id}
@@ -263,7 +230,7 @@ async def get_asset_route(request: web.Request) -> web.Response:
             USER_MANAGER.get_request_user_id(request),
         )
         return _build_error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(payload.model_dump(mode="json", exclude_none=True), status=200)
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}/content")
@@ -345,31 +312,32 @@ async def create_asset_from_hash_route(request: web.Request) -> web.Response:
             400, "INVALID_JSON", "Request body must be valid JSON."
         )
 
-    # Derive name from hash if not provided
-    name = body.name
-    if name is None:
-        name = body.hash.split(":", 1)[1] if ":" in body.hash else body.hash
-
     result = create_from_hash(
         hash_str=body.hash,
-        name=name,
+        name=body.name,
         tags=body.tags,
         user_metadata=body.user_metadata,
         owner_id=USER_MANAGER.get_request_user_id(request),
-        mime_type=body.mime_type,
-        preview_id=body.preview_id,
     )
     if result is None:
         return _build_error_response(
             404, "ASSET_NOT_FOUND", f"Asset content {body.hash} does not exist"
         )
 
-    asset = _build_asset_response(result)
     payload_out = schemas_out.AssetCreated(
-        **asset.model_dump(),
+        id=result.ref.id,
+        name=result.ref.name,
+        asset_hash=result.asset.hash,
+        size=int(result.asset.size_bytes),
+        mime_type=result.asset.mime_type,
+        tags=result.tags,
+        user_metadata=result.ref.user_metadata or {},
+        preview_id=result.ref.preview_id,
+        created_at=result.ref.created_at,
+        last_access_time=result.ref.last_access_time,
         created_new=result.created_new,
     )
-    return web.json_response(payload_out.model_dump(mode="json", exclude_none=True), status=201)
+    return web.json_response(payload_out.model_dump(mode="json"), status=201)
 
 
 @ROUTES.post("/api/assets")
@@ -390,8 +358,6 @@ async def upload_asset(request: web.Request) -> web.Response:
                 "name": parsed.provided_name,
                 "user_metadata": parsed.user_metadata_raw,
                 "hash": parsed.provided_hash,
-                "mime_type": parsed.provided_mime_type,
-                "preview_id": parsed.provided_preview_id,
             }
         )
     except ValidationError as ve:
@@ -420,8 +386,6 @@ async def upload_asset(request: web.Request) -> web.Response:
                 tags=spec.tags,
                 user_metadata=spec.user_metadata or {},
                 owner_id=owner_id,
-                mime_type=spec.mime_type,
-                preview_id=spec.preview_id,
             )
             if result is None:
                 delete_temp_file_if_exists(parsed.tmp_path)
@@ -446,8 +410,6 @@ async def upload_asset(request: web.Request) -> web.Response:
                 client_filename=parsed.file_client_name,
                 owner_id=owner_id,
                 expected_hash=spec.hash,
-                mime_type=spec.mime_type,
-                preview_id=spec.preview_id,
             )
     except AssetValidationError as e:
         delete_temp_file_if_exists(parsed.tmp_path)
@@ -466,13 +428,21 @@ async def upload_asset(request: web.Request) -> web.Response:
         logging.exception("upload_asset failed for owner_id=%s", owner_id)
         return _build_error_response(500, "INTERNAL", "Unexpected server error.")
 
-    asset = _build_asset_response(result)
-    payload_out = schemas_out.AssetCreated(
-        **asset.model_dump(),
+    payload = schemas_out.AssetCreated(
+        id=result.ref.id,
+        name=result.ref.name,
+        asset_hash=result.asset.hash,
+        size=int(result.asset.size_bytes),
+        mime_type=result.asset.mime_type,
+        tags=result.tags,
+        user_metadata=result.ref.user_metadata or {},
+        preview_id=result.ref.preview_id,
+        created_at=result.ref.created_at,
+        last_access_time=result.ref.last_access_time,
         created_new=result.created_new,
     )
     status = 201 if result.created_new else 200
-    return web.json_response(payload_out.model_dump(mode="json", exclude_none=True), status=status)
+    return web.json_response(payload.model_dump(mode="json"), status=status)
 
 
 @ROUTES.put(f"/api/assets/{{id:{UUID_RE}}}")
@@ -494,9 +464,15 @@ async def update_asset_route(request: web.Request) -> web.Response:
             name=body.name,
             user_metadata=body.user_metadata,
             owner_id=USER_MANAGER.get_request_user_id(request),
-            preview_id=body.preview_id,
         )
-        payload = _build_asset_response(result)
+        payload = schemas_out.AssetUpdated(
+            id=result.ref.id,
+            name=result.ref.name,
+            asset_hash=result.asset.hash if result.asset else None,
+            tags=result.tags,
+            user_metadata=result.ref.user_metadata or {},
+            updated_at=result.ref.updated_at,
+        )
     except PermissionError as pe:
         return _build_error_response(403, "FORBIDDEN", str(pe), {"id": reference_id})
     except ValueError as ve:
@@ -510,7 +486,7 @@ async def update_asset_route(request: web.Request) -> web.Response:
             USER_MANAGER.get_request_user_id(request),
         )
         return _build_error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(payload.model_dump(mode="json", exclude_none=True), status=200)
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}")
@@ -579,7 +555,7 @@ async def get_tags(request: web.Request) -> web.Response:
     payload = schemas_out.TagsList(
         tags=tags, total=total, has_more=(query.offset + len(tags)) < total
     )
-    return web.json_response(payload.model_dump(mode="json", exclude_none=True))
+    return web.json_response(payload.model_dump(mode="json"))
 
 
 @ROUTES.post(f"/api/assets/{{id:{UUID_RE}}}/tags")
@@ -627,7 +603,7 @@ async def add_asset_tags(request: web.Request) -> web.Response:
         )
         return _build_error_response(500, "INTERNAL", "Unexpected server error.")
 
-    return web.json_response(payload.model_dump(mode="json", exclude_none=True), status=200)
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}/tags")
@@ -674,29 +650,7 @@ async def delete_asset_tags(request: web.Request) -> web.Response:
         )
         return _build_error_response(500, "INTERNAL", "Unexpected server error.")
 
-    return web.json_response(payload.model_dump(mode="json", exclude_none=True), status=200)
-
-
-@ROUTES.get("/api/assets/tags/refine")
-@_require_assets_feature_enabled
-async def get_tags_refine(request: web.Request) -> web.Response:
-    """GET request to get tag histogram for filtered assets."""
-    query_dict = get_query_dict(request)
-    try:
-        q = schemas_in.TagsRefineQuery.model_validate(query_dict)
-    except ValidationError as ve:
-        return _build_validation_error_response("INVALID_QUERY", ve)
-
-    tag_counts = list_tag_histogram(
-        owner_id=USER_MANAGER.get_request_user_id(request),
-        include_tags=q.include_tags,
-        exclude_tags=q.exclude_tags,
-        name_contains=q.name_contains,
-        metadata_filter=q.metadata_filter,
-        limit=q.limit,
-    )
-    payload = schemas_out.TagHistogram(tag_counts=tag_counts)
-    return web.json_response(payload.model_dump(mode="json", exclude_none=True), status=200)
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.post("/api/assets/seed")

app/assets/api/schemas_in.py

@@ -45,8 +45,6 @@ class ParsedUpload:
     user_metadata_raw: str | None
     provided_hash: str | None
     provided_hash_exists: bool | None
-    provided_mime_type: str | None = None
-    provided_preview_id: str | None = None
 
 
 class ListAssetsQuery(BaseModel):
@@ -100,17 +98,11 @@ class ListAssetsQuery(BaseModel):
 class UpdateAssetBody(BaseModel):
     name: str | None = None
     user_metadata: dict[str, Any] | None = None
-    preview_id: str | None = None  # references an asset_reference id, not an asset id
 
     @model_validator(mode="after")
     def _validate_at_least_one_field(self):
-        if all(
-            v is None
-            for v in (self.name, self.user_metadata, self.preview_id)
-        ):
-            raise ValueError(
-                "Provide at least one of: name, user_metadata, preview_id."
-            )
+        if self.name is None and self.user_metadata is None:
+            raise ValueError("Provide at least one of: name, user_metadata.")
         return self
 
 
@@ -118,11 +110,9 @@ class CreateFromHashBody(BaseModel):
     model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
 
     hash: str
-    name: str | None = None
+    name: str
     tags: list[str] = Field(default_factory=list)
     user_metadata: dict[str, Any] = Field(default_factory=dict)
-    mime_type: str | None = None
-    preview_id: str | None = None  # references an asset_reference id, not an asset id
 
     @field_validator("hash")
     @classmethod
@@ -148,44 +138,6 @@ class CreateFromHashBody(BaseModel):
         return []
 
 
-class TagsRefineQuery(BaseModel):
-    include_tags: list[str] = Field(default_factory=list)
-    exclude_tags: list[str] = Field(default_factory=list)
-    name_contains: str | None = None
-    metadata_filter: dict[str, Any] | None = None
-    limit: conint(ge=1, le=1000) = 100
-
-    @field_validator("include_tags", "exclude_tags", mode="before")
-    @classmethod
-    def _split_csv_tags(cls, v):
-        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)
 
@@ -234,25 +186,21 @@ class TagsRemove(TagsAdd):
 class UploadAssetSpec(BaseModel):
     """Upload Asset operation.
 
-    - tags: optional list; if provided, first is root ('models'|'input'|'output');
+    - tags: ordered; first is root ('models'|'input'|'output');
             if root == 'models', second must be a valid category
     - name: display name
     - user_metadata: arbitrary JSON object (optional)
     - hash: optional canonical 'blake3:<hex>' for validation / fast-path
-    - mime_type: optional MIME type override
-    - preview_id: optional asset_reference ID for preview
 
     Files are stored using the content hash as filename stem.
     """
 
     model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
 
-    tags: list[str] = Field(default_factory=list)
+    tags: list[str] = Field(..., min_length=1)
     name: str | None = Field(default=None, max_length=512, description="Display Name")
     user_metadata: dict[str, Any] = Field(default_factory=dict)
     hash: str | None = Field(default=None)
-    mime_type: str | None = Field(default=None)
-    preview_id: str | None = Field(default=None)  # references an asset_reference id
 
     @field_validator("hash", mode="before")
     @classmethod
@@ -331,7 +279,7 @@ class UploadAssetSpec(BaseModel):
     @model_validator(mode="after")
     def _validate_order(self):
         if not self.tags:
-            raise ValueError("at least one tag is required for uploads")
+            raise ValueError("tags must be provided and non-empty")
         root = self.tags[0]
         if root not in {"models", "input", "output"}:
             raise ValueError("first tag must be one of: models, input, output")

app/assets/api/schemas_out.py

@@ -4,10 +4,7 @@ from typing import Any
 from pydantic import BaseModel, ConfigDict, Field, field_serializer
 
 
-class Asset(BaseModel):
-    """API view of an asset. Maps to DB ``AssetReference`` joined with its ``Asset`` blob;
-    ``id`` here is the AssetReference id, not the content-addressed Asset id."""
-
+class AssetSummary(BaseModel):
     id: str
     name: str
     asset_hash: str | None = None
@@ -15,14 +12,8 @@ class Asset(BaseModel):
     mime_type: str | None = None
     tags: list[str] = Field(default_factory=list)
     preview_url: str | None = None
-    preview_id: str | None = None  # references an asset_reference id, not an asset id
-    user_metadata: dict[str, Any] = Field(default_factory=dict)
-    is_immutable: bool = False
-    metadata: dict[str, Any] | None = None
-    job_id: str | None = None
-    prompt_id: str | None = None  # deprecated: use job_id
-    created_at: datetime
-    updated_at: datetime
+    created_at: datetime | None = None
+    updated_at: datetime | None = None
     last_access_time: datetime | None = None
 
     model_config = ConfigDict(from_attributes=True)
@@ -32,16 +23,50 @@ class Asset(BaseModel):
         return v.isoformat() if v else None
 
 
-class AssetCreated(Asset):
-    created_new: bool
-
-
 class AssetsList(BaseModel):
-    assets: list[Asset]
+    assets: list[AssetSummary]
     total: int
     has_more: bool
 
 
+class AssetUpdated(BaseModel):
+    id: str
+    name: str
+    asset_hash: str | None = None
+    tags: list[str] = Field(default_factory=list)
+    user_metadata: dict[str, Any] = Field(default_factory=dict)
+    updated_at: datetime | None = None
+
+    model_config = ConfigDict(from_attributes=True)
+
+    @field_serializer("updated_at")
+    def _serialize_updated_at(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+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 _serialize_datetime(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+class AssetCreated(AssetDetail):
+    created_new: bool
+
+
 class TagUsage(BaseModel):
     name: str
     count: int
@@ -66,7 +91,3 @@ class TagsRemove(BaseModel):
     removed: list[str] = Field(default_factory=list)
     not_present: list[str] = Field(default_factory=list)
     total_tags: list[str] = Field(default_factory=list)
-
-
-class TagHistogram(BaseModel):
-    tag_counts: dict[str, int]

app/assets/api/upload.py

@@ -52,8 +52,6 @@ async def parse_multipart_upload(
     user_metadata_raw: str | None = None
     provided_hash: str | None = None
     provided_hash_exists: bool | None = None
-    provided_mime_type: str | None = None
-    provided_preview_id: str | None = None
 
     file_written = 0
     tmp_path: str | None = None
@@ -130,16 +128,6 @@ async def parse_multipart_upload(
             provided_name = (await field.text()) or None
         elif fname == "user_metadata":
             user_metadata_raw = (await field.text()) or None
-        elif fname == "id":
-            raise UploadError(
-                400,
-                "UNSUPPORTED_FIELD",
-                "Client-provided 'id' is not supported. Asset IDs are assigned by the server.",
-            )
-        elif fname == "mime_type":
-            provided_mime_type = ((await field.text()) or "").strip() or None
-        elif fname == "preview_id":
-            provided_preview_id = ((await field.text()) or "").strip() or None
 
     if not file_present and not (provided_hash and provided_hash_exists):
         raise UploadError(
@@ -164,8 +152,6 @@ async def parse_multipart_upload(
         user_metadata_raw=user_metadata_raw,
         provided_hash=provided_hash,
         provided_hash_exists=provided_hash_exists,
-        provided_mime_type=provided_mime_type,
-        provided_preview_id=provided_preview_id,
     )
 
 

app/assets/database/models.py

@@ -45,7 +45,13 @@ class Asset(Base):
         passive_deletes=True,
     )
 
-    # preview_id on AssetReference is a self-referential FK to asset_references.id
+    preview_of: Mapped[list[AssetReference]] = relationship(
+        "AssetReference",
+        back_populates="preview_asset",
+        primaryjoin=lambda: Asset.id == foreign(AssetReference.preview_id),
+        foreign_keys=lambda: [AssetReference.preview_id],
+        viewonly=True,
+    )
 
