Add TypedDict types to scanner and bulk_ingest

Amp-Thread-ID: https://ampcode.com/threads/T-019c2af9-4d41-73e9-b38d-78d06bc28a3f
Co-authored-by: Amp <amp@ampcode.com>

.coderabbit.yaml

@@ -1,127 +0,0 @@
-# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
-language: "en-US"
-early_access: false
-tone_instructions: "Only comment on issues introduced by this PR's changes. Do not flag pre-existing problems in moved, re-indented, or reformatted code."
-
-reviews:
-  profile: "chill"
-  request_changes_workflow: false
-  high_level_summary: false
-  poem: false
-  review_status: false
-  review_details: false
-  commit_status: true
-  collapse_walkthrough: true
-  changed_files_summary: false
-  sequence_diagrams: false
-  estimate_code_review_effort: false
-  assess_linked_issues: false
-  related_issues: false
-  related_prs: false
-  suggested_labels: false
-  auto_apply_labels: false
-  suggested_reviewers: false
-  auto_assign_reviewers: false
-  in_progress_fortune: false
-  enable_prompt_for_ai_agents: true
-
-  path_filters:
-    - "!comfy_api_nodes/apis/**"
-    - "!**/generated/*.pyi"
-    - "!.ci/**"
-    - "!script_examples/**"
-    - "!**/__pycache__/**"
-    - "!**/*.ipynb"
-    - "!**/*.png"
-    - "!**/*.bat"
-
-  path_instructions:
-    - path: "**"
-      instructions: |
-        IMPORTANT: Only comment on issues directly introduced by this PR's code changes.
-        Do NOT flag pre-existing issues in code that was merely moved, re-indented,
-        de-indented, or reformatted without logic changes. If code appears in the diff
-        only due to whitespace or structural reformatting (e.g., removing a `with:` block),
-        treat it as unchanged. Contributors should not feel obligated to address
-        pre-existing issues outside the scope of their contribution.
-    - path: "comfy/**"
-      instructions: |
-        Core ML/diffusion engine. Focus on:
-        - Backward compatibility (breaking changes affect all custom nodes)
-        - Memory management and GPU resource handling
-        - Performance implications in hot paths
-        - Thread safety for concurrent execution
-    - path: "comfy_api_nodes/**"
-      instructions: |
-        Third-party API integration nodes. Focus on:
-        - No hardcoded API keys or secrets
-        - Proper error handling for API failures (timeouts, rate limits, auth errors)
-        - Correct Pydantic model usage
-        - Security of user data passed to external APIs
-    - path: "comfy_extras/**"
-      instructions: |
-        Community-contributed extra nodes. Focus on:
-        - Consistency with node patterns (INPUT_TYPES, RETURN_TYPES, FUNCTION, CATEGORY)
-        - No breaking changes to existing node interfaces
-    - path: "comfy_execution/**"
-      instructions: |
-        Execution engine (graph execution, caching, jobs). Focus on:
-        - Caching correctness
-        - Concurrent execution safety
-        - Graph validation edge cases
-    - path: "nodes.py"
-      instructions: |
-        Core node definitions (2500+ lines). Focus on:
-        - Backward compatibility of NODE_CLASS_MAPPINGS
-        - Consistency of INPUT_TYPES return format
-    - path: "alembic_db/**"
-      instructions: |
-        Database migrations. Focus on:
-        - Migration safety and rollback support
-        - Data preservation during schema changes
-
-  auto_review:
-    enabled: true
-    auto_incremental_review: true
-    drafts: false
-    ignore_title_keywords:
-      - "WIP"
-      - "DO NOT REVIEW"
-      - "DO NOT MERGE"
-
-  finishing_touches:
-    docstrings:
-      enabled: false
-    unit_tests:
-      enabled: false
-
-  tools:
-    ruff:
-      enabled: false
-    pylint:
-      enabled: false
-    flake8:
-      enabled: false
-    gitleaks:
-      enabled: true
-    shellcheck:
-      enabled: false
-    markdownlint:
-      enabled: false
-    yamllint:
-      enabled: false
-    languagetool:
-      enabled: false
-    github-checks:
-      enabled: true
-      timeout_ms: 90000
-    ast-grep:
-      essential_rules: true
-
-chat:
-  auto_reply: true
-
-knowledge_base:
-  opt_out: false
-  learnings:
-    scope: "auto"

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -16,7 +16,7 @@ body:
 
         ## Very Important
 
-        Please make sure that you post ALL your ComfyUI logs in the bug report **even if there is no crash**. Just paste everything. The startup log (everything before "To see the GUI go to: ...") contains critical information to developers trying to help. For a performance issue or crash, paste everything from "got prompt" to the end, including the crash. More is better - always. A bug report without logs will likely be ignored.
+        Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored.
   - type: checkboxes
     id: custom-nodes-test
     attributes:

.github/workflows/release-webhook.yml

@@ -7,8 +7,6 @@ on:
 jobs:
   send-webhook:
     runs-on: ubuntu-latest
-    env:
-      DESKTOP_REPO_DISPATCH_TOKEN: ${{ secrets.DESKTOP_REPO_DISPATCH_TOKEN }}
     steps:
       - name: Send release webhook
         env:
@@ -108,37 +106,3 @@ jobs:
             --fail --silent --show-error
           
           echo "✅ Release webhook sent successfully"
-
-      - name: Send repository dispatch to desktop
-        env:
-          DISPATCH_TOKEN: ${{ env.DESKTOP_REPO_DISPATCH_TOKEN }}
-          RELEASE_TAG: ${{ github.event.release.tag_name }}
-          RELEASE_URL: ${{ github.event.release.html_url }}
-        run: |
-          set -euo pipefail
-
-          if [ -z "${DISPATCH_TOKEN:-}" ]; then
-            echo "::error::DESKTOP_REPO_DISPATCH_TOKEN is required but not set."
-            exit 1
-          fi
-
-          PAYLOAD="$(jq -n \
-            --arg release_tag "$RELEASE_TAG" \
-            --arg release_url "$RELEASE_URL" \
-            '{
-              event_type: "comfyui_release_published",
-              client_payload: {
-                release_tag: $release_tag,
-                release_url: $release_url
-              }
-            }')"
-
-          curl -fsSL \
-            -X POST \
-            -H "Accept: application/vnd.github+json" \
-            -H "Content-Type: application/json" \
-            -H "Authorization: Bearer ${DISPATCH_TOKEN}" \
-            https://api.github.com/repos/Comfy-Org/desktop/dispatches \
-            -d "$PAYLOAD"
-
-          echo "✅ Dispatched ComfyUI release ${RELEASE_TAG} to Comfy-Org/desktop"

.gitignore

@@ -11,7 +11,7 @@ extra_model_paths.yaml
 /.vs
 .vscode/
 .idea/
-venv*/
+venv/
 .venv/
 /web/extensions/*
 !/web/extensions/logging.js.example

README.md

@@ -189,6 +189,8 @@ The portable above currently comes with python 3.13 and pytorch cuda 13.0. Updat
 
 [Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
 
+[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z).
+
 [Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 
 #### How do I share models between another UI and ComfyUI?
@@ -225,11 +227,11 @@ Put your VAE in: models/vae
 
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4```
 
-This is the command to install the nightly with ROCm 7.2 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.1 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1```
 
 
 ### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only.

app/assets/api/routes.py

@@ -1,19 +1,36 @@
 import logging
-import uuid
-import urllib.parse
 import os
-import contextlib
-from aiohttp import web
+import urllib.parse
+import uuid
+from typing import Any
 
+from aiohttp import web
 from pydantic import ValidationError
 
-import app.assets.manager as manager
-from app import user_manager
-from app.assets.api import schemas_in
-from app.assets.helpers import get_query_dict
-from app.assets.scanner import seed_assets
-
 import folder_paths
+from app import user_manager
+from app.assets.api import schemas_in, schemas_out
+from app.assets.api.schemas_in import (
+    AssetValidationError,
+    UploadError,
+)
+from app.assets.api.upload import parse_multipart_upload
+from app.assets.scanner import seed_assets as scanner_seed_assets
+from app.assets.services import (
+    DependencyMissingError,
+    HashMismatchError,
+    apply_tags,
+    asset_exists,
+    create_from_hash,
+    delete_asset_reference,
+    get_asset_detail,
+    list_assets_page,
+    list_tags,
+    remove_tags,
+    resolve_asset_for_download,
+    update_asset_metadata,
+    upload_from_temp_path,
+)
 
 ROUTES = web.RouteTableDef()
 USER_MANAGER: user_manager.UserManager | None = None
@@ -21,36 +38,78 @@ USER_MANAGER: user_manager.UserManager | None = None
 # UUID regex (canonical hyphenated form, case-insensitive)
 UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}"
 
+
+def get_query_dict(request: web.Request) -> dict[str, Any]:
+    """
+    Gets a dictionary of query parameters from the request.
+
+    'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic.
+    """
+    query_dict = {
+        key: request.query.getall(key)
+        if len(request.query.getall(key)) > 1
+        else request.query.get(key)
+        for key in request.query.keys()
+    }
+    return query_dict
+
+
 # Note to any custom node developers reading this code:
 # The assets system is not yet fully implemented, do not rely on the code in /app/assets remaining the same.
 
-def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None:
+
+def register_assets_system(
+    app: web.Application, user_manager_instance: user_manager.UserManager
+) -> None:
     global USER_MANAGER
     USER_MANAGER = user_manager_instance
     app.add_routes(ROUTES)
 
-def _error_response(status: int, code: str, message: str, details: dict | None = None) -> web.Response:
-    return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status)
+
+def _build_error_response(
+    status: int, code: str, message: str, details: dict | None = None
+) -> web.Response:
+    return web.json_response(
+        {"error": {"code": code, "message": message, "details": details or {}}},
+        status=status,
+    )
 
 
-def _validation_error_response(code: str, ve: ValidationError) -> web.Response:
-    return _error_response(400, code, "Validation failed.", {"errors": ve.json()})
+def _build_validation_error_response(code: str, ve: ValidationError) -> web.Response:
+    return _build_error_response(400, code, "Validation failed.", {"errors": ve.json()})
+
+
+def _validate_sort_field(requested: str | None) -> str:
+    if not requested:
+        return "created_at"
+    v = requested.lower()
+    if v in {"name", "created_at", "updated_at", "size", "last_access_time"}:
+        return v
+    return "created_at"
 
 
 @ROUTES.head("/api/assets/hash/{hash}")
 async def head_asset_by_hash(request: web.Request) -> web.Response:
     hash_str = request.match_info.get("hash", "").strip().lower()
     if not hash_str or ":" not in hash_str:
-        return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+        return _build_error_response(
+            400, "INVALID_HASH", "hash must be like 'blake3:<hex>'"
+        )
     algo, digest = hash_str.split(":", 1)
-    if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"):
-        return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-    exists = manager.asset_exists(asset_hash=hash_str)
+    if (
+        algo != "blake3"
+        or not digest
+        or any(c for c in digest if c not in "0123456789abcdef")
+    ):
+        return _build_error_response(
+            400, "INVALID_HASH", "hash must be like 'blake3:<hex>'"
+        )
+    exists = asset_exists(hash_str)
     return web.Response(status=200 if exists else 404)
 
 
 @ROUTES.get("/api/assets")
-async def list_assets(request: web.Request) -> web.Response:
+async def list_assets_route(request: web.Request) -> web.Response:
     """
     GET request to list assets.
     """
@@ -58,66 +117,124 @@ async def list_assets(request: web.Request) -> web.Response:
     try:
         q = schemas_in.ListAssetsQuery.model_validate(query_dict)
     except ValidationError as ve:
-        return _validation_error_response("INVALID_QUERY", ve)
+        return _build_validation_error_response("INVALID_QUERY", ve)
 
-    payload = manager.list_assets(
+    sort = _validate_sort_field(q.sort)
+    order = (
+        "desc"
+        if (q.order or "desc").lower() not in {"asc", "desc"}
+        else q.order.lower()
+    )
+
+    result = list_assets_page(
+        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,
         offset=q.offset,
-        sort=q.sort,
-        order=q.order,
-        owner_id=USER_MANAGER.get_request_user_id(request),
+        sort=sort,
+        order=order,
+    )
+
+    summaries = [
+        schemas_out.AssetSummary(
+            id=item.info.id,
+            name=item.info.name,
+            asset_hash=item.asset.hash if item.asset else None,
+            size=int(item.asset.size_bytes)
+            if item.asset and item.asset.size_bytes
+            else None,
+            mime_type=item.asset.mime_type if item.asset else None,
+            tags=item.tags,
+            created_at=item.info.created_at,
+            updated_at=item.info.updated_at,
+            last_access_time=item.info.last_access_time,
+        )
+        for item in result.items
+    ]
+
+    payload = schemas_out.AssetsList(
+        assets=summaries,
+        total=result.total,
+        has_more=(q.offset + len(summaries)) < result.total,
     )
     return web.json_response(payload.model_dump(mode="json", exclude_none=True))
 
 
 @ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}")
-async def get_asset(request: web.Request) -> web.Response:
+async def get_asset_route(request: web.Request) -> web.Response:
     """
     GET request to get an asset's info as JSON.
     """
     asset_info_id = str(uuid.UUID(request.match_info["id"]))
     try:
-        result = manager.get_asset(
+        result = get_asset_detail(
             asset_info_id=asset_info_id,
             owner_id=USER_MANAGER.get_request_user_id(request),
         )
+        if not result:
+            return _build_error_response(
+                404,
+                "ASSET_NOT_FOUND",
+                f"AssetInfo {asset_info_id} not found",
+                {"id": asset_info_id},
+            )
+
+        payload = schemas_out.AssetDetail(
+            id=result.info.id,
+            name=result.info.name,
+            asset_hash=result.asset.hash if result.asset else None,
+            size=int(result.asset.size_bytes)
+            if result.asset and result.asset.size_bytes is not None
+            else None,
+            mime_type=result.asset.mime_type if result.asset else None,
+            tags=result.tags,
+            user_metadata=result.info.user_metadata or {},
+            preview_id=result.info.preview_id,
+            created_at=result.info.created_at,
+            last_access_time=result.info.last_access_time,
+        )
     except ValueError as e:
-        return _error_response(404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id})
+        return _build_error_response(
+            404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id}
+        )
     except Exception:
         logging.exception(
             "get_asset failed for asset_info_id=%s, owner_id=%s",
             asset_info_id,
             USER_MANAGER.get_request_user_id(request),
         )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(result.model_dump(mode="json"), status=200)
+        return _build_error_response(500, "INTERNAL", "Unexpected server error.")
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}/content")
 async def download_asset_content(request: web.Request) -> web.Response:
-    # question: do we need disposition? could we just stick with one of these?
     disposition = request.query.get("disposition", "attachment").lower().strip()
     if disposition not in {"inline", "attachment"}:
         disposition = "attachment"
 
     try:
-        abs_path, content_type, filename = manager.resolve_asset_content_for_download(
+        result = resolve_asset_for_download(
             asset_info_id=str(uuid.UUID(request.match_info["id"])),
             owner_id=USER_MANAGER.get_request_user_id(request),
         )
+        abs_path = result.abs_path
+        content_type = result.content_type
+        filename = result.download_name
     except ValueError as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve))
+        return _build_error_response(404, "ASSET_NOT_FOUND", str(ve))
     except NotImplementedError as nie:
-        return _error_response(501, "BACKEND_UNSUPPORTED", str(nie))
+        return _build_error_response(501, "BACKEND_UNSUPPORTED", str(nie))
     except FileNotFoundError:
-        return _error_response(404, "FILE_NOT_FOUND", "Underlying file not found on disk.")
+        return _build_error_response(
+            404, "FILE_NOT_FOUND", "Underlying file not found on disk."
+        )
 
     quoted = (filename or "").replace("\r", "").replace("\n", "").replace('"', "'")
-    cd = f'{disposition}; filename="{quoted}"; filename*=UTF-8\'\'{urllib.parse.quote(filename)}'
+    cd = f"{disposition}; filename=\"{quoted}\"; filename*=UTF-8''{urllib.parse.quote(filename)}"
 
     file_size = os.path.getsize(abs_path)
     logging.info(
@@ -129,7 +246,7 @@ async def download_asset_content(request: web.Request) -> web.Response:
         filename,
     )
 
-    async def file_sender():
+    async def stream_file_chunks():
         chunk_size = 64 * 1024
         with open(abs_path, "rb") as f:
             while True:
@@ -139,7 +256,7 @@ async def download_asset_content(request: web.Request) -> web.Response:
                 yield chunk
 
     return web.Response(
-        body=file_sender(),
+        body=stream_file_chunks(),
         content_type=content_type,
         headers={
             "Content-Disposition": cd,
@@ -149,16 +266,18 @@ async def download_asset_content(request: web.Request) -> web.Response:
 
 
 @ROUTES.post("/api/assets/from-hash")
-async def create_asset_from_hash(request: web.Request) -> web.Response:
+async def create_asset_from_hash_route(request: web.Request) -> web.Response:
     try:
         payload = await request.json()
         body = schemas_in.CreateFromHashBody.model_validate(payload)
     except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
+        return _build_validation_error_response("INVALID_BODY", ve)
     except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+        return _build_error_response(
+            400, "INVALID_JSON", "Request body must be valid JSON."
+        )
 
-    result = manager.create_asset_from_hash(
+    result = create_from_hash(
         hash_str=body.hash,
         name=body.name,
         tags=body.tags,
@@ -166,228 +285,191 @@ async def create_asset_from_hash(request: web.Request) -> web.Response:
         owner_id=USER_MANAGER.get_request_user_id(request),
     )
     if result is None:
-        return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {body.hash} does not exist")
-    return web.json_response(result.model_dump(mode="json"), status=201)
+        return _build_error_response(
+            404, "ASSET_NOT_FOUND", f"Asset content {body.hash} does not exist"
+        )
+
+    payload_out = schemas_out.AssetCreated(
+        id=result.info.id,
+        name=result.info.name,
+        asset_hash=result.asset.hash,
+        size=int(result.asset.size_bytes) if result.asset.size_bytes else None,
+        mime_type=result.asset.mime_type,
+        tags=result.tags,
+        user_metadata=result.info.user_metadata or {},
+        preview_id=result.info.preview_id,
+        created_at=result.info.created_at,
+        last_access_time=result.info.last_access_time,
+        created_new=result.created_new,
+    )
+    return web.json_response(payload_out.model_dump(mode="json"), status=201)
+
+
+def _delete_temp_file_if_exists(path: str | None) -> None:
+    if path and os.path.exists(path):
+        try:
+            os.remove(path)
+        except Exception:
+            pass
 
 
 @ROUTES.post("/api/assets")
 async def upload_asset(request: web.Request) -> web.Response:
     """Multipart/form-data endpoint for Asset uploads."""
-    if not (request.content_type or "").lower().startswith("multipart/"):
-        return _error_response(415, "UNSUPPORTED_MEDIA_TYPE", "Use multipart/form-data for uploads.")
-
-    reader = await request.multipart()
-
-    file_present = False
-    file_client_name: str | None = None
-    tags_raw: list[str] = []
-    provided_name: str | None = None
-    user_metadata_raw: str | None = None
-    provided_hash: str | None = None
-    provided_hash_exists: bool | None = None
-
-    file_written = 0
-    tmp_path: str | None = None
-    while True:
-        field = await reader.next()
-        if field is None:
-            break
-
-        fname = getattr(field, "name", "") or ""
-
-        if fname == "hash":
-            try:
-                s = ((await field.text()) or "").strip().lower()
-            except Exception:
-                return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-
-            if s:
-                if ":" not in s:
-                    return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-                algo, digest = s.split(":", 1)
-                if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"):
-                    return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-                provided_hash = f"{algo}:{digest}"
-                try:
-                    provided_hash_exists = manager.asset_exists(asset_hash=provided_hash)
-                except Exception:
-                    provided_hash_exists = None  # do not fail the whole request here
-
-        elif fname == "file":
-            file_present = True
-            file_client_name = (field.filename or "").strip()
-
-            if provided_hash and provided_hash_exists is True:
-                # If client supplied a hash that we know exists, drain but do not write to disk
-                try:
-                    while True:
-                        chunk = await field.read_chunk(8 * 1024 * 1024)
-                        if not chunk:
-                            break
-                        file_written += len(chunk)
-                except Exception:
-                    return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive uploaded file.")
-                continue  # Do not create temp file; we will create AssetInfo from the existing content
-
-            # Otherwise, store to temp for hashing/ingest
-            uploads_root = os.path.join(folder_paths.get_temp_directory(), "uploads")
-            unique_dir = os.path.join(uploads_root, uuid.uuid4().hex)
-            os.makedirs(unique_dir, exist_ok=True)
-            tmp_path = os.path.join(unique_dir, ".upload.part")
-
-            try:
-                with open(tmp_path, "wb") as f:
-                    while True:
-                        chunk = await field.read_chunk(8 * 1024 * 1024)
-                        if not chunk:
-                            break
-                        f.write(chunk)
-                        file_written += len(chunk)
-            except Exception:
-                try:
-                    if os.path.exists(tmp_path or ""):
-                        os.remove(tmp_path)
-                finally:
-                    return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive and store uploaded file.")
-        elif fname == "tags":
-            tags_raw.append((await field.text()) or "")
-        elif fname == "name":
-            provided_name = (await field.text()) or None
-        elif fname == "user_metadata":
-            user_metadata_raw = (await field.text()) or None
-
-    # If client did not send file, and we are not doing a from-hash fast path -> error
-    if not file_present and not (provided_hash and provided_hash_exists):
-        return _error_response(400, "MISSING_FILE", "Form must include a 'file' part or a known 'hash'.")
-
-    if file_present and file_written == 0 and not (provided_hash and provided_hash_exists):
-        # Empty upload is only acceptable if we are fast-pathing from existing hash
-        try:
-            if tmp_path and os.path.exists(tmp_path):
-                os.remove(tmp_path)
-        finally:
-            return _error_response(400, "EMPTY_UPLOAD", "Uploaded file is empty.")
-
     try:
-        spec = schemas_in.UploadAssetSpec.model_validate({
-            "tags": tags_raw,
-            "name": provided_name,
-            "user_metadata": user_metadata_raw,
-            "hash": provided_hash,
-        })
-    except ValidationError as ve:
-        try:
-            if tmp_path and os.path.exists(tmp_path):
-                os.remove(tmp_path)
-        finally:
-            return _validation_error_response("INVALID_BODY", ve)
-
-    # Validate models category against configured folders (consistent with previous behavior)
-    if spec.tags and spec.tags[0] == "models":
-        if len(spec.tags) < 2 or spec.tags[1] not in folder_paths.folder_names_and_paths:
-            if tmp_path and os.path.exists(tmp_path):
-                os.remove(tmp_path)
-            return _error_response(
-                400, "INVALID_BODY", f"unknown models category '{spec.tags[1] if len(spec.tags) >= 2 else ''}'"
-            )
+        parsed = await parse_multipart_upload(request, check_hash_exists=asset_exists)
+    except UploadError as e:
+        return _build_error_response(e.status, e.code, e.message)
 
     owner_id = USER_MANAGER.get_request_user_id(request)
 
-    # Fast path: if a valid provided hash exists, create AssetInfo without writing anything
-    if spec.hash and provided_hash_exists is True:
-        try:
-            result = manager.create_asset_from_hash(
+    try:
+        spec = schemas_in.UploadAssetSpec.model_validate(
+            {
+                "tags": parsed.tags_raw,
+                "name": parsed.provided_name,
+                "user_metadata": parsed.user_metadata_raw,
+                "hash": parsed.provided_hash,
+            }
+        )
+    except ValidationError as ve:
+        _delete_temp_file_if_exists(parsed.tmp_path)
+        return _build_error_response(
+            400, "INVALID_BODY", f"Validation failed: {ve.json()}"
+        )
+
+    if spec.tags and spec.tags[0] == "models":
+        if (
+            len(spec.tags) < 2
+            or spec.tags[1] not in folder_paths.folder_names_and_paths
+        ):
+            _delete_temp_file_if_exists(parsed.tmp_path)
+            category = spec.tags[1] if len(spec.tags) >= 2 else ""
+            return _build_error_response(
+                400, "INVALID_BODY", f"unknown models category '{category}'"
+            )
+
+    try:
+        # Fast path: if a valid provided hash exists, create AssetInfo without writing anything
+        if spec.hash and parsed.provided_hash_exists is True:
+            result = create_from_hash(
                 hash_str=spec.hash,
                 name=spec.name or (spec.hash.split(":", 1)[1]),
                 tags=spec.tags,
                 user_metadata=spec.user_metadata or {},
                 owner_id=owner_id,
             )
-        except Exception:
-            logging.exception("create_asset_from_hash failed for hash=%s, owner_id=%s", spec.hash, owner_id)
-            return _error_response(500, "INTERNAL", "Unexpected server error.")
+            if result is None:
+                _delete_temp_file_if_exists(parsed.tmp_path)
+                return _build_error_response(
+                    404, "ASSET_NOT_FOUND", f"Asset content {spec.hash} does not exist"
+                )
+            _delete_temp_file_if_exists(parsed.tmp_path)
+        else:
+            # Otherwise, we must have a temp file path to ingest
+            if not parsed.tmp_path or not os.path.exists(parsed.tmp_path):
+                return _build_error_response(
+                    404,
+                    "ASSET_NOT_FOUND",
+                    "Provided hash not found and no file uploaded.",
+                )
 
-        if result is None:
-            return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {spec.hash} does not exist")
-
-        # Drain temp if we accidentally saved (e.g., hash field came after file)
-        if tmp_path and os.path.exists(tmp_path):
-            with contextlib.suppress(Exception):
-                os.remove(tmp_path)
-
-        status = 200 if (not result.created_new) else 201
-        return web.json_response(result.model_dump(mode="json"), status=status)
-
-    # Otherwise, we must have a temp file path to ingest
-    if not tmp_path or not os.path.exists(tmp_path):
-        # The only case we reach here without a temp file is: client sent a hash that does not exist and no file
-        return _error_response(404, "ASSET_NOT_FOUND", "Provided hash not found and no file uploaded.")
-
-    try:
-        created = manager.upload_asset_from_temp_path(
-            spec,
-            temp_path=tmp_path,
-            client_filename=file_client_name,
-            owner_id=owner_id,
-            expected_asset_hash=spec.hash,
-        )
-        status = 201 if created.created_new else 200
-        return web.json_response(created.model_dump(mode="json"), status=status)
-    except ValueError as e:
-        if tmp_path and os.path.exists(tmp_path):
-            os.remove(tmp_path)
-        msg = str(e)
-        if "HASH_MISMATCH" in msg or msg.strip().upper() == "HASH_MISMATCH":
-            return _error_response(
-                400,
-                "HASH_MISMATCH",
-                "Uploaded file hash does not match provided hash.",
+            result = upload_from_temp_path(
+                temp_path=parsed.tmp_path,
+                name=spec.name,
+                tags=spec.tags,
+                user_metadata=spec.user_metadata or {},
+                client_filename=parsed.file_client_name,
+                owner_id=owner_id,
+                expected_hash=spec.hash,
             )
-        return _error_response(400, "BAD_REQUEST", "Invalid inputs.")
+    except AssetValidationError as e:
+        _delete_temp_file_if_exists(parsed.tmp_path)
+        return _build_error_response(400, e.code, str(e))
+    except ValueError as e:
+        _delete_temp_file_if_exists(parsed.tmp_path)
+        return _build_error_response(400, "BAD_REQUEST", str(e))
+    except HashMismatchError as e:
+        _delete_temp_file_if_exists(parsed.tmp_path)
+        return _build_error_response(400, "HASH_MISMATCH", str(e))
+    except DependencyMissingError as e:
+        _delete_temp_file_if_exists(parsed.tmp_path)
+        return _build_error_response(503, "DEPENDENCY_MISSING", e.message)
     except Exception:
-        if tmp_path and os.path.exists(tmp_path):
-            os.remove(tmp_path)
-        logging.exception("upload_asset_from_temp_path failed for tmp_path=%s, owner_id=%s", tmp_path, owner_id)
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
+        _delete_temp_file_if_exists(parsed.tmp_path)
+        logging.exception("upload_asset failed for owner_id=%s", owner_id)
+        return _build_error_response(500, "INTERNAL", "Unexpected server error.")
+
+    payload = schemas_out.AssetCreated(
+        id=result.info.id,
+        name=result.info.name,
+        asset_hash=result.asset.hash,
+        size=int(result.asset.size_bytes) if result.asset.size_bytes else None,
+        mime_type=result.asset.mime_type,
+        tags=result.tags,
+        user_metadata=result.info.user_metadata or {},
+        preview_id=result.info.preview_id,
+        created_at=result.info.created_at,
+        last_access_time=result.info.last_access_time,
+        created_new=result.created_new,
+    )
+    status = 201 if result.created_new else 200
+    return web.json_response(payload.model_dump(mode="json"), status=status)
 
 
 @ROUTES.put(f"/api/assets/{{id:{UUID_RE}}}")
-async def update_asset(request: web.Request) -> web.Response:
+async def update_asset_route(request: web.Request) -> web.Response:
     asset_info_id = str(uuid.UUID(request.match_info["id"]))
     try:
         body = schemas_in.UpdateAssetBody.model_validate(await request.json())
     except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
+        return _build_validation_error_response("INVALID_BODY", ve)
     except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+        return _build_error_response(
+            400, "INVALID_JSON", "Request body must be valid JSON."
+        )
 
     try:
-        result = manager.update_asset(
+        result = update_asset_metadata(
             asset_info_id=asset_info_id,
             name=body.name,
             user_metadata=body.user_metadata,
             owner_id=USER_MANAGER.get_request_user_id(request),
         )
+        payload = schemas_out.AssetUpdated(
+            id=result.info.id,
+            name=result.info.name,
+            asset_hash=result.asset.hash if result.asset else None,
+            tags=result.tags,
+            user_metadata=result.info.user_metadata or {},
+            updated_at=result.info.updated_at,
+        )
     except (ValueError, PermissionError) as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
+        return _build_error_response(
+            404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id}
+        )
     except Exception:
         logging.exception(
             "update_asset failed for asset_info_id=%s, owner_id=%s",
             asset_info_id,
             USER_MANAGER.get_request_user_id(request),
         )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(result.model_dump(mode="json"), status=200)
+        return _build_error_response(500, "INTERNAL", "Unexpected server error.")
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}")
-async def delete_asset(request: web.Request) -> web.Response:
+async def delete_asset_route(request: web.Request) -> web.Response:
     asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    delete_content = request.query.get("delete_content")
-    delete_content = True if delete_content is None else delete_content.lower() not in {"0", "false", "no"}
+    delete_content_param = request.query.get("delete_content")
+    delete_content = (
+        True
+        if delete_content_param is None
+        else delete_content_param.lower() not in {"0", "false", "no"}
+    )
 
     try:
-        deleted = manager.delete_asset_reference(
+        deleted = delete_asset_reference(
             asset_info_id=asset_info_id,
             owner_id=USER_MANAGER.get_request_user_id(request),
             delete_content_if_orphan=delete_content,
@@ -398,10 +480,12 @@ async def delete_asset(request: web.Request) -> web.Response:
             asset_info_id,
             USER_MANAGER.get_request_user_id(request),
         )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
+        return _build_error_response(500, "INTERNAL", "Unexpected server error.")
 
     if not deleted:
-        return _error_response(404, "ASSET_NOT_FOUND", f"AssetInfo {asset_info_id} not found.")
+        return _build_error_response(
+            404, "ASSET_NOT_FOUND", f"AssetInfo {asset_info_id} not found."
+        )
     return web.Response(status=204)
 
 
@@ -416,11 +500,17 @@ async def get_tags(request: web.Request) -> web.Response:
         query = schemas_in.TagsListQuery.model_validate(query_map)
     except ValidationError as e:
         return web.json_response(
-            {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": e.errors()}},
+            {
+                "error": {
+                    "code": "INVALID_QUERY",
+                    "message": "Invalid query parameters",
+                    "details": e.errors(),
+                }
+            },
             status=400,
         )
 
-    result = manager.list_tags(
+    rows, total = list_tags(
         prefix=query.prefix,
         limit=query.limit,
         offset=query.offset,
@@ -428,72 +518,108 @@ async def get_tags(request: web.Request) -> web.Response:
         include_zero=query.include_zero,
         owner_id=USER_MANAGER.get_request_user_id(request),
     )
-    return web.json_response(result.model_dump(mode="json"))
+
+    tags = [
+        schemas_out.TagUsage(name=name, count=count, type=tag_type)
+        for (name, tag_type, count) in rows
+    ]
+    payload = schemas_out.TagsList(
+        tags=tags, total=total, has_more=(query.offset + len(tags)) < total
+    )
+    return web.json_response(payload.model_dump(mode="json"))
 
 
 @ROUTES.post(f"/api/assets/{{id:{UUID_RE}}}/tags")
 async def add_asset_tags(request: web.Request) -> web.Response:
     asset_info_id = str(uuid.UUID(request.match_info["id"]))
     try:
-        payload = await request.json()
-        data = schemas_in.TagsAdd.model_validate(payload)
+        json_payload = await request.json()
+        data = schemas_in.TagsAdd.model_validate(json_payload)
     except ValidationError as ve:
-        return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags add.", {"errors": ve.errors()})
+        return _build_error_response(
+            400,
+            "INVALID_BODY",
+            "Invalid JSON body for tags add.",
+            {"errors": ve.errors()},
+        )
     except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+        return _build_error_response(
+            400, "INVALID_JSON", "Request body must be valid JSON."
+        )
 
     try:
-        result = manager.add_tags_to_asset(
+        result = apply_tags(
             asset_info_id=asset_info_id,
             tags=data.tags,
             origin="manual",
             owner_id=USER_MANAGER.get_request_user_id(request),
         )
+        payload = schemas_out.TagsAdd(
+            added=result.added,
+            already_present=result.already_present,
+            total_tags=result.total_tags,
+        )
     except (ValueError, PermissionError) as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
+        return _build_error_response(
+            404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id}
+        )
     except Exception:
         logging.exception(
             "add_tags_to_asset failed for asset_info_id=%s, owner_id=%s",
             asset_info_id,
             USER_MANAGER.get_request_user_id(request),
         )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
+        return _build_error_response(500, "INTERNAL", "Unexpected server error.")
 
-    return web.json_response(result.model_dump(mode="json"), status=200)
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}/tags")
 async def delete_asset_tags(request: web.Request) -> web.Response:
     asset_info_id = str(uuid.UUID(request.match_info["id"]))
     try:
-        payload = await request.json()
-        data = schemas_in.TagsRemove.model_validate(payload)
+        json_payload = await request.json()
+        data = schemas_in.TagsRemove.model_validate(json_payload)
     except ValidationError as ve:
-        return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags remove.", {"errors": ve.errors()})
+        return _build_error_response(
+            400,
+            "INVALID_BODY",
+            "Invalid JSON body for tags remove.",
+            {"errors": ve.errors()},
+        )
     except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+        return _build_error_response(
+            400, "INVALID_JSON", "Request body must be valid JSON."
+        )
 
     try:
-        result = manager.remove_tags_from_asset(
+        result = remove_tags(
             asset_info_id=asset_info_id,
             tags=data.tags,
             owner_id=USER_MANAGER.get_request_user_id(request),
         )
+        payload = schemas_out.TagsRemove(
+            removed=result.removed,
+            not_present=result.not_present,
+            total_tags=result.total_tags,
+        )
     except ValueError as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
+        return _build_error_response(
+            404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id}
+        )
     except Exception:
         logging.exception(
             "remove_tags_from_asset failed for asset_info_id=%s, owner_id=%s",
             asset_info_id,
             USER_MANAGER.get_request_user_id(request),
         )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
+        return _build_error_response(500, "INTERNAL", "Unexpected server error.")
 
-    return web.json_response(result.model_dump(mode="json"), status=200)
+    return web.json_response(payload.model_dump(mode="json"), status=200)
 
 
 @ROUTES.post("/api/assets/seed")
-async def seed_assets_endpoint(request: web.Request) -> web.Response:
+async def seed_assets(request: web.Request) -> web.Response:
     """Trigger asset seeding for specified roots (models, input, output)."""
     try:
         payload = await request.json()
@@ -503,12 +629,12 @@ async def seed_assets_endpoint(request: web.Request) -> web.Response:
 
     valid_roots = [r for r in roots if r in ("models", "input", "output")]
     if not valid_roots:
-        return _error_response(400, "INVALID_BODY", "No valid roots specified")
+        return _build_error_response(400, "INVALID_BODY", "No valid roots specified")
 
     try:
-        seed_assets(tuple(valid_roots))
+        scanner_seed_assets(tuple(valid_roots))
     except Exception:
-        logging.exception("seed_assets failed for roots=%s", valid_roots)
-        return _error_response(500, "INTERNAL", "Seed operation failed")
+        logging.exception("scanner_seed_assets failed for roots=%s", valid_roots)
+        return _build_error_response(500, "INTERNAL", "Seed operation failed")
 
     return web.json_response({"seeded": valid_roots}, status=200)

app/assets/api/schemas_in.py

@@ -1,4 +1,5 @@
 import json
+from dataclasses import dataclass
 from typing import Any, Literal
 
 from pydantic import (
@@ -10,6 +11,61 @@ from pydantic import (
     model_validator,
 )
 
+
+class UploadError(Exception):
+    """Error during upload parsing with HTTP status and code (used in HTTP layer only)."""
+
+    def __init__(self, status: int, code: str, message: str):
+        super().__init__(message)
+        self.status = status
+        self.code = code
+        self.message = message
+
+
+class AssetValidationError(Exception):
+    """Validation error in asset processing (invalid tags, metadata, etc.)."""
+
+    def __init__(self, code: str, message: str):
+        super().__init__(message)
+        self.code = code
+
+
+class AssetNotFoundError(Exception):
+    """Asset or asset content not found."""
+
+    def __init__(self, message: str):
+        super().__init__(message)
+
+
+class HashMismatchError(Exception):
+    """Uploaded file hash does not match provided hash."""
+
+    pass
+
+
+class DependencyMissingError(Exception):
+    """A required dependency is not installed."""
+
+    def __init__(self, message: str):
+        super().__init__(message)
+        self.message = message
+
+
+@dataclass
+class ParsedUpload:
+    """Result of parsing a multipart upload request."""
+
+    file_present: bool
+    file_written: int
+    file_client_name: str | None
+    tmp_path: str | None
+    tags_raw: list[str]
+    provided_name: str | None
+    user_metadata_raw: str | None
+    provided_hash: str | None
+    provided_hash_exists: bool | None
+
+
 class ListAssetsQuery(BaseModel):
     include_tags: list[str] = Field(default_factory=list)
     exclude_tags: list[str] = Field(default_factory=list)
@@ -21,7 +77,9 @@ class ListAssetsQuery(BaseModel):
     limit: conint(ge=1, le=500) = 20
     offset: conint(ge=0) = 0
 
-    sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at"
+    sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = (
+        "created_at"
+    )
     order: Literal["asc", "desc"] = "desc"
 
     @field_validator("include_tags", "exclude_tags", mode="before")
@@ -61,7 +119,7 @@ class UpdateAssetBody(BaseModel):
     user_metadata: dict[str, Any] | None = None
 
     @model_validator(mode="after")
-    def _at_least_one(self):
+    def _validate_at_least_one_field(self):
         if self.name is None and self.user_metadata is None:
             raise ValueError("Provide at least one of: name, user_metadata.")
         return self
@@ -90,7 +148,7 @@ class CreateFromHashBody(BaseModel):
 
     @field_validator("tags", mode="before")
     @classmethod
-    def _tags_norm(cls, v):
+    def _normalize_tags_field(cls, v):
         if v is None:
             return []
         if isinstance(v, list):
@@ -163,6 +221,7 @@ class UploadAssetSpec(BaseModel):
     Files created via this endpoint are stored on disk using the **content hash** as the filename stem
     and the original extension is preserved when available.
     """
+
     model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
 
     tags: list[str] = Field(..., min_length=1)
@@ -260,5 +319,7 @@ class UploadAssetSpec(BaseModel):
             raise ValueError("first tag must be one of: models, input, output")
         if root == "models":
             if len(self.tags) < 2:
-                raise ValueError("models uploads require a category tag as the second tag")
+                raise ValueError(
+                    "models uploads require a category tag as the second tag"
+                )
         return self

app/assets/api/schemas_out.py

@@ -19,7 +19,7 @@ class AssetSummary(BaseModel):
     model_config = ConfigDict(from_attributes=True)
 
     @field_serializer("created_at", "updated_at", "last_access_time")
-    def _ser_dt(self, v: datetime | None, _info):
+    def _serialize_datetime(self, v: datetime | None, _info):
         return v.isoformat() if v else None
 
 
@@ -40,7 +40,7 @@ class AssetUpdated(BaseModel):
     model_config = ConfigDict(from_attributes=True)
 
     @field_serializer("updated_at")
-    def _ser_updated(self, v: datetime | None, _info):
+    def _serialize_updated_at(self, v: datetime | None, _info):
         return v.isoformat() if v else None
 
 
@@ -59,7 +59,7 @@ class AssetDetail(BaseModel):
     model_config = ConfigDict(from_attributes=True)
 
     @field_serializer("created_at", "last_access_time")
-    def _ser_dt(self, v: datetime | None, _info):
+    def _serialize_datetime(self, v: datetime | None, _info):
         return v.isoformat() if v else None
 
 

app/assets/api/upload.py

@@ -0,0 +1,165 @@
+import os
+import uuid
+
+from aiohttp import web
+
+import folder_paths
+from app.assets.api.schemas_in import ParsedUpload, UploadError
+
+
+def normalize_and_validate_hash(s: str) -> str:
+    """
+    Validate and normalize a hash string.
+
+    Returns canonical 'blake3:<hex>' or raises UploadError.
+    """
+    s = s.strip().lower()
+    if not s:
+        raise UploadError(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+    if ":" not in s:
+        raise UploadError(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+    algo, digest = s.split(":", 1)
+    if (
+        algo != "blake3"
+        or not digest
+        or any(c for c in digest if c not in "0123456789abcdef")
+    ):
+        raise UploadError(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+    return f"{algo}:{digest}"
+
+
+async def parse_multipart_upload(
+    request: web.Request,
+    check_hash_exists: callable,
+) -> ParsedUpload:
+    """
+    Parse a multipart/form-data upload request.
+
+    Args:
+        request: The aiohttp request
+        check_hash_exists: Callable(hash_str) -> bool to check if a hash exists
+
+    Returns:
+        ParsedUpload with parsed fields and temp file path
+
+    Raises:
+        UploadError: On validation or I/O errors
+    """
+    if not (request.content_type or "").lower().startswith("multipart/"):
+        raise UploadError(
+            415, "UNSUPPORTED_MEDIA_TYPE", "Use multipart/form-data for uploads."
+        )
+
+    reader = await request.multipart()
+
+    file_present = False
+    file_client_name: str | None = None
+    tags_raw: list[str] = []
+    provided_name: str | None = None
+    user_metadata_raw: str | None = None
+    provided_hash: str | None = None
+    provided_hash_exists: bool | None = None
+
+    file_written = 0
+    tmp_path: str | None = None
+
+    while True:
+        field = await reader.next()
+        if field is None:
+            break
+
+        fname = getattr(field, "name", "") or ""
+
+        if fname == "hash":
+            try:
+                s = ((await field.text()) or "").strip().lower()
+            except Exception:
+                raise UploadError(
+                    400, "INVALID_HASH", "hash must be like 'blake3:<hex>'"
+                )
+
+            if s:
+                provided_hash = normalize_and_validate_hash(s)
+                try:
+                    provided_hash_exists = check_hash_exists(provided_hash)
+                except Exception:
+                    provided_hash_exists = None  # do not fail the whole request here
+
+        elif fname == "file":
+            file_present = True
+            file_client_name = (field.filename or "").strip()
+
+            if provided_hash and provided_hash_exists is True:
+                # If client supplied a hash that we know exists, drain but do not write to disk
+                try:
+                    while True:
+                        chunk = await field.read_chunk(8 * 1024 * 1024)
+                        if not chunk:
+                            break
+                        file_written += len(chunk)
+                except Exception:
+                    raise UploadError(
+                        500, "UPLOAD_IO_ERROR", "Failed to receive uploaded file."
+                    )
+                continue
+
+            uploads_root = os.path.join(folder_paths.get_temp_directory(), "uploads")
+            unique_dir = os.path.join(uploads_root, uuid.uuid4().hex)
+            os.makedirs(unique_dir, exist_ok=True)
+            tmp_path = os.path.join(unique_dir, ".upload.part")
+
+            try:
+                with open(tmp_path, "wb") as f:
+                    while True:
+                        chunk = await field.read_chunk(8 * 1024 * 1024)
+                        if not chunk:
+                            break
+                        f.write(chunk)
+                        file_written += len(chunk)
+            except Exception:
+                _delete_temp_file_if_exists(tmp_path)
+                raise UploadError(
+                    500, "UPLOAD_IO_ERROR", "Failed to receive and store uploaded file."
+                )
+
+        elif fname == "tags":
+            tags_raw.append((await field.text()) or "")
+        elif fname == "name":
+            provided_name = (await field.text()) or None
+        elif fname == "user_metadata":
+            user_metadata_raw = (await field.text()) or None
+
+    if not file_present and not (provided_hash and provided_hash_exists):
+        raise UploadError(
+            400, "MISSING_FILE", "Form must include a 'file' part or a known 'hash'."
+        )
+
+    if (
+        file_present
+        and file_written == 0
+        and not (provided_hash and provided_hash_exists)
+    ):
+        _delete_temp_file_if_exists(tmp_path)
+        raise UploadError(400, "EMPTY_UPLOAD", "Uploaded file is empty.")
+
+    return ParsedUpload(
+        file_present=file_present,
+        file_written=file_written,
+        file_client_name=file_client_name,
+        tmp_path=tmp_path,
+        tags_raw=tags_raw,
+        provided_name=provided_name,
+        user_metadata_raw=user_metadata_raw,
+        provided_hash=provided_hash,
+        provided_hash_exists=provided_hash_exists,
+    )
+
+
+def _delete_temp_file_if_exists(tmp_path: str | None) -> None:
+    """Safely remove a temp file if it exists."""
+    if tmp_path:
+        try:
+            if os.path.exists(tmp_path):
+                os.remove(tmp_path)
+        except Exception:
+            pass

app/assets/database/bulk_ops.py

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

app/assets/database/models.py

@@ -2,8 +2,8 @@ from __future__ import annotations
 
 import uuid
 from datetime import datetime
-
 from typing import Any
+
 from sqlalchemy import (
     JSON,
     BigInteger,
@@ -20,19 +20,21 @@ from sqlalchemy import (
 )
 from sqlalchemy.orm import Mapped, foreign, mapped_column, relationship
 
-from app.assets.helpers import utcnow
-from app.database.models import to_dict, Base
+from app.assets.helpers import get_utc_now
+from app.database.models import Base, to_dict
 
 
 class Asset(Base):
     __tablename__ = "assets"
 
-    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    id: Mapped[str] = mapped_column(
+        String(36), primary_key=True, default=lambda: str(uuid.uuid4())
+    )
     hash: Mapped[str | None] = mapped_column(String(256), nullable=True)
     size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0)
     mime_type: Mapped[str | None] = mapped_column(String(255))
     created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=False), nullable=False, default=utcnow
+        DateTime(timezone=False), nullable=False, default=get_utc_now
     )
 
     infos: Mapped[list[AssetInfo]] = relationship(
@@ -75,7 +77,9 @@ class AssetCacheState(Base):
     __tablename__ = "asset_cache_state"
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
-    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False)
+    asset_id: Mapped[str] = mapped_column(
+        String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False
+    )
     file_path: Mapped[str] = mapped_column(Text, nullable=False)
     mtime_ns: Mapped[int | None] = mapped_column(BigInteger, nullable=True)
     needs_verify: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
@@ -85,7 +89,9 @@ class AssetCacheState(Base):
     __table_args__ = (
         Index("ix_asset_cache_state_file_path", "file_path"),
         Index("ix_asset_cache_state_asset_id", "asset_id"),
-        CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
+        CheckConstraint(
+            "(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"
+        ),
         UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
     )
 
@@ -99,15 +105,29 @@ class AssetCacheState(Base):
 class AssetInfo(Base):
     __tablename__ = "assets_info"
 
-    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    id: Mapped[str] = mapped_column(
+        String(36), primary_key=True, default=lambda: str(uuid.uuid4())
+    )
     owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="")
     name: Mapped[str] = mapped_column(String(512), nullable=False)
-    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False)
-    preview_id: Mapped[str | None] = mapped_column(String(36), ForeignKey("assets.id", ondelete="SET NULL"))
-    user_metadata: Mapped[dict[str, Any] | None] = mapped_column(JSON(none_as_null=True))
-    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
-    updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
-    last_access_time: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+    asset_id: Mapped[str] = mapped_column(
+        String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False
+    )
+    preview_id: Mapped[str | None] = mapped_column(
+        String(36), ForeignKey("assets.id", ondelete="SET NULL")
+    )
+    user_metadata: Mapped[dict[str, Any] | None] = mapped_column(
+        JSON(none_as_null=True)
+    )
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=get_utc_now
+    )
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=get_utc_now
+    )
+    last_access_time: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=get_utc_now
+    )
 
     asset: Mapped[Asset] = relationship(
         "Asset",
@@ -143,7 +163,9 @@ class AssetInfo(Base):
     )
 
     __table_args__ = (
-        UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
+        UniqueConstraint(
+            "asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"
+        ),
         Index("ix_assets_info_owner_name", "owner_id", "name"),
         Index("ix_assets_info_owner_id", "owner_id"),
         Index("ix_assets_info_asset_id", "asset_id"),
@@ -196,7 +218,7 @@ class AssetInfoTag(Base):
     )
     origin: Mapped[str] = mapped_column(String(32), nullable=False, default="manual")
     added_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=False), nullable=False, default=utcnow
+        DateTime(timezone=False), nullable=False, default=get_utc_now
     )
 
     asset_info: Mapped[AssetInfo] = relationship(back_populates="tag_links")
@@ -225,9 +247,7 @@ class Tag(Base):
         overlaps="asset_info_links,tag_links,tags,asset_info",
     )
 
-    __table_args__ = (
-        Index("ix_tags_tag_type", "tag_type"),
-    )
+    __table_args__ = (Index("ix_tags_tag_type", "tag_type"),)
 
     def __repr__(self) -> str:
         return f"<Tag {self.name}>"