     __table_args__ = (
         Index("uq_assets_hash", "hash", unique=True),
@@ -85,15 +91,11 @@ class AssetReference(Base):
     owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="")
     name: Mapped[str] = mapped_column(String(512), nullable=False)
     preview_id: Mapped[str | None] = mapped_column(
-        String(36), ForeignKey("asset_references.id", ondelete="SET NULL")
+        String(36), ForeignKey("assets.id", ondelete="SET NULL")
     )
     user_metadata: Mapped[dict[str, Any] | None] = mapped_column(
         JSON(none_as_null=True)
     )
-    system_metadata: Mapped[dict[str, Any] | None] = mapped_column(
-        JSON(none_as_null=True), nullable=True, default=None
-    )
-    job_id: Mapped[str | None] = mapped_column(String(36), nullable=True, default=None)
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=False), nullable=False, default=get_utc_now
     )
@@ -113,10 +115,10 @@ class AssetReference(Base):
         foreign_keys=[asset_id],
         lazy="selectin",
     )
-    preview_ref: Mapped[AssetReference | None] = relationship(
-        "AssetReference",
+    preview_asset: Mapped[Asset | None] = relationship(
+        "Asset",
+        back_populates="preview_of",
         foreign_keys=[preview_id],
-        remote_side=lambda: [AssetReference.id],
     )
 
     metadata_entries: Mapped[list[AssetReferenceMeta]] = relationship(
@@ -150,7 +152,6 @@ class AssetReference(Base):
         Index("ix_asset_references_created_at", "created_at"),
         Index("ix_asset_references_last_access_time", "last_access_time"),
         Index("ix_asset_references_deleted_at", "deleted_at"),
-        Index("ix_asset_references_preview_id", "preview_id"),
         Index("ix_asset_references_owner_name", "owner_id", "name"),
         CheckConstraint(
             "(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_ar_mtime_nonneg"
@@ -191,10 +192,6 @@ class AssetReferenceMeta(Base):
         Index("ix_asset_reference_meta_key_val_str", "key", "val_str"),
         Index("ix_asset_reference_meta_key_val_num", "key", "val_num"),
         Index("ix_asset_reference_meta_key_val_bool", "key", "val_bool"),
-        CheckConstraint(
-            "val_str IS NOT NULL OR val_num IS NOT NULL OR val_bool IS NOT NULL OR val_json IS NOT NULL",
-            name="has_value",
-        ),
     )
 
 

app/assets/database/queries/__init__.py

@@ -31,21 +31,16 @@ from app.assets.database.queries.asset_reference import (
     get_unenriched_references,
     get_unreferenced_unhashed_asset_ids,
     insert_reference,
-    list_all_file_paths_by_asset_id,
     list_references_by_asset_id,
     list_references_page,
     mark_references_missing_outside_prefixes,
-    rebuild_metadata_projection,
-    reference_exists,
     reference_exists_for_asset_id,
     restore_references_by_paths,
     set_reference_metadata,
     set_reference_preview,
-    set_reference_system_metadata,
     soft_delete_reference_by_id,
     update_reference_access_time,
     update_reference_name,
-    update_is_missing_by_asset_id,
     update_reference_timestamps,
     update_reference_updated_at,
     upsert_reference,
@@ -59,7 +54,6 @@ from app.assets.database.queries.tags import (
     bulk_insert_tags_and_meta,
     ensure_tags_exist,
     get_reference_tags,
-    list_tag_counts_for_filtered_assets,
     list_tags_with_usage,
     remove_missing_tag_for_asset_id,
     remove_tags_from_reference,
@@ -103,26 +97,20 @@ __all__ = [
     "get_unenriched_references",
     "get_unreferenced_unhashed_asset_ids",
     "insert_reference",
-    "list_all_file_paths_by_asset_id",
     "list_references_by_asset_id",
     "list_references_page",
-    "list_tag_counts_for_filtered_assets",
     "list_tags_with_usage",
     "mark_references_missing_outside_prefixes",
     "reassign_asset_references",
-    "rebuild_metadata_projection",
-    "reference_exists",
     "reference_exists_for_asset_id",
     "remove_missing_tag_for_asset_id",
     "remove_tags_from_reference",
     "restore_references_by_paths",
     "set_reference_metadata",
     "set_reference_preview",
-    "set_reference_system_metadata",
     "soft_delete_reference_by_id",
     "set_reference_tags",
     "update_asset_hash_and_mime",
-    "update_is_missing_by_asset_id",
     "update_reference_access_time",
     "update_reference_name",
     "update_reference_timestamps",

app/assets/database/queries/asset.py

@@ -69,7 +69,7 @@ def upsert_asset(
         if asset.size_bytes != int(size_bytes) and int(size_bytes) > 0:
             asset.size_bytes = int(size_bytes)
             changed = True
-        if mime_type and not asset.mime_type:
+        if mime_type and asset.mime_type != mime_type:
             asset.mime_type = mime_type
             changed = True
         if changed:
@@ -118,7 +118,7 @@ def update_asset_hash_and_mime(
         return False
     if asset_hash is not None:
         asset.hash = asset_hash
-    if mime_type is not None and not asset.mime_type:
+    if mime_type is not None:
         asset.mime_type = mime_type
     return True
 

app/assets/database/queries/asset_reference.py

@@ -10,7 +10,7 @@ from decimal import Decimal
 from typing import NamedTuple, Sequence
 
 import sqlalchemy as sa
-from sqlalchemy import delete, select
+from sqlalchemy import delete, exists, select
 from sqlalchemy.dialects import sqlite
 from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session, noload
@@ -24,14 +24,12 @@ from app.assets.database.models import (
 )
 from app.assets.database.queries.common import (
     MAX_BIND_PARAMS,
-    apply_metadata_filter,
-    apply_tag_filters,
     build_prefix_like_conditions,
     build_visible_owner_clause,
     calculate_rows_per_statement,
     iter_chunks,
 )
-from app.assets.helpers import escape_sql_like_string, get_utc_now
+from app.assets.helpers import escape_sql_like_string, get_utc_now, normalize_tags
 
 
 def _check_is_scalar(v):
@@ -46,6 +44,15 @@ def _check_is_scalar(v):
 
 def _scalar_to_row(key: str, ordinal: int, value) -> dict:
     """Convert a scalar value to a typed projection row."""
+    if value is None:
+        return {
+            "key": key,
+            "ordinal": ordinal,
+            "val_str": None,
+            "val_num": None,
+            "val_bool": None,
+            "val_json": None,
+        }
     if isinstance(value, bool):
         return {"key": key, "ordinal": ordinal, "val_bool": bool(value)}
     if isinstance(value, (int, float, Decimal)):
@@ -59,19 +66,96 @@ def _scalar_to_row(key: str, ordinal: int, value) -> dict:
 def convert_metadata_to_rows(key: str, value) -> list[dict]:
     """Turn a metadata key/value into typed projection rows."""
     if value is None:
-        return []
+        return [_scalar_to_row(key, 0, None)]
 
     if _check_is_scalar(value):
         return [_scalar_to_row(key, 0, value)]
 
     if isinstance(value, list):
         if all(_check_is_scalar(x) for x in value):
-            return [_scalar_to_row(key, i, x) for i, x in enumerate(value) if x is not None]
-        return [{"key": key, "ordinal": i, "val_json": x} for i, x in enumerate(value) if x is not None]
+            return [_scalar_to_row(key, i, x) for i, x in enumerate(value)]
+        return [{"key": key, "ordinal": i, "val_json": x} for i, x in enumerate(value)]
 
     return [{"key": key, "ordinal": 0, "val_json": value}]
 
 
+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(
+                    (AssetReferenceTag.asset_reference_id == AssetReference.id)
+                    & (AssetReferenceTag.tag_name == tag_name)
+                )
+            )
+
+    if exclude_tags:
+        stmt = stmt.where(
+            ~exists().where(
+                (AssetReferenceTag.asset_reference_id == AssetReference.id)
+                & (AssetReferenceTag.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_reference_meta projection table."""
+    if not metadata_filter:
+        return stmt
+
+    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
+        return sa.exists().where(
+            AssetReferenceMeta.asset_reference_id == AssetReference.id,
+            AssetReferenceMeta.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(
+                    AssetReferenceMeta.asset_reference_id == AssetReference.id,
+                    AssetReferenceMeta.key == key,
+                )
+            )
+            null_row = _exists_for_pred(
+                key,
+                AssetReferenceMeta.val_json.is_(None),
+                AssetReferenceMeta.val_str.is_(None),
+                AssetReferenceMeta.val_num.is_(None),
+                AssetReferenceMeta.val_bool.is_(None),
+            )
+            return sa.or_(no_row_for_key, null_row)
+
+        if isinstance(value, bool):
+            return _exists_for_pred(key, AssetReferenceMeta.val_bool == bool(value))
+        if isinstance(value, (int, float, Decimal)):
+            num = value if isinstance(value, Decimal) else Decimal(str(value))
+            return _exists_for_pred(key, AssetReferenceMeta.val_num == num)
+        if isinstance(value, str):
+            return _exists_for_pred(key, AssetReferenceMeta.val_str == value)
+        return _exists_for_pred(key, AssetReferenceMeta.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 get_reference_by_id(
@@ -128,21 +212,6 @@ def reference_exists_for_asset_id(
     return session.execute(q).first() is not None
 
 
-def reference_exists(
-    session: Session,
-    reference_id: str,
-) -> bool:
-    """Return True if a reference with the given ID exists (not soft-deleted)."""
-    q = (
-        select(sa.literal(True))
-        .select_from(AssetReference)
-        .where(AssetReference.id == reference_id)
-        .where(AssetReference.deleted_at.is_(None))
-        .limit(1)
-    )
-    return session.execute(q).first() is not None
-
-
 def insert_reference(
     session: Session,
     asset_id: str,
@@ -267,8 +336,8 @@ def list_references_page(
         escaped, esc = escape_sql_like_string(name_contains)
         base = base.where(AssetReference.name.ilike(f"%{escaped}%", escape=esc))
 
-    base = apply_tag_filters(base, include_tags, exclude_tags)
-    base = apply_metadata_filter(base, metadata_filter)
+    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()
@@ -297,8 +366,8 @@ def list_references_page(
         count_stmt = count_stmt.where(
             AssetReference.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)
+    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)
     refs = session.execute(base).unique().scalars().all()
@@ -310,7 +379,7 @@ def list_references_page(
             select(AssetReferenceTag.asset_reference_id, Tag.name)
             .join(Tag, Tag.name == AssetReferenceTag.tag_name)
             .where(AssetReferenceTag.asset_reference_id.in_(id_list))
-            .order_by(AssetReferenceTag.tag_name.asc())
+            .order_by(AssetReferenceTag.added_at)
         )
         for ref_id, tag_name in rows.all():
             tag_map[ref_id].append(tag_name)
@@ -423,42 +492,6 @@ def update_reference_updated_at(
     )
 
 
-def rebuild_metadata_projection(session: Session, ref: AssetReference) -> None:
-    """Delete and rebuild AssetReferenceMeta rows from merged system+user metadata.
-
-    The merged dict is ``{**system_metadata, **user_metadata}`` so user keys
-    override system keys of the same name.
-    """
-    session.execute(
-        delete(AssetReferenceMeta).where(
-            AssetReferenceMeta.asset_reference_id == ref.id
-        )
-    )
-    session.flush()
-
-    merged = {**(ref.system_metadata or {}), **(ref.user_metadata or {})}
-    if not merged:
-        return
-
-    rows: list[AssetReferenceMeta] = []
-    for k, v in merged.items():
-        for r in convert_metadata_to_rows(k, v):
-            rows.append(
-                AssetReferenceMeta(
-                    asset_reference_id=ref.id,
-                    key=r["key"],
-                    ordinal=int(r["ordinal"]),
-                    val_str=r.get("val_str"),
-                    val_num=r.get("val_num"),
-                    val_bool=r.get("val_bool"),
-                    val_json=r.get("val_json"),
-                )
-            )
-    if rows:
-        session.add_all(rows)
-        session.flush()
-
-
 def set_reference_metadata(
     session: Session,
     reference_id: str,
@@ -472,24 +505,33 @@ def set_reference_metadata(
     ref.updated_at = get_utc_now()
     session.flush()
 
-    rebuild_metadata_projection(session, ref)
-
-
-def set_reference_system_metadata(
-    session: Session,
-    reference_id: str,
-    system_metadata: dict | None = None,
-) -> None:
-    """Set system_metadata on a reference and rebuild the merged projection."""
-    ref = session.get(AssetReference, reference_id)
-    if not ref:
-        raise ValueError(f"AssetReference {reference_id} not found")
-
-    ref.system_metadata = system_metadata or {}
-    ref.updated_at = get_utc_now()
+    session.execute(
+        delete(AssetReferenceMeta).where(
+            AssetReferenceMeta.asset_reference_id == reference_id
+        )
+    )
     session.flush()
 
-    rebuild_metadata_projection(session, ref)
+    if not user_metadata:
+        return
+
+    rows: list[AssetReferenceMeta] = []
+    for k, v in user_metadata.items():
+        for r in convert_metadata_to_rows(k, v):
+            rows.append(
+                AssetReferenceMeta(
+                    asset_reference_id=reference_id,
+                    key=r["key"],
+                    ordinal=int(r["ordinal"]),
+                    val_str=r.get("val_str"),
+                    val_num=r.get("val_num"),
+                    val_bool=r.get("val_bool"),
+                    val_json=r.get("val_json"),
+                )
+            )
+    if rows:
+        session.add_all(rows)
+        session.flush()
 
 
 def delete_reference_by_id(
@@ -529,19 +571,19 @@ def soft_delete_reference_by_id(
 def set_reference_preview(
     session: Session,
     reference_id: str,
-    preview_reference_id: str | None = None,
+    preview_asset_id: str | None = None,
 ) -> None:
     """Set or clear preview_id and bump updated_at. Raises on unknown IDs."""
     ref = session.get(AssetReference, reference_id)
     if not ref:
         raise ValueError(f"AssetReference {reference_id} not found")
 
-    if preview_reference_id is None:
+    if preview_asset_id is None:
         ref.preview_id = None
     else:
-        if not session.get(AssetReference, preview_reference_id):
-            raise ValueError(f"Preview AssetReference {preview_reference_id} not found")
-        ref.preview_id = preview_reference_id
+        if not session.get(Asset, preview_asset_id):
+            raise ValueError(f"Preview Asset {preview_asset_id} not found")
+        ref.preview_id = preview_asset_id
 
     ref.updated_at = get_utc_now()
     session.flush()
@@ -567,8 +609,6 @@ def list_references_by_asset_id(
         session.execute(
             select(AssetReference)
             .where(AssetReference.asset_id == asset_id)
-            .where(AssetReference.is_missing == False)  # noqa: E712
-            .where(AssetReference.deleted_at.is_(None))
             .order_by(AssetReference.id.asc())
         )
         .scalars()
@@ -576,25 +616,6 @@ def list_references_by_asset_id(
     )
 
 
-def list_all_file_paths_by_asset_id(
-    session: Session,
-    asset_id: str,
-) -> list[str]:
-    """Return every file_path for an asset, including soft-deleted/missing refs.
-
-    Used for orphan cleanup where all on-disk files must be removed.
-    """
-    return list(
-        session.execute(
-            select(AssetReference.file_path)
-            .where(AssetReference.asset_id == asset_id)
-            .where(AssetReference.file_path.isnot(None))
-        )
-        .scalars()
-        .all()
-    )
-
-
 def upsert_reference(
     session: Session,
     asset_id: str,
@@ -834,22 +855,6 @@ def bulk_update_is_missing(
     return total
 
 
-def update_is_missing_by_asset_id(
-    session: Session, asset_id: str, value: bool
-) -> int:
-    """Set is_missing flag for ALL references belonging to an asset.
-
-    Returns: Number of rows updated
-    """
-    result = session.execute(
-        sa.update(AssetReference)
-        .where(AssetReference.asset_id == asset_id)
-        .where(AssetReference.deleted_at.is_(None))
-        .values(is_missing=value)
-    )
-    return result.rowcount
-
-
 def delete_references_by_ids(session: Session, reference_ids: list[str]) -> int:
     """Delete references by their IDs.
 

app/assets/database/queries/common.py

@@ -1,14 +1,12 @@
 """Shared utilities for database query modules."""
 
 import os
-from decimal import Decimal
-from typing import Iterable, Sequence
+from typing import Iterable
 
 import sqlalchemy as sa
-from sqlalchemy import exists
 
-from app.assets.database.models import AssetReference, AssetReferenceMeta, AssetReferenceTag
-from app.assets.helpers import escape_sql_like_string, normalize_tags
+from app.assets.database.models import AssetReference
+from app.assets.helpers import escape_sql_like_string
 
 MAX_BIND_PARAMS = 800
 
@@ -54,74 +52,3 @@ def build_prefix_like_conditions(
         escaped, esc = escape_sql_like_string(base)
         conds.append(AssetReference.file_path.like(escaped + "%", escape=esc))
     return conds
-
-
-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(
-                    (AssetReferenceTag.asset_reference_id == AssetReference.id)
-                    & (AssetReferenceTag.tag_name == tag_name)
-                )
-            )
-
-    if exclude_tags:
-        stmt = stmt.where(
-            ~exists().where(
-                (AssetReferenceTag.asset_reference_id == AssetReference.id)
-                & (AssetReferenceTag.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_reference_meta projection table."""
-    if not metadata_filter:
-        return stmt
-
-    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
-        return sa.exists().where(
-            AssetReferenceMeta.asset_reference_id == AssetReference.id,
-            AssetReferenceMeta.key == key,
-            *preds,
-        )
-
-    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
-        if value is None:
-            return sa.not_(
-                sa.exists().where(
-                    AssetReferenceMeta.asset_reference_id == AssetReference.id,
-                    AssetReferenceMeta.key == key,
-                )
-            )
-
-        if isinstance(value, bool):
-            return _exists_for_pred(key, AssetReferenceMeta.val_bool == bool(value))
-        if isinstance(value, (int, float, Decimal)):
-            num = value if isinstance(value, Decimal) else Decimal(str(value))
-            return _exists_for_pred(key, AssetReferenceMeta.val_num == num)
-        if isinstance(value, str):
-            return _exists_for_pred(key, AssetReferenceMeta.val_str == value)
-        return _exists_for_pred(key, AssetReferenceMeta.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

app/assets/database/queries/tags.py

@@ -8,15 +8,12 @@ from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 
 from app.assets.database.models import (
-    Asset,
     AssetReference,
     AssetReferenceMeta,
     AssetReferenceTag,
     Tag,
 )
 from app.assets.database.queries.common import (
-    apply_metadata_filter,
-    apply_tag_filters,
     build_visible_owner_clause,
     iter_row_chunks,
 )
@@ -75,9 +72,9 @@ def get_reference_tags(session: Session, reference_id: str) -> list[str]:
         tag_name
         for (tag_name,) in (
             session.execute(
-                select(AssetReferenceTag.tag_name)
-                .where(AssetReferenceTag.asset_reference_id == reference_id)
-                .order_by(AssetReferenceTag.tag_name.asc())
+                select(AssetReferenceTag.tag_name).where(
+                    AssetReferenceTag.asset_reference_id == reference_id
+                )
             )
         ).all()
     ]
@@ -120,7 +117,7 @@ def set_reference_tags(
         )
         session.flush()
 
-    return SetTagsResult(added=sorted(to_add), removed=sorted(to_remove), total=sorted(desired))
+    return SetTagsResult(added=to_add, removed=to_remove, total=desired)
 
 
 def add_tags_to_reference(
@@ -275,12 +272,6 @@ def list_tags_with_usage(
         .select_from(AssetReferenceTag)
         .join(AssetReference, AssetReference.id == AssetReferenceTag.asset_reference_id)
         .where(build_visible_owner_clause(owner_id))
-        .where(
-            sa.or_(
-                AssetReference.is_missing == False,  # noqa: E712
-                AssetReferenceTag.tag_name == "missing",
-            )
-        )
         .where(AssetReference.deleted_at.is_(None))
         .group_by(AssetReferenceTag.tag_name)
         .subquery()
@@ -317,12 +308,6 @@ def list_tags_with_usage(
             select(AssetReferenceTag.tag_name)
             .join(AssetReference, AssetReference.id == AssetReferenceTag.asset_reference_id)
             .where(build_visible_owner_clause(owner_id))
-            .where(
-                sa.or_(
-                    AssetReference.is_missing == False,  # noqa: E712
-                    AssetReferenceTag.tag_name == "missing",
-                )
-            )
             .where(AssetReference.deleted_at.is_(None))
             .group_by(AssetReferenceTag.tag_name)
         )
@@ -335,53 +320,6 @@ def list_tags_with_usage(
     return rows_norm, int(total or 0)
 
 
-def list_tag_counts_for_filtered_assets(
-    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 = 100,
-) -> dict[str, int]:
-    """Return tag counts for assets matching the given filters.
-
-    Uses the same filtering logic as list_references_page but returns
-    {tag_name: count} instead of paginated references.
-    """
-    # Build a subquery of matching reference IDs
-    ref_sq = (
-        select(AssetReference.id)
-        .join(Asset, Asset.id == AssetReference.asset_id)
-        .where(build_visible_owner_clause(owner_id))
-        .where(AssetReference.is_missing == False)  # noqa: E712
-        .where(AssetReference.deleted_at.is_(None))
-    )
-
-    if name_contains:
-        escaped, esc = escape_sql_like_string(name_contains)
-        ref_sq = ref_sq.where(AssetReference.name.ilike(f"%{escaped}%", escape=esc))
-
-    ref_sq = apply_tag_filters(ref_sq, include_tags, exclude_tags)
-    ref_sq = apply_metadata_filter(ref_sq, metadata_filter)
-    ref_sq = ref_sq.subquery()
-
-    # Count tags across those references
-    q = (
-        select(
-            AssetReferenceTag.tag_name,
-            func.count(AssetReferenceTag.asset_reference_id).label("cnt"),
-        )
-        .where(AssetReferenceTag.asset_reference_id.in_(select(ref_sq.c.id)))
-        .group_by(AssetReferenceTag.tag_name)
-        .order_by(func.count(AssetReferenceTag.asset_reference_id).desc(), AssetReferenceTag.tag_name.asc())
-        .limit(limit)
-    )
-
-    rows = session.execute(q).all()
-    return {tag_name: int(cnt) for tag_name, cnt in rows}
-
-
 def bulk_insert_tags_and_meta(
     session: Session,
     tag_rows: list[dict],

app/assets/scanner.py

@@ -18,7 +18,7 @@ from app.assets.database.queries import (
     mark_references_missing_outside_prefixes,
     reassign_asset_references,
     remove_missing_tag_for_asset_id,
-    set_reference_system_metadata,
+    set_reference_metadata,
     update_asset_hash_and_mime,
 )
 from app.assets.services.bulk_ingest import (
@@ -490,8 +490,8 @@ def enrich_asset(
             logging.warning("Failed to hash %s: %s", file_path, e)
 
     if extract_metadata and metadata:
-        system_metadata = metadata.to_user_metadata()
-        set_reference_system_metadata(session, reference_id, system_metadata)
+        user_metadata = metadata.to_user_metadata()
+        set_reference_metadata(session, reference_id, user_metadata)
 
     if full_hash:
         existing = get_asset_by_hash(session, full_hash)

app/assets/services/asset_management.py

@@ -16,12 +16,10 @@ from app.assets.database.queries import (
     get_reference_by_id,
     get_reference_with_owner_check,
     list_references_page,
-    list_all_file_paths_by_asset_id,
     list_references_by_asset_id,
     set_reference_metadata,
     set_reference_preview,
     set_reference_tags,
-    update_asset_hash_and_mime,
     update_reference_access_time,
     update_reference_name,
     update_reference_updated_at,
@@ -69,8 +67,6 @@ def update_asset_metadata(
     user_metadata: UserMetadata = None,
     tag_origin: str = "manual",
     owner_id: str = "",
-    mime_type: str | None = None,
-    preview_id: str | None = None,
 ) -> AssetDetailResult:
     with create_session() as session:
         ref = get_reference_with_owner_check(session, reference_id, owner_id)
@@ -107,21 +103,6 @@ def update_asset_metadata(
             )
             touched = True
 
-        if mime_type is not None:
-            updated = update_asset_hash_and_mime(
-                session, asset_id=ref.asset_id, mime_type=mime_type
-            )
-            if updated:
-                touched = True
-
-        if preview_id is not None:
-            set_reference_preview(
-                session,
-                reference_id=reference_id,
-                preview_reference_id=preview_id,
-            )
-            touched = True
-
         if touched and user_metadata is None:
             update_reference_updated_at(session, reference_id=reference_id)
 
@@ -178,9 +159,11 @@ def delete_asset_reference(
             session.commit()
             return True
 
-        # Orphaned asset - gather ALL file paths (including
-        # soft-deleted / missing refs) so their on-disk files get cleaned up.
-        file_paths = list_all_file_paths_by_asset_id(session, asset_id=asset_id)
+        # Orphaned asset - delete it and its files
+        refs = list_references_by_asset_id(session, asset_id=asset_id)
+        file_paths = [
+            r.file_path for r in (refs or []) if getattr(r, "file_path", None)
+        ]
         # Also include the just-deleted file path
         if file_path:
             file_paths.append(file_path)
@@ -202,7 +185,7 @@ def delete_asset_reference(
 
 def set_asset_preview(
     reference_id: str,
-    preview_reference_id: str | None = None,
+    preview_asset_id: str | None = None,
     owner_id: str = "",
 ) -> AssetDetailResult:
     with create_session() as session:
@@ -211,7 +194,7 @@ def set_asset_preview(
         set_reference_preview(
             session,
             reference_id=reference_id,
-            preview_reference_id=preview_reference_id,
+            preview_asset_id=preview_asset_id,
         )
 
         result = fetch_reference_asset_and_tags(
@@ -280,47 +263,6 @@ def list_assets_page(
         return ListAssetsResult(items=items, total=total)
 
 
-def resolve_hash_to_path(
-    asset_hash: str,
-    owner_id: str = "",
-) -> DownloadResolutionResult | None:
-    """Resolve a blake3 hash to an on-disk file path.
-
-    Only references visible to *owner_id* are considered (owner-less
-    references are always visible).
-
-    Returns a DownloadResolutionResult with abs_path, content_type, and
-    download_name, or None if no asset or live path is found.
-    """
-    with create_session() as session:
-        asset = queries_get_asset_by_hash(session, asset_hash)
-        if not asset:
-            return None
-        refs = list_references_by_asset_id(session, asset_id=asset.id)
-        visible = [
-            r for r in refs
-            if r.owner_id == "" or r.owner_id == owner_id
-        ]
-        abs_path = select_best_live_path(visible)
-        if not abs_path:
-            return None
-        display_name = os.path.basename(abs_path)
-        for ref in visible:
-            if ref.file_path == abs_path and ref.name:
-                display_name = ref.name
-                break
-        ctype = (
-            asset.mime_type
-            or mimetypes.guess_type(display_name)[0]
-            or "application/octet-stream"
-        )
-    return DownloadResolutionResult(
-        abs_path=abs_path,
-        content_type=ctype,
-        download_name=display_name,
-    )
-
-
 def resolve_asset_for_download(
     reference_id: str,
     owner_id: str = "",

app/assets/services/ingest.py

@@ -11,14 +11,13 @@ from app.assets.database.queries import (
     add_tags_to_reference,
     fetch_reference_and_asset,
     get_asset_by_hash,
+    get_existing_asset_ids,
     get_reference_by_file_path,
     get_reference_tags,
     get_or_create_reference,
-    reference_exists,
     remove_missing_tag_for_asset_id,
     set_reference_metadata,
     set_reference_tags,
-    update_asset_hash_and_mime,
     upsert_asset,
     upsert_reference,
     validate_tags_exist,
@@ -27,7 +26,6 @@ from app.assets.helpers import normalize_tags
 from app.assets.services.file_utils import get_size_and_mtime_ns
 from app.assets.services.path_utils import (
     compute_relative_filename,
-    get_name_and_tags_from_asset_path,
     resolve_destination_from_tags,
     validate_path_within_base,
 )
@@ -67,7 +65,7 @@ def _ingest_file_from_path(
 
     with create_session() as session:
         if preview_id:
-            if not reference_exists(session, preview_id):
+            if preview_id not in get_existing_asset_ids(session, [preview_id]):
                 preview_id = None
 
         asset, asset_created, asset_updated = upsert_asset(
@@ -137,8 +135,6 @@ def _register_existing_asset(
     tags: list[str] | None = None,
     tag_origin: str = "manual",
     owner_id: str = "",
-    mime_type: str | None = None,
-    preview_id: str | None = None,
 ) -> RegisterAssetResult:
     user_metadata = user_metadata or {}
 
@@ -147,25 +143,14 @@ def _register_existing_asset(
         if not asset:
             raise ValueError(f"No asset with hash {asset_hash}")
 
-        if mime_type and not asset.mime_type:
-            update_asset_hash_and_mime(session, asset_id=asset.id, mime_type=mime_type)
-
-        if preview_id:
-            if not reference_exists(session, preview_id):
-                preview_id = None
-
         ref, ref_created = get_or_create_reference(
             session,
             asset_id=asset.id,
             owner_id=owner_id,
             name=name,
-            preview_id=preview_id,
         )
 
         if not ref_created:
-            if preview_id and ref.preview_id != preview_id:
-                ref.preview_id = preview_id
-
             tag_names = get_reference_tags(session, reference_id=ref.id)
             result = RegisterAssetResult(
                 ref=extract_reference_data(ref),
@@ -257,8 +242,6 @@ def upload_from_temp_path(
     client_filename: str | None = None,
     owner_id: str = "",
     expected_hash: str | None = None,
-    mime_type: str | None = None,
-    preview_id: str | None = None,
 ) -> UploadResult:
     try:
         digest, _ = hashing.compute_blake3_hash(temp_path)
@@ -287,8 +270,6 @@ def upload_from_temp_path(
             tags=tags or [],
             tag_origin="manual",
             owner_id=owner_id,
-            mime_type=mime_type,
-            preview_id=preview_id,
         )
         return UploadResult(
             ref=result.ref,
@@ -310,7 +291,7 @@ def upload_from_temp_path(
     dest_abs = os.path.abspath(os.path.join(dest_dir, hashed_basename))
     validate_path_within_base(dest_abs, base_dir)
 
-    content_type = mime_type or (
+    content_type = (
         mimetypes.guess_type(os.path.basename(src_for_ext), strict=False)[0]
         or mimetypes.guess_type(hashed_basename, strict=False)[0]
         or "application/octet-stream"
@@ -334,7 +315,7 @@ def upload_from_temp_path(
         mime_type=content_type,
         info_name=_sanitize_filename(name or client_filename, fallback=digest),
         owner_id=owner_id,
-        preview_id=preview_id,
+        preview_id=None,
         user_metadata=user_metadata or {},
         tags=tags,
         tag_origin="manual",
@@ -361,99 +342,30 @@ def upload_from_temp_path(
     )
 
 
-def register_file_in_place(
-    abs_path: str,
-    name: str,
-    tags: list[str],
-    owner_id: str = "",
-    mime_type: str | None = None,
-) -> UploadResult:
-    """Register an already-saved file in the asset database without moving it.
-
-    Tags are derived from the filesystem path (root category + subfolder names),
-    merged with any caller-provided tags, matching the behavior of the scanner.
-    If the path is not under a known root, only the caller-provided tags are used.
-    """
-    try:
-        _, path_tags = get_name_and_tags_from_asset_path(abs_path)
-    except ValueError:
-        path_tags = []
-    merged_tags = normalize_tags([*path_tags, *tags])
-
-    try:
-        digest, _ = hashing.compute_blake3_hash(abs_path)
-    except ImportError as e:
-        raise DependencyMissingError(str(e))
-    except Exception as e:
-        raise RuntimeError(f"failed to hash file: {e}")
-    asset_hash = "blake3:" + digest
-
-    size_bytes, mtime_ns = get_size_and_mtime_ns(abs_path)
-    content_type = mime_type or (
-        mimetypes.guess_type(abs_path, strict=False)[0]
-        or "application/octet-stream"
-    )
-
-    ingest_result = _ingest_file_from_path(
-        abs_path=abs_path,
-        asset_hash=asset_hash,
-        size_bytes=size_bytes,
-        mtime_ns=mtime_ns,
-        mime_type=content_type,
-        info_name=_sanitize_filename(name, fallback=digest),
-        owner_id=owner_id,
-        tags=merged_tags,
-        tag_origin="upload",
-        require_existing_tags=False,
-    )
-    reference_id = ingest_result.reference_id
-    if not reference_id:
-        raise RuntimeError("failed to create asset reference")
-
-    with create_session() as session:
-        pair = fetch_reference_and_asset(
-            session, reference_id=reference_id, owner_id=owner_id
-        )
-        if not pair:
-            raise RuntimeError("inconsistent DB state after ingest")
-        ref, asset = pair
-        tag_names = get_reference_tags(session, reference_id=ref.id)
-
-    return UploadResult(
-        ref=extract_reference_data(ref),
-        asset=extract_asset_data(asset),
-        tags=tag_names,
-        created_new=ingest_result.asset_created,
-    )
-
-
 def create_from_hash(
     hash_str: str,
     name: str,
     tags: list[str] | None = None,
     user_metadata: dict | None = None,
     owner_id: str = "",
-    mime_type: str | None = None,
-    preview_id: str | None = None,
 ) -> UploadResult | None:
     canonical = hash_str.strip().lower()
 