app/assets/database/queries.py

@@ -1,976 +0,0 @@
-import os
-import logging
-import sqlalchemy as sa
-from collections import defaultdict
-from datetime import datetime
-from typing import Iterable, Any
-from sqlalchemy import select, delete, exists, func
-from sqlalchemy.dialects import sqlite
-from sqlalchemy.exc import IntegrityError
-from sqlalchemy.orm import Session, contains_eager, noload
-from app.assets.database.models import Asset, AssetInfo, AssetCacheState, AssetInfoMeta, AssetInfoTag, Tag
-from app.assets.helpers import (
-    compute_relative_filename, escape_like_prefix, normalize_tags, project_kv, utcnow
-)
-from typing import Sequence
-
-
-def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement:
-    """Build owner visibility predicate for reads. Owner-less rows are visible to everyone."""
-    owner_id = (owner_id or "").strip()
-    if owner_id == "":
-        return AssetInfo.owner_id == ""
-    return AssetInfo.owner_id.in_(["", owner_id])
-
-
-def pick_best_live_path(states: Sequence[AssetCacheState]) -> str:
-    """
-    Return the best on-disk path among cache states:
-      1) Prefer a path that exists with needs_verify == False (already verified).
-      2) Otherwise, pick the first path that exists.
-      3) Otherwise return empty string.
-    """
-    alive = [s for s in states if getattr(s, "file_path", None) and os.path.isfile(s.file_path)]
-    if not alive:
-        return ""
-    for s in alive:
-        if not getattr(s, "needs_verify", False):
-            return s.file_path
-    return alive[0].file_path
-
-
-def apply_tag_filters(
-    stmt: sa.sql.Select,
-    include_tags: Sequence[str] | None = None,
-    exclude_tags: Sequence[str] | None = None,
-) -> sa.sql.Select:
-    """include_tags: every tag must be present; exclude_tags: none may be present."""
-    include_tags = normalize_tags(include_tags)
-    exclude_tags = normalize_tags(exclude_tags)
-
-    if include_tags:
-        for tag_name in include_tags:
-            stmt = stmt.where(
-                exists().where(
-                    (AssetInfoTag.asset_info_id == AssetInfo.id)
-                    & (AssetInfoTag.tag_name == tag_name)
-                )
-            )
-
-    if exclude_tags:
-        stmt = stmt.where(
-            ~exists().where(
-                (AssetInfoTag.asset_info_id == AssetInfo.id)
-                & (AssetInfoTag.tag_name.in_(exclude_tags))
-            )
-        )
-    return stmt
-
-
-def apply_metadata_filter(
-    stmt: sa.sql.Select,
-    metadata_filter: dict | None = None,
-) -> sa.sql.Select:
-    """Apply filters using asset_info_meta projection table."""
-    if not metadata_filter:
-        return stmt
-
-    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
-        return sa.exists().where(
-            AssetInfoMeta.asset_info_id == AssetInfo.id,
-            AssetInfoMeta.key == key,
-            *preds,
-        )
-
-    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
-        if value is None:
-            no_row_for_key = sa.not_(
-                sa.exists().where(
-                    AssetInfoMeta.asset_info_id == AssetInfo.id,
-                    AssetInfoMeta.key == key,
-                )
-            )
-            null_row = _exists_for_pred(
-                key,
-                AssetInfoMeta.val_json.is_(None),
-                AssetInfoMeta.val_str.is_(None),
-                AssetInfoMeta.val_num.is_(None),
-                AssetInfoMeta.val_bool.is_(None),
-            )
-            return sa.or_(no_row_for_key, null_row)
-
-        if isinstance(value, bool):
-            return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value))
-        if isinstance(value, (int, float)):
-            from decimal import Decimal
-            num = value if isinstance(value, Decimal) else Decimal(str(value))
-            return _exists_for_pred(key, AssetInfoMeta.val_num == num)
-        if isinstance(value, str):
-            return _exists_for_pred(key, AssetInfoMeta.val_str == value)
-        return _exists_for_pred(key, AssetInfoMeta.val_json == value)
-
-    for k, v in metadata_filter.items():
-        if isinstance(v, list):
-            ors = [_exists_clause_for_value(k, elem) for elem in v]
-            if ors:
-                stmt = stmt.where(sa.or_(*ors))
-        else:
-            stmt = stmt.where(_exists_clause_for_value(k, v))
-    return stmt
-
-
-def asset_exists_by_hash(
-    session: Session,
-    *,
-    asset_hash: str,
-) -> bool:
-    """
-    Check if an asset with a given hash exists in database.
-    """
-    row = (
-        session.execute(
-            select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1)
-        )
-    ).first()
-    return row is not None
-
-
-def asset_info_exists_for_asset_id(
-    session: Session,
-    *,
-    asset_id: str,
-) -> bool:
-    q = (
-        select(sa.literal(True))
-        .select_from(AssetInfo)
-        .where(AssetInfo.asset_id == asset_id)
-        .limit(1)
-    )
-    return (session.execute(q)).first() is not None
-
-
-def get_asset_by_hash(
-    session: Session,
-    *,
-    asset_hash: str,
-) -> Asset | None:
-    return (
-        session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
-    ).scalars().first()
-
-
-def get_asset_info_by_id(
-    session: Session,
-    *,
-    asset_info_id: str,
-) -> AssetInfo | None:
-    return session.get(AssetInfo, asset_info_id)
-
-
-def list_asset_infos_page(
-    session: Session,
-    owner_id: str = "",
-    include_tags: Sequence[str] | None = None,
-    exclude_tags: Sequence[str] | None = None,
-    name_contains: str | None = None,
-    metadata_filter: dict | None = None,
-    limit: int = 20,
-    offset: int = 0,
-    sort: str = "created_at",
-    order: str = "desc",
-) -> tuple[list[AssetInfo], dict[str, list[str]], int]:
-    base = (
-        select(AssetInfo)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags))
-        .where(visible_owner_clause(owner_id))
-    )
-
-    if name_contains:
-        escaped, esc = escape_like_prefix(name_contains)
-        base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
-
-    base = apply_tag_filters(base, include_tags, exclude_tags)
-    base = apply_metadata_filter(base, metadata_filter)
-
-    sort = (sort or "created_at").lower()
-    order = (order or "desc").lower()
-    sort_map = {
-        "name": AssetInfo.name,
-        "created_at": AssetInfo.created_at,
-        "updated_at": AssetInfo.updated_at,
-        "last_access_time": AssetInfo.last_access_time,
-        "size": Asset.size_bytes,
-    }
-    sort_col = sort_map.get(sort, AssetInfo.created_at)
-    sort_exp = sort_col.desc() if order == "desc" else sort_col.asc()
-
-    base = base.order_by(sort_exp).limit(limit).offset(offset)
-
-    count_stmt = (
-        select(sa.func.count())
-        .select_from(AssetInfo)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .where(visible_owner_clause(owner_id))
-    )
-    if name_contains:
-        escaped, esc = escape_like_prefix(name_contains)
-        count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
-    count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags)
-    count_stmt = apply_metadata_filter(count_stmt, metadata_filter)
-
-    total = int((session.execute(count_stmt)).scalar_one() or 0)
-
-    infos = (session.execute(base)).unique().scalars().all()
-
-    id_list: list[str] = [i.id for i in infos]
-    tag_map: dict[str, list[str]] = defaultdict(list)
-    if id_list:
-        rows = session.execute(
-            select(AssetInfoTag.asset_info_id, Tag.name)
-            .join(Tag, Tag.name == AssetInfoTag.tag_name)
-            .where(AssetInfoTag.asset_info_id.in_(id_list))
-            .order_by(AssetInfoTag.added_at)
-        )
-        for aid, tag_name in rows.all():
-            tag_map[aid].append(tag_name)
-
-    return infos, tag_map, total
-
-
-def fetch_asset_info_asset_and_tags(
-    session: Session,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> tuple[AssetInfo, Asset, list[str]] | None:
-    stmt = (
-        select(AssetInfo, Asset, Tag.name)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True)
-        .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True)
-        .where(
-            AssetInfo.id == asset_info_id,
-            visible_owner_clause(owner_id),
-        )
-        .options(noload(AssetInfo.tags))
-        .order_by(Tag.name.asc())
-    )
-
-    rows = (session.execute(stmt)).all()
-    if not rows:
-        return None
-
-    first_info, first_asset, _ = rows[0]
-    tags: list[str] = []
-    seen: set[str] = set()
-    for _info, _asset, tag_name in rows:
-        if tag_name and tag_name not in seen:
-            seen.add(tag_name)
-            tags.append(tag_name)
-    return first_info, first_asset, tags
-
-
-def fetch_asset_info_and_asset(
-    session: Session,
-    *,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> tuple[AssetInfo, Asset] | None:
-    stmt = (
-        select(AssetInfo, Asset)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .where(
-            AssetInfo.id == asset_info_id,
-            visible_owner_clause(owner_id),
-        )
-        .limit(1)
-        .options(noload(AssetInfo.tags))
-    )
-    row = session.execute(stmt)
-    pair = row.first()
-    if not pair:
-        return None
-    return pair[0], pair[1]
-
-def list_cache_states_by_asset_id(
-    session: Session, *, asset_id: str
-) -> Sequence[AssetCacheState]:
-    return (
-        session.execute(
-            select(AssetCacheState)
-            .where(AssetCacheState.asset_id == asset_id)
-            .order_by(AssetCacheState.id.asc())
-        )
-    ).scalars().all()
-
-
-def touch_asset_info_by_id(
-    session: Session,
-    *,
-    asset_info_id: str,
-    ts: datetime | None = None,
-    only_if_newer: bool = True,
-) -> None:
-    ts = ts or utcnow()
-    stmt = sa.update(AssetInfo).where(AssetInfo.id == asset_info_id)
-    if only_if_newer:
-        stmt = stmt.where(
-            sa.or_(AssetInfo.last_access_time.is_(None), AssetInfo.last_access_time < ts)
-        )
-    session.execute(stmt.values(last_access_time=ts))
-
-
-def create_asset_info_for_existing_asset(
-    session: Session,
-    *,
-    asset_hash: str,
-    name: str,
-    user_metadata: dict | None = None,
-    tags: Sequence[str] | None = None,
-    tag_origin: str = "manual",
-    owner_id: str = "",
-) -> AssetInfo:
-    """Create or return an existing AssetInfo for an Asset identified by asset_hash."""
-    now = utcnow()
-    asset = get_asset_by_hash(session, asset_hash=asset_hash)
-    if not asset:
-        raise ValueError(f"Unknown asset hash {asset_hash}")
-
-    info = AssetInfo(
-        owner_id=owner_id,
-        name=name,
-        asset_id=asset.id,
-        preview_id=None,
-        created_at=now,
-        updated_at=now,
-        last_access_time=now,
-    )
-    try:
-        with session.begin_nested():
-            session.add(info)
-            session.flush()
-    except IntegrityError:
-        existing = (
-            session.execute(
-                select(AssetInfo)
-                .options(noload(AssetInfo.tags))
-                .where(
-                    AssetInfo.asset_id == asset.id,
-                    AssetInfo.name == name,
-                    AssetInfo.owner_id == owner_id,
-                )
-                .limit(1)
-            )
-        ).unique().scalars().first()
-        if not existing:
-            raise RuntimeError("AssetInfo upsert failed to find existing row after conflict.")
-        return existing
-
-    # metadata["filename"] hack
-    new_meta = dict(user_metadata or {})
-    computed_filename = None
-    try:
-        p = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=asset.id))
-        if p:
-            computed_filename = compute_relative_filename(p)
-    except Exception:
-        computed_filename = None
-    if computed_filename:
-        new_meta["filename"] = computed_filename
-    if new_meta:
-        replace_asset_info_metadata_projection(
-            session,
-            asset_info_id=info.id,
-            user_metadata=new_meta,
-        )
-
-    if tags is not None:
-        set_asset_info_tags(
-            session,
-            asset_info_id=info.id,
-            tags=tags,
-            origin=tag_origin,
-        )
-    return info
-
-
-def set_asset_info_tags(
-    session: Session,
-    *,
-    asset_info_id: str,
-    tags: Sequence[str],
-    origin: str = "manual",
-) -> dict:
-    desired = normalize_tags(tags)
-
-    current = set(
-        tag_name for (tag_name,) in (
-            session.execute(select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id))
-        ).all()
-    )
-
-    to_add = [t for t in desired if t not in current]
-    to_remove = [t for t in current if t not in desired]
-
-    if to_add:
-        ensure_tags_exist(session, to_add, tag_type="user")
-        session.add_all([
-            AssetInfoTag(asset_info_id=asset_info_id, tag_name=t, origin=origin, added_at=utcnow())
-            for t in to_add
-        ])
-        session.flush()
-
-    if to_remove:
-        session.execute(
-            delete(AssetInfoTag)
-            .where(AssetInfoTag.asset_info_id == asset_info_id, AssetInfoTag.tag_name.in_(to_remove))
-        )
-        session.flush()
-
-    return {"added": to_add, "removed": to_remove, "total": desired}
-
-
-def replace_asset_info_metadata_projection(
-    session: Session,
-    *,
-    asset_info_id: str,
-    user_metadata: dict | None = None,
-) -> None:
-    info = session.get(AssetInfo, asset_info_id)
-    if not info:
-        raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    info.user_metadata = user_metadata or {}
-    info.updated_at = utcnow()
-    session.flush()
-
-    session.execute(delete(AssetInfoMeta).where(AssetInfoMeta.asset_info_id == asset_info_id))
-    session.flush()
-
-    if not user_metadata:
-        return
-
-    rows: list[AssetInfoMeta] = []
-    for k, v in user_metadata.items():
-        for r in project_kv(k, v):
-            rows.append(
-                AssetInfoMeta(
-                    asset_info_id=asset_info_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 ingest_fs_asset(
-    session: Session,
-    *,
-    asset_hash: str,
-    abs_path: str,
-    size_bytes: int,
-    mtime_ns: int,
-    mime_type: str | None = None,
-    info_name: str | None = None,
-    owner_id: str = "",
-    preview_id: str | None = None,
-    user_metadata: dict | None = None,
-    tags: Sequence[str] = (),
-    tag_origin: str = "manual",
-    require_existing_tags: bool = False,
-) -> dict:
-    """
-    Idempotently upsert:
-      - Asset by content hash (create if missing)
-      - AssetCacheState(file_path) pointing to asset_id
-      - Optionally AssetInfo + tag links and metadata projection
-    Returns flags and ids.
-    """
-    locator = os.path.abspath(abs_path)
-    now = utcnow()
-
-    if preview_id:
-        if not session.get(Asset, preview_id):
-            preview_id = None
-
-    out: dict[str, Any] = {
-        "asset_created": False,
-        "asset_updated": False,
-        "state_created": False,
-        "state_updated": False,
-        "asset_info_id": None,
-    }
-
-    # 1) Asset by hash
-    asset = (
-        session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
-    ).scalars().first()
-    if not asset:
-        vals = {
-            "hash": asset_hash,
-            "size_bytes": int(size_bytes),
-            "mime_type": mime_type,
-            "created_at": now,
-        }
-        res = session.execute(
-            sqlite.insert(Asset)
-            .values(**vals)
-            .on_conflict_do_nothing(index_elements=[Asset.hash])
-        )
-        if int(res.rowcount or 0) > 0:
-            out["asset_created"] = True
-        asset = (
-            session.execute(
-                select(Asset).where(Asset.hash == asset_hash).limit(1)
-            )
-        ).scalars().first()
-        if not asset:
-            raise RuntimeError("Asset row not found after upsert.")
-    else:
-        changed = False
-        if asset.size_bytes != int(size_bytes) and int(size_bytes) > 0:
-            asset.size_bytes = int(size_bytes)
-            changed = True
-        if mime_type and asset.mime_type != mime_type:
-            asset.mime_type = mime_type
-            changed = True
-        if changed:
-            out["asset_updated"] = True
-
-    # 2) AssetCacheState upsert by file_path (unique)
-    vals = {
-        "asset_id": asset.id,
-        "file_path": locator,
-        "mtime_ns": int(mtime_ns),
-    }
-    ins = (
-        sqlite.insert(AssetCacheState)
-        .values(**vals)
-        .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
-    )
-
-    res = session.execute(ins)
-    if int(res.rowcount or 0) > 0:
-        out["state_created"] = True
-    else:
-        upd = (
-            sa.update(AssetCacheState)
-            .where(AssetCacheState.file_path == locator)
-            .where(
-                sa.or_(
-                    AssetCacheState.asset_id != asset.id,
-                    AssetCacheState.mtime_ns.is_(None),
-                    AssetCacheState.mtime_ns != int(mtime_ns),
-                )
-            )
-            .values(asset_id=asset.id, mtime_ns=int(mtime_ns))
-        )
-        res2 = session.execute(upd)
-        if int(res2.rowcount or 0) > 0:
-            out["state_updated"] = True
-
-    # 3) Optional AssetInfo + tags + metadata
-    if info_name:
-        try:
-            with session.begin_nested():
-                info = AssetInfo(
-                    owner_id=owner_id,
-                    name=info_name,
-                    asset_id=asset.id,
-                    preview_id=preview_id,
-                    created_at=now,
-                    updated_at=now,
-                    last_access_time=now,
-                )
-                session.add(info)
-                session.flush()
-                out["asset_info_id"] = info.id
-        except IntegrityError:
-            pass
-
-        existing_info = (
-            session.execute(
-                select(AssetInfo)
-                .where(
-                    AssetInfo.asset_id == asset.id,
-                    AssetInfo.name == info_name,
-                    (AssetInfo.owner_id == owner_id),
-                )
-                .limit(1)
-            )
-        ).unique().scalar_one_or_none()
-        if not existing_info:
-            raise RuntimeError("Failed to update or insert AssetInfo.")
-
-        if preview_id and existing_info.preview_id != preview_id:
-            existing_info.preview_id = preview_id
-
-        existing_info.updated_at = now
-        if existing_info.last_access_time < now:
-            existing_info.last_access_time = now
-        session.flush()
-        out["asset_info_id"] = existing_info.id
-
-        norm = [t.strip().lower() for t in (tags or []) if (t or "").strip()]
-        if norm and out["asset_info_id"] is not None:
-            if not require_existing_tags:
-                ensure_tags_exist(session, norm, tag_type="user")
-
-            existing_tag_names = set(
-                name for (name,) in (session.execute(select(Tag.name).where(Tag.name.in_(norm)))).all()
-            )
-            missing = [t for t in norm if t not in existing_tag_names]
-            if missing and require_existing_tags:
-                raise ValueError(f"Unknown tags: {missing}")
-
-            existing_links = set(
-                tag_name
-                for (tag_name,) in (
-                    session.execute(
-                        select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == out["asset_info_id"])
-                    )
-                ).all()
-            )
-            to_add = [t for t in norm if t in existing_tag_names and t not in existing_links]
-            if to_add:
-                session.add_all(
-                    [
-                        AssetInfoTag(
-                            asset_info_id=out["asset_info_id"],
-                            tag_name=t,
-                            origin=tag_origin,
-                            added_at=now,
-                        )
-                        for t in to_add
-                    ]
-                )
-                session.flush()
-
-        # metadata["filename"] hack
-        if out["asset_info_id"] is not None:
-            primary_path = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=asset.id))
-            computed_filename = compute_relative_filename(primary_path) if primary_path else None
-
-            current_meta = existing_info.user_metadata or {}
-            new_meta = dict(current_meta)
-            if user_metadata is not None:
-                for k, v in user_metadata.items():
-                    new_meta[k] = v
-            if computed_filename:
-                new_meta["filename"] = computed_filename
-
-            if new_meta != current_meta:
-                replace_asset_info_metadata_projection(
-                    session,
-                    asset_info_id=out["asset_info_id"],
-                    user_metadata=new_meta,
-                )
-
-    try:
-        remove_missing_tag_for_asset_id(session, asset_id=asset.id)
-    except Exception:
-        logging.exception("Failed to clear 'missing' tag for asset %s", asset.id)
-    return out
-
-
-def update_asset_info_full(
-    session: Session,
-    *,
-    asset_info_id: str,
-    name: str | None = None,
-    tags: Sequence[str] | None = None,
-    user_metadata: dict | None = None,
-    tag_origin: str = "manual",
-    asset_info_row: Any = None,
-) -> AssetInfo:
-    if not asset_info_row:
-        info = session.get(AssetInfo, asset_info_id)
-        if not info:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-    else:
-        info = asset_info_row
-
-    touched = False
-    if name is not None and name != info.name:
-        info.name = name
-        touched = True
-
-    computed_filename = None
-    try:
-        p = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=info.asset_id))
-        if p:
-            computed_filename = compute_relative_filename(p)
-    except Exception:
-        computed_filename = None
-
-    if user_metadata is not None:
-        new_meta = dict(user_metadata)
-        if computed_filename:
-            new_meta["filename"] = computed_filename
-        replace_asset_info_metadata_projection(
-            session, asset_info_id=asset_info_id, user_metadata=new_meta
-        )
-        touched = True
-    else:
-        if computed_filename:
-            current_meta = info.user_metadata or {}
-            if current_meta.get("filename") != computed_filename:
-                new_meta = dict(current_meta)
-                new_meta["filename"] = computed_filename
-                replace_asset_info_metadata_projection(
-                    session, asset_info_id=asset_info_id, user_metadata=new_meta
-                )
-                touched = True
-
-    if tags is not None:
-        set_asset_info_tags(
-            session,
-            asset_info_id=asset_info_id,
-            tags=tags,
-            origin=tag_origin,
-        )
-        touched = True
-
-    if touched and user_metadata is None:
-        info.updated_at = utcnow()
-        session.flush()
-
-    return info
-
-
-def delete_asset_info_by_id(
-    session: Session,
-    *,
-    asset_info_id: str,
-    owner_id: str,
-) -> bool:
-    stmt = sa.delete(AssetInfo).where(
-        AssetInfo.id == asset_info_id,
-        visible_owner_clause(owner_id),
-    )
-    return int((session.execute(stmt)).rowcount or 0) > 0
-
-
-def list_tags_with_usage(
-    session: Session,
-    prefix: str | None = None,
-    limit: int = 100,
-    offset: int = 0,
-    include_zero: bool = True,
-    order: str = "count_desc",
-    owner_id: str = "",
-) -> tuple[list[tuple[str, str, int]], int]:
-    counts_sq = (
-        select(
-            AssetInfoTag.tag_name.label("tag_name"),
-            func.count(AssetInfoTag.asset_info_id).label("cnt"),
-        )
-        .select_from(AssetInfoTag)
-        .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id)
-        .where(visible_owner_clause(owner_id))
-        .group_by(AssetInfoTag.tag_name)
-        .subquery()
-    )
-
-    q = (
-        select(
-            Tag.name,
-            Tag.tag_type,
-            func.coalesce(counts_sq.c.cnt, 0).label("count"),
-        )
-        .select_from(Tag)
-        .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True)
-    )
-
-    if prefix:
-        escaped, esc = escape_like_prefix(prefix.strip().lower())
-        q = q.where(Tag.name.like(escaped + "%", escape=esc))
-
-    if not include_zero:
-        q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0)
-
-    if order == "name_asc":
-        q = q.order_by(Tag.name.asc())
-    else:
-        q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc())
-
-    total_q = select(func.count()).select_from(Tag)
-    if prefix:
-        escaped, esc = escape_like_prefix(prefix.strip().lower())
-        total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc))
-    if not include_zero:
-        total_q = total_q.where(
-            Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name))
-        )
-
-    rows = (session.execute(q.limit(limit).offset(offset))).all()
-    total = (session.execute(total_q)).scalar_one()
-
-    rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows]
-    return rows_norm, int(total or 0)
-
-
-def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None:
-    wanted = normalize_tags(list(names))
-    if not wanted:
-        return
-    rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))]
-    ins = (
-        sqlite.insert(Tag)
-        .values(rows)
-        .on_conflict_do_nothing(index_elements=[Tag.name])
-    )
-    session.execute(ins)
-
-
-def get_asset_tags(session: Session, *, asset_info_id: str) -> list[str]:
-    return [
-        tag_name for (tag_name,) in (
-            session.execute(
-                select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
-            )
-        ).all()
-    ]
-
-
-def add_tags_to_asset_info(
-    session: Session,
-    *,
-    asset_info_id: str,
-    tags: Sequence[str],
-    origin: str = "manual",
-    create_if_missing: bool = True,
-    asset_info_row: Any = None,
-) -> dict:
-    if not asset_info_row:
-        info = session.get(AssetInfo, asset_info_id)
-        if not info:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    norm = normalize_tags(tags)
-    if not norm:
-        total = get_asset_tags(session, asset_info_id=asset_info_id)
-        return {"added": [], "already_present": [], "total_tags": total}
-
-    if create_if_missing:
-        ensure_tags_exist(session, norm, tag_type="user")
-
-    current = {
-        tag_name
-        for (tag_name,) in (
-            session.execute(
-                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
-            )
-        ).all()
-    }
-
-    want = set(norm)
-    to_add = sorted(want - current)
-
-    if to_add:
-        with session.begin_nested() as nested:
-            try:
-                session.add_all(
-                    [
-                        AssetInfoTag(
-                            asset_info_id=asset_info_id,
-                            tag_name=t,
-                            origin=origin,
-                            added_at=utcnow(),
-                        )
-                        for t in to_add
-                    ]
-                )
-                session.flush()
-            except IntegrityError:
-                nested.rollback()
-
-    after = set(get_asset_tags(session, asset_info_id=asset_info_id))
-    return {
-        "added": sorted(((after - current) & want)),
-        "already_present": sorted(want & current),
-        "total_tags": sorted(after),
-    }
-
-
-def remove_tags_from_asset_info(
-    session: Session,
-    *,
-    asset_info_id: str,
-    tags: Sequence[str],
-) -> dict:
-    info = session.get(AssetInfo, asset_info_id)
-    if not info:
-        raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    norm = normalize_tags(tags)
-    if not norm:
-        total = get_asset_tags(session, asset_info_id=asset_info_id)
-        return {"removed": [], "not_present": [], "total_tags": total}
-
-    existing = {
-        tag_name
-        for (tag_name,) in (
-            session.execute(
-                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
-            )
-        ).all()
-    }
-
-    to_remove = sorted(set(t for t in norm if t in existing))
-    not_present = sorted(set(t for t in norm if t not in existing))
-
-    if to_remove:
-        session.execute(
-            delete(AssetInfoTag)
-            .where(
-                AssetInfoTag.asset_info_id == asset_info_id,
-                AssetInfoTag.tag_name.in_(to_remove),
-            )
-        )
-        session.flush()
-
-    total = get_asset_tags(session, asset_info_id=asset_info_id)
-    return {"removed": to_remove, "not_present": not_present, "total_tags": total}
-
-
-def remove_missing_tag_for_asset_id(
-    session: Session,
-    *,
-    asset_id: str,
-) -> None:
-    session.execute(
-        sa.delete(AssetInfoTag).where(
-            AssetInfoTag.asset_info_id.in_(sa.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)),
-            AssetInfoTag.tag_name == "missing",
-        )
-    )
-
-
-def set_asset_info_preview(
-    session: Session,
-    *,
-    asset_info_id: str,
-    preview_asset_id: str | None = None,
-) -> None:
-    """Set or clear preview_id and bump updated_at. Raises on unknown IDs."""
-    info = session.get(AssetInfo, asset_info_id)
-    if not info:
-        raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    if preview_asset_id is None:
-        info.preview_id = None
-    else:
-        # validate preview asset exists
-        if not session.get(Asset, preview_asset_id):
-            raise ValueError(f"Preview Asset {preview_asset_id} not found")
-        info.preview_id = preview_asset_id
-
-    info.updated_at = utcnow()
-    session.flush()

app/assets/database/queries/__init__.py

@@ -0,0 +1,99 @@
+from app.assets.database.queries.asset import (
+    asset_exists_by_hash,
+    bulk_insert_assets,
+    get_asset_by_hash,
+    upsert_asset,
+)
+from app.assets.database.queries.asset_info import (
+    asset_info_exists_for_asset_id,
+    bulk_insert_asset_infos_ignore_conflicts,
+    delete_asset_info_by_id,
+    fetch_asset_info_and_asset,
+    fetch_asset_info_asset_and_tags,
+    get_asset_info_by_id,
+    get_asset_info_ids_by_ids,
+    get_or_create_asset_info,
+    insert_asset_info,
+    list_asset_infos_page,
+    set_asset_info_metadata,
+    set_asset_info_preview,
+    update_asset_info_access_time,
+    update_asset_info_name,
+    update_asset_info_timestamps,
+    update_asset_info_updated_at,
+)
+from app.assets.database.queries.cache_state import (
+    CacheStateRow,
+    bulk_insert_cache_states_ignore_conflicts,
+    bulk_set_needs_verify,
+    delete_assets_by_ids,
+    delete_cache_states_by_ids,
+    delete_cache_states_outside_prefixes,
+    delete_orphaned_seed_asset,
+    get_cache_states_by_paths_and_asset_ids,
+    get_cache_states_for_prefixes,
+    get_orphaned_seed_asset_ids,
+    list_cache_states_by_asset_id,
+    upsert_cache_state,
+)
+from app.assets.database.queries.tags import (
+    AddTagsDict,
+    RemoveTagsDict,
+    SetTagsDict,
+    add_missing_tag_for_asset_id,
+    add_tags_to_asset_info,
+    bulk_insert_tags_and_meta,
+    ensure_tags_exist,
+    get_asset_tags,
+    list_tags_with_usage,
+    remove_missing_tag_for_asset_id,
+    remove_tags_from_asset_info,
+    set_asset_info_tags,
+)
+
+__all__ = [
+    "AddTagsDict",
+    "CacheStateRow",
+    "RemoveTagsDict",
+    "SetTagsDict",
+    "add_missing_tag_for_asset_id",
+    "add_tags_to_asset_info",
+    "asset_exists_by_hash",
+    "asset_info_exists_for_asset_id",
+    "bulk_insert_asset_infos_ignore_conflicts",
+    "bulk_insert_assets",
+    "bulk_insert_cache_states_ignore_conflicts",
+    "bulk_insert_tags_and_meta",
+    "bulk_set_needs_verify",
+    "delete_asset_info_by_id",
+    "delete_assets_by_ids",
+    "delete_cache_states_by_ids",
+    "delete_cache_states_outside_prefixes",
+    "delete_orphaned_seed_asset",
+    "ensure_tags_exist",
+    "fetch_asset_info_and_asset",
+    "fetch_asset_info_asset_and_tags",
+    "get_asset_by_hash",
+    "get_asset_info_by_id",
+    "get_asset_info_ids_by_ids",
+    "get_asset_tags",
+    "get_cache_states_by_paths_and_asset_ids",
+    "get_cache_states_for_prefixes",
+    "get_or_create_asset_info",
+    "get_orphaned_seed_asset_ids",
+    "insert_asset_info",
+    "list_asset_infos_page",
+    "list_cache_states_by_asset_id",
+    "list_tags_with_usage",
+    "remove_missing_tag_for_asset_id",
+    "remove_tags_from_asset_info",
+    "set_asset_info_metadata",
+    "set_asset_info_preview",
+    "set_asset_info_tags",
+    "update_asset_info_access_time",
+    "update_asset_info_name",
+    "update_asset_info_timestamps",
+    "update_asset_info_updated_at",
+    "upsert_asset",
+    "upsert_cache_state",
+]

app/assets/database/queries/asset.py

@@ -0,0 +1,90 @@
+import sqlalchemy as sa
+from sqlalchemy import select
+from sqlalchemy.dialects import sqlite
+from sqlalchemy.orm import Session
+
+from app.assets.database.models import Asset
+from app.assets.database.queries.common import calculate_rows_per_statement, iter_chunks
+
+
+def asset_exists_by_hash(
+    session: Session,
+    asset_hash: str,
+) -> bool:
+    """
+    Check if an asset with a given hash exists in database.
+    """
+    row = (
+        session.execute(
+            select(sa.literal(True))
+            .select_from(Asset)
+            .where(Asset.hash == asset_hash)
+            .limit(1)
+        )
+    ).first()
+    return row is not None
+
+
+def get_asset_by_hash(
+    session: Session,
+    asset_hash: str,
+) -> Asset | None:
+    return (
+        (session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1)))
+        .scalars()
+        .first()
+    )
+
+
+def upsert_asset(
+    session: Session,
+    asset_hash: str,
+    size_bytes: int,
+    mime_type: str | None = None,
+) -> tuple[Asset, bool, bool]:
+    """Upsert an Asset by hash. Returns (asset, created, updated)."""
+    vals = {"hash": asset_hash, "size_bytes": int(size_bytes)}
+    if mime_type:
+        vals["mime_type"] = mime_type
+
+    ins = (
+        sqlite.insert(Asset)
+        .values(**vals)
+        .on_conflict_do_nothing(index_elements=[Asset.hash])
+    )
+    res = session.execute(ins)
+    created = int(res.rowcount or 0) > 0
+
+    asset = (
+        session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
+        .scalars()
+        .first()
+    )
+    if not asset:
+        raise RuntimeError("Asset row not found after upsert.")
+
+    updated = False
+    if not created:
+        changed = False
+        if asset.size_bytes != int(size_bytes) and int(size_bytes) > 0:
+            asset.size_bytes = int(size_bytes)
+            changed = True
+        if mime_type and asset.mime_type != mime_type:
+            asset.mime_type = mime_type
+            changed = True
+        if changed:
+            updated = True
+
+    return asset, created, updated
+
+
+def bulk_insert_assets(
+    session: Session,
+    rows: list[dict],
+) -> None:
+    """Bulk insert Asset rows. Each dict should have: id, hash, size_bytes, mime_type, created_at."""
+    if not rows:
+        return
+    ins = sqlite.insert(Asset)
+    for chunk in iter_chunks(rows, calculate_rows_per_statement(5)):
+        session.execute(ins, chunk)

app/assets/database/queries/asset_info.py

@@ -0,0 +1,527 @@
+from collections import defaultdict
+from datetime import datetime
+from decimal import Decimal
+from typing import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import delete, exists, select
+from sqlalchemy.dialects import sqlite
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.orm import Session, contains_eager, noload
+
+from app.assets.database.models import (
+    Asset,
+    AssetInfo,
+    AssetInfoMeta,
+    AssetInfoTag,
+    Tag,
+)
+from app.assets.database.queries.common import (
+    MAX_BIND_PARAMS,
+    build_visible_owner_clause,
+    calculate_rows_per_statement,
+    iter_chunks,
+)
+from app.assets.helpers import escape_sql_like_string, get_utc_now, normalize_tags
+
+
+def _check_is_scalar(v):
+    if v is None:
+        return True
+    if isinstance(v, bool):
+        return True
+    if isinstance(v, (int, float, Decimal, str)):
+        return True
+    return False
+
+
+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)):
+        num = value if isinstance(value, Decimal) else Decimal(str(value))
+        return {"key": key, "ordinal": ordinal, "val_num": num}
+    if isinstance(value, str):
+        return {"key": key, "ordinal": ordinal, "val_str": value}
+    return {"key": key, "ordinal": ordinal, "val_json": value}
+
+
+def convert_metadata_to_rows(key: str, value) -> list[dict]:
+    """
+    Turn a metadata key/value into typed projection rows.
+    Returns list[dict] with keys:
+      key, ordinal, and one of val_str / val_num / val_bool / val_json (others None)
+    """
+    if value is None:
+        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)]
+        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(
+                    (AssetInfoTag.asset_info_id == AssetInfo.id)
+                    & (AssetInfoTag.tag_name == tag_name)
+                )
+            )
+
+    if exclude_tags:
+        stmt = stmt.where(
+            ~exists().where(
+                (AssetInfoTag.asset_info_id == AssetInfo.id)
+                & (AssetInfoTag.tag_name.in_(exclude_tags))
+            )
+        )
+    return stmt
+
+
+def _apply_metadata_filter(
+    stmt: sa.sql.Select,
+    metadata_filter: dict | None = None,
+) -> sa.sql.Select:
+    """Apply filters using asset_info_meta projection table."""
+    if not metadata_filter:
+        return stmt
+
+    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
+        return sa.exists().where(
+            AssetInfoMeta.asset_info_id == AssetInfo.id,
+            AssetInfoMeta.key == key,
+            *preds,
+        )
+
+    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
+        if value is None:
+            no_row_for_key = sa.not_(
+                sa.exists().where(
+                    AssetInfoMeta.asset_info_id == AssetInfo.id,
+                    AssetInfoMeta.key == key,
+                )
+            )
+            null_row = _exists_for_pred(
+                key,
+                AssetInfoMeta.val_json.is_(None),
+                AssetInfoMeta.val_str.is_(None),
+                AssetInfoMeta.val_num.is_(None),
+                AssetInfoMeta.val_bool.is_(None),
+            )
+            return sa.or_(no_row_for_key, null_row)
+
+        if isinstance(value, bool):
+            return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value))
+        if isinstance(value, (int, float)):
+            num = value if isinstance(value, Decimal) else Decimal(str(value))
+            return _exists_for_pred(key, AssetInfoMeta.val_num == num)
+        if isinstance(value, str):
+            return _exists_for_pred(key, AssetInfoMeta.val_str == value)
+        return _exists_for_pred(key, AssetInfoMeta.val_json == value)
+
+    for k, v in metadata_filter.items():
+        if isinstance(v, list):
+            ors = [_exists_clause_for_value(k, elem) for elem in v]
+            if ors:
+                stmt = stmt.where(sa.or_(*ors))
+        else:
+            stmt = stmt.where(_exists_clause_for_value(k, v))
+    return stmt
+
+
+def asset_info_exists_for_asset_id(
+    session: Session,
+    asset_id: str,
+) -> bool:
+    q = (
+        select(sa.literal(True))
+        .select_from(AssetInfo)
+        .where(AssetInfo.asset_id == asset_id)
+        .limit(1)
+    )
+    return (session.execute(q)).first() is not None
+
+
+def get_asset_info_by_id(
+    session: Session,
+    asset_info_id: str,
+) -> AssetInfo | None:
+    return session.get(AssetInfo, asset_info_id)
+
+
+def insert_asset_info(
+    session: Session,
+    asset_id: str,
+    owner_id: str,
+    name: str,
+    preview_id: str | None = None,
+) -> AssetInfo | None:
+    """Insert a new AssetInfo. Returns None if unique constraint violated."""
+    now = get_utc_now()
+    try:
+        with session.begin_nested():
+            info = AssetInfo(
+                owner_id=owner_id,
+                name=name,
+                asset_id=asset_id,
+                preview_id=preview_id,
+                created_at=now,
+                updated_at=now,
+                last_access_time=now,
+            )
+            session.add(info)
+            session.flush()
+            return info
+    except IntegrityError:
+        return None
+
+
+def get_or_create_asset_info(
+    session: Session,
+    asset_id: str,
+    owner_id: str,
+    name: str,
+    preview_id: str | None = None,
+) -> tuple[AssetInfo, bool]:
+    """Get existing or create new AssetInfo. Returns (info, created)."""
+    info = insert_asset_info(
+        session,
+        asset_id=asset_id,
+        owner_id=owner_id,
+        name=name,
+        preview_id=preview_id,
+    )
+    if info:
+        return info, True
+
+    existing = (
+        session.execute(
+            select(AssetInfo)
+            .where(
+                AssetInfo.asset_id == asset_id,
+                AssetInfo.name == name,
+                AssetInfo.owner_id == owner_id,
+            )
+            .limit(1)
+        )
+        .unique()
+        .scalar_one_or_none()
+    )
+    if not existing:
+        raise RuntimeError("Failed to find AssetInfo after insert conflict.")
+    return existing, False
+
+
+def update_asset_info_timestamps(
+    session: Session,
+    asset_info: AssetInfo,
+    preview_id: str | None = None,
+) -> None:
+    """Update timestamps and optionally preview_id on existing AssetInfo."""
+    now = get_utc_now()
+    if preview_id and asset_info.preview_id != preview_id:
+        asset_info.preview_id = preview_id
+    asset_info.updated_at = now
+    if asset_info.last_access_time < now:
+        asset_info.last_access_time = now
+    session.flush()
+
+
+def list_asset_infos_page(
+    session: Session,
+    owner_id: str = "",
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+    name_contains: str | None = None,
+    metadata_filter: dict | None = None,
+    limit: int = 20,
+    offset: int = 0,
+    sort: str = "created_at",
+    order: str = "desc",
+) -> tuple[list[AssetInfo], dict[str, list[str]], int]:
+    base = (
+        select(AssetInfo)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags))
+        .where(build_visible_owner_clause(owner_id))
+    )
+
+    if name_contains:
+        escaped, esc = escape_sql_like_string(name_contains)
+        base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
+
+    base = _apply_tag_filters(base, include_tags, exclude_tags)
+    base = _apply_metadata_filter(base, metadata_filter)
+
+    sort = (sort or "created_at").lower()
+    order = (order or "desc").lower()
+    sort_map = {
+        "name": AssetInfo.name,
+        "created_at": AssetInfo.created_at,
+        "updated_at": AssetInfo.updated_at,
+        "last_access_time": AssetInfo.last_access_time,
+        "size": Asset.size_bytes,
+    }
+    sort_col = sort_map.get(sort, AssetInfo.created_at)
+    sort_exp = sort_col.desc() if order == "desc" else sort_col.asc()
+
+    base = base.order_by(sort_exp).limit(limit).offset(offset)
+
+    count_stmt = (
+        select(sa.func.count())
+        .select_from(AssetInfo)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .where(build_visible_owner_clause(owner_id))
+    )
+    if name_contains:
+        escaped, esc = escape_sql_like_string(name_contains)
+        count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
+    count_stmt = _apply_tag_filters(count_stmt, include_tags, exclude_tags)
+    count_stmt = _apply_metadata_filter(count_stmt, metadata_filter)
+
+    total = int((session.execute(count_stmt)).scalar_one() or 0)
+
+    infos = (session.execute(base)).unique().scalars().all()
+
+    id_list: list[str] = [i.id for i in infos]
+    tag_map: dict[str, list[str]] = defaultdict(list)
+    if id_list:
+        rows = session.execute(
+            select(AssetInfoTag.asset_info_id, Tag.name)
+            .join(Tag, Tag.name == AssetInfoTag.tag_name)
+            .where(AssetInfoTag.asset_info_id.in_(id_list))
+            .order_by(AssetInfoTag.added_at)
+        )
+        for aid, tag_name in rows.all():
+            tag_map[aid].append(tag_name)
+
+    return infos, tag_map, total
+
+
+def fetch_asset_info_asset_and_tags(
+    session: Session,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> tuple[AssetInfo, Asset, list[str]] | None:
+    stmt = (
+        select(AssetInfo, Asset, Tag.name)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True)
+        .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True)
+        .where(
+            AssetInfo.id == asset_info_id,
+            build_visible_owner_clause(owner_id),
+        )
+        .options(noload(AssetInfo.tags))
+        .order_by(Tag.name.asc())
+    )
+
+    rows = (session.execute(stmt)).all()
+    if not rows:
+        return None
+
+    first_info, first_asset, _ = rows[0]
+    tags: list[str] = []
+    seen: set[str] = set()
+    for _info, _asset, tag_name in rows:
+        if tag_name and tag_name not in seen:
+            seen.add(tag_name)
+            tags.append(tag_name)
+    return first_info, first_asset, tags
+
+
+def fetch_asset_info_and_asset(
+    session: Session,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> tuple[AssetInfo, Asset] | None:
+    stmt = (
+        select(AssetInfo, Asset)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .where(
+            AssetInfo.id == asset_info_id,
+            build_visible_owner_clause(owner_id),
+        )
+        .limit(1)
+        .options(noload(AssetInfo.tags))
+    )
+    row = session.execute(stmt)
+    pair = row.first()
+    if not pair:
+        return None
+    return pair[0], pair[1]
+
+
+def update_asset_info_access_time(
+    session: Session,
+    asset_info_id: str,
+    ts: datetime | None = None,
+    only_if_newer: bool = True,
+) -> None:
+    ts = ts or get_utc_now()
+    stmt = sa.update(AssetInfo).where(AssetInfo.id == asset_info_id)
+    if only_if_newer:
+        stmt = stmt.where(
+            sa.or_(
+                AssetInfo.last_access_time.is_(None), AssetInfo.last_access_time < ts
+            )
+        )
+    session.execute(stmt.values(last_access_time=ts))
+
+
+def update_asset_info_name(
+    session: Session,
+    asset_info_id: str,
+    name: str,
+) -> None:
+    """Update the name of an AssetInfo."""
+    now = get_utc_now()
+    session.execute(
+        sa.update(AssetInfo)
+        .where(AssetInfo.id == asset_info_id)
+        .values(name=name, updated_at=now)
+    )
+
+
+def update_asset_info_updated_at(
+    session: Session,
+    asset_info_id: str,
+    ts: datetime | None = None,
+) -> None:
+    """Update the updated_at timestamp of an AssetInfo."""
+    ts = ts or get_utc_now()
+    session.execute(
+        sa.update(AssetInfo).where(AssetInfo.id == asset_info_id).values(updated_at=ts)
+    )
+
+
+def set_asset_info_metadata(
+    session: Session,
+    asset_info_id: str,
+    user_metadata: dict | None = None,
+) -> None:
+    info = session.get(AssetInfo, asset_info_id)
+    if not info:
+        raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    info.user_metadata = user_metadata or {}
+    info.updated_at = get_utc_now()
+    session.flush()
+
+    session.execute(
+        delete(AssetInfoMeta).where(AssetInfoMeta.asset_info_id == asset_info_id)
+    )
+    session.flush()
+
+    if not user_metadata:
+        return
+
+    rows: list[AssetInfoMeta] = []
+    for k, v in user_metadata.items():
+        for r in convert_metadata_to_rows(k, v):
+            rows.append(
+                AssetInfoMeta(
+                    asset_info_id=asset_info_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_asset_info_by_id(
+    session: Session,
+    asset_info_id: str,
+    owner_id: str,
+) -> bool:
+    stmt = sa.delete(AssetInfo).where(
+        AssetInfo.id == asset_info_id,
+        build_visible_owner_clause(owner_id),
+    )
+    return int((session.execute(stmt)).rowcount or 0) > 0
+
+
+def set_asset_info_preview(
+    session: Session,
+    asset_info_id: str,
+    preview_asset_id: str | None = None,
+) -> None:
+    """Set or clear preview_id and bump updated_at. Raises on unknown IDs."""
+    info = session.get(AssetInfo, asset_info_id)
+    if not info:
+        raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    if preview_asset_id is None:
+        info.preview_id = None
+    else:
+        if not session.get(Asset, preview_asset_id):
+            raise ValueError(f"Preview Asset {preview_asset_id} not found")
+        info.preview_id = preview_asset_id
+
+    info.updated_at = get_utc_now()
+    session.flush()
+
+
+def bulk_insert_asset_infos_ignore_conflicts(
+    session: Session,
+    rows: list[dict],
+) -> None:
+    """Bulk insert AssetInfo rows with ON CONFLICT DO NOTHING.
+
+    Each dict should have: id, owner_id, name, asset_id, preview_id,
+    user_metadata, created_at, updated_at, last_access_time
+    """
+    if not rows:
+        return
+    ins = sqlite.insert(AssetInfo).on_conflict_do_nothing(
+        index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name]
+    )
+    for chunk in iter_chunks(rows, calculate_rows_per_statement(9)):
+        session.execute(ins, chunk)
+
+
+def get_asset_info_ids_by_ids(
+    session: Session,
+    info_ids: list[str],
+) -> set[str]:
+    """Query to find which AssetInfo IDs exist in the database."""
+    if not info_ids:
+        return set()
+
+    found: set[str] = set()
+    for chunk in iter_chunks(info_ids, MAX_BIND_PARAMS):
+        result = session.execute(select(AssetInfo.id).where(AssetInfo.id.in_(chunk)))
+        found.update(result.scalars().all())
+    return found