-    try:
-        result = _register_existing_asset(
-            asset_hash=canonical,
-            name=_sanitize_filename(
-                name, fallback=canonical.split(":", 1)[1] if ":" in canonical else canonical
-            ),
-            user_metadata=user_metadata or {},
-            tags=tags or [],
-            tag_origin="manual",
-            owner_id=owner_id,
-            mime_type=mime_type,
-            preview_id=preview_id,
-        )
-    except ValueError:
-        logging.warning("create_from_hash: no asset found for hash %s", canonical)
-        return None
+    with create_session() as session:
+        asset = get_asset_by_hash(session, asset_hash=canonical)
+        if not asset:
+            return None
+
+    result = _register_existing_asset(
+        asset_hash=canonical,
+        name=_sanitize_filename(
+            name, fallback=canonical.split(":", 1)[1] if ":" in canonical else canonical
+        ),
+        user_metadata=user_metadata or {},
+        tags=tags or [],
+        tag_origin="manual",
+        owner_id=owner_id,
+    )
 
     return UploadResult(
         ref=result.ref,

app/assets/services/schemas.py

@@ -25,9 +25,7 @@ class ReferenceData:
     preview_id: str | None
     created_at: datetime
     updated_at: datetime
-    system_metadata: dict[str, Any] | None = None
-    job_id: str | None = None
-    last_access_time: datetime | None = None
+    last_access_time: datetime | None
 
 
 @dataclass(frozen=True)
@@ -95,8 +93,6 @@ def extract_reference_data(ref: AssetReference) -> ReferenceData:
         file_path=ref.file_path,
         user_metadata=ref.user_metadata,
         preview_id=ref.preview_id,
-        system_metadata=ref.system_metadata,
-        job_id=ref.job_id,
         created_at=ref.created_at,
         updated_at=ref.updated_at,
         last_access_time=ref.last_access_time,

app/assets/services/tagging.py

@@ -1,5 +1,3 @@
-from typing import Sequence
-
 from app.assets.database.queries import (
     AddTagsResult,
     RemoveTagsResult,
@@ -8,7 +6,6 @@ from app.assets.database.queries import (
     list_tags_with_usage,
     remove_tags_from_reference,
 )
-from app.assets.database.queries.tags import list_tag_counts_for_filtered_assets
 from app.assets.services.schemas import TagUsage
 from app.database.db import create_session
 
@@ -76,23 +73,3 @@ def list_tags(
         )
 
     return [TagUsage(name, tag_type, count) for name, tag_type, count in rows], total
-
-
-def list_tag_histogram(
-    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 = 100,
-) -> dict[str, int]:
-    with create_session() as session:
-        return list_tag_counts_for_filtered_assets(
-            session,
-            owner_id=owner_id,
-            include_tags=include_tags,
-            exclude_tags=exclude_tags,
-            name_contains=name_contains,
-            metadata_filter=metadata_filter,
-            limit=limit,
-        )

app/database/models.py

@@ -1,18 +1,9 @@
 from typing import Any
 from datetime import datetime
-from sqlalchemy import MetaData
 from sqlalchemy.orm import DeclarativeBase
 
-NAMING_CONVENTION = {
-    "ix": "ix_%(table_name)s_%(column_0_N_name)s",
-    "uq": "uq_%(table_name)s_%(column_0_N_name)s",
-    "ck": "ck_%(table_name)s_%(constraint_name)s",
-    "fk": "fk_%(table_name)s_%(column_0_name)s_%(referred_table_name)s",
-    "pk": "pk_%(table_name)s",
-}
-
 class Base(DeclarativeBase):
-    metadata = MetaData(naming_convention=NAMING_CONVENTION)
+    pass
 
 def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]:
     fields = obj.__table__.columns.keys()

app/user_manager.py

@@ -6,7 +6,6 @@ import uuid
 import glob
 import shutil
 import logging
-import tempfile
 from aiohttp import web
 from urllib import parse
 from comfy.cli_args import args
@@ -378,15 +377,8 @@ class UserManager():
             try:
                 body = await request.read()
 
-                dir_name = os.path.dirname(path)
-                fd, tmp_path = tempfile.mkstemp(dir=dir_name)
-                try:
-                    with os.fdopen(fd, "wb") as f:
-                        f.write(body)
-                    os.replace(tmp_path, path)
-                except:
-                    os.unlink(tmp_path)
-                    raise
+                with open(path, "wb") as f:
+                    f.write(body)
             except OSError as e:
                 logging.warning(f"Error saving file '{path}': {e}")
                 return web.Response(

comfy/cli_args.py

@@ -149,7 +149,6 @@ parser.add_argument("--reserve-vram", type=float, default=None, help="Set the am
 parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.")
 parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.")
 parser.add_argument("--disable-dynamic-vram", action="store_true", help="Disable dynamic VRAM and use estimate based model loading.")
-parser.add_argument("--enable-dynamic-vram", action="store_true", help="Enable dynamic VRAM on systems where it's not enabled by default.")
 
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 
@@ -263,6 +262,4 @@ else:
     args.fast = set(args.fast)
 
 def enables_dynamic_vram():
-    if args.enable_dynamic_vram:
-        return True
     return not args.disable_dynamic_vram and not args.highvram and not args.gpu_only and not args.novram and not args.cpu

comfy/context_windows.py

@@ -93,50 +93,6 @@ class IndexListCallbacks:
         return {}
 
 
-def slice_cond(cond_value, window: IndexListContextWindow, x_in: torch.Tensor, device, temporal_dim: int, temporal_scale: int=1, temporal_offset: int=0, retain_index_list: list[int]=[]):
-    if not (hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor)):
-        return None
-    cond_tensor = cond_value.cond
-    if temporal_dim >= cond_tensor.ndim:
-        return None
-
-    cond_size = cond_tensor.size(temporal_dim)
-
-    if temporal_scale == 1:
-        expected_size = x_in.size(window.dim) - temporal_offset
-        if cond_size != expected_size:
-            return None
-
-    if temporal_offset == 0 and temporal_scale == 1:
-        sliced = window.get_tensor(cond_tensor, device, dim=temporal_dim, retain_index_list=retain_index_list)
-        return cond_value._copy_with(sliced)
-
-    # skip leading latent positions that have no corresponding conditioning (e.g. reference frames)
-    if temporal_offset > 0:
-        indices = [i - temporal_offset for i in window.index_list[temporal_offset:]]
-        indices = [i for i in indices if 0 <= i]
-    else:
-        indices = list(window.index_list)
-
-    if not indices:
-        return None
-
-    if temporal_scale > 1:
-        scaled = []
-        for i in indices:
-            for k in range(temporal_scale):
-                si = i * temporal_scale + k
-                if si < cond_size:
-                    scaled.append(si)
-        indices = scaled
-        if not indices:
-            return None
-
-    idx = tuple([slice(None)] * temporal_dim + [indices])
-    sliced = cond_tensor[idx].to(device)
-    return cond_value._copy_with(sliced)
-
-
 @dataclass
 class ContextSchedule:
     name: str
@@ -221,17 +177,10 @@ class IndexListContextHandler(ContextHandlerABC):
                                     new_cond_item[cond_key] = result
                                     handled = True
                                     break
-                            if not handled and self._model is not None:
-                                result = self._model.resize_cond_for_context_window(
-                                    cond_key, cond_value, window, x_in, device,
-                                    retain_index_list=self.cond_retain_index_list)
-                                if result is not None:
-                                    new_cond_item[cond_key] = result
-                                    handled = True
                             if handled:
                                 continue
                             if isinstance(cond_value, torch.Tensor):
-                                if (self.dim < cond_value.ndim and cond_value.size(self.dim) == x_in.size(self.dim)) or \
+                                if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
                                    (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):
                                     new_cond_item[cond_key] = window.get_tensor(cond_value, device)
                             # Handle audio_embed (temporal dim is 1)
@@ -275,7 +224,6 @@ class IndexListContextHandler(ContextHandlerABC):
         return context_windows
 
     def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
-        self._model = model
         self.set_step(timestep, model_options)
         context_windows = self.get_context_windows(model, x_in, model_options)
         enumerated_context_windows = list(enumerate(context_windows))

comfy/ldm/cascade/stage_a.py

@@ -136,7 +136,16 @@ class ResBlock(nn.Module):
             ops.Linear(c_hidden, c),
         )
 
-        self.gammas = nn.Parameter(torch.zeros(6), requires_grad=False)
+        self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True)
+
+        # Init weights
+        def _basic_init(module):
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+
+        self.apply(_basic_init)
 
     def _norm(self, x, norm):
         return norm(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)

comfy/ldm/hunyuan3dv2_1/hunyuandit.py

@@ -343,7 +343,6 @@ class CrossAttention(nn.Module):
             k.reshape(b, s2, self.num_heads * self.head_dim),
             v,
             heads=self.num_heads,
-            low_precision_attention=False,
         )
 
         out = self.out_proj(x)
@@ -413,7 +412,6 @@ class Attention(nn.Module):
             key.reshape(B, N, self.num_heads * self.head_dim),
             value,
             heads=self.num_heads,
-            low_precision_attention=False,
         )
 
         x = self.out_proj(x)

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -23,11 +23,6 @@ class CausalConv3d(nn.Module):
         self.in_channels = in_channels
         self.out_channels = out_channels
 
-        if isinstance(stride, int):
-            self.time_stride = stride
-        else:
-            self.time_stride = stride[0]
-
         kernel_size = (kernel_size, kernel_size, kernel_size)
         self.time_kernel_size = kernel_size[0]
 
@@ -63,25 +58,16 @@ class CausalConv3d(nn.Module):
         pieces = [ cached, x ]
         if is_end and not causal:
             pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
-        input_length = sum([piece.shape[2] for piece in pieces])
-        cache_length = (self.time_kernel_size - self.time_stride) + ((input_length - self.time_kernel_size) % self.time_stride)
 
         needs_caching = not is_end
-        if needs_caching and cache_length == 0:
-            self.temporal_cache_state[tid] = (x[:, :, :0, :, :], False)
+        if needs_caching and x.shape[2] >= self.time_kernel_size - 1:
             needs_caching = False
-        if needs_caching and x.shape[2] >= cache_length:
-            needs_caching = False
-            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
 
         x = torch.cat(pieces, dim=2)
-        del pieces
-        del cached
 
         if needs_caching:
-            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
-        elif is_end:
-            self.temporal_cache_state[tid] = (None, True)
+            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, :, :]
 

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -233,7 +233,10 @@ class Encoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def _forward_chunk(self, sample: torch.FloatTensor) -> Optional[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)
         sample = self.conv_in(sample)
 
         checkpoint_fn = (
@@ -244,14 +247,10 @@ class Encoder(nn.Module):
 
         for down_block in self.down_blocks:
             sample = checkpoint_fn(down_block)(sample)
-            if sample is None or sample.shape[2] == 0:
-                return None
 
         sample = self.conv_norm_out(sample)
         sample = self.conv_act(sample)
         sample = self.conv_out(sample)
-        if sample is None or sample.shape[2] == 0:
-            return None
 
         if self.latent_log_var == "uniform":
             last_channel = sample[:, -1:, ...]
@@ -283,35 +282,9 @@ class Encoder(nn.Module):
 
         return sample
 
-    def forward_orig(self, sample: torch.FloatTensor, device=None) -> torch.FloatTensor:
-        r"""The forward method of the `Encoder` class."""
-
-        max_chunk_size = get_max_chunk_size(sample.device if device is None else device) * 2  # encoder is more memory-efficient than decoder
-        frame_size = sample[:, :, :1, :, :].numel() * sample.element_size()
-        frame_size = int(frame_size * (self.conv_in.out_channels / self.conv_in.in_channels))
-
-        outputs = []
-        samples = [sample[:, :, :1, :, :]]
-        if sample.shape[2] > 1:
-            chunk_t = max(2, max_chunk_size // frame_size)
-            if chunk_t < 4:
-                chunk_t = 2
-            elif chunk_t < 8:
-                chunk_t = 4
-            else:
-                chunk_t = (chunk_t // 8) * 8
-            samples += list(torch.split(sample[:, :, 1:, :, :], chunk_t, dim=2))
-        for chunk_idx, chunk in enumerate(samples):
-            if chunk_idx == len(samples) - 1:
-                mark_conv3d_ended(self)
-            chunk = patchify(chunk, patch_size_hw=self.patch_size, patch_size_t=1).to(device=device)
-            output = self._forward_chunk(chunk)
-            if output is not None:
-                outputs.append(output)
-
-        return torch_cat_if_needed(outputs, dim=2)
-
     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:
@@ -324,23 +297,7 @@ class Encoder(nn.Module):
                     module.temporal_cache_state.pop(tid, None)
 
 
-MIN_VRAM_FOR_CHUNK_SCALING = 6 * 1024 ** 3
-MAX_VRAM_FOR_CHUNK_SCALING = 24 * 1024 ** 3
-MIN_CHUNK_SIZE = 32 * 1024 ** 2
-MAX_CHUNK_SIZE = 128 * 1024 ** 2
-
-def get_max_chunk_size(device: torch.device) -> int:
-    total_memory = comfy.model_management.get_total_memory(dev=device)
-
-    if total_memory <= MIN_VRAM_FOR_CHUNK_SCALING:
-        return MIN_CHUNK_SIZE
-    if total_memory >= MAX_VRAM_FOR_CHUNK_SCALING:
-        return MAX_CHUNK_SIZE
-
-    interp = (total_memory - MIN_VRAM_FOR_CHUNK_SCALING) / (
-        MAX_VRAM_FOR_CHUNK_SCALING - MIN_VRAM_FOR_CHUNK_SCALING
-    )
-    return int(MIN_CHUNK_SIZE + interp * (MAX_CHUNK_SIZE - MIN_CHUNK_SIZE))
+MAX_CHUNK_SIZE=(128 * 1024 ** 2)
 
 class Decoder(nn.Module):
     r"""
@@ -500,17 +457,6 @@ class Decoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-        # Precompute output scale factors: (channels, (t_scale, h_scale, w_scale), t_offset)
-        ts, hs, ws, to = 1, 1, 1, 0
-        for block in self.up_blocks:
-            if isinstance(block, DepthToSpaceUpsample):
-                ts *= block.stride[0]
-                hs *= block.stride[1]
-                ws *= block.stride[2]
-                if block.stride[0] > 1:
-                    to = to * block.stride[0] + 1
-        self._output_scale = (out_channels // (patch_size ** 2), (ts, hs * patch_size, ws * patch_size), to)
-
         self.timestep_conditioning = timestep_conditioning
 
         if timestep_conditioning:
@@ -532,62 +478,11 @@ class Decoder(nn.Module):
             )
 
 
-    def decode_output_shape(self, input_shape):
-        c, (ts, hs, ws), to = self._output_scale
-        return (input_shape[0], c, input_shape[2] * ts - to, input_shape[3] * hs, input_shape[4] * ws)
-
-    def run_up(self, idx, sample_ref, ended, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size):
-        sample = sample_ref[0]
-        sample_ref[0] = None
-        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:
-                sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
-                t = sample.shape[2]
-                output_buffer[:, :, output_offset[0]:output_offset[0] + t].copy_(sample)
-                output_offset[0] += t
-            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
-
-        if num_chunks == 1:
-            # when we are not chunking, detach our x so the callee can free it as soon as they are done
-            next_sample_ref = [sample]
-            del sample
-            self.run_up(idx + 1, next_sample_ref, ended, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
-            return
-        else:
-            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
-
-            for chunk_idx, sample1 in enumerate(samples):
-                self.run_up(idx + 1, [sample1], ended and chunk_idx == len(samples) - 1, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
-
+    # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
     def forward_orig(
         self,
         sample: torch.FloatTensor,
         timestep: Optional[torch.Tensor] = None,
-        output_buffer: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
         r"""The forward method of the `Decoder` class."""
         batch_size = sample.shape[0]
@@ -602,7 +497,6 @@ class Decoder(nn.Module):
         )
 
         timestep_shift_scale = None
-        scaled_timestep = None
         if self.timestep_conditioning:
             assert (
                 timestep is not None
@@ -630,18 +524,48 @@ class Decoder(nn.Module):
             )
             timestep_shift_scale = ada_values.unbind(dim=1)
 
-        if output_buffer is None:
-            output_buffer = torch.empty(
-                self.decode_output_shape(sample.shape),
-                dtype=sample.dtype, device=comfy.model_management.intermediate_device(),
-            )
-        output_offset = [0]
+        output = []
 
-        max_chunk_size = get_max_chunk_size(sample.device)
+        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.to(comfy.model_management.intermediate_device()))
+                return
 
-        self.run_up(0, [sample], True, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
+            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)
 
-        return output_buffer
+            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:
@@ -765,25 +689,12 @@ class SpaceToDepthDownsample(nn.Module):
             causal=True,
             spatial_padding_mode=spatial_padding_mode,
         )
-        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True):
-        tid = threading.get_ident()
-        cached, pad_first, cached_x, cached_input = self.temporal_cache_state.get(tid, (None, True, None, None))
-        if cached_input is not None:
-            x = torch_cat_if_needed([cached_input, x], dim=2)
-            cached_input = None
-
-        if self.stride[0] == 2 and pad_first:
+        if self.stride[0] == 2:
             x = torch.cat(
                 [x[:, :, :1, :, :], x], dim=2
             )  # duplicate first frames for padding
-            pad_first = False
-
-        if x.shape[2] < self.stride[0]:
-            cached_input = x
-            self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
-            return None
 
         # skip connection
         x_in = rearrange(
@@ -798,26 +709,15 @@ class SpaceToDepthDownsample(nn.Module):
 
         # conv
         x = self.conv(x, causal=causal)
-        if self.stride[0] == 2 and x.shape[2] == 1:
-            if cached_x is not None:
-                x = torch_cat_if_needed([cached_x, x], dim=2)
-                cached_x = None
-            else:
-                cached_x = x
-                x = None
+        x = rearrange(
+            x,
+            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
 
-        if x is not None:
-            x = rearrange(
-                x,
-                "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
-                p1=self.stride[0],
-                p2=self.stride[1],
-                p3=self.stride[2],
-            )
-
-        cached = add_exchange_cache(x, cached, x_in, dim=2)
-
-        self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
+        x = x + x_in
 
         return x
 
@@ -1150,8 +1050,6 @@ class processor(nn.Module):
         return (x - self.get_buffer("mean-of-means").view(1, -1, 1, 1, 1).to(x)) / self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)
 
 class VideoVAE(nn.Module):
-    comfy_has_chunked_io = True
-
     def __init__(self, version=0, config=None):
         super().__init__()
 
@@ -1294,15 +1192,14 @@ class VideoVAE(nn.Module):
             }
         return config
 
-    def encode(self, x, device=None):
-        x = x[:, :, :max(1, 1 + ((x.shape[2] - 1) // 8) * 8), :, :]
-        means, logvar = torch.chunk(self.encoder(x, device=device), 2, dim=1)
+    def encode(self, x):
+        frames_count = x.shape[2]
+        if ((frames_count - 1) % 8) != 0:
+            raise ValueError("Invalid number of frames: Encode input must have 1 + 8 * x frames (e.g., 1, 9, 17, ...). Please check your input.")
+        means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
         return self.per_channel_statistics.normalize(means)
 
-    def decode_output_shape(self, input_shape):
-        return self.decoder.decode_output_shape(input_shape)
-
-    def decode(self, x, output_buffer=None):
+    def decode(self, x):
         if self.timestep_conditioning: #TODO: seed
             x = torch.randn_like(x) * self.decode_noise_scale + (1.0 - self.decode_noise_scale) * x
-        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=self.decode_timestep, output_buffer=output_buffer)
+        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=self.decode_timestep)

comfy/ldm/wan/vae.py

@@ -99,7 +99,7 @@ class Resample(nn.Module):
         else:
             self.resample = nn.Identity()
 
-    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
         b, c, t, h, w = x.size()
         if self.mode == 'upsample3d':
             if feat_cache is not None:
@@ -109,7 +109,22 @@ class Resample(nn.Module):
                     feat_idx[0] += 1
                 else:
 
-                    cache_x = x[:, :, -CACHE_T:, :, :]
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
                     if feat_cache[idx] == 'Rep':
                         x = self.time_conv(x)
                     else:
@@ -130,24 +145,19 @@ class Resample(nn.Module):
             if feat_cache is not None:
                 idx = feat_idx[0]
                 if feat_cache[idx] is None:
-                    feat_cache[idx] = x
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
                 else:
 
-                    cache_x = x[:, :, -1:, :, :]
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    # if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx]!='Rep':
+                    #     # cache last frame of last two chunk
+                    #     cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
                     x = self.time_conv(
                         torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
                     feat_cache[idx] = cache_x
-
-                    deferred_x = feat_cache[idx + 1]
-                    if deferred_x is not None:
-                        x = torch.cat([deferred_x, x], 2)
-                        feat_cache[idx + 1] = None
-
-                    if x.shape[2] == 1 and not final:
-                        feat_cache[idx + 1] = x
-                        x = None
-
-                feat_idx[0] += 2
+                    feat_idx[0] += 1
         return x
 
 
@@ -167,12 +177,19 @@ class ResidualBlock(nn.Module):
         self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
             if in_dim != out_dim else nn.Identity()
 
-    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
         old_x = x
         for layer in self.residual:
             if isinstance(layer, CausalConv3d) and feat_cache is not None:
                 idx = feat_idx[0]
-                cache_x = x[:, :, -CACHE_T:, :, :]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
                 x = layer(x, cache_list=feat_cache, cache_idx=idx)
                 feat_cache[idx] = cache_x
                 feat_idx[0] += 1
@@ -196,7 +213,7 @@ class AttentionBlock(nn.Module):
         self.proj = ops.Conv2d(dim, dim, 1)
         self.optimized_attention = vae_attention()
 
-    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+    def forward(self, x):
         identity = x
         b, c, t, h, w = x.size()
         x = rearrange(x, 'b c t h w -> (b t) c h w')
@@ -266,10 +283,17 @@ class Encoder3d(nn.Module):
             RMS_norm(out_dim, images=False), nn.SiLU(),
             CausalConv3d(out_dim, z_dim, 3, padding=1))
 
-    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
         if feat_cache is not None:
             idx = feat_idx[0]
-            cache_x = x[:, :, -CACHE_T:, :, :]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
             x = self.conv1(x, feat_cache[idx])
             feat_cache[idx] = cache_x
             feat_idx[0] += 1
@@ -279,16 +303,14 @@ class Encoder3d(nn.Module):
         ## downsamples
         for layer in self.downsamples:
             if feat_cache is not None:
-                x = layer(x, feat_cache, feat_idx, final=final)
-                if x is None:
-                    return None
+                x = layer(x, feat_cache, feat_idx)
             else:
                 x = layer(x)
 
         ## middle
         for layer in self.middle:
-            if feat_cache is not None:
-                x = layer(x, feat_cache, feat_idx, final=final)
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
             else:
                 x = layer(x)
 
@@ -296,7 +318,14 @@ class Encoder3d(nn.Module):
         for layer in self.head:
             if isinstance(layer, CausalConv3d) and feat_cache is not None:
                 idx = feat_idx[0]
-                cache_x = x[:, :, -CACHE_T:, :, :]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
                 x = layer(x, feat_cache[idx])
                 feat_cache[idx] = cache_x
                 feat_idx[0] += 1
@@ -360,48 +389,18 @@ class Decoder3d(nn.Module):
             RMS_norm(out_dim, images=False), nn.SiLU(),
             CausalConv3d(out_dim, output_channels, 3, padding=1))
 
-    def run_up(self, layer_idx, x_ref, feat_cache, feat_idx, out_chunks):
-        x = x_ref[0]
-        x_ref[0] = None
-        if layer_idx >= len(self.upsamples):
-            for layer in self.head:
-                if isinstance(layer, CausalConv3d) and feat_cache is not None:
-                    cache_x = x[:, :, -CACHE_T:, :, :]
-                    x = layer(x, feat_cache[feat_idx[0]])
-                    feat_cache[feat_idx[0]] = cache_x
-                    feat_idx[0] += 1
-                else:
-                    x = layer(x)
-            out_chunks.append(x)
-            return
-
-        layer = self.upsamples[layer_idx]
-        if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 1:
-            for frame_idx in range(x.shape[2]):
-                self.run_up(
-                    layer_idx,
-                    [x[:, :, frame_idx:frame_idx + 1, :, :]],
-                    feat_cache,
-                    feat_idx.copy(),
-                    out_chunks,
-                )
-            del x
-            return
-
-        if feat_cache is not None:
-            x = layer(x, feat_cache, feat_idx)
-        else:
-            x = layer(x)
-
-        next_x_ref = [x]
-        del x
-        self.run_up(layer_idx + 1, next_x_ref, feat_cache, feat_idx, out_chunks)
-
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         ## conv1
         if feat_cache is not None:
             idx = feat_idx[0]
-            cache_x = x[:, :, -CACHE_T:, :, :]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
             x = self.conv1(x, feat_cache[idx])
             feat_cache[idx] = cache_x
             feat_idx[0] += 1
@@ -410,21 +409,42 @@ class Decoder3d(nn.Module):
 
         ## middle
         for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
             if feat_cache is not None:
                 x = layer(x, feat_cache, feat_idx)
             else:
                 x = layer(x)
 
-        out_chunks = []
-
-        self.run_up(0, [x], feat_cache, feat_idx, out_chunks)
-        return out_chunks
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
 
 
-def count_cache_layers(model):
+def count_conv3d(model):
     count = 0
     for m in model.modules():
-        if isinstance(m, CausalConv3d) or (isinstance(m, Resample) and m.mode == 'downsample3d'):
+        if isinstance(m, CausalConv3d):
             count += 1
     return count
 
@@ -462,12 +482,11 @@ class WanVAE(nn.Module):
         conv_idx = [0]
         ## cache
         t = x.shape[2]
-        t = 1 + ((t - 1) // 4) * 4
-        iter_ = 1 + (t - 1) // 2
+        iter_ = 1 + (t - 1) // 4
         feat_map = None
         if iter_ > 1:
-            feat_map = [None] * count_cache_layers(self.encoder)
-        ## 对encode输入的x，按时间拆分为1、2、2、2....(总帧数先按4N+1向下取整)
+            feat_map = [None] * count_conv3d(self.encoder)
+        ## 对encode输入的x，按时间拆分为1、4、4、4....
         for i in range(iter_):
             conv_idx = [0]
             if i == 0:
@@ -477,23 +496,20 @@ class WanVAE(nn.Module):
                     feat_idx=conv_idx)
             else:
                 out_ = self.encoder(
-                    x[:, :, 1 + 2 * (i - 1):1 + 2 * i, :, :],
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
                     feat_cache=feat_map,
-                    feat_idx=conv_idx,
-                    final=(i == (iter_ - 1)))
-                if out_ is None:
-                    continue
+                    feat_idx=conv_idx)
                 out = torch.cat([out, out_], 2)
-
         mu, log_var = self.conv1(out).chunk(2, dim=1)
         return mu
 
     def decode(self, z):
+        conv_idx = [0]
         # z: [b,c,t,h,w]
-        iter_ = 1 + z.shape[2] // 2
+        iter_ = z.shape[2]
         feat_map = None
         if iter_ > 1:
-            feat_map = [None] * count_cache_layers(self.decoder)
+            feat_map = [None] * count_conv3d(self.decoder)
         x = self.conv2(z)
         for i in range(iter_):
             conv_idx = [0]
@@ -504,8 +520,8 @@ class WanVAE(nn.Module):
                     feat_idx=conv_idx)
             else:
                 out_ = self.decoder(
-                    x[:, :, 1 + 2 * (i - 1):1 + 2 * i, :, :],
+                    x[:, :, i:i + 1, :, :],
                     feat_cache=feat_map,
                     feat_idx=conv_idx)
-                out += out_
-        return torch.cat(out, 2)
+                out = torch.cat([out, out_], 2)
+        return out

comfy/memory_management.py

@@ -39,10 +39,7 @@ def read_tensor_file_slice_into(tensor, destination):
     if (destination.device.type != "cpu"
             or file_obj is None
             or threading.get_ident() != info.thread_id
-            or destination.numel() * destination.element_size() < info.size
-            or tensor.numel() * tensor.element_size() != info.size
-            or tensor.storage_offset() != 0
-            or not tensor.is_contiguous()):
+            or destination.numel() * destination.element_size() < info.size):
         return False
 
     if info.size == 0:

comfy/model_base.py

@@ -285,12 +285,6 @@ class BaseModel(torch.nn.Module):
             return data
         return None
 
-    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
-        """Override in subclasses to handle model-specific cond slicing for context windows.
-        Return a sliced cond object, or None to fall through to default handling.
-        Use comfy.context_windows.slice_cond() for common cases."""
-        return None
-
     def extra_conds(self, **kwargs):
         out = {}
         concat_cond = self.concat_cond(**kwargs)
@@ -1381,12 +1375,6 @@ class WAN21_Vace(WAN21):
         out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
         return out
 
-    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
-        if cond_key == "vace_context":
-            import comfy.context_windows
-            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=3, retain_index_list=retain_index_list)
-        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
-
 class WAN21_Camera(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.CameraWanModel)
@@ -1439,12 +1427,6 @@ class WAN21_HuMo(WAN21):
 
         return out
 
-    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
-        if cond_key == "audio_embed":
-            import comfy.context_windows
-            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=1)
-        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
-
 class WAN22_Animate(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model_animate.AnimateWanModel)
@@ -1462,14 +1444,6 @@ class WAN22_Animate(WAN21):
             out['pose_latents'] = comfy.conds.CONDRegular(self.process_latent_in(pose_latents))
         return out
 
-    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
-        import comfy.context_windows
-        if cond_key == "face_pixel_values":
-            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_scale=4, temporal_offset=1)
-        if cond_key == "pose_latents":
-            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_offset=1)
-        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
-
 class WAN22_S2V(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
@@ -1506,12 +1480,6 @@ class WAN22_S2V(WAN21):
             out['reference_motion'] = reference_motion.shape
         return out
 
-    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
-        if cond_key == "audio_embed":
-            import comfy.context_windows
-            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=1)
-        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
-
 class WAN22(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)

comfy/model_management.py

@@ -541,7 +541,6 @@ class LoadedModel:
         if model.parent is not None:
             self._parent_model = weakref.ref(model.parent)
             self._patcher_finalizer = weakref.finalize(model, self._switch_parent)
-            self._patcher_finalizer.atexit = False
 
     def _switch_parent(self):
         model = self._parent_model()
@@ -588,7 +587,6 @@ class LoadedModel:
 
         self.real_model = weakref.ref(real_model)
         self.model_finalizer = weakref.finalize(real_model, cleanup_models)
-        self.model_finalizer.atexit = False
         return real_model
 
     def should_reload_model(self, force_patch_weights=False):

comfy/ops.py

@@ -776,71 +776,6 @@ from .quant_ops import (
 )
 