app/assets/database/queries/cache_state.py

@@ -0,0 +1,280 @@
+import os
+from typing import NamedTuple, Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import select
+from sqlalchemy.dialects import sqlite
+from sqlalchemy.orm import Session
+
+from app.assets.database.models import Asset, AssetCacheState, AssetInfo
+from app.assets.database.queries.common import (
+    MAX_BIND_PARAMS,
+    calculate_rows_per_statement,
+    iter_chunks,
+)
+from app.assets.helpers import escape_sql_like_string
+
+
+class CacheStateRow(NamedTuple):
+    """Row from cache state query with joined asset data."""
+
+    state_id: int
+    file_path: str
+    mtime_ns: int | None
+    needs_verify: bool
+    asset_id: str
+    asset_hash: str | None
+    size_bytes: int
+
+
+def list_cache_states_by_asset_id(
+    session: Session, *, asset_id: str
+) -> Sequence[AssetCacheState]:
+    return (
+        (
+            session.execute(
+                select(AssetCacheState)
+                .where(AssetCacheState.asset_id == asset_id)
+                .order_by(AssetCacheState.id.asc())
+            )
+        )
+        .scalars()
+        .all()
+    )
+
+
+def upsert_cache_state(
+    session: Session,
+    asset_id: str,
+    file_path: str,
+    mtime_ns: int,
+) -> tuple[bool, bool]:
+    """Upsert a cache state by file_path. Returns (created, updated)."""
+    vals = {
+        "asset_id": asset_id,
+        "file_path": file_path,
+        "mtime_ns": int(mtime_ns),
+    }
+    ins = (
+        sqlite.insert(AssetCacheState)
+        .values(**vals)
+        .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
+    )
+    res = session.execute(ins)
+    created = int(res.rowcount or 0) > 0
+
+    if created:
+        return True, False
+
+    upd = (
+        sa.update(AssetCacheState)
+        .where(AssetCacheState.file_path == file_path)
+        .where(
+            sa.or_(
+                AssetCacheState.asset_id != asset_id,
+                AssetCacheState.mtime_ns.is_(None),
+                AssetCacheState.mtime_ns != int(mtime_ns),
+            )
+        )
+        .values(asset_id=asset_id, mtime_ns=int(mtime_ns))
+    )
+    res2 = session.execute(upd)
+    updated = int(res2.rowcount or 0) > 0
+    return False, updated
+
+
+def delete_cache_states_outside_prefixes(
+    session: Session, valid_prefixes: list[str]
+) -> int:
+    """Delete cache states with file_path not matching any of the valid prefixes.
+
+    Args:
+        session: Database session
+        valid_prefixes: List of absolute directory prefixes that are valid
+
+    Returns:
+        Number of cache states deleted
+    """
+    if not valid_prefixes:
+        return 0
+
+    def make_prefix_condition(prefix: str):
+        base = prefix if prefix.endswith(os.sep) else prefix + os.sep
+        escaped, esc = escape_sql_like_string(base)
+        return AssetCacheState.file_path.like(escaped + "%", escape=esc)
+
+    matches_valid_prefix = sa.or_(*[make_prefix_condition(p) for p in valid_prefixes])
+    result = session.execute(sa.delete(AssetCacheState).where(~matches_valid_prefix))
+    return result.rowcount
+
+
+def get_orphaned_seed_asset_ids(session: Session) -> list[str]:
+    """Get IDs of seed assets (hash is None) with no remaining cache states.
+
+    Returns:
+        List of asset IDs that are orphaned
+    """
+    orphan_subq = (
+        sa.select(Asset.id)
+        .outerjoin(AssetCacheState, AssetCacheState.asset_id == Asset.id)
+        .where(Asset.hash.is_(None), AssetCacheState.id.is_(None))
+    )
+    return [row[0] for row in session.execute(orphan_subq).all()]
+
+
+def delete_assets_by_ids(session: Session, asset_ids: list[str]) -> int:
+    """Delete assets and their AssetInfos by ID.
+
+    Args:
+        session: Database session
+        asset_ids: List of asset IDs to delete
+
+    Returns:
+        Number of assets deleted
+    """
+    if not asset_ids:
+        return 0
+    session.execute(sa.delete(AssetInfo).where(AssetInfo.asset_id.in_(asset_ids)))
+    result = session.execute(sa.delete(Asset).where(Asset.id.in_(asset_ids)))
+    return result.rowcount
+
+
+def get_cache_states_for_prefixes(
+    session: Session,
+    prefixes: list[str],
+) -> list[CacheStateRow]:
+    """Get all cache states with paths matching any of the given prefixes.
+
+    Args:
+        session: Database session
+        prefixes: List of absolute directory prefixes to match
+
+    Returns:
+        List of cache state rows with joined asset data, ordered by asset_id, state_id
+    """
+    if not prefixes:
+        return []
+
+    conds = []
+    for p in prefixes:
+        base = os.path.abspath(p)
+        if not base.endswith(os.sep):
+            base += os.sep
+        escaped, esc = escape_sql_like_string(base)
+        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
+
+    rows = session.execute(
+        sa.select(
+            AssetCacheState.id,
+            AssetCacheState.file_path,
+            AssetCacheState.mtime_ns,
+            AssetCacheState.needs_verify,
+            AssetCacheState.asset_id,
+            Asset.hash,
+            Asset.size_bytes,
+        )
+        .join(Asset, Asset.id == AssetCacheState.asset_id)
+        .where(sa.or_(*conds))
+        .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
+    ).all()
+
+    return [
+        CacheStateRow(
+            state_id=row[0],
+            file_path=row[1],
+            mtime_ns=row[2],
+            needs_verify=row[3],
+            asset_id=row[4],
+            asset_hash=row[5],
+            size_bytes=int(row[6] or 0),
+        )
+        for row in rows
+    ]
+
+
+def bulk_set_needs_verify(session: Session, state_ids: list[int], value: bool) -> int:
+    """Set needs_verify flag for multiple cache states.
+
+    Returns: Number of rows updated
+    """
+    if not state_ids:
+        return 0
+    result = session.execute(
+        sa.update(AssetCacheState)
+        .where(AssetCacheState.id.in_(state_ids))
+        .values(needs_verify=value)
+    )
+    return result.rowcount
+
+
+def delete_cache_states_by_ids(session: Session, state_ids: list[int]) -> int:
+    """Delete cache states by their IDs.
+
+    Returns: Number of rows deleted
+    """
+    if not state_ids:
+        return 0
+    result = session.execute(
+        sa.delete(AssetCacheState).where(AssetCacheState.id.in_(state_ids))
+    )
+    return result.rowcount
+
+
+def delete_orphaned_seed_asset(session: Session, asset_id: str) -> bool:
+    """Delete a seed asset (hash is None) and its AssetInfos.
+
+    Returns: True if asset was deleted, False if not found
+    """
+    session.execute(sa.delete(AssetInfo).where(AssetInfo.asset_id == asset_id))
+    asset = session.get(Asset, asset_id)
+    if asset:
+        session.delete(asset)
+        return True
+    return False
+
+
+def bulk_insert_cache_states_ignore_conflicts(
+    session: Session,
+    rows: list[dict],
+) -> None:
+    """Bulk insert cache state rows with ON CONFLICT DO NOTHING on file_path.
+
+    Each dict should have: asset_id, file_path, mtime_ns
+    """
+    if not rows:
+        return
+    ins = sqlite.insert(AssetCacheState).on_conflict_do_nothing(
+        index_elements=[AssetCacheState.file_path]
+    )
+    for chunk in iter_chunks(rows, calculate_rows_per_statement(3)):
+        session.execute(ins, chunk)
+
+
+def get_cache_states_by_paths_and_asset_ids(
+    session: Session,
+    path_to_asset: dict[str, str],
+) -> set[str]:
+    """Query cache states to find paths where our asset_id won the insert.
+
+    Args:
+        path_to_asset: Mapping of file_path -> asset_id we tried to insert
+
+    Returns:
+        Set of file_paths where our asset_id is present
+    """
+    if not path_to_asset:
+        return set()
+
+    paths = list(path_to_asset.keys())
+    winners: set[str] = set()
+
+    for chunk in iter_chunks(paths, MAX_BIND_PARAMS):
+        result = session.execute(
+            select(AssetCacheState.file_path).where(
+                AssetCacheState.file_path.in_(chunk),
+                AssetCacheState.asset_id.in_([path_to_asset[p] for p in chunk]),
+            )
+        )
+        winners.update(result.scalars().all())
+
+    return winners

app/assets/database/queries/common.py

@@ -0,0 +1,37 @@
+"""Shared utilities for database query modules."""
+
+from typing import Iterable
+
+import sqlalchemy as sa
+
+from app.assets.database.models import AssetInfo
+
+MAX_BIND_PARAMS = 800
+
+
+def calculate_rows_per_statement(cols: int) -> int:
+    """Calculate how many rows can fit in one statement given column count."""
+    return max(1, MAX_BIND_PARAMS // max(1, cols))
+
+
+def iter_chunks(seq, n: int):
+    """Yield successive n-sized chunks from seq."""
+    for i in range(0, len(seq), n):
+        yield seq[i : i + n]
+
+
+def iter_row_chunks(rows: list[dict], cols_per_row: int) -> Iterable[list[dict]]:
+    """Yield chunks of rows sized to fit within bind param limits."""
+    if not rows:
+        return []
+    rows_per_stmt = max(1, MAX_BIND_PARAMS // max(1, cols_per_row))
+    for i in range(0, len(rows), rows_per_stmt):
+        yield rows[i : i + rows_per_stmt]
+
+
+def build_visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement:
+    """Build owner visibility predicate for reads. Owner-less rows are visible to everyone."""
+    owner_id = (owner_id or "").strip()
+    if owner_id == "":
+        return AssetInfo.owner_id == ""
+    return AssetInfo.owner_id.in_(["", owner_id])

app/assets/database/queries/tags.py

@@ -0,0 +1,349 @@
+from typing import Iterable, Sequence, TypedDict
+
+import sqlalchemy as sa
+from sqlalchemy import delete, func, select
+from sqlalchemy.dialects import sqlite
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.orm import Session
+
+from app.assets.database.models import AssetInfo, AssetInfoMeta, AssetInfoTag, Tag
+from app.assets.database.queries.common import (
+    build_visible_owner_clause,
+    iter_row_chunks,
+)
+from app.assets.helpers import escape_sql_like_string, get_utc_now, normalize_tags
+
+
+class AddTagsDict(TypedDict):
+    added: list[str]
+    already_present: list[str]
+    total_tags: list[str]
+
+
+class RemoveTagsDict(TypedDict):
+    removed: list[str]
+    not_present: list[str]
+    total_tags: list[str]
+
+
+class SetTagsDict(TypedDict):
+    added: list[str]
+    removed: list[str]
+    total: list[str]
+
+
+def ensure_tags_exist(
+    session: Session, names: Iterable[str], tag_type: str = "user"
+) -> None:
+    wanted = normalize_tags(list(names))
+    if not wanted:
+        return
+    rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))]
+    ins = (
+        sqlite.insert(Tag)
+        .values(rows)
+        .on_conflict_do_nothing(index_elements=[Tag.name])
+    )
+    session.execute(ins)
+
+
+def get_asset_tags(session: Session, asset_info_id: str) -> list[str]:
+    return [
+        tag_name
+        for (tag_name,) in (
+            session.execute(
+                select(AssetInfoTag.tag_name).where(
+                    AssetInfoTag.asset_info_id == asset_info_id
+                )
+            )
+        ).all()
+    ]
+
+
+def set_asset_info_tags(
+    session: Session,
+    asset_info_id: str,
+    tags: Sequence[str],
+    origin: str = "manual",
+) -> SetTagsDict:
+    desired = normalize_tags(tags)
+
+    current = set(
+        tag_name
+        for (tag_name,) in (
+            session.execute(
+                select(AssetInfoTag.tag_name).where(
+                    AssetInfoTag.asset_info_id == asset_info_id
+                )
+            )
+        ).all()
+    )
+
+    to_add = [t for t in desired if t not in current]
+    to_remove = [t for t in current if t not in desired]
+
+    if to_add:
+        ensure_tags_exist(session, to_add, tag_type="user")
+        session.add_all(
+            [
+                AssetInfoTag(
+                    asset_info_id=asset_info_id,
+                    tag_name=t,
+                    origin=origin,
+                    added_at=get_utc_now(),
+                )
+                for t in to_add
+            ]
+        )
+        session.flush()
+
+    if to_remove:
+        session.execute(
+            delete(AssetInfoTag).where(
+                AssetInfoTag.asset_info_id == asset_info_id,
+                AssetInfoTag.tag_name.in_(to_remove),
+            )
+        )
+        session.flush()
+
+    return {"added": to_add, "removed": to_remove, "total": desired}
+
+
+def add_tags_to_asset_info(
+    session: Session,
+    asset_info_id: str,
+    tags: Sequence[str],
+    origin: str = "manual",
+    create_if_missing: bool = True,
+    asset_info_row: AssetInfo | None = None,
+) -> AddTagsDict:
+    if not asset_info_row:
+        info = session.get(AssetInfo, asset_info_id)
+        if not info:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    norm = normalize_tags(tags)
+    if not norm:
+        total = get_asset_tags(session, asset_info_id=asset_info_id)
+        return {"added": [], "already_present": [], "total_tags": total}
+
+    if create_if_missing:
+        ensure_tags_exist(session, norm, tag_type="user")
+
+    current = {
+        tag_name
+        for (tag_name,) in (
+            session.execute(
+                sa.select(AssetInfoTag.tag_name).where(
+                    AssetInfoTag.asset_info_id == asset_info_id
+                )
+            )
+        ).all()
+    }
+
+    want = set(norm)
+    to_add = sorted(want - current)
+
+    if to_add:
+        with session.begin_nested() as nested:
+            try:
+                session.add_all(
+                    [
+                        AssetInfoTag(
+                            asset_info_id=asset_info_id,
+                            tag_name=t,
+                            origin=origin,
+                            added_at=get_utc_now(),
+                        )
+                        for t in to_add
+                    ]
+                )
+                session.flush()
+            except IntegrityError:
+                nested.rollback()
+
+    after = set(get_asset_tags(session, asset_info_id=asset_info_id))
+    return {
+        "added": sorted(((after - current) & want)),
+        "already_present": sorted(want & current),
+        "total_tags": sorted(after),
+    }
+
+
+def remove_tags_from_asset_info(
+    session: Session,
+    asset_info_id: str,
+    tags: Sequence[str],
+) -> RemoveTagsDict:
+    info = session.get(AssetInfo, asset_info_id)
+    if not info:
+        raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    norm = normalize_tags(tags)
+    if not norm:
+        total = get_asset_tags(session, asset_info_id=asset_info_id)
+        return {"removed": [], "not_present": [], "total_tags": total}
+
+    existing = {
+        tag_name
+        for (tag_name,) in (
+            session.execute(
+                sa.select(AssetInfoTag.tag_name).where(
+                    AssetInfoTag.asset_info_id == asset_info_id
+                )
+            )
+        ).all()
+    }
+
+    to_remove = sorted(set(t for t in norm if t in existing))
+    not_present = sorted(set(t for t in norm if t not in existing))
+
+    if to_remove:
+        session.execute(
+            delete(AssetInfoTag).where(
+                AssetInfoTag.asset_info_id == asset_info_id,
+                AssetInfoTag.tag_name.in_(to_remove),
+            )
+        )
+        session.flush()
+
+    total = get_asset_tags(session, asset_info_id=asset_info_id)
+    return {"removed": to_remove, "not_present": not_present, "total_tags": total}
+
+
+def add_missing_tag_for_asset_id(
+    session: Session,
+    asset_id: str,
+    origin: str = "automatic",
+) -> None:
+    select_rows = (
+        sa.select(
+            AssetInfo.id.label("asset_info_id"),
+            sa.literal("missing").label("tag_name"),
+            sa.literal(origin).label("origin"),
+            sa.literal(get_utc_now()).label("added_at"),
+        )
+        .where(AssetInfo.asset_id == asset_id)
+        .where(
+            sa.not_(
+                sa.exists().where(
+                    (AssetInfoTag.asset_info_id == AssetInfo.id)
+                    & (AssetInfoTag.tag_name == "missing")
+                )
+            )
+        )
+    )
+    session.execute(
+        sqlite.insert(AssetInfoTag)
+        .from_select(
+            ["asset_info_id", "tag_name", "origin", "added_at"],
+            select_rows,
+        )
+        .on_conflict_do_nothing(
+            index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name]
+        )
+    )
+
+
+def remove_missing_tag_for_asset_id(
+    session: Session,
+    asset_id: str,
+) -> None:
+    session.execute(
+        sa.delete(AssetInfoTag).where(
+            AssetInfoTag.asset_info_id.in_(
+                sa.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)
+            ),
+            AssetInfoTag.tag_name == "missing",
+        )
+    )
+
+
+def list_tags_with_usage(
+    session: Session,
+    prefix: str | None = None,
+    limit: int = 100,
+    offset: int = 0,
+    include_zero: bool = True,
+    order: str = "count_desc",
+    owner_id: str = "",
+) -> tuple[list[tuple[str, str, int]], int]:
+    counts_sq = (
+        select(
+            AssetInfoTag.tag_name.label("tag_name"),
+            func.count(AssetInfoTag.asset_info_id).label("cnt"),
+        )
+        .select_from(AssetInfoTag)
+        .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id)
+        .where(build_visible_owner_clause(owner_id))
+        .group_by(AssetInfoTag.tag_name)
+        .subquery()
+    )
+
+    q = (
+        select(
+            Tag.name,
+            Tag.tag_type,
+            func.coalesce(counts_sq.c.cnt, 0).label("count"),
+        )
+        .select_from(Tag)
+        .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True)
+    )
+
+    if prefix:
+        escaped, esc = escape_sql_like_string(prefix.strip().lower())
+        q = q.where(Tag.name.like(escaped + "%", escape=esc))
+
+    if not include_zero:
+        q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0)
+
+    if order == "name_asc":
+        q = q.order_by(Tag.name.asc())
+    else:
+        q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc())
+
+    total_q = select(func.count()).select_from(Tag)
+    if prefix:
+        escaped, esc = escape_sql_like_string(prefix.strip().lower())
+        total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc))
+    if not include_zero:
+        total_q = total_q.where(
+            Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name))
+        )
+
+    rows = (session.execute(q.limit(limit).offset(offset))).all()
+    total = (session.execute(total_q)).scalar_one()
+
+    rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows]
+    return rows_norm, int(total or 0)
+
+
+def bulk_insert_tags_and_meta(
+    session: Session,
+    tag_rows: list[dict],
+    meta_rows: list[dict],
+) -> None:
+    """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING.
+
+    Args:
+        session: Database session
+        tag_rows: List of dicts with keys: asset_info_id, tag_name, origin, added_at
+        meta_rows: List of dicts with keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json
+    """
+    if tag_rows:
+        ins_tags = sqlite.insert(AssetInfoTag).on_conflict_do_nothing(
+            index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name]
+        )
+        for chunk in iter_row_chunks(tag_rows, cols_per_row=4):
+            session.execute(ins_tags, chunk)
+
+    if meta_rows:
+        ins_meta = sqlite.insert(AssetInfoMeta).on_conflict_do_nothing(
+            index_elements=[
+                AssetInfoMeta.asset_info_id,
+                AssetInfoMeta.key,
+                AssetInfoMeta.ordinal,
+            ]
+        )
+        for chunk in iter_row_chunks(meta_rows, cols_per_row=7):
+            session.execute(ins_meta, chunk)

app/assets/database/tags.py

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

app/assets/hashing.py

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

app/assets/helpers.py

@@ -1,52 +1,36 @@
-import contextlib
 import os
-from decimal import Decimal
-from aiohttp import web
 from datetime import datetime, timezone
-from pathlib import Path
-from typing import Literal, Any
-
-import folder_paths
+from typing import Literal, Sequence
 
 
-RootType = Literal["models", "input", "output"]
-ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output")
-
-def get_query_dict(request: web.Request) -> dict[str, Any]:
+def select_best_live_path(states: Sequence) -> str:
     """
-    Gets a dictionary of query parameters from the request.
-
-    'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic.
+    Return the best on-disk path among cache states:
+      1) Prefer a path that exists with needs_verify == False (already verified).
+      2) Otherwise, pick the first path that exists.
+      3) Otherwise return empty string.
     """
-    query_dict = {
-        key: request.query.getall(key) if len(request.query.getall(key)) > 1 else request.query.get(key)
-        for key in request.query.keys()
-    }
-    return query_dict
+    alive = [
+        s
+        for s in states
+        if getattr(s, "file_path", None) and os.path.isfile(s.file_path)
+    ]
+    if not alive:
+        return ""
+    for s in alive:
+        if not getattr(s, "needs_verify", False):
+            return s.file_path
+    return alive[0].file_path
 
-def list_tree(base_dir: str) -> list[str]:
-    out: list[str] = []
-    base_abs = os.path.abspath(base_dir)
-    if not os.path.isdir(base_abs):
-        return out
-    for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False):
-        for name in filenames:
-            out.append(os.path.abspath(os.path.join(dirpath, name)))
-    return out
 
-def prefixes_for_root(root: RootType) -> list[str]:
-    if root == "models":
-        bases: list[str] = []
-        for _bucket, paths in get_comfy_models_folders():
-            bases.extend(paths)
-        return [os.path.abspath(p) for p in bases]
-    if root == "input":
-        return [os.path.abspath(folder_paths.get_input_directory())]
-    if root == "output":
-        return [os.path.abspath(folder_paths.get_output_directory())]
-    return []
+ALLOWED_ROOTS: tuple[Literal["models", "input", "output"], ...] = (
+    "models",
+    "input",
+    "output",
+)
 
-def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]:
+
+def escape_sql_like_string(s: str, escape: str = "!") -> tuple[str, str]:
     """Escapes %, _ and the escape char itself in a LIKE prefix.
     Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like().
     """
@@ -54,173 +38,11 @@ def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]:
     s = s.replace("%", escape + "%").replace("_", escape + "_")  # escape LIKE wildcards
     return s, escape
 
-def fast_asset_file_check(
-    *,
-    mtime_db: int | None,
-    size_db: int | None,
-    stat_result: os.stat_result,
-) -> bool:
-    if mtime_db is None:
-        return False
-    actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000))
-    if int(mtime_db) != int(actual_mtime_ns):
-        return False
-    sz = int(size_db or 0)
-    if sz > 0:
-        return int(stat_result.st_size) == sz
-    return True
 
-def utcnow() -> datetime:
+def get_utc_now() -> datetime:
     """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC."""
     return datetime.now(timezone.utc).replace(tzinfo=None)
 
-def get_comfy_models_folders() -> list[tuple[str, list[str]]]:
-    """Build a list of (folder_name, base_paths[]) categories that are configured for model locations.
-
-    We trust `folder_paths.folder_names_and_paths` and include a category if
-    *any* of its base paths lies under the Comfy `models_dir`.
-    """
-    targets: list[tuple[str, list[str]]] = []
-    models_root = os.path.abspath(folder_paths.models_dir)
-    for name, values in folder_paths.folder_names_and_paths.items():
-        paths, _exts = values[0], values[1]  # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI
-        if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths):
-            targets.append((name, paths))
-    return targets
-
-def resolve_destination_from_tags(tags: list[str]) -> tuple[str, list[str]]:
-    """Validates and maps tags -> (base_dir, subdirs_for_fs)"""
-    root = tags[0]
-    if root == "models":
-        if len(tags) < 2:
-            raise ValueError("at least two tags required for model asset")
-        try:
-            bases = folder_paths.folder_names_and_paths[tags[1]][0]
-        except KeyError:
-            raise ValueError(f"unknown model category '{tags[1]}'")
-        if not bases:
-            raise ValueError(f"no base path configured for category '{tags[1]}'")
-        base_dir = os.path.abspath(bases[0])
-        raw_subdirs = tags[2:]
-    else:
-        base_dir = os.path.abspath(
-            folder_paths.get_input_directory() if root == "input" else folder_paths.get_output_directory()
-        )
-        raw_subdirs = tags[1:]
-    for i in raw_subdirs:
-        if i in (".", ".."):
-            raise ValueError("invalid path component in tags")
-
-    return base_dir, raw_subdirs if raw_subdirs else []
-
-def ensure_within_base(candidate: str, base: str) -> None:
-    cand_abs = os.path.abspath(candidate)
-    base_abs = os.path.abspath(base)
-    try:
-        if os.path.commonpath([cand_abs, base_abs]) != base_abs:
-            raise ValueError("destination escapes base directory")
-    except Exception:
-        raise ValueError("invalid destination path")
-
-def compute_relative_filename(file_path: str) -> str | None:
-    """
-    Return the model's path relative to the last well-known folder (the model category),
-    using forward slashes, eg:
-      /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors"
-      /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors"
-
-    For non-model paths, returns None.
-    NOTE: this is a temporary helper, used only for initializing metadata["filename"] field.
-    """
-    try:
-        root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path)
-    except ValueError:
-        return None
-
-    p = Path(rel_path)
-    parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)]
-    if not parts:
-        return None
-
-    if root_category == "models":
-        # parts[0] is the category ("checkpoints", "vae", etc) – drop it
-        inside = parts[1:] if len(parts) > 1 else [parts[0]]
-        return "/".join(inside)
-    return "/".join(parts)  # input/output: keep all parts
-
-def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]:
-    """Given an absolute or relative file path, determine which root category the path belongs to:
-      - 'input' if the file resides under `folder_paths.get_input_directory()`
-      - 'output' if the file resides under `folder_paths.get_output_directory()`
-      - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()`
-
-    Returns:
-        (root_category, relative_path_inside_that_root)
-        For 'models', the relative path is prefixed with the category name:
-            e.g. ('models', 'vae/test/sub/ae.safetensors')
-
-    Raises:
-        ValueError: if the path does not belong to input, output, or configured model bases.
-    """
-    fp_abs = os.path.abspath(file_path)
-
-    def _is_within(child: str, parent: str) -> bool:
-        try:
-            return os.path.commonpath([child, parent]) == parent
-        except Exception:
-            return False
-
-    def _rel(child: str, parent: str) -> str:
-        return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep)
-
-    # 1) input
-    input_base = os.path.abspath(folder_paths.get_input_directory())
-    if _is_within(fp_abs, input_base):
-        return "input", _rel(fp_abs, input_base)
-
-    # 2) output
-    output_base = os.path.abspath(folder_paths.get_output_directory())
-    if _is_within(fp_abs, output_base):
-        return "output", _rel(fp_abs, output_base)
-
-    # 3) models (check deepest matching base to avoid ambiguity)
-    best: tuple[int, str, str] | None = None  # (base_len, bucket, rel_inside_bucket)
-    for bucket, bases in get_comfy_models_folders():
-        for b in bases:
-            base_abs = os.path.abspath(b)
-            if not _is_within(fp_abs, base_abs):
-                continue
-            cand = (len(base_abs), bucket, _rel(fp_abs, base_abs))
-            if best is None or cand[0] > best[0]:
-                best = cand
-
-    if best is not None:
-        _, bucket, rel_inside = best
-        combined = os.path.join(bucket, rel_inside)
-        return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep)
-
-    raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}")
-
-def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:
-    """Return a tuple (name, tags) derived from a filesystem path.
-
-    Semantics:
-      - Root category is determined by `get_relative_to_root_category_path_of_asset`.
-      - The returned `name` is the base filename with extension from the relative path.
-      - The returned `tags` are:
-            [root_category] + parent folders of the relative path (in order)
-        For 'models', this means:
-            file '/.../ModelsDir/vae/test_tag/ae.safetensors'
-            -> root_category='models', some_path='vae/test_tag/ae.safetensors'
-            -> name='ae.safetensors', tags=['models', 'vae', 'test_tag']
-
-    Raises:
-        ValueError: if the path does not belong to input, output, or configured model bases.
-    """
-    root_category, some_path = get_relative_to_root_category_path_of_asset(file_path)
-    p = Path(some_path)
-    parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)]
-    return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts])))
 
 def normalize_tags(tags: list[str] | None) -> list[str]:
     """
@@ -229,84 +51,3 @@ def normalize_tags(tags: list[str] | None) -> list[str]:
       - Removing duplicates.
     """
     return [t.strip().lower() for t in (tags or []) if (t or "").strip()]
-
-def collect_models_files() -> list[str]:
-    out: list[str] = []
-    for folder_name, bases in get_comfy_models_folders():
-        rel_files = folder_paths.get_filename_list(folder_name) or []
-        for rel_path in rel_files:
-            abs_path = folder_paths.get_full_path(folder_name, rel_path)
-            if not abs_path:
-                continue
-            abs_path = os.path.abspath(abs_path)
-            allowed = False
-            for b in bases:
-                base_abs = os.path.abspath(b)
-                with contextlib.suppress(Exception):
-                    if os.path.commonpath([abs_path, base_abs]) == base_abs:
-                        allowed = True
-                        break
-            if allowed:
-                out.append(abs_path)
-    return out
-
-def is_scalar(v):
-    if v is None:
-        return True
-    if isinstance(v, bool):
-        return True
-    if isinstance(v, (int, float, Decimal, str)):
-        return True
-    return False
-
-def project_kv(key: str, value):
-    """
-    Turn a metadata key/value into typed projection rows.
-    Returns list[dict] with keys:
-      key, ordinal, and one of val_str / val_num / val_bool / val_json (others None)
-    """
-    rows: list[dict] = []
-
-    def _null_row(ordinal: int) -> dict:
-        return {
-            "key": key, "ordinal": ordinal,
-            "val_str": None, "val_num": None, "val_bool": None, "val_json": None
-        }
-
-    if value is None:
-        rows.append(_null_row(0))
-        return rows
-
-    if is_scalar(value):
-        if isinstance(value, bool):
-            rows.append({"key": key, "ordinal": 0, "val_bool": bool(value)})
-        elif isinstance(value, (int, float, Decimal)):
-            num = value if isinstance(value, Decimal) else Decimal(str(value))
-            rows.append({"key": key, "ordinal": 0, "val_num": num})
-        elif isinstance(value, str):
-            rows.append({"key": key, "ordinal": 0, "val_str": value})
-        else:
-            rows.append({"key": key, "ordinal": 0, "val_json": value})
-        return rows
-
-    if isinstance(value, list):
-        if all(is_scalar(x) for x in value):
-            for i, x in enumerate(value):
-                if x is None:
-                    rows.append(_null_row(i))
-                elif isinstance(x, bool):
-                    rows.append({"key": key, "ordinal": i, "val_bool": bool(x)})
-                elif isinstance(x, (int, float, Decimal)):
-                    num = x if isinstance(x, Decimal) else Decimal(str(x))
-                    rows.append({"key": key, "ordinal": i, "val_num": num})
-                elif isinstance(x, str):
-                    rows.append({"key": key, "ordinal": i, "val_str": x})
-                else:
-                    rows.append({"key": key, "ordinal": i, "val_json": x})
-            return rows
-        for i, x in enumerate(value):
-            rows.append({"key": key, "ordinal": i, "val_json": x})
-        return rows
-
-    rows.append({"key": key, "ordinal": 0, "val_json": value})
-    return rows