 
-class QuantLinearFunc(torch.autograd.Function):
-    """Custom autograd function for quantized linear: quantized forward, compute_dtype backward.
-    Handles any input rank by flattening to 2D for matmul and restoring shape after.
-    """
-
-    @staticmethod
-    def forward(ctx, input_float, weight, bias, layout_type, input_scale, compute_dtype):
-        input_shape = input_float.shape
-        inp = input_float.detach().flatten(0, -2)  # zero-cost view to 2D
-
-        # Quantize input (same as inference path)
-        if layout_type is not None:
-            q_input = QuantizedTensor.from_float(inp, layout_type, scale=input_scale)
-        else:
-            q_input = inp
-
-        w = weight.detach() if weight.requires_grad else weight
-        b = bias.detach() if bias is not None and bias.requires_grad else bias
-
-        output = torch.nn.functional.linear(q_input, w, b)
-
-        # Restore original input shape
-        if len(input_shape) > 2:
-            output = output.unflatten(0, input_shape[:-1])
-
-        ctx.save_for_backward(input_float, weight)
-        ctx.input_shape = input_shape
-        ctx.has_bias = bias is not None
-        ctx.compute_dtype = compute_dtype
-        ctx.weight_requires_grad = weight.requires_grad
-
-        return output
-
-    @staticmethod
-    @torch.autograd.function.once_differentiable
-    def backward(ctx, grad_output):
-        input_float, weight = ctx.saved_tensors
-        compute_dtype = ctx.compute_dtype
-        grad_2d = grad_output.flatten(0, -2).to(compute_dtype)
-
-        # Dequantize weight to compute dtype for backward matmul
-        if isinstance(weight, QuantizedTensor):
-            weight_f = weight.dequantize().to(compute_dtype)
-        else:
-            weight_f = weight.to(compute_dtype)
-
-        # grad_input = grad_output @ weight
-        grad_input = torch.mm(grad_2d, weight_f)
-        if len(ctx.input_shape) > 2:
-            grad_input = grad_input.unflatten(0, ctx.input_shape[:-1])
-
-        # grad_weight (only if weight requires grad, typically frozen for quantized training)
-        grad_weight = None
-        if ctx.weight_requires_grad:
-            input_f = input_float.flatten(0, -2).to(compute_dtype)
-            grad_weight = torch.mm(grad_2d.t(), input_f)
-
-        # grad_bias
-        grad_bias = None
-        if ctx.has_bias:
-            grad_bias = grad_2d.sum(dim=0)
-
-        return grad_input, grad_weight, grad_bias, None, None, None
-
-
 def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]):
     class MixedPrecisionOps(manual_cast):
         _quant_config = quant_config
@@ -1035,37 +970,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 #If cast needs to apply lora, it should be done in the compute dtype
                 compute_dtype = input.dtype
 
-                _use_quantized = (
-                    getattr(self, 'layout_type', None) is not None and
+                if (getattr(self, 'layout_type', None) is not None and
                     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
-                )
-
-                # Training path: quantized forward with compute_dtype backward via autograd function
-                if (input.requires_grad and _use_quantized):
-
-                    weight, bias, offload_stream = cast_bias_weight(
-                        self,
-                        input,
-                        offloadable=True,
-                        compute_dtype=compute_dtype,
-                        want_requant=True
-                    )
-
-                    scale = getattr(self, 'input_scale', None)
-                    if scale is not None:
-                        scale = comfy.model_management.cast_to_device(scale, input.device, None)
-
-                    output = QuantLinearFunc.apply(
-                        input, weight, bias, self.layout_type, scale, compute_dtype
-                    )
-
-                    uncast_bias_weight(self, weight, bias, offload_stream)
-                    return output
-
-                # Inference path (unchanged)
-                if _use_quantized:
+                    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
@@ -1113,10 +1021,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 for key, param in self._parameters.items():
                     if param is None:
                         continue
-                    p = fn(param)
-                    if p.is_inference():
-                        p = p.clone()
-                    self.register_parameter(key, torch.nn.Parameter(p, requires_grad=False))
+                    self.register_parameter(key, torch.nn.Parameter(fn(param), requires_grad=False))
                 for key, buf in self._buffers.items():
                     if buf is not None:
                         self._buffers[key] = fn(buf)

comfy/sample.py

@@ -64,10 +64,10 @@ def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative
     sampler = comfy.samplers.KSampler(model, steps=steps, device=model.load_device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options)
 
     samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
-    samples = samples.to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
+    samples = samples.to(comfy.model_management.intermediate_device())
     return samples
 
 def sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=None, callback=None, disable_pbar=False, seed=None):
     samples = comfy.samplers.sample(model, noise, positive, negative, cfg, model.load_device, sampler, sigmas, model_options=model.model_options, latent_image=latent_image, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
-    samples = samples.to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
+    samples = samples.to(comfy.model_management.intermediate_device())
     return samples

comfy/sd.py

@@ -455,7 +455,7 @@ class VAE:
         self.output_channels = 3
         self.pad_channel_value = None
         self.process_input = lambda image: image * 2.0 - 1.0
-        self.process_output = lambda image: image.add_(1.0).div_(2.0).clamp_(0.0, 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]
         self.disable_offload = False
         self.not_video = False
@@ -951,23 +951,12 @@ class VAE:
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
-            # Pre-allocate output for VAEs that support direct buffer writes
-            preallocated = False
-            if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
-                pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype())
-                preallocated = True
-
             for x in range(0, samples_in.shape[0], batch_number):
-                samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype)
-                if preallocated:
-                    self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options)
-                else:
-                    out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True)
-                    if pixel_samples is None:
-                        pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
-                    pixel_samples[x:x+batch_number].copy_(out)
-                    del out
-                self.process_output(pixel_samples[x:x+batch_number])
+                samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).to(dtype=self.vae_output_dtype()))
+                if pixel_samples is None:
+                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
+                pixel_samples[x:x+batch_number] = out
         except Exception as e:
             model_management.raise_non_oom(e)
             logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
@@ -978,7 +967,6 @@ class VAE:
             do_tile = True
 
         if do_tile:
-            comfy.model_management.soft_empty_cache()
             dims = samples_in.ndim - 2
             if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
@@ -1039,13 +1027,8 @@ class VAE:
             batch_number = max(1, batch_number)
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
-                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype)
-                if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
-                    out = self.first_stage_model.encode(pixels_in, device=self.device)
-                else:
-                    pixels_in = pixels_in.to(self.device)
-                    out = self.first_stage_model.encode(pixels_in)
-                out = out.to(self.output_device).to(dtype=self.vae_output_dtype())
+                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
+                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
                     samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out
@@ -1060,7 +1043,6 @@ class VAE:
             do_tile = True
 
         if do_tile:
-            comfy.model_management.soft_empty_cache()
             if self.latent_dim == 3:
                 tile = 256
                 overlap = tile // 4

comfy/sd1_clip.py

@@ -46,7 +46,7 @@ class ClipTokenWeightEncoder:
         out, pooled = o[:2]
 
         if pooled is not None:
-            first_pooled = pooled[0:1].to(device=model_management.intermediate_device())
+            first_pooled = pooled[0:1].to(model_management.intermediate_device())
         else:
             first_pooled = pooled
 
@@ -63,16 +63,16 @@ class ClipTokenWeightEncoder:
             output.append(z)
 
         if (len(output) == 0):
-            r = (out[-1:].to(device=model_management.intermediate_device()), first_pooled)
+            r = (out[-1:].to(model_management.intermediate_device()), first_pooled)
         else:
-            r = (torch.cat(output, dim=-2).to(device=model_management.intermediate_device()), first_pooled)
+            r = (torch.cat(output, dim=-2).to(model_management.intermediate_device()), first_pooled)
 
         if len(o) > 2:
             extra = {}
             for k in o[2]:
                 v = o[2][k]
                 if k == "attention_mask":
-                    v = v[:sections].flatten().unsqueeze(dim=0).to(device=model_management.intermediate_device())
+                    v = v[:sections].flatten().unsqueeze(dim=0).to(model_management.intermediate_device())
                 extra[k] = v
 
             r = r + (extra,)

comfy/utils.py

@@ -897,10 +897,6 @@ def set_attr(obj, attr, value):
     return prev
 
 def set_attr_param(obj, attr, value):
-    # Clone inference tensors (created under torch.inference_mode) since
-    # their version counter is frozen and nn.Parameter() cannot wrap them.
-    if (not torch.is_inference_mode_enabled()) and value.is_inference():
-        value = value.clone()
     return set_attr(obj, attr, torch.nn.Parameter(value, requires_grad=False))
 
 def set_attr_buffer(obj, attr, value):
@@ -1135,8 +1131,8 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
                 pbar.update(1)
             continue
 
-        out = output[b:b+1].zero_()
-        out_div = torch.zeros([s.shape[0], 1] + mult_list_upscale(s.shape[2:]), device=output_device)
+        out = torch.zeros([s.shape[0], out_channels] + mult_list_upscale(s.shape[2:]), device=output_device)
+        out_div = torch.zeros([s.shape[0], out_channels] + mult_list_upscale(s.shape[2:]), device=output_device)
 
         positions = [range(0, s.shape[d+2] - overlap[d], tile[d] - overlap[d]) if s.shape[d+2] > tile[d] else [0] for d in range(dims)]
 
@@ -1151,7 +1147,7 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
                 upscaled.append(round(get_pos(d, pos)))
 
             ps = function(s_in).to(output_device)
-            mask = torch.ones([1, 1] + list(ps.shape[2:]), device=output_device)
+            mask = torch.ones_like(ps)
 
             for d in range(2, dims + 2):
                 feather = round(get_scale(d - 2, overlap[d - 2]))
@@ -1174,7 +1170,7 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
             if pbar is not None:
                 pbar.update(1)
 
-        out.div_(out_div)
+        output[b:b+1] = out/out_div
     return output
 
 def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):

comfy_api/latest/__init__.py

@@ -15,7 +15,6 @@ from comfy_execution.progress import get_progress_state, PreviewImageTuple
 from PIL import Image
 from comfy.cli_args import args
 import numpy as np
-import os
 
 
 class ComfyAPI_latest(ComfyAPIBase):
@@ -26,7 +25,6 @@ class ComfyAPI_latest(ComfyAPIBase):
         super().__init__()
         self.node_replacement = self.NodeReplacement()
         self.execution = self.Execution()
-        self.environment = self.Environment()
         self.caching = self.Caching()
 
     class NodeReplacement(ProxiedSingleton):
@@ -87,27 +85,6 @@ class ComfyAPI_latest(ComfyAPIBase):
                 image=to_display,
             )
 
-    class Environment(ProxiedSingleton):
-        """
-        Query the current execution environment.
-
-        Managed deployments set the ``COMFY_EXECUTION_ENVIRONMENT`` env var
-        so custom nodes can adapt their behaviour at runtime.
-
-        Example::
-
-            from comfy_api.latest import api
-
-            env = api.environment.get()  # "local" | "cloud" | "remote"
-        """
-
-        _VALID = {"local", "cloud", "remote"}
-
-        async def get(self) -> str:
-            """Return the execution environment: ``"local"``, ``"cloud"``, or ``"remote"``."""
-            value = os.environ.get("COMFY_EXECUTION_ENVIRONMENT", "local").lower().strip()
-            return value if value in self._VALID else "local"
-
     class Caching(ProxiedSingleton):
         """
         External cache provider API for sharing cached node outputs

comfy_api_nodes/apis/gemini.py

@@ -67,7 +67,6 @@ class GeminiPart(BaseModel):
     inlineData: GeminiInlineData | None = Field(None)
     fileData: GeminiFileData | None = Field(None)
     text: str | None = Field(None)
-    thought: bool | None = Field(None)
 
 
 class GeminiTextPart(BaseModel):

comfy_api_nodes/nodes_bytedance.py

@@ -47,10 +47,6 @@ SEEDREAM_MODELS = {
 BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"
 BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"  # + /{task_id}
 
-DEPRECATED_MODELS = {"seedance-1-0-lite-t2v-250428", "seedance-1-0-lite-i2v-250428"}
-
-logger = logging.getLogger(__name__)
-
 
 def get_image_url_from_response(response: ImageTaskCreationResponse) -> str:
     if response.error:
@@ -139,7 +135,6 @@ class ByteDanceImageNode(IO.ComfyNode):
             price_badge=IO.PriceBadge(
                 expr="""{"type":"usd","usd":0.03}""",
             ),
-            is_deprecated=True,
         )
 
     @classmethod
@@ -947,7 +942,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
         ]
         return await process_video_task(
             cls,
-            payload=Image2VideoTaskCreationRequest(model=model, content=x, generate_audio=None),
+            payload=Image2VideoTaskCreationRequest(model=model, content=x),
             estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))),
         )
 
@@ -957,12 +952,6 @@ async def process_video_task(
     payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest,
     estimated_duration: int | None,
 ) -> IO.NodeOutput:
-    if payload.model in DEPRECATED_MODELS:
-        logger.warning(
-            "Model '%s' is deprecated and will be deactivated on May 13, 2026. "
-            "Please switch to a newer model. Recommended: seedance-1-0-pro-fast-251015.",
-            payload.model,
-        )
     initial_response = await sync_op(
         cls,
         ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),

comfy_api_nodes/nodes_gemini.py

@@ -63,7 +63,7 @@ GEMINI_IMAGE_2_PRICE_BADGE = IO.PriceBadge(
       $m := widgets.model;
       $r := widgets.resolution;
       $isFlash := $contains($m, "nano banana 2");
-      $flashPrices := {"1k": 0.0696, "2k": 0.1014, "4k": 0.154};
+      $flashPrices := {"1k": 0.0696, "2k": 0.0696, "4k": 0.123};
       $proPrices := {"1k": 0.134, "2k": 0.134, "4k": 0.24};
       $prices := $isFlash ? $flashPrices : $proPrices;
       {"type":"usd","usd": $lookup($prices, $r), "format":{"suffix":"/Image","approximate":true}}
@@ -188,12 +188,10 @@ def get_text_from_response(response: GeminiGenerateContentResponse) -> str:
     return "\n".join([part.text for part in parts])
 