app/assets/manager.py

@@ -1,516 +0,0 @@
-import os
-import mimetypes
-import contextlib
-from typing import Sequence
-
-from app.database.db import create_session
-from app.assets.api import schemas_out, schemas_in
-from app.assets.database.queries import (
-    asset_exists_by_hash,
-    asset_info_exists_for_asset_id,
-    get_asset_by_hash,
-    get_asset_info_by_id,
-    fetch_asset_info_asset_and_tags,
-    fetch_asset_info_and_asset,
-    create_asset_info_for_existing_asset,
-    touch_asset_info_by_id,
-    update_asset_info_full,
-    delete_asset_info_by_id,
-    list_cache_states_by_asset_id,
-    list_asset_infos_page,
-    list_tags_with_usage,
-    get_asset_tags,
-    add_tags_to_asset_info,
-    remove_tags_from_asset_info,
-    pick_best_live_path,
-    ingest_fs_asset,
-    set_asset_info_preview,
-)
-from app.assets.helpers import resolve_destination_from_tags, ensure_within_base
-from app.assets.database.models import Asset
-
-
-def _safe_sort_field(requested: str | None) -> str:
-    if not requested:
-        return "created_at"
-    v = requested.lower()
-    if v in {"name", "created_at", "updated_at", "size", "last_access_time"}:
-        return v
-    return "created_at"
-
-
-def _get_size_mtime_ns(path: str) -> tuple[int, int]:
-    st = os.stat(path, follow_symlinks=True)
-    return st.st_size, getattr(st, "st_mtime_ns", int(st.st_mtime * 1_000_000_000))
-
-
-def _safe_filename(name: str | None, fallback: str) -> str:
-    n = os.path.basename((name or "").strip() or fallback)
-    if n:
-        return n
-    return fallback
-
-
-def asset_exists(*, asset_hash: str) -> bool:
-    """
-    Check if an asset with a given hash exists in database.
-    """
-    with create_session() as session:
-        return asset_exists_by_hash(session, asset_hash=asset_hash)
-
-
-def list_assets(
-    *,
-    include_tags: Sequence[str] | None = None,
-    exclude_tags: Sequence[str] | None = None,
-    name_contains: str | None = None,
-    metadata_filter: dict | None = None,
-    limit: int = 20,
-    offset: int = 0,
-    sort: str = "created_at",
-    order: str = "desc",
-    owner_id: str = "",
-) -> schemas_out.AssetsList:
-    sort = _safe_sort_field(sort)
-    order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower()
-
-    with create_session() as session:
-        infos, tag_map, total = list_asset_infos_page(
-            session,
-            owner_id=owner_id,
-            include_tags=include_tags,
-            exclude_tags=exclude_tags,
-            name_contains=name_contains,
-            metadata_filter=metadata_filter,
-            limit=limit,
-            offset=offset,
-            sort=sort,
-            order=order,
-        )
-
-    summaries: list[schemas_out.AssetSummary] = []
-    for info in infos:
-        asset = info.asset
-        tags = tag_map.get(info.id, [])
-        summaries.append(
-            schemas_out.AssetSummary(
-                id=info.id,
-                name=info.name,
-                asset_hash=asset.hash if asset else None,
-                size=int(asset.size_bytes) if asset else None,
-                mime_type=asset.mime_type if asset else None,
-                tags=tags,
-                created_at=info.created_at,
-                updated_at=info.updated_at,
-                last_access_time=info.last_access_time,
-            )
-        )
-
-    return schemas_out.AssetsList(
-        assets=summaries,
-        total=total,
-        has_more=(offset + len(summaries)) < total,
-    )
-
-
-def get_asset(
-    *,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> schemas_out.AssetDetail:
-    with create_session() as session:
-        res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not res:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        info, asset, tag_names = res
-        preview_id = info.preview_id
-
-    return schemas_out.AssetDetail(
-        id=info.id,
-        name=info.name,
-        asset_hash=asset.hash if asset else None,
-        size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
-        mime_type=asset.mime_type if asset else None,
-        tags=tag_names,
-        user_metadata=info.user_metadata or {},
-        preview_id=preview_id,
-        created_at=info.created_at,
-        last_access_time=info.last_access_time,
-    )
-
-
-def resolve_asset_content_for_download(
-    *,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> tuple[str, str, str]:
-    with create_session() as session:
-        pair = fetch_asset_info_and_asset(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not pair:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-        info, asset = pair
-        states = list_cache_states_by_asset_id(session, asset_id=asset.id)
-        abs_path = pick_best_live_path(states)
-        if not abs_path:
-            raise FileNotFoundError
-
-        touch_asset_info_by_id(session, asset_info_id=asset_info_id)
-        session.commit()
-
-        ctype = asset.mime_type or mimetypes.guess_type(info.name or abs_path)[0] or "application/octet-stream"
-        download_name = info.name or os.path.basename(abs_path)
-        return abs_path, ctype, download_name
-
-
-def upload_asset_from_temp_path(
-    spec: schemas_in.UploadAssetSpec,
-    *,
-    temp_path: str,
-    client_filename: str | None = None,
-    owner_id: str = "",
-    expected_asset_hash: str | None = None,
-) -> schemas_out.AssetCreated:
-    """
-    Create new asset or update existing asset from a temporary file path.
-    """
-    try:
-        # NOTE: blake3 is not required right now, so this will fail if blake3 is not installed in local environment
-        import app.assets.hashing as hashing
-        digest = hashing.blake3_hash(temp_path)
-    except Exception as e:
-        raise RuntimeError(f"failed to hash uploaded file: {e}")
-    asset_hash = "blake3:" + digest
-
-    if expected_asset_hash and asset_hash != expected_asset_hash.strip().lower():
-        raise ValueError("HASH_MISMATCH")
-
-    with create_session() as session:
-        existing = get_asset_by_hash(session, asset_hash=asset_hash)
-        if existing is not None:
-            with contextlib.suppress(Exception):
-                if temp_path and os.path.exists(temp_path):
-                    os.remove(temp_path)
-
-            display_name = _safe_filename(spec.name or (client_filename or ""), fallback=digest)
-            info = create_asset_info_for_existing_asset(
-                session,
-                asset_hash=asset_hash,
-                name=display_name,
-                user_metadata=spec.user_metadata or {},
-                tags=spec.tags or [],
-                tag_origin="manual",
-                owner_id=owner_id,
-            )
-            tag_names = get_asset_tags(session, asset_info_id=info.id)
-            session.commit()
-
-            return schemas_out.AssetCreated(
-                id=info.id,
-                name=info.name,
-                asset_hash=existing.hash,
-                size=int(existing.size_bytes) if existing.size_bytes is not None else None,
-                mime_type=existing.mime_type,
-                tags=tag_names,
-                user_metadata=info.user_metadata or {},
-                preview_id=info.preview_id,
-                created_at=info.created_at,
-                last_access_time=info.last_access_time,
-                created_new=False,
-            )
-
-    base_dir, subdirs = resolve_destination_from_tags(spec.tags)
-    dest_dir = os.path.join(base_dir, *subdirs) if subdirs else base_dir
-    os.makedirs(dest_dir, exist_ok=True)
-
-    src_for_ext = (client_filename or spec.name or "").strip()
-    _ext = os.path.splitext(os.path.basename(src_for_ext))[1] if src_for_ext else ""
-    ext = _ext if 0 < len(_ext) <= 16 else ""
-    hashed_basename = f"{digest}{ext}"
-    dest_abs = os.path.abspath(os.path.join(dest_dir, hashed_basename))
-    ensure_within_base(dest_abs, base_dir)
-
-    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"
-    )
-
-    try:
-        os.replace(temp_path, dest_abs)
-    except Exception as e:
-        raise RuntimeError(f"failed to move uploaded file into place: {e}")
-
-    try:
-        size_bytes, mtime_ns = _get_size_mtime_ns(dest_abs)
-    except OSError as e:
-        raise RuntimeError(f"failed to stat destination file: {e}")
-
-    with create_session() as session:
-        result = ingest_fs_asset(
-            session,
-            asset_hash=asset_hash,
-            abs_path=dest_abs,
-            size_bytes=size_bytes,
-            mtime_ns=mtime_ns,
-            mime_type=content_type,
-            info_name=_safe_filename(spec.name or (client_filename or ""), fallback=digest),
-            owner_id=owner_id,
-            preview_id=None,
-            user_metadata=spec.user_metadata or {},
-            tags=spec.tags,
-            tag_origin="manual",
-            require_existing_tags=False,
-        )
-        info_id = result["asset_info_id"]
-        if not info_id:
-            raise RuntimeError("failed to create asset metadata")
-
-        pair = fetch_asset_info_and_asset(session, asset_info_id=info_id, owner_id=owner_id)
-        if not pair:
-            raise RuntimeError("inconsistent DB state after ingest")
-        info, asset = pair
-        tag_names = get_asset_tags(session, asset_info_id=info.id)
-        created_result = schemas_out.AssetCreated(
-            id=info.id,
-            name=info.name,
-            asset_hash=asset.hash,
-            size=int(asset.size_bytes),
-            mime_type=asset.mime_type,
-            tags=tag_names,
-            user_metadata=info.user_metadata or {},
-            preview_id=info.preview_id,
-            created_at=info.created_at,
-            last_access_time=info.last_access_time,
-            created_new=result["asset_created"],
-        )
-        session.commit()
-
-    return created_result
-
-
-def update_asset(
-    *,
-    asset_info_id: str,
-    name: str | None = None,
-    tags: list[str] | None = None,
-    user_metadata: dict | None = None,
-    owner_id: str = "",
-) -> schemas_out.AssetUpdated:
-    with create_session() as session:
-        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-
-        info = update_asset_info_full(
-            session,
-            asset_info_id=asset_info_id,
-            name=name,
-            tags=tags,
-            user_metadata=user_metadata,
-            tag_origin="manual",
-            asset_info_row=info_row,
-        )
-
-        tag_names = get_asset_tags(session, asset_info_id=asset_info_id)
-        result = schemas_out.AssetUpdated(
-            id=info.id,
-            name=info.name,
-            asset_hash=info.asset.hash if info.asset else None,
-            tags=tag_names,
-            user_metadata=info.user_metadata or {},
-            updated_at=info.updated_at,
-        )
-        session.commit()
-
-    return result
-
-
-def set_asset_preview(
-    *,
-    asset_info_id: str,
-    preview_asset_id: str | None = None,
-    owner_id: str = "",
-) -> schemas_out.AssetDetail:
-    with create_session() as session:
-        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-
-        set_asset_info_preview(
-            session,
-            asset_info_id=asset_info_id,
-            preview_asset_id=preview_asset_id,
-        )
-
-        res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not res:
-            raise RuntimeError("State changed during preview update")
-        info, asset, tags = res
-        result = schemas_out.AssetDetail(
-            id=info.id,
-            name=info.name,
-            asset_hash=asset.hash if asset else None,
-            size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
-            mime_type=asset.mime_type if asset else None,
-            tags=tags,
-            user_metadata=info.user_metadata or {},
-            preview_id=info.preview_id,
-            created_at=info.created_at,
-            last_access_time=info.last_access_time,
-        )
-        session.commit()
-
-    return result
-
-
-def delete_asset_reference(*, asset_info_id: str, owner_id: str, delete_content_if_orphan: bool = True) -> bool:
-    with create_session() as session:
-        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        asset_id = info_row.asset_id if info_row else None
-        deleted = delete_asset_info_by_id(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not deleted:
-            session.commit()
-            return False
-
-        if not delete_content_if_orphan or not asset_id:
-            session.commit()
-            return True
-
-        still_exists = asset_info_exists_for_asset_id(session, asset_id=asset_id)
-        if still_exists:
-            session.commit()
-            return True
-
-        states = list_cache_states_by_asset_id(session, asset_id=asset_id)
-        file_paths = [s.file_path for s in (states or []) if getattr(s, "file_path", None)]
-
-        asset_row = session.get(Asset, asset_id)
-        if asset_row is not None:
-            session.delete(asset_row)
-
-        session.commit()
-        for p in file_paths:
-            with contextlib.suppress(Exception):
-                if p and os.path.isfile(p):
-                    os.remove(p)
-    return True
-
-
-def create_asset_from_hash(
-    *,
-    hash_str: str,
-    name: str,
-    tags: list[str] | None = None,
-    user_metadata: dict | None = None,
-    owner_id: str = "",
-) -> schemas_out.AssetCreated | None:
-    canonical = hash_str.strip().lower()
-    with create_session() as session:
-        asset = get_asset_by_hash(session, asset_hash=canonical)
-        if not asset:
-            return None
-
-        info = create_asset_info_for_existing_asset(
-            session,
-            asset_hash=canonical,
-            name=_safe_filename(name, fallback=canonical.split(":", 1)[1]),
-            user_metadata=user_metadata or {},
-            tags=tags or [],
-            tag_origin="manual",
-            owner_id=owner_id,
-        )
-        tag_names = get_asset_tags(session, asset_info_id=info.id)
-        result = schemas_out.AssetCreated(
-            id=info.id,
-            name=info.name,
-            asset_hash=asset.hash,
-            size=int(asset.size_bytes),
-            mime_type=asset.mime_type,
-            tags=tag_names,
-            user_metadata=info.user_metadata or {},
-            preview_id=info.preview_id,
-            created_at=info.created_at,
-            last_access_time=info.last_access_time,
-            created_new=False,
-        )
-        session.commit()
-
-    return result
-
-
-def add_tags_to_asset(
-    *,
-    asset_info_id: str,
-    tags: list[str],
-    origin: str = "manual",
-    owner_id: str = "",
-) -> schemas_out.TagsAdd:
-    with create_session() as session:
-        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-        data = add_tags_to_asset_info(
-            session,
-            asset_info_id=asset_info_id,
-            tags=tags,
-            origin=origin,
-            create_if_missing=True,
-            asset_info_row=info_row,
-        )
-        session.commit()
-    return schemas_out.TagsAdd(**data)
-
-
-def remove_tags_from_asset(
-    *,
-    asset_info_id: str,
-    tags: list[str],
-    owner_id: str = "",
-) -> schemas_out.TagsRemove:
-    with create_session() as session:
-        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-
-        data = remove_tags_from_asset_info(
-            session,
-            asset_info_id=asset_info_id,
-            tags=tags,
-        )
-        session.commit()
-    return schemas_out.TagsRemove(**data)
-
-
-def list_tags(
-    prefix: str | None = None,
-    limit: int = 100,
-    offset: int = 0,
-    order: str = "count_desc",
-    include_zero: bool = True,
-    owner_id: str = "",
-) -> schemas_out.TagsList:
-    limit = max(1, min(1000, limit))
-    offset = max(0, offset)
-
-    with create_session() as session:
-        rows, total = list_tags_with_usage(
-            session,
-            prefix=prefix,
-            limit=limit,
-            offset=offset,
-            include_zero=include_zero,
-            order=order,
-            owner_id=owner_id,
-        )
-
-    tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows]
-    return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total)

app/assets/scanner.py

@@ -1,263 +1,318 @@
 import contextlib
-import time
 import logging
 import os
-import sqlalchemy
+import time
+from typing import Literal, TypedDict
 
 import folder_paths
-from app.database.db import create_session, dependencies_available
-from app.assets.helpers import (
-    collect_models_files, compute_relative_filename, fast_asset_file_check, get_name_and_tags_from_asset_path,
-    list_tree,prefixes_for_root, escape_like_prefix,
-    RootType
+from app.assets.database.queries import (
+    add_missing_tag_for_asset_id,
+    bulk_set_needs_verify,
+    delete_cache_states_by_ids,
+    delete_orphaned_seed_asset,
+    ensure_tags_exist,
+    get_cache_states_for_prefixes,
+    remove_missing_tag_for_asset_id,
 )
-from app.assets.database.tags import add_missing_tag_for_asset_id, ensure_tags_exist, remove_missing_tag_for_asset_id
-from app.assets.database.bulk_ops import seed_from_paths_batch
-from app.assets.database.models import Asset, AssetCacheState, AssetInfo
+from app.assets.services.bulk_ingest import (
+    SeedAssetSpec,
+    batch_insert_seed_assets,
+    prune_orphaned_assets,
+)
+from app.assets.services.file_utils import (
+    get_mtime_ns,
+    list_files_recursively,
+    verify_file_unchanged,
+)
+from app.assets.services.path_utils import (
+    compute_relative_filename,
+    get_comfy_models_folders,
+    get_name_and_tags_from_asset_path,
+)
+from app.database.db import create_session, dependencies_available
+
+
+class _StateInfo(TypedDict):
+    sid: int
+    fp: str
+    exists: bool
+    fast_ok: bool
+    needs_verify: bool
+
+
+class _AssetAccumulator(TypedDict):
+    hash: str | None
+    size_db: int
+    states: list[_StateInfo]
+
+RootType = Literal["models", "input", "output"]
+
+
+def get_prefixes_for_root(root: RootType) -> list[str]:
+    if root == "models":
+        bases: list[str] = []
+        for _bucket, paths in get_comfy_models_folders():
+            bases.extend(paths)
+        return [os.path.abspath(p) for p in bases]
+    if root == "input":
+        return [os.path.abspath(folder_paths.get_input_directory())]
+    if root == "output":
+        return [os.path.abspath(folder_paths.get_output_directory())]
+    return []
+
+
+def collect_models_files() -> list[str]:
+    out: list[str] = []
+    for folder_name, bases in get_comfy_models_folders():
+        rel_files = folder_paths.get_filename_list(folder_name) or []
+        for rel_path in rel_files:
+            abs_path = folder_paths.get_full_path(folder_name, rel_path)
+            if not abs_path:
+                continue
+            abs_path = os.path.abspath(abs_path)
+            allowed = False
+            for b in bases:
+                base_abs = os.path.abspath(b)
+                with contextlib.suppress(Exception):
+                    if os.path.commonpath([abs_path, base_abs]) == base_abs:
+                        allowed = True
+                        break
+            if allowed:
+                out.append(abs_path)
+    return out
+
+
+def sync_cache_states_with_filesystem(
+    session,
+    root: RootType,
+    collect_existing_paths: bool = False,
+    update_missing_tags: bool = False,
+) -> set[str] | None:
+    """Reconcile cache states with filesystem for a root.
+
+    - Toggle needs_verify per state using fast mtime/size check
+    - For hashed assets with at least one fast-ok state in this root: delete stale missing states
+    - For seed assets with all states missing: delete Asset and its AssetInfos
+    - Optionally add/remove 'missing' tags based on fast-ok in this root
+    - Optionally return surviving absolute paths
+
+    Args:
+        session: Database session
+        root: Root type to scan
+        collect_existing_paths: If True, return set of surviving file paths
+        update_missing_tags: If True, update 'missing' tags based on file status
+
+    Returns:
+        Set of surviving absolute paths if collect_existing_paths=True, else None
+    """
+    prefixes = get_prefixes_for_root(root)
+    if not prefixes:
+        return set() if collect_existing_paths else None
+
+    rows = get_cache_states_for_prefixes(session, prefixes)
+
+    by_asset: dict[str, _AssetAccumulator] = {}
+    for row in rows:
+        acc = by_asset.get(row.asset_id)
+        if acc is None:
+            acc = {"hash": row.asset_hash, "size_db": row.size_bytes, "states": []}
+            by_asset[row.asset_id] = acc
+
+        fast_ok = False
+        try:
+            exists = True
+            fast_ok = verify_file_unchanged(
+                mtime_db=row.mtime_ns,
+                size_db=acc["size_db"],
+                stat_result=os.stat(row.file_path, follow_symlinks=True),
+            )
+        except FileNotFoundError:
+            exists = False
+        except OSError:
+            exists = False
+
+        acc["states"].append(
+            {
+                "sid": row.state_id,
+                "fp": row.file_path,
+                "exists": exists,
+                "fast_ok": fast_ok,
+                "needs_verify": row.needs_verify,
+            }
+        )
+
+    to_set_verify: list[int] = []
+    to_clear_verify: list[int] = []
+    stale_state_ids: list[int] = []
+    survivors: set[str] = set()
+
+    for aid, acc in by_asset.items():
+        a_hash = acc["hash"]
+        states = acc["states"]
+        any_fast_ok = any(s["fast_ok"] for s in states)
+        all_missing = all(not s["exists"] for s in states)
+
+        for s in states:
+            if not s["exists"]:
+                continue
+            if s["fast_ok"] and s["needs_verify"]:
+                to_clear_verify.append(s["sid"])
+            if not s["fast_ok"] and not s["needs_verify"]:
+                to_set_verify.append(s["sid"])
+
+        if a_hash is None:
+            if states and all_missing:
+                delete_orphaned_seed_asset(session, aid)
+            else:
+                for s in states:
+                    if s["exists"]:
+                        survivors.add(os.path.abspath(s["fp"]))
+            continue
+
+        if any_fast_ok:
+            for s in states:
+                if not s["exists"]:
+                    stale_state_ids.append(s["sid"])
+            if update_missing_tags:
+                with contextlib.suppress(Exception):
+                    remove_missing_tag_for_asset_id(session, asset_id=aid)
+        elif update_missing_tags:
+            with contextlib.suppress(Exception):
+                add_missing_tag_for_asset_id(session, asset_id=aid, origin="automatic")
+
+        for s in states:
+            if s["exists"]:
+                survivors.add(os.path.abspath(s["fp"]))
+
+    delete_cache_states_by_ids(session, stale_state_ids)
+    bulk_set_needs_verify(session, to_set_verify, value=True)
+    bulk_set_needs_verify(session, to_clear_verify, value=False)
+
+    return survivors if collect_existing_paths else None
+
+
+def _sync_root_safely(root: RootType) -> set[str]:
+    """Sync a single root's cache states with the filesystem.
+
+    Returns survivors (existing paths) or empty set on failure.
+    """
+    try:
+        with create_session() as sess:
+            survivors = sync_cache_states_with_filesystem(
+                sess,
+                root,
+                collect_existing_paths=True,
+                update_missing_tags=True,
+            )
+            sess.commit()
+            return survivors or set()
+    except Exception as e:
+        logging.exception("fast DB scan failed for %s: %s", root, e)
+        return set()
+
+
+def _prune_orphans_safely(prefixes: list[str]) -> int:
+    """Prune orphaned assets outside the given prefixes.
+
+    Returns count pruned or 0 on failure.
+    """
+    try:
+        with create_session() as sess:
+            count = prune_orphaned_assets(sess, prefixes)
+            sess.commit()
+            return count
+    except Exception as e:
+        logging.exception("orphan pruning failed: %s", e)
+        return 0
+
+
+def _collect_paths_for_roots(roots: tuple[RootType, ...]) -> list[str]:
+    """Collect all file paths for the given roots."""
+    paths: list[str] = []
+    if "models" in roots:
+        paths.extend(collect_models_files())
+    if "input" in roots:
+        paths.extend(list_files_recursively(folder_paths.get_input_directory()))
+    if "output" in roots:
+        paths.extend(list_files_recursively(folder_paths.get_output_directory()))
+    return paths
+
+
+def _build_asset_specs(
+    paths: list[str],
+    existing_paths: set[str],
+) -> tuple[list[SeedAssetSpec], set[str], int]:
+    """Build asset specs from paths, returning (specs, tag_pool, skipped_count)."""
+    specs: list[SeedAssetSpec] = []
+    tag_pool: set[str] = set()
+    skipped = 0
+
+    for p in paths:
+        abs_p = os.path.abspath(p)
+        if abs_p in existing_paths:
+            skipped += 1
+            continue
+        try:
+            stat_p = os.stat(abs_p, follow_symlinks=False)
+        except OSError:
+            continue
+        if not stat_p.st_size:
+            continue
+        name, tags = get_name_and_tags_from_asset_path(abs_p)
+        specs.append(
+            {
+                "abs_path": abs_p,
+                "size_bytes": stat_p.st_size,
+                "mtime_ns": get_mtime_ns(stat_p),
+                "info_name": name,
+                "tags": tags,
+                "fname": compute_relative_filename(abs_p),
+            }
+        )
+        tag_pool.update(tags)
+
+    return specs, tag_pool, skipped
+
+
+def _insert_asset_specs(specs: list[SeedAssetSpec], tag_pool: set[str]) -> int:
+    """Insert asset specs into database, returning count of created infos."""
+    if not specs:
+        return 0
+    with create_session() as sess:
+        if tag_pool:
+            ensure_tags_exist(sess, tag_pool, tag_type="user")
+        result = batch_insert_seed_assets(sess, specs=specs, owner_id="")
+        sess.commit()
+        return result.inserted_infos
 
 
 def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> None:
-    """
-    Scan the given roots and seed the assets into the database.
-    """
+    """Scan the given roots and seed the assets into the database."""
     if not dependencies_available():
         if enable_logging:
             logging.warning("Database dependencies not available, skipping assets scan")
         return
+
     t_start = time.perf_counter()
-    created = 0
-    skipped_existing = 0
-    orphans_pruned = 0
-    paths: list[str] = []
-    try:
-        existing_paths: set[str] = set()
-        for r in roots:
-            try:
-                survivors: set[str] = _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True)
-                if survivors:
-                    existing_paths.update(survivors)
-            except Exception as e:
-                logging.exception("fast DB scan failed for %s: %s", r, e)
 
-        try:
-            orphans_pruned = _prune_orphaned_assets(roots)
-        except Exception as e:
-            logging.exception("orphan pruning failed: %s", e)
+    existing_paths: set[str] = set()
+    for r in roots:
+        existing_paths.update(_sync_root_safely(r))
 
-        if "models" in roots:
-            paths.extend(collect_models_files())
-        if "input" in roots:
-            paths.extend(list_tree(folder_paths.get_input_directory()))
-        if "output" in roots:
-            paths.extend(list_tree(folder_paths.get_output_directory()))
+    all_prefixes = [os.path.abspath(p) for r in roots for p in get_prefixes_for_root(r)]
+    orphans_pruned = _prune_orphans_safely(all_prefixes)
 
-        specs: list[dict] = []
-        tag_pool: set[str] = set()
-        for p in paths:
-            abs_p = os.path.abspath(p)
-            if abs_p in existing_paths:
-                skipped_existing += 1
-                continue
-            try:
-                stat_p = os.stat(abs_p, follow_symlinks=False)
-            except OSError:
-                continue
-            # skip empty files
-            if not stat_p.st_size:
-                continue
-            name, tags = get_name_and_tags_from_asset_path(abs_p)
-            specs.append(
-                {
-                    "abs_path": abs_p,
-                    "size_bytes": stat_p.st_size,
-                    "mtime_ns": getattr(stat_p, "st_mtime_ns", int(stat_p.st_mtime * 1_000_000_000)),
-                    "info_name": name,
-                    "tags": tags,
-                    "fname": compute_relative_filename(abs_p),
-                }
-            )
-            for t in tags:
-                tag_pool.add(t)
-        # if no file specs, nothing to do
-        if not specs:
-            return
-        with create_session() as sess:
-            if tag_pool:
-                ensure_tags_exist(sess, tag_pool, tag_type="user")
+    paths = _collect_paths_for_roots(roots)
+    specs, tag_pool, skipped_existing = _build_asset_specs(paths, existing_paths)
+    created = _insert_asset_specs(specs, tag_pool)
 
-            result = seed_from_paths_batch(sess, specs=specs, owner_id="")
-            created += result["inserted_infos"]
-            sess.commit()
-    finally:
-        if enable_logging:
-            logging.info(
-                "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, orphans_pruned=%d, total_seen=%d)",
-                roots,
-                time.perf_counter() - t_start,
-                created,
-                skipped_existing,
-                orphans_pruned,
-                len(paths),
-            )
-
-
-def _prune_orphaned_assets(roots: tuple[RootType, ...]) -> int:
-    """Prune cache states outside configured prefixes, then delete orphaned seed assets."""
-    all_prefixes = [os.path.abspath(p) for r in roots for p in prefixes_for_root(r)]
-    if not all_prefixes:
-        return 0
-
-    def make_prefix_condition(prefix: str):
-        base = prefix if prefix.endswith(os.sep) else prefix + os.sep
-        escaped, esc = escape_like_prefix(base)
-        return AssetCacheState.file_path.like(escaped + "%", escape=esc)
-
-    matches_valid_prefix = sqlalchemy.or_(*[make_prefix_condition(p) for p in all_prefixes])
-
-    orphan_subq = (
-        sqlalchemy.select(Asset.id)
-        .outerjoin(AssetCacheState, AssetCacheState.asset_id == Asset.id)
-        .where(Asset.hash.is_(None), AssetCacheState.id.is_(None))
-    ).scalar_subquery()
-
-    with create_session() as sess:
-        sess.execute(sqlalchemy.delete(AssetCacheState).where(~matches_valid_prefix))
-        sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id.in_(orphan_subq)))
-        result = sess.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(orphan_subq)))
-        sess.commit()
-        return result.rowcount
-
-
-def _fast_db_consistency_pass(
-    root: RootType,
-    *,
-    collect_existing_paths: bool = False,
-    update_missing_tags: bool = False,
-) -> set[str] | None:
-    """Fast DB+FS pass for a root:
-      - Toggle needs_verify per state using fast check
-      - For hashed assets with at least one fast-ok state in this root: delete stale missing states
-      - For seed assets with all states missing: delete Asset and its AssetInfos
-      - Optionally add/remove 'missing' tags based on fast-ok in this root
-      - Optionally return surviving absolute paths
-    """
-    prefixes = prefixes_for_root(root)
-    if not prefixes:
-        return set() if collect_existing_paths else None
-
-    conds = []
-    for p in prefixes:
-        base = os.path.abspath(p)
-        if not base.endswith(os.sep):
-            base += os.sep
-        escaped, esc = escape_like_prefix(base)
-        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
-
-    with create_session() as sess:
-        rows = (
-            sess.execute(
-                sqlalchemy.select(
-                    AssetCacheState.id,
-                    AssetCacheState.file_path,
-                    AssetCacheState.mtime_ns,
-                    AssetCacheState.needs_verify,
-                    AssetCacheState.asset_id,
-                    Asset.hash,
-                    Asset.size_bytes,
-                )
-                .join(Asset, Asset.id == AssetCacheState.asset_id)
-                .where(sqlalchemy.or_(*conds))
-                .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
-            )
-        ).all()
-
-        by_asset: dict[str, dict] = {}
-        for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows:
-            acc = by_asset.get(aid)
-            if acc is None:
-                acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []}
-                by_asset[aid] = acc
-
-            fast_ok = False
-            try:
-                exists = True
-                fast_ok = fast_asset_file_check(
-                    mtime_db=mtime_db,
-                    size_db=acc["size_db"],
-                    stat_result=os.stat(fp, follow_symlinks=True),
-                )
-            except FileNotFoundError:
-                exists = False
-            except OSError:
-                exists = False
-
-            acc["states"].append({
-                "sid": sid,
-                "fp": fp,
-                "exists": exists,
-                "fast_ok": fast_ok,
-                "needs_verify": bool(needs_verify),
-            })
-
-        to_set_verify: list[int] = []
-        to_clear_verify: list[int] = []
-        stale_state_ids: list[int] = []
-        survivors: set[str] = set()
-
-        for aid, acc in by_asset.items():
-            a_hash = acc["hash"]
-            states = acc["states"]
-            any_fast_ok = any(s["fast_ok"] for s in states)
-            all_missing = all(not s["exists"] for s in states)
-
-            for s in states:
-                if not s["exists"]:
-                    continue
-                if s["fast_ok"] and s["needs_verify"]:
-                    to_clear_verify.append(s["sid"])
-                if not s["fast_ok"] and not s["needs_verify"]:
-                    to_set_verify.append(s["sid"])
-
-            if a_hash is None:
-                if states and all_missing:  # remove seed Asset completely, if no valid AssetCache exists
-                    sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id == aid))
-                    asset = sess.get(Asset, aid)
-                    if asset:
-                        sess.delete(asset)
-                else:
-                    for s in states:
-                        if s["exists"]:
-                            survivors.add(os.path.abspath(s["fp"]))
-                continue
-
-            if any_fast_ok:  # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records
-                for s in states:
-                    if not s["exists"]:
-                        stale_state_ids.append(s["sid"])
-                if update_missing_tags:
-                    with contextlib.suppress(Exception):
-                        remove_missing_tag_for_asset_id(sess, asset_id=aid)
-            elif update_missing_tags:
-                with contextlib.suppress(Exception):
-                    add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic")
-
-            for s in states:
-                if s["exists"]:
-                    survivors.add(os.path.abspath(s["fp"]))
-
-        if stale_state_ids:
-            sess.execute(sqlalchemy.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids)))
-        if to_set_verify:
-            sess.execute(
-                sqlalchemy.update(AssetCacheState)
-                .where(AssetCacheState.id.in_(to_set_verify))
-                .values(needs_verify=True)
-            )
-        if to_clear_verify:
-            sess.execute(
-                sqlalchemy.update(AssetCacheState)
-                .where(AssetCacheState.id.in_(to_clear_verify))
-                .values(needs_verify=False)
-            )
-        sess.commit()
-        return survivors if collect_existing_paths else None
+    if enable_logging:
+        logging.info(
+            "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, orphans_pruned=%d, total_seen=%d)",
+            roots,
+            time.perf_counter() - t_start,
+            created,
+            skipped_existing,
+            orphans_pruned,
+            len(paths),
+        )

app/assets/services/__init__.py

@@ -0,0 +1,91 @@
+from app.assets.services.asset_management import (
+    asset_exists,
+    delete_asset_reference,
+    get_asset_by_hash,
+    get_asset_detail,
+    list_assets_page,
+    resolve_asset_for_download,
+    set_asset_preview,
+    update_asset_metadata,
+)
+from app.assets.services.bulk_ingest import (
+    BulkInsertResult,
+    batch_insert_seed_assets,
+    prune_orphaned_assets,
+)
+from app.assets.services.file_utils import (
+    get_mtime_ns,
+    get_size_and_mtime_ns,
+    list_files_recursively,
+    verify_file_unchanged,
+)
+from app.assets.services.ingest import (
+    DependencyMissingError,
+    HashMismatchError,
+    create_from_hash,
+    ingest_file_from_path,
+    register_existing_asset,
+    upload_from_temp_path,
+)
+from app.assets.services.schemas import (
+    AddTagsResult,
+    AssetData,
+    AssetDetailResult,
+    AssetInfoData,
+    AssetSummaryData,
+    DownloadResolutionResult,
+    IngestResult,
+    ListAssetsResult,
+    RegisterAssetResult,
+    RemoveTagsResult,
+    SetTagsResult,
+    TagUsage,
+    UploadResult,
+    UserMetadata,
+)
+from app.assets.services.tagging import (
+    apply_tags,
+    list_tags,
+    remove_tags,
+)
+
+__all__ = [
+    "AddTagsResult",
+    "AssetData",
+    "AssetDetailResult",
+    "AssetInfoData",
+    "AssetSummaryData",
+    "BulkInsertResult",
+    "DependencyMissingError",
+    "DownloadResolutionResult",
+    "HashMismatchError",
+    "IngestResult",
+    "ListAssetsResult",
+    "RegisterAssetResult",
+    "RemoveTagsResult",
+    "SetTagsResult",
+    "TagUsage",
+    "UploadResult",
+    "UserMetadata",
+    "apply_tags",
+    "asset_exists",
+    "batch_insert_seed_assets",
+    "create_from_hash",
+    "delete_asset_reference",
+    "get_asset_by_hash",
+    "get_asset_detail",
+    "get_mtime_ns",
+    "get_size_and_mtime_ns",
+    "ingest_file_from_path",
+    "list_assets_page",
+    "list_files_recursively",
+    "list_tags",
+    "prune_orphaned_assets",
+    "register_existing_asset",
+    "remove_tags",
+    "resolve_asset_for_download",
+    "set_asset_preview",
+    "update_asset_metadata",
+    "upload_from_temp_path",
+    "verify_file_unchanged",
+]

app/assets/services/asset_management.py

@@ -0,0 +1,290 @@
+import contextlib
+import mimetypes
+import os
+from typing import Sequence
+
+
+from app.assets.database.models import Asset
+from app.assets.database.queries import (
+    asset_exists_by_hash,
+    asset_info_exists_for_asset_id,
+    delete_asset_info_by_id,
+    fetch_asset_info_and_asset,
+    fetch_asset_info_asset_and_tags,
+    get_asset_by_hash as queries_get_asset_by_hash,
+    get_asset_info_by_id,
+    list_asset_infos_page,
+    list_cache_states_by_asset_id,
+    set_asset_info_metadata,
+    set_asset_info_preview,
+    set_asset_info_tags,
+    update_asset_info_access_time,
+    update_asset_info_name,
+    update_asset_info_updated_at,
+)
+from app.assets.helpers import select_best_live_path
+from app.assets.services.path_utils import compute_filename_for_asset
+from app.assets.services.schemas import (
+    AssetData,
+    AssetDetailResult,
+    AssetSummaryData,
+    DownloadResolutionResult,
+    ListAssetsResult,
+    UserMetadata,
+    extract_asset_data,
+    extract_info_data,
+)
+from app.database.db import create_session
+
+
+def get_asset_detail(
+    asset_info_id: str,
+    owner_id: str = "",
+) -> AssetDetailResult | None:
+    with create_session() as session:
+        result = fetch_asset_info_asset_and_tags(
+            session,
+            asset_info_id=asset_info_id,
+            owner_id=owner_id,
+        )
+        if not result:
+            return None
+
+        info, asset, tags = result
+        return AssetDetailResult(
+            info=extract_info_data(info),
+            asset=extract_asset_data(asset),
+            tags=tags,
+        )
+
+
+def update_asset_metadata(
+    asset_info_id: str,
+    name: str | None = None,
+    tags: Sequence[str] | None = None,
+    user_metadata: UserMetadata = None,
+    tag_origin: str = "manual",
+    owner_id: str = "",
+) -> AssetDetailResult:
+    with create_session() as session:
+        info = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info.owner_id and info.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        touched = False
+        if name is not None and name != info.name:
+            update_asset_info_name(session, asset_info_id=asset_info_id, name=name)
+            touched = True
+
+        computed_filename = compute_filename_for_asset(session, info.asset_id)
+
+        new_meta: dict | None = None
+        if user_metadata is not None:
+            new_meta = dict(user_metadata)
+        elif computed_filename:
+            current_meta = info.user_metadata or {}
+            if current_meta.get("filename") != computed_filename:
+                new_meta = dict(current_meta)
+
+        if new_meta is not None:
+            if computed_filename:
+                new_meta["filename"] = computed_filename
+            set_asset_info_metadata(
+                session, asset_info_id=asset_info_id, user_metadata=new_meta
+            )
+            touched = True
+
+        if tags is not None:
+            set_asset_info_tags(
+                session,
+                asset_info_id=asset_info_id,
+                tags=tags,
+                origin=tag_origin,
+            )
+            touched = True
+
+        if touched and user_metadata is None:
+            update_asset_info_updated_at(session, asset_info_id=asset_info_id)
+
+        result = fetch_asset_info_asset_and_tags(
+            session,
+            asset_info_id=asset_info_id,
+            owner_id=owner_id,
+        )
+        if not result:
+            raise RuntimeError("State changed during update")
+
+        info, asset, tag_list = result
+        detail = AssetDetailResult(
+            info=extract_info_data(info),
+            asset=extract_asset_data(asset),
+            tags=tag_list,
+        )
+        session.commit()
+
+        return detail
+
+
+def delete_asset_reference(
+    asset_info_id: str,
+    owner_id: str,
+    delete_content_if_orphan: bool = True,
+) -> bool:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        asset_id = info_row.asset_id if info_row else None
+
+        deleted = delete_asset_info_by_id(
+            session, asset_info_id=asset_info_id, owner_id=owner_id
+        )
+        if not deleted:
+            session.commit()
+            return False
+
+        if not delete_content_if_orphan or not asset_id:
+            session.commit()
+            return True
+
+        still_exists = asset_info_exists_for_asset_id(session, asset_id=asset_id)
+        if still_exists:
+            session.commit()
+            return True
+
+        # Orphaned asset - delete it and its files
+        states = list_cache_states_by_asset_id(session, asset_id=asset_id)
+        file_paths = [
+            s.file_path for s in (states or []) if getattr(s, "file_path", None)
+        ]
+
+        asset_row = session.get(Asset, asset_id)
+        if asset_row is not None:
+            session.delete(asset_row)
+
+        session.commit()
+
+        # Delete files after commit
+        for p in file_paths:
+            with contextlib.suppress(Exception):
+                if p and os.path.isfile(p):
+                    os.remove(p)
+
+    return True
+
+
+def set_asset_preview(
+    asset_info_id: str,
+    preview_asset_id: str | None = None,
+    owner_id: str = "",
+) -> AssetDetailResult:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        set_asset_info_preview(
+            session,
+            asset_info_id=asset_info_id,
+            preview_asset_id=preview_asset_id,
+        )
+
+        result = fetch_asset_info_asset_and_tags(
+            session, asset_info_id=asset_info_id, owner_id=owner_id
+        )
+        if not result:
+            raise RuntimeError("State changed during preview update")
+
+        info, asset, tags = result
+        detail = AssetDetailResult(
+            info=extract_info_data(info),
+            asset=extract_asset_data(asset),
+            tags=tags,
+        )
+        session.commit()
+
+        return detail
+
+
+def asset_exists(asset_hash: str) -> bool:
+    with create_session() as session:
+        return asset_exists_by_hash(session, asset_hash=asset_hash)
+
+
+def get_asset_by_hash(asset_hash: str) -> AssetData | None:
+    with create_session() as session:
+        asset = queries_get_asset_by_hash(session, asset_hash=asset_hash)
+        return extract_asset_data(asset)
+
+
+def list_assets_page(
+    owner_id: str = "",
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+    name_contains: str | None = None,
+    metadata_filter: dict | None = None,
+    limit: int = 20,
+    offset: int = 0,
+    sort: str = "created_at",
+    order: str = "desc",
+) -> ListAssetsResult:
+    with create_session() as session:
+        infos, tag_map, total = list_asset_infos_page(
+            session,
+            owner_id=owner_id,
+            include_tags=include_tags,
+            exclude_tags=exclude_tags,
+            name_contains=name_contains,
+            metadata_filter=metadata_filter,
+            limit=limit,
+            offset=offset,
+            sort=sort,
+            order=order,
+        )
+
+        items: list[AssetSummaryData] = []
+        for info in infos:
+            items.append(
+                AssetSummaryData(
+                    info=extract_info_data(info),
+                    asset=extract_asset_data(info.asset),
+                    tags=tag_map.get(info.id, []),
+                )
+            )
+
+        return ListAssetsResult(items=items, total=total)
+
+
+def resolve_asset_for_download(
+    asset_info_id: str,
+    owner_id: str = "",
+) -> DownloadResolutionResult:
+    with create_session() as session:
+        pair = fetch_asset_info_and_asset(
+            session, asset_info_id=asset_info_id, owner_id=owner_id
+        )
+        if not pair:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+        info, asset = pair
+        states = list_cache_states_by_asset_id(session, asset_id=asset.id)
+        abs_path = select_best_live_path(states)
+        if not abs_path:
+            raise FileNotFoundError
+
+        update_asset_info_access_time(session, asset_info_id=asset_info_id)
+        session.commit()
+
+        ctype = (
+            asset.mime_type
+            or mimetypes.guess_type(info.name or abs_path)[0]
+            or "application/octet-stream"
+        )
+        download_name = info.name or os.path.basename(abs_path)
+        return DownloadResolutionResult(
+            abs_path=abs_path,
+            content_type=ctype,
+            download_name=download_name,
+        )

app/assets/services/bulk_ingest.py

@@ -0,0 +1,203 @@
+import os
+import uuid
+from dataclasses import dataclass
+from typing import TypedDict
+
+from sqlalchemy.orm import Session
+
+
+class SeedAssetSpec(TypedDict):
+    """Spec for seeding an asset from filesystem."""
+
+    abs_path: str
+    size_bytes: int
+    mtime_ns: int
+    info_name: str
+    tags: list[str]
+    fname: str
+
+from app.assets.database.queries import (
+    bulk_insert_asset_infos_ignore_conflicts,
+    bulk_insert_assets,
+    bulk_insert_cache_states_ignore_conflicts,
+    bulk_insert_tags_and_meta,
+    delete_assets_by_ids,
+    delete_cache_states_outside_prefixes,
+    get_asset_info_ids_by_ids,
+    get_cache_states_by_paths_and_asset_ids,
+    get_orphaned_seed_asset_ids,
+)
+from app.assets.helpers import get_utc_now
+
+
+@dataclass
+class BulkInsertResult:
+    """Result of bulk asset insertion."""
+
+    inserted_infos: int
+    won_states: int
+    lost_states: int
+
+
+def batch_insert_seed_assets(
+    session: Session,
+    specs: list[SeedAssetSpec],
+    owner_id: str = "",
+) -> BulkInsertResult:
+    """Seed assets from filesystem specs in batch.
+
+    Each spec is a dict with keys:
+      - abs_path: str
+      - size_bytes: int
+      - mtime_ns: int
+      - info_name: str
+      - tags: list[str]
+      - fname: Optional[str]
+
+    This function orchestrates:
+    1. Insert seed Assets (hash=NULL)
+    2. Claim cache states with ON CONFLICT DO NOTHING
+    3. Query to find winners (paths where our asset_id was inserted)
+    4. Delete Assets for losers (path already claimed by another asset)
+    5. Insert AssetInfo for winners
+    6. Insert tags and metadata for successfully inserted AssetInfos
+
+    Returns:
+        BulkInsertResult with inserted_infos, won_states, lost_states
+    """
+    if not specs:
+        return BulkInsertResult(inserted_infos=0, won_states=0, lost_states=0)
+
+    now = get_utc_now()
+    asset_rows: list[dict] = []
+    state_rows: list[dict] = []
+    path_to_asset: dict[str, str] = {}
+    asset_to_info: dict[str, dict] = {}
+    path_list: list[str] = []
+
+    for sp in specs:
+        ap = os.path.abspath(sp["abs_path"])
+        aid = str(uuid.uuid4())
+        iid = str(uuid.uuid4())
+        path_list.append(ap)
+        path_to_asset[ap] = aid
+
+        asset_rows.append(
+            {
+                "id": aid,
+                "hash": None,
+                "size_bytes": sp["size_bytes"],
+                "mime_type": None,
+                "created_at": now,
+            }
+        )
+        state_rows.append(
+            {
+                "asset_id": aid,
+                "file_path": ap,
+                "mtime_ns": sp["mtime_ns"],
+            }
+        )
+        asset_to_info[aid] = {
+            "id": iid,
+            "owner_id": owner_id,
+            "name": sp["info_name"],
+            "asset_id": aid,
+            "preview_id": None,
+            "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None,
+            "created_at": now,
+            "updated_at": now,
+            "last_access_time": now,
+            "_tags": sp["tags"],
+            "_filename": sp["fname"],
+        }
+
+    bulk_insert_assets(session, asset_rows)
+    bulk_insert_cache_states_ignore_conflicts(session, state_rows)
+    winners_by_path = get_cache_states_by_paths_and_asset_ids(session, path_to_asset)
+
+    all_paths_set = set(path_list)
+    losers_by_path = all_paths_set - winners_by_path
+    lost_assets = [path_to_asset[p] for p in losers_by_path]
+
+    if lost_assets:
+        delete_assets_by_ids(session, lost_assets)
+
+    if not winners_by_path:
+        return BulkInsertResult(
+            inserted_infos=0,
+            won_states=0,
+            lost_states=len(losers_by_path),
+        )
+
+    winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path]
+    db_info_rows = [
+        {
+            "id": row["id"],
+            "owner_id": row["owner_id"],
+            "name": row["name"],
+            "asset_id": row["asset_id"],
+            "preview_id": row["preview_id"],
+            "user_metadata": row["user_metadata"],
+            "created_at": row["created_at"],
+            "updated_at": row["updated_at"],
+            "last_access_time": row["last_access_time"],
+        }
+        for row in winner_info_rows
+    ]
+    bulk_insert_asset_infos_ignore_conflicts(session, db_info_rows)
+
+    all_info_ids = [row["id"] for row in winner_info_rows]
+    inserted_info_ids = get_asset_info_ids_by_ids(session, all_info_ids)
+
+    tag_rows: list[dict] = []
+    meta_rows: list[dict] = []
+    if inserted_info_ids:
+        for row in winner_info_rows:
+            iid = row["id"]
+            if iid not in inserted_info_ids:
+                continue
+            for t in row["_tags"]:
+                tag_rows.append(
+                    {
+                        "asset_info_id": iid,
+                        "tag_name": t,
+                        "origin": "automatic",
+                        "added_at": now,
+                    }
+                )
+            if row["_filename"]:
+                meta_rows.append(
+                    {
+                        "asset_info_id": iid,
+                        "key": "filename",
+                        "ordinal": 0,
+                        "val_str": row["_filename"],
+                        "val_num": None,
+                        "val_bool": None,
+                        "val_json": None,
+                    }
+                )
+
+    bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows)
+
+    return BulkInsertResult(
+        inserted_infos=len(inserted_info_ids),
+        won_states=len(winners_by_path),
+        lost_states=len(losers_by_path),
+    )
+
+
+def prune_orphaned_assets(session: Session, valid_prefixes: list[str]) -> int:
+    """Prune cache states outside valid prefixes, then delete orphaned seed assets.
+
+    Args:
+        session: Database session
+        valid_prefixes: List of absolute directory prefixes that are valid
+
+    Returns:
+        Number of orphaned assets deleted
+    """
+    delete_cache_states_outside_prefixes(session, valid_prefixes)
+    orphan_ids = get_orphaned_seed_asset_ids(session)
+    return delete_assets_by_ids(session, orphan_ids)

app/assets/services/file_utils.py

@@ -0,0 +1,49 @@
+import os
+
+
+def get_mtime_ns(stat_result: os.stat_result) -> int:
+    """Extract mtime in nanoseconds from a stat result."""
+    return getattr(
+        stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000)
+    )
+
+
+def get_size_and_mtime_ns(path: str, follow_symlinks: bool = True) -> tuple[int, int]:
+    """Get file size in bytes and mtime in nanoseconds."""
+    st = os.stat(path, follow_symlinks=follow_symlinks)
+    return st.st_size, get_mtime_ns(st)
+
+
+def verify_file_unchanged(
+    mtime_db: int | None,
+    size_db: int | None,
+    stat_result: os.stat_result,
+) -> bool:
+    """Check if a file is unchanged based on mtime and size.
+
+    Returns True if the file's mtime and size match the database values.
+    Returns False if mtime_db is None or values don't match.
+    """
+    if mtime_db is None:
+        return False
+    actual_mtime_ns = get_mtime_ns(stat_result)
+    if int(mtime_db) != int(actual_mtime_ns):
+        return False
+    sz = int(size_db or 0)
+    if sz > 0:
+        return int(stat_result.st_size) == sz
+    return True
+
+
+def list_files_recursively(base_dir: str) -> list[str]:
+    """Recursively list all files in a directory."""
+    out: list[str] = []
+    base_abs = os.path.abspath(base_dir)
+    if not os.path.isdir(base_abs):
+        return out
+    for dirpath, _subdirs, filenames in os.walk(
+        base_abs, topdown=True, followlinks=False
+    ):
+        for name in filenames:
+            out.append(os.path.abspath(os.path.join(dirpath, name)))
+    return out

app/assets/services/hashing.py

@@ -0,0 +1,54 @@
+import asyncio
+import os
+from typing import IO
+
+from blake3 import blake3
+
+DEFAULT_CHUNK = 8 * 1024 * 1024
+
+
+def compute_blake3_hash(
+    fp: str | IO[bytes],
+    chunk_size: int = DEFAULT_CHUNK,
+) -> str:
+    if hasattr(fp, "read"):
+        return _hash_file_obj(fp, chunk_size)
+
+    with open(os.fspath(fp), "rb") as f:
+        return _hash_file_obj(f, chunk_size)
+
+
+async def compute_compute_blake3_hash_async(
+    fp: str | IO[bytes],
+    chunk_size: int = DEFAULT_CHUNK,
+) -> str:
+    if hasattr(fp, "read"):
+        return await asyncio.to_thread(compute_blake3_hash, fp, chunk_size)
+
+    def _worker() -> str:
+        with open(os.fspath(fp), "rb") as f:
+            return _hash_file_obj(f, chunk_size)
+
+    return await asyncio.to_thread(_worker)
+
+
+def _hash_file_obj(file_obj: IO, chunk_size: int = DEFAULT_CHUNK) -> str:
+    if chunk_size <= 0:
+        chunk_size = DEFAULT_CHUNK
+
+    orig_pos = file_obj.tell()
+
+    try:
+        if orig_pos != 0:
+            file_obj.seek(0)
+
+        h = blake3()
+        while True:
+            chunk = file_obj.read(chunk_size)
+            if not chunk:
+                break
+            h.update(chunk)
+        return h.hexdigest()
+    finally:
+        if orig_pos != 0:
+            file_obj.seek(orig_pos)

app/assets/services/ingest.py

@@ -0,0 +1,388 @@
+import contextlib
+import logging
+import mimetypes
+import os
+from typing import Sequence
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+import app.assets.services.hashing as hashing
+from app.assets.database.models import Asset, AssetInfo, Tag
+from app.assets.database.queries import (
+    add_tags_to_asset_info,
+    fetch_asset_info_and_asset,
+    get_asset_by_hash,
+    get_asset_tags,
+    get_or_create_asset_info,
+    remove_missing_tag_for_asset_id,
+    set_asset_info_metadata,
+    set_asset_info_tags,
+    update_asset_info_timestamps,
+    upsert_asset,
+    upsert_cache_state,
+)
+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_filename_for_asset,
+    resolve_destination_from_tags,
+    validate_path_within_base,
+)
+from app.assets.services.schemas import (
+    IngestResult,
+    RegisterAssetResult,
+    UploadResult,
+    UserMetadata,
+    extract_asset_data,
+    extract_info_data,
+)
+from app.database.db import create_session
+
+
+def ingest_file_from_path(
+    abs_path: str,
+    asset_hash: str,
+    size_bytes: int,
+    mtime_ns: int,
+    mime_type: str | None = None,
+    info_name: str | None = None,
+    owner_id: str = "",
+    preview_id: str | None = None,
+    user_metadata: UserMetadata = None,
+    tags: Sequence[str] = (),
+    tag_origin: str = "manual",
+    require_existing_tags: bool = False,
+) -> IngestResult:
+    locator = os.path.abspath(abs_path)
+
+    asset_created = False
+    asset_updated = False
+    state_created = False
+    state_updated = False
+    asset_info_id: str | None = None
+
+    with create_session() as session:
+        if preview_id:
+            if not session.get(Asset, preview_id):
+                preview_id = None
+
+        asset, asset_created, asset_updated = upsert_asset(
+            session,
+            asset_hash=asset_hash,
+            size_bytes=size_bytes,
+            mime_type=mime_type,
+        )
+
+        state_created, state_updated = upsert_cache_state(
+            session,
+            asset_id=asset.id,
+            file_path=locator,
+            mtime_ns=mtime_ns,
+        )
+
+        if info_name:
+            info, info_created = get_or_create_asset_info(
+                session,
+                asset_id=asset.id,
+                owner_id=owner_id,
+                name=info_name,
+                preview_id=preview_id,
+            )
+            if info_created:
+                asset_info_id = info.id
+            else:
+                update_asset_info_timestamps(
+                    session, asset_info=info, preview_id=preview_id
+                )
+                asset_info_id = info.id
+
+            norm = normalize_tags(list(tags))
+            if norm and asset_info_id:
+                if require_existing_tags:
+                    _validate_tags_exist(session, norm)
+                add_tags_to_asset_info(
+                    session,
+                    asset_info_id=asset_info_id,
+                    tags=norm,
+                    origin=tag_origin,
+                    create_if_missing=not require_existing_tags,
+                )
+
+            if asset_info_id:
+                _update_metadata_with_filename(
+                    session,
+                    asset_info_id=asset_info_id,
+                    asset_id=asset.id,
+                    info=info,
+                    user_metadata=user_metadata,
+                )
+
+        try:
+            remove_missing_tag_for_asset_id(session, asset_id=asset.id)
+        except Exception:
+            logging.exception("Failed to clear 'missing' tag for asset %s", asset.id)
+
+        session.commit()
+
+    return IngestResult(
+        asset_created=asset_created,
+        asset_updated=asset_updated,
+        state_created=state_created,
+        state_updated=state_updated,
+        asset_info_id=asset_info_id,
+    )
+
+
+def register_existing_asset(
+    asset_hash: str,
+    name: str,
+    user_metadata: UserMetadata = None,
+    tags: list[str] | None = None,
+    tag_origin: str = "manual",
+    owner_id: str = "",
+) -> RegisterAssetResult:
+    with create_session() as session:
+        asset = get_asset_by_hash(session, asset_hash=asset_hash)
+        if not asset:
+            raise ValueError(f"No asset with hash {asset_hash}")
+
+        info, info_created = get_or_create_asset_info(
+            session,
+            asset_id=asset.id,
+            owner_id=owner_id,
+            name=name,
+            preview_id=None,
+        )
+
+        if not info_created:
+            tag_names = get_asset_tags(session, asset_info_id=info.id)
+            result = RegisterAssetResult(
+                info=extract_info_data(info),
+                asset=extract_asset_data(asset),
+                tags=tag_names,
+                created=False,
+            )
+            session.commit()
+            return result
+
+        new_meta = dict(user_metadata or {})
+        computed_filename = compute_filename_for_asset(session, asset.id)
+        if computed_filename:
+            new_meta["filename"] = computed_filename
+
+        if new_meta:
+            set_asset_info_metadata(
+                session,
+                asset_info_id=info.id,
+                user_metadata=new_meta,
+            )
+
+        if tags is not None:
+            set_asset_info_tags(
+                session,
+                asset_info_id=info.id,
+                tags=tags,
+                origin=tag_origin,
+            )
+
+        tag_names = get_asset_tags(session, asset_info_id=info.id)
+        session.refresh(info)
+        result = RegisterAssetResult(
+            info=extract_info_data(info),
+            asset=extract_asset_data(asset),
+            tags=tag_names,
+            created=True,
+        )
+        session.commit()
+
+        return result
+
+
+def _validate_tags_exist(session: Session, tags: list[str]) -> None:
+    existing_tag_names = set(
+        name
+        for (name,) in session.execute(select(Tag.name).where(Tag.name.in_(tags))).all()
+    )
+    missing = [t for t in tags if t not in existing_tag_names]
+    if missing:
+        raise ValueError(f"Unknown tags: {missing}")
+
+
+def _update_metadata_with_filename(
+    session: Session,
+    asset_info_id: str,
+    asset_id: str,
+    info: AssetInfo,
+    user_metadata: UserMetadata,
+) -> None:
+    computed_filename = compute_filename_for_asset(session, asset_id)
+
+    current_meta = info.user_metadata or {}
+    new_meta = dict(current_meta)
+    if user_metadata:
+        for k, v in user_metadata.items():
+            new_meta[k] = v
+    if computed_filename:
+        new_meta["filename"] = computed_filename
+
+    if new_meta != current_meta:
+        set_asset_info_metadata(
+            session,
+            asset_info_id=asset_info_id,
+            user_metadata=new_meta,
+        )
+
+
+def _sanitize_filename(name: str | None, fallback: str) -> str:
+    n = os.path.basename((name or "").strip() or fallback)
+    return n if n else fallback
+
+
+class HashMismatchError(Exception):
+    pass
+
+
+class DependencyMissingError(Exception):
+    def __init__(self, message: str):
+        self.message = message
+        super().__init__(message)
+
+
+def upload_from_temp_path(
+    temp_path: str,
+    name: str | None = None,
+    tags: list[str] | None = None,
+    user_metadata: dict | None = None,
+    client_filename: str | None = None,
+    owner_id: str = "",
+    expected_hash: str | None = None,
+) -> UploadResult:
+    try:
+        digest = hashing.compute_blake3_hash(temp_path)
+    except ImportError as e:
+        raise DependencyMissingError(str(e))
+    except Exception as e:
+        raise RuntimeError(f"failed to hash uploaded file: {e}")
+    asset_hash = "blake3:" + digest
+
+    if expected_hash and asset_hash != expected_hash.strip().lower():
+        raise HashMismatchError("Uploaded file hash does not match provided hash.")
+
+    with create_session() as session:
+        existing = get_asset_by_hash(session, asset_hash=asset_hash)
+
+    if existing is not None:
+        with contextlib.suppress(Exception):
+            if temp_path and os.path.exists(temp_path):
+                os.remove(temp_path)
+
+        display_name = _sanitize_filename(name or client_filename, fallback=digest)
+        result = register_existing_asset(
+            asset_hash=asset_hash,
+            name=display_name,
+            user_metadata=user_metadata or {},
+            tags=tags or [],
+            tag_origin="manual",
+            owner_id=owner_id,
+        )
+        return UploadResult(
+            info=result.info,
+            asset=result.asset,
+            tags=result.tags,
+            created_new=False,
+        )
+
+    base_dir, subdirs = resolve_destination_from_tags(tags)
+    dest_dir = os.path.join(base_dir, *subdirs) if subdirs else base_dir
+    os.makedirs(dest_dir, exist_ok=True)
+
+    src_for_ext = (client_filename or name or "").strip()
+    _ext = os.path.splitext(os.path.basename(src_for_ext))[1] if src_for_ext else ""
+    ext = _ext if 0 < len(_ext) <= 16 else ""
+    hashed_basename = f"{digest}{ext}"
+    dest_abs = os.path.abspath(os.path.join(dest_dir, hashed_basename))
+    validate_path_within_base(dest_abs, base_dir)
+
+    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"
+    )
+
+    try:
+        os.replace(temp_path, dest_abs)
+    except Exception as e:
+        raise RuntimeError(f"failed to move uploaded file into place: {e}")
+
+    try:
+        size_bytes, mtime_ns = get_size_and_mtime_ns(dest_abs)
+    except OSError as e:
+        raise RuntimeError(f"failed to stat destination file: {e}")
+
+    ingest_result = ingest_file_from_path(
+        asset_hash=asset_hash,
+        abs_path=dest_abs,
+        size_bytes=size_bytes,
+        mtime_ns=mtime_ns,
+        mime_type=content_type,
+        info_name=_sanitize_filename(name or client_filename, fallback=digest),
+        owner_id=owner_id,
+        preview_id=None,
+        user_metadata=user_metadata or {},
+        tags=tags,
+        tag_origin="manual",
+        require_existing_tags=False,
+    )
+    info_id = ingest_result.asset_info_id
+    if not info_id:
+        raise RuntimeError("failed to create asset metadata")
+
+    with create_session() as session:
+        pair = fetch_asset_info_and_asset(
+            session, asset_info_id=info_id, owner_id=owner_id
+        )
+        if not pair:
+            raise RuntimeError("inconsistent DB state after ingest")
+        info, asset = pair
+        tag_names = get_asset_tags(session, asset_info_id=info.id)
+
+    return UploadResult(
+        info=extract_info_data(info),
+        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 = "",
+) -> UploadResult | None:
+    canonical = hash_str.strip().lower()
+
+    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(
+        info=result.info,
+        asset=result.asset,
+        tags=result.tags,
+        created_new=False,
+    )