 
-async def get_image_from_response(response: GeminiGenerateContentResponse, thought: bool = False) -> Input.Image:
+async def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image:
     image_tensors: list[Input.Image] = []
     parts = get_parts_by_type(response, "image/*")
     for part in parts:
-        if (part.thought is True) != thought:
-            continue
         if part.inlineData:
             image_data = base64.b64decode(part.inlineData.data)
             returned_image = bytesio_to_image_tensor(BytesIO(image_data))
@@ -933,11 +931,6 @@ class GeminiNanoBanana2(IO.ComfyNode):
             outputs=[
                 IO.Image.Output(),
                 IO.String.Output(),
-                IO.Image.Output(
-                    display_name="thought_image",
-                    tooltip="First image from the model's thinking process. "
-                    "Only available with thinking_level HIGH and IMAGE+TEXT modality.",
-                ),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -999,11 +992,7 @@ class GeminiNanoBanana2(IO.ComfyNode):
             response_model=GeminiGenerateContentResponse,
             price_extractor=calculate_tokens_price,
         )
-        return IO.NodeOutput(
-            await get_image_from_response(response),
-            get_text_from_response(response),
-            await get_image_from_response(response, thought=True),
-        )
+        return IO.NodeOutput(await get_image_from_response(response), get_text_from_response(response))
 
 
 class GeminiExtension(ComfyExtension):

comfy_extras/nodes_context_windows.py

@@ -27,8 +27,8 @@ class ContextWindowsManualNode(io.ComfyNode):
                 io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
                 io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."),
                 io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."),
-                io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."),
-                io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."),
+                #io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."),
+                #io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."),
             ],
             outputs=[
                 io.Model.Output(tooltip="The model with context windows applied during sampling."),

comfy_extras/nodes_train.py

@@ -15,7 +15,6 @@ import comfy.sampler_helpers
 import comfy.sd
 import comfy.utils
 import comfy.model_management
-from comfy.cli_args import args, PerformanceFeature
 import comfy_extras.nodes_custom_sampler
 import folder_paths
 import node_helpers
@@ -139,7 +138,6 @@ class TrainSampler(comfy.samplers.Sampler):
         training_dtype=torch.bfloat16,
         real_dataset=None,
         bucket_latents=None,
-        use_grad_scaler=False,
     ):
         self.loss_fn = loss_fn
         self.optimizer = optimizer
@@ -154,8 +152,6 @@ class TrainSampler(comfy.samplers.Sampler):
         self.bucket_latents: list[torch.Tensor] | None = (
             bucket_latents  # list of (Bi, C, Hi, Wi)
         )
-        # GradScaler for fp16 training
-        self.grad_scaler = torch.amp.GradScaler() if use_grad_scaler else None
         # Precompute bucket offsets and weights for sampling
         if bucket_latents is not None:
             self._init_bucket_data(bucket_latents)
@@ -208,13 +204,10 @@ class TrainSampler(comfy.samplers.Sampler):
                 batch_sigmas.requires_grad_(True),
                 **batch_extra_args,
             )
-            loss = self.loss_fn(x0_pred.float(), x0.float())
+            loss = self.loss_fn(x0_pred, x0)
         if bwd:
             bwd_loss = loss / self.grad_acc
-            if self.grad_scaler is not None:
-                self.grad_scaler.scale(bwd_loss).backward()
-            else:
-                bwd_loss.backward()
+            bwd_loss.backward()
         return loss
 
     def _generate_batch_sigmas(self, model_wrap, batch_size, device):
@@ -314,10 +307,7 @@ class TrainSampler(comfy.samplers.Sampler):
             )
             total_loss += loss
         total_loss = total_loss / self.grad_acc / len(indicies)
-        if self.grad_scaler is not None:
-            self.grad_scaler.scale(total_loss).backward()
-        else:
-            total_loss.backward()
+        total_loss.backward()
         if self.loss_callback:
             self.loss_callback(total_loss.item())
         pbar.set_postfix({"loss": f"{total_loss.item():.4f}"})
@@ -358,18 +348,12 @@ class TrainSampler(comfy.samplers.Sampler):
                 self._train_step_multires_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar)
 
             if (i + 1) % self.grad_acc == 0:
-                if self.grad_scaler is not None:
-                    self.grad_scaler.unscale_(self.optimizer)
                 for param_groups in self.optimizer.param_groups:
                     for param in param_groups["params"]:
                         if param.grad is None:
                             continue
                         param.grad.data = param.grad.data.to(param.data.dtype)
-                if self.grad_scaler is not None:
-                    self.grad_scaler.step(self.optimizer)
-                    self.grad_scaler.update()
-                else:
-                    self.optimizer.step()
+                self.optimizer.step()
                 self.optimizer.zero_grad()
             ui_pbar.update(1)
         torch.cuda.empty_cache()
@@ -1020,9 +1004,9 @@ class TrainLoraNode(io.ComfyNode):
                 ),
                 io.Combo.Input(
                     "training_dtype",
-                    options=["bf16", "fp32", "none"],
+                    options=["bf16", "fp32"],
                     default="bf16",
-                    tooltip="The dtype to use for training. 'none' preserves the model's native compute dtype instead of overriding it. For fp16 models, GradScaler is automatically enabled.",
+                    tooltip="The dtype to use for training.",
                 ),
                 io.Combo.Input(
                     "lora_dtype",
@@ -1051,7 +1035,7 @@ class TrainLoraNode(io.ComfyNode):
                 io.Boolean.Input(
                     "offloading",
                     default=False,
-                    tooltip="Offload model weights to CPU during training to save GPU memory.",
+                    tooltip="Offload the Model to RAM. Requires Bypass Mode.",
                 ),
                 io.Combo.Input(
                     "existing_lora",
@@ -1136,32 +1120,22 @@ class TrainLoraNode(io.ComfyNode):
 
         # Setup model and dtype
         mp = model.clone()
-        use_grad_scaler = False
-        if training_dtype != "none":
-            dtype = node_helpers.string_to_torch_dtype(training_dtype)
-            mp.set_model_compute_dtype(dtype)
-        else:
-            # Detect model's native dtype for autocast
-            model_dtype = mp.model.get_dtype()
-            if model_dtype == torch.float16:
-                dtype = torch.float16
-                use_grad_scaler = True
-                # Warn about fp16 accumulation instability during training
-                if PerformanceFeature.Fp16Accumulation in args.fast:
-                    logging.warning(
-                        "WARNING: FP16 model detected with fp16_accumulation enabled. "
-                        "This combination can be numerically unstable during training and may cause NaN values. "
-                        "Suggested fixes: 1) Set training_dtype to 'bf16', or 2) Disable fp16_accumulation (remove from --fast flags)."
-                    )
-            else:
-                # For fp8, bf16, or other dtypes, use bf16 autocast
-                dtype = torch.bfloat16
+        dtype = node_helpers.string_to_torch_dtype(training_dtype)
         lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype)
+        mp.set_model_compute_dtype(dtype)
+
+        if mp.is_dynamic():
+            if not bypass_mode:
+                logging.info("Training MP is Dynamic - forcing bypass mode. Start comfy with --highvram to force weight diff mode")
+                bypass_mode = True
+            offloading = True
+        elif offloading:
+            if not bypass_mode:
+                logging.info("Training Offload selected - forcing bypass mode. Set bypass = True to remove this message")
 
         # Prepare latents and compute counts
-        latents_dtype = dtype if dtype not in (None,) else torch.bfloat16
         latents, num_images, multi_res = _prepare_latents_and_count(
-            latents, latents_dtype, bucket_mode
+            latents, dtype, bucket_mode
         )
 
         # Validate and expand conditioning
@@ -1227,7 +1201,6 @@ class TrainLoraNode(io.ComfyNode):
                     seed=seed,
                     training_dtype=dtype,
                     bucket_latents=latents,
-                    use_grad_scaler=use_grad_scaler,
                 )
             else:
                 train_sampler = TrainSampler(
@@ -1240,7 +1213,6 @@ class TrainLoraNode(io.ComfyNode):
                     seed=seed,
                     training_dtype=dtype,
                     real_dataset=latents if multi_res else None,
-                    use_grad_scaler=use_grad_scaler,
                 )
 
             # Setup guider
@@ -1365,7 +1337,7 @@ class SaveLoRA(io.ComfyNode):
                 io.Int.Input(
                     "steps",
                     optional=True,
-                    tooltip="Optional: The number of steps the LoRA has been trained for, used to name the saved file.",
+                    tooltip="Optional: The number of steps to LoRA has been trained for, used to name the saved file.",
                 ),
             ],
             outputs=[],

main.py

@@ -206,8 +206,8 @@ import hook_breaker_ac10a0
 import comfy.memory_management
 import comfy.model_patcher
 
-if args.enable_dynamic_vram or (enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl()):
-    if (not args.enable_dynamic_vram) and (comfy.model_management.torch_version_numeric < (2, 8)):
+if enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl():
+    if comfy.model_management.torch_version_numeric < (2, 8):
         logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
     elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
         if args.verbose == 'DEBUG':

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b6
+comfyui_manager==4.1b5

nodes.py

@@ -952,7 +952,7 @@ class UNETLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "unet_name": (folder_paths.get_filename_list("diffusion_models"), ),
-                              "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e4m3fn_fast", "fp8_e5m2"], {"advanced": True})
+                              "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e4m3fn_fast", "fp8_e5m2"],)
                              }}
     RETURN_TYPES = ("MODEL",)
     FUNCTION = "load_unet"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.41.21
-comfyui-workflow-templates==0.9.26
+comfyui-frontend-package==1.41.20
+comfyui-workflow-templates==0.9.21
 comfyui-embedded-docs==0.4.3
 torch
 torchsde

server.py

@@ -35,8 +35,6 @@ from app.frontend_management import FrontendManager, parse_version
 from comfy_api.internal import _ComfyNodeInternal
 from app.assets.seeder import asset_seeder
 from app.assets.api.routes import register_assets_routes
-from app.assets.services.ingest import register_file_in_place
-from app.assets.services.asset_management import resolve_hash_to_path
 
 from app.user_manager import UserManager
 from app.model_manager import ModelFileManager
@@ -421,24 +419,7 @@ class PromptServer():
                         with open(filepath, "wb") as f:
                             f.write(image.file.read())
 
-                resp = {"name" : filename, "subfolder": subfolder, "type": image_upload_type}
-
-                if args.enable_assets:
-                    try:
-                        tag = image_upload_type if image_upload_type in ("input", "output") else "input"
-                        result = register_file_in_place(abs_path=filepath, name=filename, tags=[tag])
-                        resp["asset"] = {
-                            "id": result.ref.id,
-                            "name": result.ref.name,
-                            "asset_hash": result.asset.hash,
-                            "size": result.asset.size_bytes,
-                            "mime_type": result.asset.mime_type,
-                            "tags": result.tags,
-                        }
-                    except Exception:
-                        logging.warning("Failed to register uploaded image as asset", exc_info=True)
-
-                return web.json_response(resp)
+                return web.json_response({"name" : filename, "subfolder": subfolder, "type": image_upload_type})
             else:
                 return web.Response(status=400)
 
@@ -498,43 +479,30 @@ class PromptServer():
         async def view_image(request):
             if "filename" in request.rel_url.query:
                 filename = request.rel_url.query["filename"]
+                filename, output_dir = folder_paths.annotated_filepath(filename)
 
-                # The frontend's LoadImage combo widget uses asset_hash values
-                # (e.g. "blake3:...") as widget values. When litegraph renders the
-                # node preview, it constructs /view?filename=<asset_hash>, so this
-                # endpoint must resolve blake3 hashes to their on-disk file paths.
-                if filename.startswith("blake3:"):
-                    owner_id = self.user_manager.get_request_user_id(request)
-                    result = resolve_hash_to_path(filename, owner_id=owner_id)
-                    if result is None:
-                        return web.Response(status=404)
-                    file, filename, resolved_content_type = result.abs_path, result.download_name, result.content_type
-                else:
-                    resolved_content_type = None
-                    filename, output_dir = folder_paths.annotated_filepath(filename)
+                if not filename:
+                    return web.Response(status=400)
 
-                    if not filename:
-                        return web.Response(status=400)
+                # validation for security: prevent accessing arbitrary path
+                if filename[0] == '/' or '..' in filename:
+                    return web.Response(status=400)
 
-                    # validation for security: prevent accessing arbitrary path
-                    if filename[0] == '/' or '..' in filename:
-                        return web.Response(status=400)
+                if output_dir is None:
+                    type = request.rel_url.query.get("type", "output")
+                    output_dir = folder_paths.get_directory_by_type(type)
 
-                    if output_dir is None:
-                        type = request.rel_url.query.get("type", "output")
-                        output_dir = folder_paths.get_directory_by_type(type)
+                if output_dir is None:
+                    return web.Response(status=400)
 
-                    if output_dir is None:
-                        return web.Response(status=400)
+                if "subfolder" in request.rel_url.query:
+                    full_output_dir = os.path.join(output_dir, request.rel_url.query["subfolder"])
+                    if os.path.commonpath((os.path.abspath(full_output_dir), output_dir)) != output_dir:
+                        return web.Response(status=403)
+                    output_dir = full_output_dir
 
-                    if "subfolder" in request.rel_url.query:
-                        full_output_dir = os.path.join(output_dir, request.rel_url.query["subfolder"])
-                        if os.path.commonpath((os.path.abspath(full_output_dir), output_dir)) != output_dir:
-                            return web.Response(status=403)
-                        output_dir = full_output_dir
-
-                    filename = os.path.basename(filename)
-                    file = os.path.join(output_dir, filename)
+                filename = os.path.basename(filename)
+                file = os.path.join(output_dir, filename)
 
                 if os.path.isfile(file):
                     if 'preview' in request.rel_url.query:
@@ -594,13 +562,8 @@ class PromptServer():
                             return web.Response(body=alpha_buffer.read(), content_type='image/png',
                                                 headers={"Content-Disposition": f"filename=\"{filename}\""})
                     else:
-                        # Use the content type from asset resolution if available,
-                        # otherwise guess from the filename.
-                        content_type = (
-                            resolved_content_type
-                            or mimetypes.guess_type(filename)[0]
-                            or 'application/octet-stream'
-                        )
+                        # Get content type from mimetype, defaulting to 'application/octet-stream'
+                        content_type = mimetypes.guess_type(filename)[0] or 'application/octet-stream'
 
                         # For security, force certain mimetypes to download instead of display
                         if content_type in {'text/html', 'text/html-sandboxed', 'application/xhtml+xml', 'text/javascript', 'text/css'}:

tests-unit/app_test/test_migrations.py

@@ -1,57 +0,0 @@
-"""Test that Alembic migrations run cleanly on a file-backed SQLite DB.
-
-This catches problems like unnamed FK constraints that prevent batch-mode
-drop_constraint from working on real SQLite files (see MB-2).
-
-Migrations 0001 and 0002 are already shipped, so we only exercise
-upgrade/downgrade for 0003+.
-"""
-
-import os
-
-import pytest
-from alembic import command
-from alembic.config import Config
-
-
-# Oldest shipped revision — we upgrade to here as a baseline and never
-# downgrade past it.
-_BASELINE = "0002_merge_to_asset_references"
-
-
-def _make_config(db_path: str) -> Config:
-    root = os.path.join(os.path.dirname(__file__), "../..")
-    config_path = os.path.abspath(os.path.join(root, "alembic.ini"))
-    scripts_path = os.path.abspath(os.path.join(root, "alembic_db"))
-
-    cfg = Config(config_path)
-    cfg.set_main_option("script_location", scripts_path)
-    cfg.set_main_option("sqlalchemy.url", f"sqlite:///{db_path}")
-    return cfg
-
-
-@pytest.fixture
-def migration_db(tmp_path):
-    """Yield an alembic Config pre-upgraded to the baseline revision."""
-    db_path = str(tmp_path / "test_migration.db")
-    cfg = _make_config(db_path)
-    command.upgrade(cfg, _BASELINE)
-    yield cfg
-
-
-def test_upgrade_to_head(migration_db):
-    """Upgrade from baseline to head must succeed on a file-backed DB."""
-    command.upgrade(migration_db, "head")
-
-
-def test_downgrade_to_baseline(migration_db):
-    """Upgrade to head then downgrade back to baseline."""
-    command.upgrade(migration_db, "head")
-    command.downgrade(migration_db, _BASELINE)
-
-
-def test_upgrade_downgrade_cycle(migration_db):
-    """Full cycle: upgrade → downgrade → upgrade again."""
-    command.upgrade(migration_db, "head")
-    command.downgrade(migration_db, _BASELINE)
-    command.upgrade(migration_db, "head")

tests-unit/assets_test/queries/test_asset.py

@@ -10,7 +10,6 @@ from app.assets.database.queries import (
     get_asset_by_hash,
     upsert_asset,
     bulk_insert_assets,
-    update_asset_hash_and_mime,
 )
 