app/assets/services/path_utils.py

@@ -0,0 +1,184 @@
+import os
+from pathlib import Path
+from typing import Literal
+
+import folder_paths
+from app.assets.helpers import normalize_tags
+
+
+def get_comfy_models_folders() -> list[tuple[str, list[str]]]:
+    """Build a list of (folder_name, base_paths[]) categories that are configured for model locations.
+
+    We trust `folder_paths.folder_names_and_paths` and include a category if
+    *any* of its base paths lies under the Comfy `models_dir`.
+    """
+    targets: list[tuple[str, list[str]]] = []
+    models_root = os.path.abspath(folder_paths.models_dir)
+    for name, values in folder_paths.folder_names_and_paths.items():
+        paths, _exts = (
+            values[0],
+            values[1],
+        )  # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI
+        if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths):
+            targets.append((name, paths))
+    return targets
+
+
+def resolve_destination_from_tags(tags: list[str]) -> tuple[str, list[str]]:
+    """Validates and maps tags -> (base_dir, subdirs_for_fs)"""
+    root = tags[0]
+    if root == "models":
+        if len(tags) < 2:
+            raise ValueError("at least two tags required for model asset")
+        try:
+            bases = folder_paths.folder_names_and_paths[tags[1]][0]
+        except KeyError:
+            raise ValueError(f"unknown model category '{tags[1]}'")
+        if not bases:
+            raise ValueError(f"no base path configured for category '{tags[1]}'")
+        base_dir = os.path.abspath(bases[0])
+        raw_subdirs = tags[2:]
+    else:
+        base_dir = os.path.abspath(
+            folder_paths.get_input_directory()
+            if root == "input"
+            else folder_paths.get_output_directory()
+        )
+        raw_subdirs = tags[1:]
+    for i in raw_subdirs:
+        if i in (".", ".."):
+            raise ValueError("invalid path component in tags")
+
+    return base_dir, raw_subdirs if raw_subdirs else []
+
+
+def validate_path_within_base(candidate: str, base: str) -> None:
+    cand_abs = os.path.abspath(candidate)
+    base_abs = os.path.abspath(base)
+    try:
+        if os.path.commonpath([cand_abs, base_abs]) != base_abs:
+            raise ValueError("destination escapes base directory")
+    except Exception:
+        raise ValueError("invalid destination path")
+
+
+def compute_relative_filename(file_path: str) -> str | None:
+    """
+    Return the model's path relative to the last well-known folder (the model category),
+    using forward slashes, eg:
+      /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors"
+      /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors"
+
+    For non-model paths, returns None.
+    NOTE: this is a temporary helper, used only for initializing metadata["filename"] field.
+    """
+    try:
+        root_category, rel_path = get_asset_category_and_relative_path(file_path)
+    except ValueError:
+        return None
+
+    p = Path(rel_path)
+    parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)]
+    if not parts:
+        return None
+
+    if root_category == "models":
+        # parts[0] is the category ("checkpoints", "vae", etc) – drop it
+        inside = parts[1:] if len(parts) > 1 else [parts[0]]
+        return "/".join(inside)
+    return "/".join(parts)  # input/output: keep all parts
+
+
+def get_asset_category_and_relative_path(
+    file_path: str,
+) -> tuple[Literal["input", "output", "models"], str]:
+    """Given an absolute or relative file path, determine which root category the path belongs to:
+      - 'input' if the file resides under `folder_paths.get_input_directory()`
+      - 'output' if the file resides under `folder_paths.get_output_directory()`
+      - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()`
+
+    Returns:
+        (root_category, relative_path_inside_that_root)
+        For 'models', the relative path is prefixed with the category name:
+            e.g. ('models', 'vae/test/sub/ae.safetensors')
+
+    Raises:
+        ValueError: if the path does not belong to input, output, or configured model bases.
+    """
+    fp_abs = os.path.abspath(file_path)
+
+    def _check_is_within(child: str, parent: str) -> bool:
+        try:
+            return os.path.commonpath([child, parent]) == parent
+        except Exception:
+            return False
+
+    def _compute_relative(child: str, parent: str) -> str:
+        return os.path.relpath(
+            os.path.join(os.sep, os.path.relpath(child, parent)), os.sep
+        )
+
+    # 1) input
+    input_base = os.path.abspath(folder_paths.get_input_directory())
+    if _check_is_within(fp_abs, input_base):
+        return "input", _compute_relative(fp_abs, input_base)
+
+    # 2) output
+    output_base = os.path.abspath(folder_paths.get_output_directory())
+    if _check_is_within(fp_abs, output_base):
+        return "output", _compute_relative(fp_abs, output_base)
+
+    # 3) models (check deepest matching base to avoid ambiguity)
+    best: tuple[int, str, str] | None = None  # (base_len, bucket, rel_inside_bucket)
+    for bucket, bases in get_comfy_models_folders():
+        for b in bases:
+            base_abs = os.path.abspath(b)
+            if not _check_is_within(fp_abs, base_abs):
+                continue
+            cand = (len(base_abs), bucket, _compute_relative(fp_abs, base_abs))
+            if best is None or cand[0] > best[0]:
+                best = cand
+
+    if best is not None:
+        _, bucket, rel_inside = best
+        combined = os.path.join(bucket, rel_inside)
+        return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep)
+
+    raise ValueError(
+        f"Path is not within input, output, or configured model bases: {file_path}"
+    )
+
+
+def compute_filename_for_asset(session, asset_id: str) -> str | None:
+    """Compute the relative filename for an asset from its best live cache state path."""
+    from app.assets.database.queries import list_cache_states_by_asset_id
+    from app.assets.helpers import select_best_live_path
+
+    primary_path = select_best_live_path(
+        list_cache_states_by_asset_id(session, asset_id=asset_id)
+    )
+    return compute_relative_filename(primary_path) if primary_path else None
+
+
+def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:
+    """Return a tuple (name, tags) derived from a filesystem path.
+
+    Semantics:
+      - Root category is determined by `get_asset_category_and_relative_path`.
+      - The returned `name` is the base filename with extension from the relative path.
+      - The returned `tags` are:
+            [root_category] + parent folders of the relative path (in order)
+        For 'models', this means:
+            file '/.../ModelsDir/vae/test_tag/ae.safetensors'
+            -> root_category='models', some_path='vae/test_tag/ae.safetensors'
+            -> name='ae.safetensors', tags=['models', 'vae', 'test_tag']
+
+    Raises:
+        ValueError: if the path does not belong to input, output, or configured model bases.
+    """
+    root_category, some_path = get_asset_category_and_relative_path(file_path)
+    p = Path(some_path)
+    parent_parts = [
+        part for part in p.parent.parts if part not in (".", "..", p.anchor)
+    ]
+    return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts])))

app/assets/services/schemas.py

@@ -0,0 +1,126 @@
+from dataclasses import dataclass
+from datetime import datetime
+from typing import Any, NamedTuple
+
+from app.assets.database.models import Asset, AssetInfo
+
+UserMetadata = dict[str, Any] | None
+
+
+@dataclass(frozen=True)
+class AssetData:
+    hash: str
+    size_bytes: int | None
+    mime_type: str | None
+
+
+@dataclass(frozen=True)
+class AssetInfoData:
+    id: str
+    name: str
+    user_metadata: UserMetadata
+    preview_id: str | None
+    created_at: datetime
+    updated_at: datetime
+    last_access_time: datetime | None
+
+
+@dataclass(frozen=True)
+class AssetDetailResult:
+    info: AssetInfoData
+    asset: AssetData | None
+    tags: list[str]
+
+
+@dataclass(frozen=True)
+class RegisterAssetResult:
+    info: AssetInfoData
+    asset: AssetData
+    tags: list[str]
+    created: bool
+
+
+@dataclass(frozen=True)
+class IngestResult:
+    asset_created: bool
+    asset_updated: bool
+    state_created: bool
+    state_updated: bool
+    asset_info_id: str | None
+
+
+@dataclass(frozen=True)
+class AddTagsResult:
+    added: list[str]
+    already_present: list[str]
+    total_tags: list[str]
+
+
+@dataclass(frozen=True)
+class RemoveTagsResult:
+    removed: list[str]
+    not_present: list[str]
+    total_tags: list[str]
+
+
+@dataclass(frozen=True)
+class SetTagsResult:
+    added: list[str]
+    removed: list[str]
+    total: list[str]
+
+
+class TagUsage(NamedTuple):
+    name: str
+    tag_type: str
+    count: int
+
+
+@dataclass(frozen=True)
+class AssetSummaryData:
+    info: AssetInfoData
+    asset: AssetData | None
+    tags: list[str]
+
+
+@dataclass(frozen=True)
+class ListAssetsResult:
+    items: list[AssetSummaryData]
+    total: int
+
+
+@dataclass(frozen=True)
+class DownloadResolutionResult:
+    abs_path: str
+    content_type: str
+    download_name: str
+
+
+@dataclass(frozen=True)
+class UploadResult:
+    info: AssetInfoData
+    asset: AssetData
+    tags: list[str]
+    created_new: bool
+
+
+def extract_info_data(info: AssetInfo) -> AssetInfoData:
+    return AssetInfoData(
+        id=info.id,
+        name=info.name,
+        user_metadata=info.user_metadata,
+        preview_id=info.preview_id,
+        created_at=info.created_at,
+        updated_at=info.updated_at,
+        last_access_time=info.last_access_time,
+    )
+
+
+def extract_asset_data(asset: Asset | None) -> AssetData | None:
+    if asset is None:
+        return None
+    return AssetData(
+        hash=asset.hash,
+        size_bytes=asset.size_bytes,
+        mime_type=asset.mime_type,
+    )

app/assets/services/tagging.py

@@ -0,0 +1,89 @@
+from app.assets.database.queries import (
+    add_tags_to_asset_info,
+    get_asset_info_by_id,
+    list_tags_with_usage,
+    remove_tags_from_asset_info,
+)
+from app.assets.services.schemas import AddTagsResult, RemoveTagsResult, TagUsage
+from app.database.db import create_session
+
+
+def apply_tags(
+    asset_info_id: str,
+    tags: list[str],
+    origin: str = "manual",
+    owner_id: str = "",
+) -> AddTagsResult:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        data = add_tags_to_asset_info(
+            session,
+            asset_info_id=asset_info_id,
+            tags=tags,
+            origin=origin,
+            create_if_missing=True,
+            asset_info_row=info_row,
+        )
+        session.commit()
+
+    return AddTagsResult(
+        added=data["added"],
+        already_present=data["already_present"],
+        total_tags=data["total_tags"],
+    )
+
+
+def remove_tags(
+    asset_info_id: str,
+    tags: list[str],
+    owner_id: str = "",
+) -> RemoveTagsResult:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        data = remove_tags_from_asset_info(
+            session,
+            asset_info_id=asset_info_id,
+            tags=tags,
+        )
+        session.commit()
+
+    return RemoveTagsResult(
+        removed=data["removed"],
+        not_present=data["not_present"],
+        total_tags=data["total_tags"],
+    )
+
+
+def list_tags(
+    prefix: str | None = None,
+    limit: int = 100,
+    offset: int = 0,
+    order: str = "count_desc",
+    include_zero: bool = True,
+    owner_id: str = "",
+) -> tuple[list[TagUsage], int]:
+    limit = max(1, min(1000, limit))
+    offset = max(0, offset)
+
+    with create_session() as session:
+        rows, total = list_tags_with_usage(
+            session,
+            prefix=prefix,
+            limit=limit,
+            offset=offset,
+            include_zero=include_zero,
+            order=order,
+            owner_id=owner_id,
+        )
+
+    return [TagUsage(name, tag_type, count) for name, tag_type, count in rows], total

app/frontend_management.py

@@ -17,7 +17,7 @@ from importlib.metadata import version
 import requests
 from typing_extensions import NotRequired
 
-from utils.install_util import get_missing_requirements_message, get_required_packages_versions
+from utils.install_util import get_missing_requirements_message, requirements_path
 
 from comfy.cli_args import DEFAULT_VERSION_STRING
 import app.logger
@@ -45,7 +45,25 @@ def get_installed_frontend_version():
 
 
 def get_required_frontend_version():
-    return get_required_packages_versions().get("comfyui-frontend-package", None)
+    """Get the required frontend version from requirements.txt."""
+    try:
+        with open(requirements_path, "r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith("comfyui-frontend-package=="):
+                    version_str = line.split("==")[-1]
+                    if not is_valid_version(version_str):
+                        logging.error(f"Invalid version format in requirements.txt: {version_str}")
+                        return None
+                    return version_str
+            logging.error("comfyui-frontend-package not found in requirements.txt")
+            return None
+    except FileNotFoundError:
+        logging.error("requirements.txt not found. Cannot determine required frontend version.")
+        return None
+    except Exception as e:
+        logging.error(f"Error reading requirements.txt: {e}")
+        return None
 
 
 def check_frontend_version():
@@ -199,7 +217,25 @@ class FrontendManager:
 
     @classmethod
     def get_required_templates_version(cls) -> str:
-        return get_required_packages_versions().get("comfyui-workflow-templates", None)
+        """Get the required workflow templates version from requirements.txt."""
+        try:
+            with open(requirements_path, "r", encoding="utf-8") as f:
+                for line in f:
+                    line = line.strip()
+                    if line.startswith("comfyui-workflow-templates=="):
+                        version_str = line.split("==")[-1]
+                        if not is_valid_version(version_str):
+                            logging.error(f"Invalid templates version format in requirements.txt: {version_str}")
+                            return None
+                        return version_str
+                logging.error("comfyui-workflow-templates not found in requirements.txt")
+                return None
+        except FileNotFoundError:
+            logging.error("requirements.txt not found. Cannot determine required templates version.")
+            return None
+        except Exception as e:
+            logging.error(f"Error reading requirements.txt: {e}")
+            return None
 
     @classmethod
     def default_frontend_path(cls) -> str:

app/node_replace_manager.py

@@ -1,107 +0,0 @@
-from __future__ import annotations
-
-from aiohttp import web
-
-from typing import TYPE_CHECKING, TypedDict
-if TYPE_CHECKING:
-    from comfy_api.latest._io_public import NodeReplace
-
-from comfy_execution.graph_utils import is_link
-import nodes
-
-class NodeStruct(TypedDict):
-    inputs: dict[str, str | int | float | bool | tuple[str, int]]
-    class_type: str
-    _meta: dict[str, str]
-
-def copy_node_struct(node_struct: NodeStruct, empty_inputs: bool = False) -> NodeStruct:
-    new_node_struct = node_struct.copy()
-    if empty_inputs:
-        new_node_struct["inputs"] = {}
-    else:
-        new_node_struct["inputs"] = node_struct["inputs"].copy()
-    new_node_struct["_meta"] = node_struct["_meta"].copy()
-    return new_node_struct
-
-
-class NodeReplaceManager:
-    """Manages node replacement registrations."""
-
-    def __init__(self):
-        self._replacements: dict[str, list[NodeReplace]] = {}
-
-    def register(self, node_replace: NodeReplace):
-        """Register a node replacement mapping."""
-        self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace)
-
-    def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None:
-        """Get replacements for an old node ID."""
-        return self._replacements.get(old_node_id)
-
-    def has_replacement(self, old_node_id: str) -> bool:
-        """Check if a replacement exists for an old node ID."""
-        return old_node_id in self._replacements
-
-    def apply_replacements(self, prompt: dict[str, NodeStruct]):
-        connections: dict[str, list[tuple[str, str, int]]] = {}
-        need_replacement: set[str] = set()
-        for node_number, node_struct in prompt.items():
-            if "class_type" not in node_struct or "inputs" not in node_struct:
-                continue
-            class_type = node_struct["class_type"]
-            # need replacement if not in NODE_CLASS_MAPPINGS and has replacement
-            if class_type not in nodes.NODE_CLASS_MAPPINGS.keys() and self.has_replacement(class_type):
-                need_replacement.add(node_number)
-            # keep track of connections
-            for input_id, input_value in node_struct["inputs"].items():
-                if is_link(input_value):
-                    conn_number = input_value[0]
-                    connections.setdefault(conn_number, []).append((node_number, input_id, input_value[1]))
-        for node_number in need_replacement:
-            node_struct = prompt[node_number]
-            class_type = node_struct["class_type"]
-            replacements = self.get_replacement(class_type)
-            if replacements is None:
-                continue
-            # just use the first replacement
-            replacement = replacements[0]
-            new_node_id = replacement.new_node_id
-            # if replacement is not a valid node, skip trying to replace it as will only cause confusion
-            if new_node_id not in nodes.NODE_CLASS_MAPPINGS.keys():
-                continue
-            # first, replace node id (class_type)
-            new_node_struct = copy_node_struct(node_struct, empty_inputs=True)
-            new_node_struct["class_type"] = new_node_id
-            # TODO: consider replacing display_name in _meta as well for error reporting purposes; would need to query node schema
-            # second, replace inputs
-            if replacement.input_mapping is not None:
-                for input_map in replacement.input_mapping:
-                    if "set_value" in input_map:
-                        new_node_struct["inputs"][input_map["new_id"]] = input_map["set_value"]
-                    elif "old_id" in input_map:
-                        new_node_struct["inputs"][input_map["new_id"]] = node_struct["inputs"][input_map["old_id"]]
-            # finalize input replacement
-            prompt[node_number] = new_node_struct
-            # third, replace outputs
-            if replacement.output_mapping is not None:
-                # re-mapping outputs requires changing the input values of nodes that receive connections from this one
-                if node_number in connections:
-                    for conns in connections[node_number]:
-                        conn_node_number, conn_input_id, old_output_idx = conns
-                        for output_map in replacement.output_mapping:
-                            if output_map["old_idx"] == old_output_idx:
-                                new_output_idx = output_map["new_idx"]
-                                previous_input = prompt[conn_node_number]["inputs"][conn_input_id]
-                                previous_input[1] = new_output_idx
-
-    def as_dict(self):
-        """Serialize all replacements to dict."""
-        return {
-            k: [v.as_dict() for v in v_list]
-            for k, v_list in self._replacements.items()
-        }
-
-    def add_routes(self, routes):
-        @routes.get("/node_replacements")
-        async def get_node_replacements(request):
-            return web.json_response(self.as_dict())

app/subgraph_manager.py

@@ -53,7 +53,7 @@ class SubgraphManager:
         return entry_id, entry
 
     async def load_entry_data(self, entry: SubgraphEntry):
-        with open(entry['path'], 'r', encoding='utf-8') as f:
+        with open(entry['path'], 'r') as f:
             entry['data'] = f.read()
         return entry
 

blueprints/.glsl/Brightness_and_Contrast_1.frag

@@ -1,44 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform float u_float0; // Brightness slider -100..100
-uniform float u_float1; // Contrast slider -100..100
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-const float MID_GRAY = 0.18;  // 18% reflectance
-
-// sRGB gamma 2.2 approximation
-vec3 srgbToLinear(vec3 c) {
-    return pow(max(c, 0.0), vec3(2.2));
-}
-
-vec3 linearToSrgb(vec3 c) {
-    return pow(max(c, 0.0), vec3(1.0/2.2));
-}
-
-float mapBrightness(float b) {
-    return clamp(b / 100.0, -1.0, 1.0);
-}
-
-float mapContrast(float c) {
-    return clamp(c / 100.0 + 1.0, 0.0, 2.0);
-}
-
-void main() {
-    vec4 orig = texture(u_image0, v_texCoord);
-
-    float brightness = mapBrightness(u_float0);
-    float contrast   = mapContrast(u_float1);
-
-    vec3 lin = srgbToLinear(orig.rgb);
-
-    lin = (lin - MID_GRAY) * contrast + brightness + MID_GRAY;
-
-    // Convert back to sRGB
-    vec3 result = linearToSrgb(clamp(lin, 0.0, 1.0));
-
-    fragColor = vec4(result, orig.a);
-}

blueprints/.glsl/Chromatic_Aberration_16.frag

@@ -1,72 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform vec2 u_resolution;
-uniform int u_int0;      // Mode
-uniform float u_float0;  // Amount (0 to 100)
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-const int MODE_LINEAR   = 0;
-const int MODE_RADIAL   = 1;
-const int MODE_BARREL   = 2;
-const int MODE_SWIRL    = 3;
-const int MODE_DIAGONAL = 4;
-
-const float AMOUNT_SCALE = 0.0005;
-const float RADIAL_MULT = 4.0;
-const float BARREL_MULT = 8.0;
-const float INV_SQRT2 = 0.70710678118;
-
-void main() {
-    vec2 uv = v_texCoord;
-    vec4 original = texture(u_image0, uv);
-
-    float amount = u_float0 * AMOUNT_SCALE;
-
-    if (amount < 0.000001) {
-        fragColor = original;
-        return;
-    }
-
-    // Aspect-corrected coordinates for circular effects
-    float aspect = u_resolution.x / u_resolution.y;
-    vec2 centered = uv - 0.5;
-    vec2 corrected = vec2(centered.x * aspect, centered.y);
-    float r = length(corrected);
-    vec2 dir = r > 0.0001 ? corrected / r : vec2(0.0);
-    vec2 offset = vec2(0.0);
-
-    if (u_int0 == MODE_LINEAR) {
-        // Horizontal shift (no aspect correction needed)
-        offset = vec2(amount, 0.0);
-    }
-    else if (u_int0 == MODE_RADIAL) {
-        // Outward from center, stronger at edges
-        offset = dir * r * amount * RADIAL_MULT;
-        offset.x /= aspect;  // Convert back to UV space
-    }
-    else if (u_int0 == MODE_BARREL) {
-        // Lens distortion simulation (r² falloff)
-        offset = dir * r * r * amount * BARREL_MULT;
-        offset.x /= aspect;  // Convert back to UV space
-    }
-    else if (u_int0 == MODE_SWIRL) {
-        // Perpendicular to radial (rotational aberration)
-        vec2 perp = vec2(-dir.y, dir.x);
-        offset = perp * r * amount * RADIAL_MULT;
-        offset.x /= aspect;  // Convert back to UV space
-    }
-    else if (u_int0 == MODE_DIAGONAL) {
-        // 45° offset (no aspect correction needed)
-        offset = vec2(amount, amount) * INV_SQRT2;
-    }
-    
-    float red = texture(u_image0, uv + offset).r;
-    float green = original.g;
-    float blue = texture(u_image0, uv - offset).b;
-    
-    fragColor = vec4(red, green, blue, original.a);
-}

blueprints/.glsl/Color_Adjustment_15.frag

@@ -1,78 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform float u_float0; // temperature (-100 to 100)
-uniform float u_float1; // tint (-100 to 100)
-uniform float u_float2; // vibrance (-100 to 100)
-uniform float u_float3; // saturation (-100 to 100)
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-const float INPUT_SCALE = 0.01;
-const float TEMP_TINT_PRIMARY = 0.3;
-const float TEMP_TINT_SECONDARY = 0.15;
-const float VIBRANCE_BOOST = 2.0;
-const float SATURATION_BOOST = 2.0;
-const float SKIN_PROTECTION = 0.5;
-const float EPSILON = 0.001;
-const vec3 LUMA_WEIGHTS = vec3(0.299, 0.587, 0.114);
-
-void main() {
-    vec4 tex = texture(u_image0, v_texCoord);
-    vec3 color = tex.rgb;
-    
-    // Scale inputs: -100/100 → -1/1
-    float temperature = u_float0 * INPUT_SCALE;
-    float tint = u_float1 * INPUT_SCALE;
-    float vibrance = u_float2 * INPUT_SCALE;
-    float saturation = u_float3 * INPUT_SCALE;
-    
-    // Temperature (warm/cool): positive = warm, negative = cool
-    color.r += temperature * TEMP_TINT_PRIMARY;
-    color.b -= temperature * TEMP_TINT_PRIMARY;
-    
-    // Tint (green/magenta): positive = green, negative = magenta
-    color.g += tint * TEMP_TINT_PRIMARY;
-    color.r -= tint * TEMP_TINT_SECONDARY;
-    color.b -= tint * TEMP_TINT_SECONDARY;
-    
-    // Single clamp after temperature/tint
-    color = clamp(color, 0.0, 1.0);
-    
-    // Vibrance with skin protection
-    if (vibrance != 0.0) {
-        float maxC = max(color.r, max(color.g, color.b));
-        float minC = min(color.r, min(color.g, color.b));
-        float sat = maxC - minC;
-        float gray = dot(color, LUMA_WEIGHTS);
-        
-        if (vibrance < 0.0) {
-            // Desaturate: -100 → gray
-            color = mix(vec3(gray), color, 1.0 + vibrance);
-        } else {
-            // Boost less saturated colors more
-            float vibranceAmt = vibrance * (1.0 - sat);
-            
-            // Branchless skin tone protection
-            float isWarmTone = step(color.b, color.g) * step(color.g, color.r);
-            float warmth = (color.r - color.b) / max(maxC, EPSILON);
-            float skinTone = isWarmTone * warmth * sat * (1.0 - sat);
-            vibranceAmt *= (1.0 - skinTone * SKIN_PROTECTION);
-            
-            color = mix(vec3(gray), color, 1.0 + vibranceAmt * VIBRANCE_BOOST);
-        }
-    }
-    
-    // Saturation
-    if (saturation != 0.0) {
-        float gray = dot(color, LUMA_WEIGHTS);
-        float satMix = saturation < 0.0
-            ? 1.0 + saturation                      // -100 → gray
-            : 1.0 + saturation * SATURATION_BOOST;  // +100 → 3x boost
-        color = mix(vec3(gray), color, satMix);
-    }
-    
-    fragColor = vec4(clamp(color, 0.0, 1.0), tex.a);
-}

blueprints/.glsl/Edge-Preserving_Blur_128.frag

@@ -1,94 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform float u_float0;   // Blur radius (0–20, default ~5)
-uniform float u_float1;   // Edge threshold (0–100, default ~30)
-uniform int u_int0;       // Step size (0/1 = every pixel, 2+ = skip pixels)
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-const int MAX_RADIUS = 20;
-const float EPSILON = 0.0001;
-
-// Perceptual luminance
-float getLuminance(vec3 rgb) {
-    return dot(rgb, vec3(0.299, 0.587, 0.114));
-}
-
-vec4 bilateralFilter(vec2 uv, vec2 texelSize, int radius,
-                     float sigmaSpatial, float sigmaColor)
-{
-    vec4 center = texture(u_image0, uv);
-    vec3 centerRGB = center.rgb;
-
-    float invSpatial2 = -0.5 / (sigmaSpatial * sigmaSpatial);
-    float invColor2   = -0.5 / (sigmaColor * sigmaColor + EPSILON);
-
-    vec3 sumRGB = vec3(0.0);
-    float sumWeight = 0.0;
-
-    int step = max(u_int0, 1);
-    float radius2 = float(radius * radius);
-
-    for (int dy = -MAX_RADIUS; dy <= MAX_RADIUS; dy++) {
-        if (dy < -radius || dy > radius) continue;
-        if (abs(dy) % step != 0) continue;
-
-        for (int dx = -MAX_RADIUS; dx <= MAX_RADIUS; dx++) {
-            if (dx < -radius || dx > radius) continue;
-            if (abs(dx) % step != 0) continue;
-
-            vec2 offset = vec2(float(dx), float(dy));
-            float dist2 = dot(offset, offset);
-            if (dist2 > radius2) continue;
-
-            vec3 sampleRGB = texture(u_image0, uv + offset * texelSize).rgb;
-
-            // Spatial Gaussian
-            float spatialWeight = exp(dist2 * invSpatial2);
-
-            // Perceptual color distance (weighted RGB)
-            vec3 diff = sampleRGB - centerRGB;
-            float colorDist = dot(diff * diff, vec3(0.299, 0.587, 0.114));
-            float colorWeight = exp(colorDist * invColor2);
-
-            float w = spatialWeight * colorWeight;
-            sumRGB += sampleRGB * w;
-            sumWeight += w;
-        }
-    }
-
-    vec3 resultRGB = sumRGB / max(sumWeight, EPSILON);
-    return vec4(resultRGB, center.a); // preserve center alpha
-}
-
-void main() {
-    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
-
-    float radiusF = clamp(u_float0, 0.0, float(MAX_RADIUS));
-    int radius = int(radiusF + 0.5);
-
-    if (radius == 0) {
-        fragColor = texture(u_image0, v_texCoord);
-        return;
-    }
-
-    // Edge threshold → color sigma
-    // Squared curve for better low-end control
-    float t = clamp(u_float1, 0.0, 100.0) / 100.0;
-    t *= t;
-    float sigmaColor = mix(0.01, 0.5, t);
-
-    // Spatial sigma tied to radius
-    float sigmaSpatial = max(radiusF * 0.75, 0.5);
-
-    fragColor = bilateralFilter(
-        v_texCoord,
-        texelSize,
-        radius,
-        sigmaSpatial,
-        sigmaColor
-    );
-}

blueprints/.glsl/Film_Grain_15.frag

@@ -1,124 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform vec2 u_resolution;
-uniform float u_float0; // grain amount      [0.0 – 1.0]   typical: 0.2–0.8
-uniform float u_float1; // grain size        [0.3 – 3.0]   lower = finer grain
-uniform float u_float2; // color amount      [0.0 – 1.0]   0 = monochrome, 1 = RGB grain
-uniform float u_float3; // luminance bias    [0.0 – 1.0]   0 = uniform, 1 = shadows only
-uniform int   u_int0;   // noise mode        [0 or 1]      0 = smooth, 1 = grainy
-
-in vec2 v_texCoord;
-layout(location = 0) out vec4 fragColor0;
-
-// High-quality integer hash (pcg-like)
-uint pcg(uint v) {
-    uint state = v * 747796405u + 2891336453u;
-    uint word = ((state >> ((state >> 28u) + 4u)) ^ state) * 277803737u;
-    return (word >> 22u) ^ word;
-}
-
-// 2D -> 1D hash input
-uint hash2d(uvec2 p) {
-    return pcg(p.x + pcg(p.y));
-}
-
-// Hash to float [0, 1]
-float hashf(uvec2 p) {
-    return float(hash2d(p)) / float(0xffffffffu);
-}
-
-// Hash to float with offset (for RGB channels)
-float hashf(uvec2 p, uint offset) {
-    return float(pcg(hash2d(p) + offset)) / float(0xffffffffu);
-}
-
-// Convert uniform [0,1] to roughly Gaussian distribution
-// Using simple approximation: average of multiple samples
-float toGaussian(uvec2 p) {
-    float sum = hashf(p, 0u) + hashf(p, 1u) + hashf(p, 2u) + hashf(p, 3u);
-    return (sum - 2.0) * 0.7;  // Centered, scaled
-}
-
-float toGaussian(uvec2 p, uint offset) {
-    float sum = hashf(p, offset) + hashf(p, offset + 1u) 
-              + hashf(p, offset + 2u) + hashf(p, offset + 3u);
-    return (sum - 2.0) * 0.7;
-}
-
-// Smooth noise with better interpolation
-float smoothNoise(vec2 p) {
-    vec2 i = floor(p);
-    vec2 f = fract(p);
-    
-    // Quintic interpolation (less banding than cubic)
-    f = f * f * f * (f * (f * 6.0 - 15.0) + 10.0);
-    
-    uvec2 ui = uvec2(i);
-    float a = toGaussian(ui);
-    float b = toGaussian(ui + uvec2(1u, 0u));
-    float c = toGaussian(ui + uvec2(0u, 1u));
-    float d = toGaussian(ui + uvec2(1u, 1u));
-    
-    return mix(mix(a, b, f.x), mix(c, d, f.x), f.y);
-}
-
-float smoothNoise(vec2 p, uint offset) {
-    vec2 i = floor(p);
-    vec2 f = fract(p);
-    
-    f = f * f * f * (f * (f * 6.0 - 15.0) + 10.0);
-    
-    uvec2 ui = uvec2(i);
-    float a = toGaussian(ui, offset);
-    float b = toGaussian(ui + uvec2(1u, 0u), offset);
-    float c = toGaussian(ui + uvec2(0u, 1u), offset);
-    float d = toGaussian(ui + uvec2(1u, 1u), offset);
-    
-    return mix(mix(a, b, f.x), mix(c, d, f.x), f.y);
-}
-
-void main() {
-    vec4 color = texture(u_image0, v_texCoord);
-    
-    // Luminance (Rec.709)
-    float luma = dot(color.rgb, vec3(0.2126, 0.7152, 0.0722));
-    
-    // Grain UV (resolution-independent)
-    vec2 grainUV = v_texCoord * u_resolution / max(u_float1, 0.01);
-    uvec2 grainPixel = uvec2(grainUV);
-    
-    float g;
-    vec3 grainRGB;
-    
-    if (u_int0 == 1) {
-        // Grainy mode: pure hash noise (no interpolation = no banding)
-        g = toGaussian(grainPixel);
-        grainRGB = vec3(
-            toGaussian(grainPixel, 100u),
-            toGaussian(grainPixel, 200u),
-            toGaussian(grainPixel, 300u)
-        );
-    } else {
-        // Smooth mode: interpolated with quintic curve
-        g = smoothNoise(grainUV);
-        grainRGB = vec3(
-            smoothNoise(grainUV, 100u),
-            smoothNoise(grainUV, 200u),
-            smoothNoise(grainUV, 300u)
-        );
-    }
-    
-    // Luminance weighting (less grain in highlights)
-    float lumWeight = mix(1.0, 1.0 - luma, clamp(u_float3, 0.0, 1.0));
-    
-    // Strength
-    float strength = u_float0 * 0.15;
-    
-    // Color vs monochrome grain
-    vec3 grainColor = mix(vec3(g), grainRGB, clamp(u_float2, 0.0, 1.0));
-    
-    color.rgb += grainColor * strength * lumWeight;
-    fragColor0 = vec4(clamp(color.rgb, 0.0, 1.0), color.a);
-}