 
@@ -143,45 +142,3 @@ class TestBulkInsertAssets:
         session.commit()
 
         assert session.query(Asset).count() == 200
-
-
-class TestMimeTypeImmutability:
-    """mime_type on Asset is write-once: set on first ingest, never overwritten."""
-
-    @pytest.mark.parametrize(
-        "initial_mime,second_mime,expected_mime",
-        [
-            ("image/png", "image/jpeg", "image/png"),
-            (None, "image/png", "image/png"),
-        ],
-        ids=["preserves_existing", "fills_null"],
-    )
-    def test_upsert_mime_immutability(self, session: Session, initial_mime, second_mime, expected_mime):
-        h = f"blake3:upsert_{initial_mime}_{second_mime}"
-        upsert_asset(session, asset_hash=h, size_bytes=100, mime_type=initial_mime)
-        session.commit()
-
-        asset, created, _ = upsert_asset(session, asset_hash=h, size_bytes=100, mime_type=second_mime)
-        assert created is False
-        assert asset.mime_type == expected_mime
-
-    @pytest.mark.parametrize(
-        "initial_mime,update_mime,update_hash,expected_mime,expected_hash",
-        [
-            (None, "image/png", None, "image/png", "blake3:upd0"),
-            ("image/png", "image/jpeg", None, "image/png", "blake3:upd1"),
-            ("image/png", "image/jpeg", "blake3:upd2_new", "image/png", "blake3:upd2_new"),
-        ],
-        ids=["fills_null", "preserves_existing", "hash_updates_mime_locked"],
-    )
-    def test_update_asset_hash_and_mime_immutability(
-        self, session: Session, initial_mime, update_mime, update_hash, expected_mime, expected_hash,
-    ):
-        h = expected_hash.removesuffix("_new")
-        asset = Asset(hash=h, size_bytes=100, mime_type=initial_mime)
-        session.add(asset)
-        session.flush()
-
-        update_asset_hash_and_mime(session, asset_id=asset.id, mime_type=update_mime, asset_hash=update_hash)
-        assert asset.mime_type == expected_mime
-        assert asset.hash == expected_hash

tests-unit/assets_test/queries/test_asset_info.py

@@ -242,24 +242,22 @@ class TestSetReferencePreview:
         asset = _make_asset(session, "hash1")
         preview_asset = _make_asset(session, "preview_hash")
         ref = _make_reference(session, asset)
-        preview_ref = _make_reference(session, preview_asset, name="preview.png")
         session.commit()
 
-        set_reference_preview(session, reference_id=ref.id, preview_reference_id=preview_ref.id)
+        set_reference_preview(session, reference_id=ref.id, preview_asset_id=preview_asset.id)
         session.commit()
 
         session.refresh(ref)
-        assert ref.preview_id == preview_ref.id
+        assert ref.preview_id == preview_asset.id
 
     def test_clears_preview(self, session: Session):
         asset = _make_asset(session, "hash1")
         preview_asset = _make_asset(session, "preview_hash")
         ref = _make_reference(session, asset)
-        preview_ref = _make_reference(session, preview_asset, name="preview.png")
-        ref.preview_id = preview_ref.id
+        ref.preview_id = preview_asset.id
         session.commit()
 
-        set_reference_preview(session, reference_id=ref.id, preview_reference_id=None)
+        set_reference_preview(session, reference_id=ref.id, preview_asset_id=None)
         session.commit()
 
         session.refresh(ref)
@@ -267,15 +265,15 @@ class TestSetReferencePreview:
 
     def test_raises_for_nonexistent_reference(self, session: Session):
         with pytest.raises(ValueError, match="not found"):
-            set_reference_preview(session, reference_id="nonexistent", preview_reference_id=None)
+            set_reference_preview(session, reference_id="nonexistent", preview_asset_id=None)
 
     def test_raises_for_nonexistent_preview(self, session: Session):
         asset = _make_asset(session, "hash1")
         ref = _make_reference(session, asset)
         session.commit()
 
-        with pytest.raises(ValueError, match="Preview AssetReference"):
-            set_reference_preview(session, reference_id=ref.id, preview_reference_id="nonexistent")
+        with pytest.raises(ValueError, match="Preview Asset"):
+            set_reference_preview(session, reference_id=ref.id, preview_asset_id="nonexistent")
 
 
 class TestInsertReference:
@@ -353,14 +351,13 @@ class TestUpdateReferenceTimestamps:
         asset = _make_asset(session, "hash1")
         preview_asset = _make_asset(session, "preview_hash")
         ref = _make_reference(session, asset)
-        preview_ref = _make_reference(session, preview_asset, name="preview.png")
         session.commit()
 
-        update_reference_timestamps(session, ref, preview_id=preview_ref.id)
+        update_reference_timestamps(session, ref, preview_id=preview_asset.id)
         session.commit()
 
         session.refresh(ref)
-        assert ref.preview_id == preview_ref.id
+        assert ref.preview_id == preview_asset.id
 
 
 class TestSetReferenceMetadata:

tests-unit/assets_test/queries/test_metadata.py

@@ -20,7 +20,6 @@ def _make_reference(
     asset: Asset,
     name: str,
     metadata: dict | None = None,
-    system_metadata: dict | None = None,
 ) -> AssetReference:
     now = get_utc_now()
     ref = AssetReference(
@@ -28,7 +27,6 @@ def _make_reference(
         name=name,
         asset_id=asset.id,
         user_metadata=metadata,
-        system_metadata=system_metadata,
         created_at=now,
         updated_at=now,
         last_access_time=now,
@@ -36,10 +34,8 @@ def _make_reference(
     session.add(ref)
     session.flush()
 
-    # Build merged projection: {**system_metadata, **user_metadata}
-    merged = {**(system_metadata or {}), **(metadata or {})}
-    if merged:
-        for key, val in merged.items():
+    if metadata:
+        for key, val in metadata.items():
             for row in convert_metadata_to_rows(key, val):
                 meta_row = AssetReferenceMeta(
                     asset_reference_id=ref.id,
@@ -186,46 +182,3 @@ class TestMetadataFilterEmptyDict:
 
         refs, _, total = list_references_page(session, metadata_filter={})
         assert total == 2
-
-
-class TestSystemMetadataProjection:
-    """Tests for system_metadata merging into the filter projection."""
-
-    def test_system_metadata_keys_are_filterable(self, session: Session):
-        """system_metadata keys should appear in the merged projection."""
-        asset = _make_asset(session, "hash1")
-        _make_reference(
-            session, asset, "with_sys",
-            system_metadata={"source": "scanner"},
-        )
-        _make_reference(session, asset, "without_sys")
-        session.commit()
-
-        refs, _, total = list_references_page(
-            session, metadata_filter={"source": "scanner"}
-        )
-        assert total == 1
-        assert refs[0].name == "with_sys"
-
-    def test_user_metadata_overrides_system_metadata(self, session: Session):
-        """user_metadata should win when both have the same key."""
-        asset = _make_asset(session, "hash1")
-        _make_reference(
-            session, asset, "overridden",
-            metadata={"origin": "user_upload"},
-            system_metadata={"origin": "auto_scan"},
-        )
-        session.commit()
-
-        # Should match the user value, not the system value
-        refs, _, total = list_references_page(
-            session, metadata_filter={"origin": "user_upload"}
-        )
-        assert total == 1
-        assert refs[0].name == "overridden"
-
-        # Should NOT match the system value (it was overridden)
-        refs, _, total = list_references_page(
-            session, metadata_filter={"origin": "auto_scan"}
-        )
-        assert total == 0

tests-unit/assets_test/services/test_asset_management.py

@@ -11,7 +11,6 @@ from app.assets.services import (
     delete_asset_reference,
     set_asset_preview,
 )
-from app.assets.services.asset_management import resolve_hash_to_path
 
 
 def _make_asset(session: Session, hash_val: str = "blake3:test", size: int = 1024) -> Asset:
@@ -220,33 +219,31 @@ class TestSetAssetPreview:
         asset = _make_asset(session, hash_val="blake3:main")
         preview_asset = _make_asset(session, hash_val="blake3:preview")
         ref = _make_reference(session, asset)
-        preview_ref = _make_reference(session, preview_asset, name="preview.png")
         ref_id = ref.id
-        preview_ref_id = preview_ref.id
+        preview_id = preview_asset.id
         session.commit()
 
         set_asset_preview(
             reference_id=ref_id,
-            preview_reference_id=preview_ref_id,
+            preview_asset_id=preview_id,
         )
 
         # Verify by re-fetching from DB
         session.expire_all()
         updated_ref = session.get(AssetReference, ref_id)
-        assert updated_ref.preview_id == preview_ref_id
+        assert updated_ref.preview_id == preview_id
 
     def test_clears_preview(self, mock_create_session, session: Session):
         asset = _make_asset(session)
         preview_asset = _make_asset(session, hash_val="blake3:preview")
         ref = _make_reference(session, asset)
-        preview_ref = _make_reference(session, preview_asset, name="preview.png")
-        ref.preview_id = preview_ref.id
+        ref.preview_id = preview_asset.id
         ref_id = ref.id
         session.commit()
 
         set_asset_preview(
             reference_id=ref_id,
-            preview_reference_id=None,
+            preview_asset_id=None,
         )
 
         # Verify by re-fetching from DB
@@ -266,45 +263,6 @@ class TestSetAssetPreview:
         with pytest.raises(PermissionError, match="not owner"):
             set_asset_preview(
                 reference_id=ref.id,
-                preview_reference_id=None,
+                preview_asset_id=None,
                 owner_id="user2",
             )
-
-
-class TestResolveHashToPath:
-    def test_returns_none_for_unknown_hash(self, mock_create_session):
-        result = resolve_hash_to_path("blake3:" + "a" * 64)
-        assert result is None
-
-    @pytest.mark.parametrize(
-        "ref_owner, query_owner, expect_found",
-        [
-            ("user1", "user1", True),
-            ("user1", "user2", False),
-            ("", "anyone", True),
-            ("", "", True),
-        ],
-        ids=[
-            "owner_sees_own_ref",
-            "other_owner_blocked",
-            "ownerless_visible_to_anyone",
-            "ownerless_visible_to_empty",
-        ],
-    )
-    def test_owner_visibility(
-        self, ref_owner, query_owner, expect_found,
-        mock_create_session, session: Session, temp_dir,
-    ):
-        f = temp_dir / "file.bin"
-        f.write_bytes(b"data")
-        asset = _make_asset(session, hash_val="blake3:" + "b" * 64)
-        ref = _make_reference(session, asset, name="file.bin", owner_id=ref_owner)
-        ref.file_path = str(f)
-        session.commit()
-
-        result = resolve_hash_to_path(asset.hash, owner_id=query_owner)
-        if expect_found:
-            assert result is not None
-            assert result.abs_path == str(f)
-        else:
-            assert result is None

tests-unit/assets_test/services/test_ingest.py

@@ -113,19 +113,11 @@ class TestIngestFileFromPath:
         file_path = temp_dir / "with_preview.bin"
         file_path.write_bytes(b"data")
 
-        # Create a preview asset and reference
+        # Create a preview asset first
         preview_asset = Asset(hash="blake3:preview", size_bytes=100)
         session.add(preview_asset)
-        session.flush()
-        from app.assets.helpers import get_utc_now
-        now = get_utc_now()
-        preview_ref = AssetReference(
-            asset_id=preview_asset.id, name="preview.png", owner_id="",
-            created_at=now, updated_at=now, last_access_time=now,
-        )
-        session.add(preview_ref)
         session.commit()
-        preview_id = preview_ref.id
+        preview_id = preview_asset.id
 
         result = _ingest_file_from_path(
             abs_path=str(file_path),

tests-unit/assets_test/services/test_tag_histogram.py

@@ -1,123 +0,0 @@
-"""Tests for list_tag_histogram service function."""
-from sqlalchemy.orm import Session
-
-from app.assets.database.models import Asset, AssetReference
-from app.assets.database.queries import ensure_tags_exist, add_tags_to_reference
-from app.assets.helpers import get_utc_now
-from app.assets.services.tagging import list_tag_histogram
-
-
-def _make_asset(session: Session, hash_val: str = "blake3:test") -> Asset:
-    asset = Asset(hash=hash_val, size_bytes=1024)
-    session.add(asset)
-    session.flush()
-    return asset
-
-
-def _make_reference(
-    session: Session,
-    asset: Asset,
-    name: str = "test",
-    owner_id: str = "",
-) -> AssetReference:
-    now = get_utc_now()
-    ref = AssetReference(
-        owner_id=owner_id,
-        name=name,
-        asset_id=asset.id,
-        created_at=now,
-        updated_at=now,
-        last_access_time=now,
-    )
-    session.add(ref)
-    session.flush()
-    return ref
-
-
-class TestListTagHistogram:
-    def test_returns_counts_for_all_tags(self, mock_create_session, session: Session):
-        ensure_tags_exist(session, ["alpha", "beta"])
-        a1 = _make_asset(session, "blake3:aaa")
-        r1 = _make_reference(session, a1, name="r1")
-        add_tags_to_reference(session, reference_id=r1.id, tags=["alpha", "beta"])
-
-        a2 = _make_asset(session, "blake3:bbb")
-        r2 = _make_reference(session, a2, name="r2")
-        add_tags_to_reference(session, reference_id=r2.id, tags=["alpha"])
-        session.commit()
-
-        result = list_tag_histogram()
-
-        assert result["alpha"] == 2
-        assert result["beta"] == 1
-
-    def test_empty_when_no_assets(self, mock_create_session, session: Session):
-        ensure_tags_exist(session, ["unused"])
-        session.commit()
-
-        result = list_tag_histogram()
-
-        assert result == {}
-
-    def test_include_tags_filter(self, mock_create_session, session: Session):
-        ensure_tags_exist(session, ["models", "loras", "input"])
-        a1 = _make_asset(session, "blake3:aaa")
-        r1 = _make_reference(session, a1, name="r1")
-        add_tags_to_reference(session, reference_id=r1.id, tags=["models", "loras"])
-
-        a2 = _make_asset(session, "blake3:bbb")
-        r2 = _make_reference(session, a2, name="r2")
-        add_tags_to_reference(session, reference_id=r2.id, tags=["input"])
-        session.commit()
-
-        result = list_tag_histogram(include_tags=["models"])
-
-        # Only r1 has "models", so only its tags appear
-        assert "models" in result
-        assert "loras" in result
-        assert "input" not in result
-
-    def test_exclude_tags_filter(self, mock_create_session, session: Session):
-        ensure_tags_exist(session, ["models", "loras", "input"])
-        a1 = _make_asset(session, "blake3:aaa")
-        r1 = _make_reference(session, a1, name="r1")
-        add_tags_to_reference(session, reference_id=r1.id, tags=["models", "loras"])
-
-        a2 = _make_asset(session, "blake3:bbb")
-        r2 = _make_reference(session, a2, name="r2")
-        add_tags_to_reference(session, reference_id=r2.id, tags=["input"])
-        session.commit()
-
-        result = list_tag_histogram(exclude_tags=["models"])
-
-        # r1 excluded, only r2's tags remain
-        assert "input" in result
-        assert "loras" not in result
-
-    def test_name_contains_filter(self, mock_create_session, session: Session):
-        ensure_tags_exist(session, ["alpha", "beta"])
-        a1 = _make_asset(session, "blake3:aaa")
-        r1 = _make_reference(session, a1, name="my_model.safetensors")
-        add_tags_to_reference(session, reference_id=r1.id, tags=["alpha"])
-
-        a2 = _make_asset(session, "blake3:bbb")
-        r2 = _make_reference(session, a2, name="picture.png")
-        add_tags_to_reference(session, reference_id=r2.id, tags=["beta"])
-        session.commit()
-
-        result = list_tag_histogram(name_contains="model")
-
-        assert "alpha" in result
-        assert "beta" not in result
-
-    def test_limit_caps_results(self, mock_create_session, session: Session):
-        tags = [f"tag{i}" for i in range(10)]
-        ensure_tags_exist(session, tags)
-        a = _make_asset(session, "blake3:aaa")
-        r = _make_reference(session, a, name="r1")
-        add_tags_to_reference(session, reference_id=r.id, tags=tags)
-        session.commit()
-
-        result = list_tag_histogram(limit=3)
-
-        assert len(result) == 3

tests-unit/assets_test/test_uploads.py

@@ -243,15 +243,6 @@ def test_upload_tags_traversal_guard(http: requests.Session, api_base: str):
     assert body["error"]["code"] in ("BAD_REQUEST", "INVALID_BODY")
 
 
-def test_upload_empty_tags_rejected(http: requests.Session, api_base: str):
-    files = {"file": ("notags.bin", b"A" * 64, "application/octet-stream")}
-    form = {"tags": json.dumps([]), "name": "notags.bin", "user_metadata": json.dumps({})}
-    r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120)
-    body = r.json()
-    assert r.status_code == 400
-    assert body["error"]["code"] == "INVALID_BODY"
-
-
 @pytest.mark.parametrize("root", ["input", "output"])
 def test_duplicate_upload_same_display_name_does_not_clobber(
     root: str,