blueprints/.glsl/Glow_30.frag

@@ -1,133 +0,0 @@
-#version 300 es
-precision mediump float;
-
-uniform sampler2D u_image0;
-uniform vec2 u_resolution;
-uniform int u_int0;      // Blend mode
-uniform int u_int1;      // Color tint
-uniform float u_float0;  // Intensity
-uniform float u_float1;  // Radius
-uniform float u_float2;  // Threshold
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-const int BLEND_ADD      = 0;
-const int BLEND_SCREEN   = 1;
-const int BLEND_SOFT     = 2;
-const int BLEND_OVERLAY  = 3;
-const int BLEND_LIGHTEN  = 4;
-
-const float GOLDEN_ANGLE = 2.39996323;
-const int MAX_SAMPLES = 48;
-const vec3 LUMA = vec3(0.299, 0.587, 0.114);
-
-float hash(vec2 p) {
-    p = fract(p * vec2(123.34, 456.21));
-    p += dot(p, p + 45.32);
-    return fract(p.x * p.y);
-}
-
-vec3 hexToRgb(int h) {
-    return vec3(
-        float((h >> 16) & 255),
-        float((h >> 8) & 255),
-        float(h & 255)
-    ) * (1.0 / 255.0);
-}
-
-vec3 blend(vec3 base, vec3 glow, int mode) {
-    if (mode == BLEND_SCREEN) {
-        return 1.0 - (1.0 - base) * (1.0 - glow);
-    }
-    if (mode == BLEND_SOFT) {
-        return mix(
-            base - (1.0 - 2.0 * glow) * base * (1.0 - base),
-            base + (2.0 * glow - 1.0) * (sqrt(base) - base),
-            step(0.5, glow)
-        );
-    }
-    if (mode == BLEND_OVERLAY) {
-        return mix(
-            2.0 * base * glow,
-            1.0 - 2.0 * (1.0 - base) * (1.0 - glow),
-            step(0.5, base)
-        );
-    }
-    if (mode == BLEND_LIGHTEN) {
-        return max(base, glow);
-    }
-    return base + glow;
-}
-
-void main() {
-    vec4 original = texture(u_image0, v_texCoord);
-    
-    float intensity = u_float0 * 0.05;
-    float radius = u_float1 * u_float1 * 0.012;
-    
-    if (intensity < 0.001 || radius < 0.1) {
-        fragColor = original;
-        return;
-    }
-    
-    float threshold = 1.0 - u_float2 * 0.01;
-    float t0 = threshold - 0.15;
-    float t1 = threshold + 0.15;
-    
-    vec2 texelSize = 1.0 / u_resolution;
-    float radius2 = radius * radius;
-    
-    float sampleScale = clamp(radius * 0.75, 0.35, 1.0);
-    int samples = int(float(MAX_SAMPLES) * sampleScale);
-    
-    float noise = hash(gl_FragCoord.xy);
-    float angleOffset = noise * GOLDEN_ANGLE;
-    float radiusJitter = 0.85 + noise * 0.3;
-    
-    float ca = cos(GOLDEN_ANGLE);
-    float sa = sin(GOLDEN_ANGLE);
-    vec2 dir = vec2(cos(angleOffset), sin(angleOffset));
-    
-    vec3 glow = vec3(0.0);
-    float totalWeight = 0.0;
-    
-    // Center tap
-    float centerMask = smoothstep(t0, t1, dot(original.rgb, LUMA));
-    glow += original.rgb * centerMask * 2.0;
-    totalWeight += 2.0;
-    
-    for (int i = 1; i < MAX_SAMPLES; i++) {
-        if (i >= samples) break;
-        
-        float fi = float(i);
-        float dist = sqrt(fi / float(samples)) * radius * radiusJitter;
-        
-        vec2 offset = dir * dist * texelSize;
-        vec3 c = texture(u_image0, v_texCoord + offset).rgb;
-        float mask = smoothstep(t0, t1, dot(c, LUMA));
-        
-        float w = 1.0 - (dist * dist) / (radius2 * 1.5);
-        w = max(w, 0.0);
-        w *= w;
-        
-        glow += c * mask * w;
-        totalWeight += w;
-        
-        dir = vec2(
-            dir.x * ca - dir.y * sa,
-            dir.x * sa + dir.y * ca
-        );
-    }
-    
-    glow *= intensity / max(totalWeight, 0.001);
-    
-    if (u_int1 > 0) {
-        glow *= hexToRgb(u_int1);
-    }
-    
-    vec3 result = blend(original.rgb, glow, u_int0);
-    result += (noise - 0.5) * (1.0 / 255.0);
-    
-    fragColor = vec4(clamp(result, 0.0, 1.0), original.a);
-}

blueprints/.glsl/Hue_and_Saturation_1.frag

@@ -1,222 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform int u_int0;      // Mode: 0=Master, 1=Reds, 2=Yellows, 3=Greens, 4=Cyans, 5=Blues, 6=Magentas, 7=Colorize
-uniform int u_int1;      // Color Space: 0=HSL, 1=HSB/HSV
-uniform float u_float0;  // Hue (-180 to 180)
-uniform float u_float1;  // Saturation (-100 to 100)
-uniform float u_float2;  // Lightness/Brightness (-100 to 100)
-uniform float u_float3;  // Overlap (0 to 100) - feathering between adjacent color ranges
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-// Color range modes
-const int MODE_MASTER   = 0;
-const int MODE_RED      = 1;
-const int MODE_YELLOW   = 2;
-const int MODE_GREEN    = 3;
-const int MODE_CYAN     = 4;
-const int MODE_BLUE     = 5;
-const int MODE_MAGENTA  = 6;
-const int MODE_COLORIZE = 7;
-
-// Color space modes
-const int COLORSPACE_HSL = 0;
-const int COLORSPACE_HSB = 1;
-
-const float EPSILON = 0.0001;
-
-//=============================================================================
-// RGB <-> HSL Conversions
-//=============================================================================
-
-vec3 rgb2hsl(vec3 c) {
-    float maxC = max(max(c.r, c.g), c.b);
-    float minC = min(min(c.r, c.g), c.b);
-    float delta = maxC - minC;
-
-    float h = 0.0;
-    float s = 0.0;
-    float l = (maxC + minC) * 0.5;
-
-    if (delta > EPSILON) {
-        s = l < 0.5
-            ? delta / (maxC + minC)
-            : delta / (2.0 - maxC - minC);
-
-        if (maxC == c.r) {
-            h = (c.g - c.b) / delta + (c.g < c.b ? 6.0 : 0.0);
-        } else if (maxC == c.g) {
-            h = (c.b - c.r) / delta + 2.0;
-        } else {
-            h = (c.r - c.g) / delta + 4.0;
-        }
-        h /= 6.0;
-    }
-
-    return vec3(h, s, l);
-}
-
-float hue2rgb(float p, float q, float t) {
-    t = fract(t);
-    if (t < 1.0/6.0) return p + (q - p) * 6.0 * t;
-    if (t < 0.5)       return q;
-    if (t < 2.0/3.0)   return p + (q - p) * (2.0/3.0 - t) * 6.0;
-    return p;
-}
-
-vec3 hsl2rgb(vec3 hsl) {
-    if (hsl.y < EPSILON) return vec3(hsl.z);
-
-    float q = hsl.z < 0.5
-        ? hsl.z * (1.0 + hsl.y)
-        : hsl.z + hsl.y - hsl.z * hsl.y;
-    float p = 2.0 * hsl.z - q;
-
-    return vec3(
-        hue2rgb(p, q, hsl.x + 1.0/3.0),
-        hue2rgb(p, q, hsl.x),
-        hue2rgb(p, q, hsl.x - 1.0/3.0)
-    );
-}
-
-vec3 rgb2hsb(vec3 c) {
-    float maxC = max(max(c.r, c.g), c.b);
-    float minC = min(min(c.r, c.g), c.b);
-    float delta = maxC - minC;
-
-    float h = 0.0;
-    float s = (maxC > EPSILON) ? delta / maxC : 0.0;
-    float b = maxC;
-
-    if (delta > EPSILON) {
-        if (maxC == c.r) {
-            h = (c.g - c.b) / delta + (c.g < c.b ? 6.0 : 0.0);
-        } else if (maxC == c.g) {
-            h = (c.b - c.r) / delta + 2.0;
-        } else {
-            h = (c.r - c.g) / delta + 4.0;
-        }
-        h /= 6.0;
-    }
-
-    return vec3(h, s, b);
-}
-
-vec3 hsb2rgb(vec3 hsb) {
-    vec3 rgb = clamp(abs(mod(hsb.x * 6.0 + vec3(0.0, 4.0, 2.0), 6.0) - 3.0) - 1.0, 0.0, 1.0);
-    return hsb.z * mix(vec3(1.0), rgb, hsb.y);
-}
-
-//=============================================================================
-// Color Range Weight Calculation
-//=============================================================================
-
-float hueDistance(float a, float b) {
-    float d = abs(a - b);
-    return min(d, 1.0 - d);
-}
-
-float getHueWeight(float hue, float center, float overlap) {
-    float baseWidth = 1.0 / 6.0;
-    float feather = baseWidth * overlap;
-
-    float d = hueDistance(hue, center);
-
-    float inner = baseWidth * 0.5;
-    float outer = inner + feather;
-
-    return 1.0 - smoothstep(inner, outer, d);
-}
-
-float getModeWeight(float hue, int mode, float overlap) {
-    if (mode == MODE_MASTER || mode == MODE_COLORIZE) return 1.0;
-
-    if (mode == MODE_RED) {
-        return max(
-            getHueWeight(hue, 0.0, overlap),
-            getHueWeight(hue, 1.0, overlap)
-        );
-    }
-
-    float center = float(mode - 1) / 6.0;
-    return getHueWeight(hue, center, overlap);
-}
-
-//=============================================================================
-// Adjustment Functions
-//=============================================================================
-
-float adjustLightness(float l, float amount) {
-    return amount > 0.0
-        ? l + (1.0 - l) * amount
-        : l + l * amount;
-}
-
-float adjustBrightness(float b, float amount) {
-    return clamp(b + amount, 0.0, 1.0);
-}
-
-float adjustSaturation(float s, float amount) {
-    return amount > 0.0
-        ? s + (1.0 - s) * amount
-        : s + s * amount;
-}
-
-vec3 colorize(vec3 rgb, float hue, float sat, float light) {
-    float lum = dot(rgb, vec3(0.299, 0.587, 0.114));
-    float l = adjustLightness(lum, light);
-
-    vec3 hsl = vec3(fract(hue), clamp(sat, 0.0, 1.0), clamp(l, 0.0, 1.0));
-    return hsl2rgb(hsl);
-}
-
-//=============================================================================
-// Main
-//=============================================================================
-
-void main() {
-    vec4 original = texture(u_image0, v_texCoord);
-
-    float hueShift   = u_float0 / 360.0;   // -180..180 -> -0.5..0.5
-    float satAmount  = u_float1 / 100.0;   // -100..100 -> -1..1
-    float lightAmount= u_float2 / 100.0;   // -100..100 -> -1..1
-    float overlap    = u_float3 / 100.0;   // 0..100 -> 0..1
-
-    vec3 result;
-
-    if (u_int0 == MODE_COLORIZE) {
-        result = colorize(original.rgb, hueShift, satAmount, lightAmount);
-        fragColor = vec4(result, original.a);
-        return;
-    }
-
-    vec3 hsx = (u_int1 == COLORSPACE_HSL)
-        ? rgb2hsl(original.rgb)
-        : rgb2hsb(original.rgb);
-
-    float weight = getModeWeight(hsx.x, u_int0, overlap);
-
-    if (u_int0 != MODE_MASTER && hsx.y < EPSILON) {
-        weight = 0.0;
-    }
-
-    if (weight > EPSILON) {
-        float h = fract(hsx.x + hueShift * weight);
-        float s = clamp(adjustSaturation(hsx.y, satAmount * weight), 0.0, 1.0);
-        float v = (u_int1 == COLORSPACE_HSL)
-            ? clamp(adjustLightness(hsx.z, lightAmount * weight), 0.0, 1.0)
-            : clamp(adjustBrightness(hsx.z, lightAmount * weight), 0.0, 1.0);
-
-        vec3 adjusted = vec3(h, s, v);
-        result = (u_int1 == COLORSPACE_HSL)
-            ? hsl2rgb(adjusted)
-            : hsb2rgb(adjusted);
-    } else {
-        result = original.rgb;
-    }
-
-    fragColor = vec4(result, original.a);
-}

blueprints/.glsl/Image_Blur_1.frag

@@ -1,111 +0,0 @@
-#version 300 es
-#pragma passes 2
-precision highp float;
-
-// Blur type constants
-const int BLUR_GAUSSIAN = 0;
-const int BLUR_BOX = 1;
-const int BLUR_RADIAL = 2;
-
-// Radial blur config
-const int RADIAL_SAMPLES = 12;
-const float RADIAL_STRENGTH = 0.0003;
-
-uniform sampler2D u_image0;
-uniform vec2 u_resolution;
-uniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)
-uniform float u_float0;  // Blur radius/amount
-uniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)
-
-in vec2 v_texCoord;
-layout(location = 0) out vec4 fragColor0;
-
-float gaussian(float x, float sigma) {
-    return exp(-(x * x) / (2.0 * sigma * sigma));
-}
-
-void main() {
-    vec2 texelSize = 1.0 / u_resolution;
-    float radius = max(u_float0, 0.0);
-
-    // Radial (angular) blur - single pass, doesn't use separable
-    if (u_int0 == BLUR_RADIAL) {
-        // Only execute on first pass
-        if (u_pass > 0) {
-            fragColor0 = texture(u_image0, v_texCoord);
-            return;
-        }
-
-        vec2 center = vec2(0.5);
-        vec2 dir = v_texCoord - center;
-        float dist = length(dir);
-
-        if (dist < 1e-4) {
-            fragColor0 = texture(u_image0, v_texCoord);
-            return;
-        }
-
-        vec4 sum = vec4(0.0);
-        float totalWeight = 0.0;
-        float angleStep = radius * RADIAL_STRENGTH;
-
-        dir /= dist;
-
-        float cosStep = cos(angleStep);
-        float sinStep = sin(angleStep);
-
-        float negAngle = -float(RADIAL_SAMPLES) * angleStep;
-        vec2 rotDir = vec2(
-            dir.x * cos(negAngle) - dir.y * sin(negAngle),
-            dir.x * sin(negAngle) + dir.y * cos(negAngle)
-        );
-
-        for (int i = -RADIAL_SAMPLES; i <= RADIAL_SAMPLES; i++) {
-            vec2 uv = center + rotDir * dist;
-            float w = 1.0 - abs(float(i)) / float(RADIAL_SAMPLES);
-            sum += texture(u_image0, uv) * w;
-            totalWeight += w;
-
-            rotDir = vec2(
-                rotDir.x * cosStep - rotDir.y * sinStep,
-                rotDir.x * sinStep + rotDir.y * cosStep
-            );
-        }
-
-        fragColor0 = sum / max(totalWeight, 0.001);
-        return;
-    }
-
-    // Separable Gaussian / Box blur
-    int samples = int(ceil(radius));
-
-    if (samples == 0) {
-        fragColor0 = texture(u_image0, v_texCoord);
-        return;
-    }
-
-    // Direction: pass 0 = horizontal, pass 1 = vertical
-    vec2 dir = (u_pass == 0) ? vec2(1.0, 0.0) : vec2(0.0, 1.0);
-
-    vec4 color = vec4(0.0);
-    float totalWeight = 0.0;
-    float sigma = radius / 2.0;
-
-    for (int i = -samples; i <= samples; i++) {
-        vec2 offset = dir * float(i) * texelSize;
-        vec4 sample_color = texture(u_image0, v_texCoord + offset);
-
-        float weight;
-        if (u_int0 == BLUR_GAUSSIAN) {
-            weight = gaussian(float(i), sigma);
-        } else {
-            // BLUR_BOX
-            weight = 1.0;
-        }
-
-        color += sample_color * weight;
-        totalWeight += weight;
-    }
-
-    fragColor0 = color / totalWeight;
-}

blueprints/.glsl/Image_Channels_23.frag

@@ -1,19 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-
-in vec2 v_texCoord;
-layout(location = 0) out vec4 fragColor0;
-layout(location = 1) out vec4 fragColor1;
-layout(location = 2) out vec4 fragColor2;
-layout(location = 3) out vec4 fragColor3;
-
-void main() {
-  vec4 color = texture(u_image0, v_texCoord);
-  // Output each channel as grayscale to separate render targets
-  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel
-  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel
-  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel
-  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel
-}

blueprints/.glsl/Image_Levels_1.frag

@@ -1,71 +0,0 @@
-#version 300 es
-precision highp float;
-
-// Levels Adjustment
-// u_int0:   channel      (0=RGB, 1=R, 2=G, 3=B)         default: 0
-// u_float0: input black  (0-255)                        default: 0
-// u_float1: input white  (0-255)                        default: 255
-// u_float2: gamma        (0.01-9.99)                    default: 1.0
-// u_float3: output black (0-255)                        default: 0
-// u_float4: output white (0-255)                        default: 255
-
-uniform sampler2D u_image0;
-uniform int u_int0;
-uniform float u_float0;
-uniform float u_float1;
-uniform float u_float2;
-uniform float u_float3;
-uniform float u_float4;
-
-in vec2 v_texCoord;
-out vec4 fragColor;
-
-vec3 applyLevels(vec3 color, float inBlack, float inWhite, float gamma, float outBlack, float outWhite) {
-    float inRange = max(inWhite - inBlack, 0.0001);
-    vec3 result = clamp((color - inBlack) / inRange, 0.0, 1.0);
-    result = pow(result, vec3(1.0 / gamma));
-    result = mix(vec3(outBlack), vec3(outWhite), result);
-    return result;
-}
-
-float applySingleChannel(float value, float inBlack, float inWhite, float gamma, float outBlack, float outWhite) {
-    float inRange = max(inWhite - inBlack, 0.0001);
-    float result = clamp((value - inBlack) / inRange, 0.0, 1.0);
-    result = pow(result, 1.0 / gamma);
-    result = mix(outBlack, outWhite, result);
-    return result;
-}
-
-void main() {
-    vec4 texColor = texture(u_image0, v_texCoord);
-    vec3 color = texColor.rgb;
-    
-    float inBlack = u_float0 / 255.0;
-    float inWhite = u_float1 / 255.0;
-    float gamma = u_float2;
-    float outBlack = u_float3 / 255.0;
-    float outWhite = u_float4 / 255.0;
-    
-    vec3 result;
-    
-    if (u_int0 == 0) {
-        result = applyLevels(color, inBlack, inWhite, gamma, outBlack, outWhite);
-    }
-    else if (u_int0 == 1) {
-        result = color;
-        result.r = applySingleChannel(color.r, inBlack, inWhite, gamma, outBlack, outWhite);
-    }
-    else if (u_int0 == 2) {
-        result = color;
-        result.g = applySingleChannel(color.g, inBlack, inWhite, gamma, outBlack, outWhite);
-    }
-    else if (u_int0 == 3) {
-        result = color;
-        result.b = applySingleChannel(color.b, inBlack, inWhite, gamma, outBlack, outWhite);
-    }
-    else {
-        result = color;
-    }
-    
-    fragColor = vec4(result, texColor.a);
-}

blueprints/.glsl/README.md

@@ -1,28 +0,0 @@
-# GLSL Shader Sources
-
-This folder contains the GLSL fragment shaders extracted from blueprint JSON files for easier editing and version control.
-
-## File Naming Convention
-
-`{Blueprint_Name}_{node_id}.frag`
-
-- **Blueprint_Name**: The JSON filename with spaces/special chars replaced by underscores
-- **node_id**: The GLSLShader node ID within the subgraph
-
-## Usage
-
-```bash
-# Extract shaders from blueprint JSONs to this folder
-python update_blueprints.py extract
-
-# Patch edited shaders back into blueprint JSONs
-python update_blueprints.py patch
-```
-
-## Workflow
-
-1. Run `extract` to pull current shaders from JSONs
-2. Edit `.frag` files
-3. Run `patch` to update the blueprint JSONs
-4. Test
-5. Commit both `.frag` files and updated JSONs

blueprints/.glsl/Sharpen_23.frag

@@ -1,28 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform vec2 u_resolution;
-uniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0
-
-in vec2 v_texCoord;
-layout(location = 0) out vec4 fragColor0;
-
-void main() {
-    vec2 texel = 1.0 / u_resolution;
-    
-    // Sample center and neighbors
-    vec4 center = texture(u_image0, v_texCoord);
-    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));
-    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));
-    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));
-    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));
-    
-    // Edge enhancement (Laplacian)
-    vec4 edges = center * 4.0 - top - bottom - left - right;
-    
-    // Add edges back scaled by strength
-    vec4 sharpened = center + edges * u_float0;
-    
-    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);
-}

blueprints/.glsl/Unsharp_Mask_26.frag

@@ -1,61 +0,0 @@
-#version 300 es
-precision highp float;
-
-uniform sampler2D u_image0;
-uniform vec2 u_resolution;
-uniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5
-uniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels
-uniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen
-
-in vec2 v_texCoord;
-layout(location = 0) out vec4 fragColor0;
-
-float gaussian(float x, float sigma) {
-    return exp(-(x * x) / (2.0 * sigma * sigma));
-}
-
-float getLuminance(vec3 color) {
-    return dot(color, vec3(0.2126, 0.7152, 0.0722));
-}
-
-void main() {
-    vec2 texel = 1.0 / u_resolution;
-    float radius = max(u_float1, 0.5);
-    float amount = u_float0;
-    float threshold = u_float2;
-
-    vec4 original = texture(u_image0, v_texCoord);
-
-    // Gaussian blur for the "unsharp" mask
-    int samples = int(ceil(radius));
-    float sigma = radius / 2.0;
-
-    vec4 blurred = vec4(0.0);
-    float totalWeight = 0.0;
-
-    for (int x = -samples; x <= samples; x++) {
-        for (int y = -samples; y <= samples; y++) {
-            vec2 offset = vec2(float(x), float(y)) * texel;
-            vec4 sample_color = texture(u_image0, v_texCoord + offset);
-
-            float dist = length(vec2(float(x), float(y)));
-            float weight = gaussian(dist, sigma);
-            blurred += sample_color * weight;
-            totalWeight += weight;
-        }
-    }
-    blurred /= totalWeight;
-
-    // Unsharp mask = original - blurred
-    vec3 mask = original.rgb - blurred.rgb;
-
-    // Luminance-based threshold with smooth falloff
-    float lumaDelta = abs(getLuminance(original.rgb) - getLuminance(blurred.rgb));
-    float thresholdScale = smoothstep(0.0, threshold, lumaDelta);
-    mask *= thresholdScale;
-
-    // Sharpen: original + mask * amount
-    vec3 sharpened = original.rgb + mask * amount;
-
-    fragColor0 = vec4(clamp(sharpened, 0.0, 1.0), original.a);
-}

blueprints/.glsl/update_blueprints.py

@@ -1,159 +0,0 @@
-#!/usr/bin/env python3
-"""
-Shader Blueprint Updater
-
-Syncs GLSL shader files between this folder and blueprint JSON files.
-
-File naming convention:
-    {Blueprint Name}_{node_id}.frag
-
-Usage:
-    python update_blueprints.py extract   # Extract shaders from JSONs to here
-    python update_blueprints.py patch     # Patch shaders back into JSONs
-    python update_blueprints.py           # Same as patch (default)
-"""
-
-import json
-import logging
-import sys
-import re
-from pathlib import Path
-
-logging.basicConfig(level=logging.INFO, format='%(message)s')
-logger = logging.getLogger(__name__)
-
-GLSL_DIR = Path(__file__).parent
-BLUEPRINTS_DIR = GLSL_DIR.parent
-
-
-def get_blueprint_files():
-    """Get all blueprint JSON files."""
-    return sorted(BLUEPRINTS_DIR.glob("*.json"))
-
-
-def sanitize_filename(name):
-    """Convert blueprint name to safe filename."""
-    return re.sub(r'[^\w\-]', '_', name)
-
-
-def extract_shaders():
-    """Extract all shaders from blueprint JSONs to this folder."""
-    extracted = 0
-    for json_path in get_blueprint_files():
-        blueprint_name = json_path.stem
-
-        try:
-            with open(json_path, 'r') as f:
-                data = json.load(f)
-        except (json.JSONDecodeError, IOError) as e:
-            logger.warning("Skipping %s: %s", json_path.name, e)
-            continue
-
-        # Find GLSLShader nodes in subgraphs
-        for subgraph in data.get('definitions', {}).get('subgraphs', []):
-            for node in subgraph.get('nodes', []):
-                if node.get('type') == 'GLSLShader':
-                    node_id = node.get('id')
-                    widgets = node.get('widgets_values', [])
-
-                    # Find shader code (first string that looks like GLSL)
-                    for widget in widgets:
-                        if isinstance(widget, str) and widget.startswith('#version'):
-                            safe_name = sanitize_filename(blueprint_name)
-                            frag_name = f"{safe_name}_{node_id}.frag"
-                            frag_path = GLSL_DIR / frag_name
-
-                            with open(frag_path, 'w') as f:
-                                f.write(widget)
-
-                            logger.info("  Extracted: %s", frag_name)
-                            extracted += 1
-                            break
-
-    logger.info("\nExtracted %d shader(s)", extracted)
-
-
-def patch_shaders():
-    """Patch shaders from this folder back into blueprint JSONs."""
-    # Build lookup: blueprint_name -> [(node_id, shader_code), ...]
-    shader_updates = {}
-
-    for frag_path in sorted(GLSL_DIR.glob("*.frag")):
-        # Parse filename: {blueprint_name}_{node_id}.frag
-        parts = frag_path.stem.rsplit('_', 1)
-        if len(parts) != 2:
-            logger.warning("Skipping %s: invalid filename format", frag_path.name)
-            continue
-
-        blueprint_name, node_id_str = parts
-
-        try:
-            node_id = int(node_id_str)
-        except ValueError:
-            logger.warning("Skipping %s: invalid node_id", frag_path.name)
-            continue
-
-        with open(frag_path, 'r') as f:
-            shader_code = f.read()
-
-        if blueprint_name not in shader_updates:
-            shader_updates[blueprint_name] = []
-        shader_updates[blueprint_name].append((node_id, shader_code))
-
-    # Apply updates to JSON files
-    patched = 0
-    for json_path in get_blueprint_files():
-        blueprint_name = sanitize_filename(json_path.stem)
-
-        if blueprint_name not in shader_updates:
-            continue
-
-        try:
-            with open(json_path, 'r') as f:
-                data = json.load(f)
-        except (json.JSONDecodeError, IOError) as e:
-            logger.error("Error reading %s: %s", json_path.name, e)
-            continue
-
-        modified = False
-        for node_id, shader_code in shader_updates[blueprint_name]:
-            # Find the node and update
-            for subgraph in data.get('definitions', {}).get('subgraphs', []):
-                for node in subgraph.get('nodes', []):
-                    if node.get('id') == node_id and node.get('type') == 'GLSLShader':
-                        widgets = node.get('widgets_values', [])
-                        if len(widgets) > 0 and widgets[0] != shader_code:
-                            widgets[0] = shader_code
-                            modified = True
-                            logger.info("  Patched: %s (node %d)", json_path.name, node_id)
-                            patched += 1
-
-        if modified:
-            with open(json_path, 'w') as f:
-                json.dump(data, f)
-
-    if patched == 0:
-        logger.info("No changes to apply.")
-    else:
-        logger.info("\nPatched %d shader(s)", patched)
-
-
-def main():
-    if len(sys.argv) < 2:
-        command = "patch"
-    else:
-        command = sys.argv[1].lower()
-
-    if command == "extract":
-        logger.info("Extracting shaders from blueprints...")
-        extract_shaders()
-    elif command in ("patch", "update", "apply"):
-        logger.info("Patching shaders into blueprints...")
-        patch_shaders()
-    else:
-        logger.info(__doc__)
-        sys.exit(1)
-
-
-if __name__ == "__main__":
-    main()

blueprints/Image Channels.json

@@ -1 +0,0 @@
-{"revision": 0, "last_node_id": 29, "last_link_id": 0, "nodes": [{"id": 29, "type": "4c9d6ea4-b912-40e5-8766-6793a9758c53", "pos": [1970, -230], "size": [180, 86], "flags": {}, "order": 5, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": null}], "outputs": [{"label": "R", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": []}, {"label": "G", "localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": []}, {"label": "B", "localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": []}, {"label": "A", "localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": []}], "title": "Image Channels", "properties": {"proxyWidgets": []}, "widgets_values": []}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "4c9d6ea4-b912-40e5-8766-6793a9758c53", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 28, "lastLinkId": 39, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Image Channels", "inputNode": {"id": -10, "bounding": [1820, -185, 120, 60]}, "outputNode": {"id": -20, "bounding": [2460, -215, 120, 120]}, "inputs": [{"id": "3522932b-2d86-4a1f-a02a-cb29f3a9d7fe", "name": "images.image0", "type": "IMAGE", "linkIds": [39], "localized_name": "images.image0", "label": "image", "pos": [1920, -165]}], "outputs": [{"id": "605cb9c3-b065-4d9b-81d2-3ec331889b2b", "name": "IMAGE0", "type": "IMAGE", "linkIds": [26], "localized_name": "IMAGE0", "label": "R", "pos": [2480, -195]}, {"id": "fb44a77e-0522-43e9-9527-82e7465b3596", "name": "IMAGE1", "type": "IMAGE", "linkIds": [27], "localized_name": "IMAGE1", "label": "G", "pos": [2480, -175]}, {"id": "81460ee6-0131-402a-874f-6bf3001fc4ff", "name": "IMAGE2", "type": "IMAGE", "linkIds": [28], "localized_name": "IMAGE2", "label": "B", "pos": [2480, -155]}, {"id": "ae690246-80d4-4951-b1d9-9306d8a77417", "name": "IMAGE3", "type": "IMAGE", "linkIds": [29], "localized_name": "IMAGE3", "label": "A", "pos": [2480, -135]}], "widgets": [], "nodes": [{"id": 23, "type": "GLSLShader", "pos": [2000, -330], "size": [400, 172], "flags": {}, "order": 0, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": 39}, {"localized_name": "fragment_shader", "name": "fragment_shader", "type": "STRING", "widget": {"name": "fragment_shader"}, "link": null}, {"localized_name": "size_mode", "name": "size_mode", "type": "COMFY_DYNAMICCOMBO_V3", "widget": {"name": "size_mode"}, "link": null}, {"label": "image1", "localized_name": "images.image1", "name": "images.image1", "shape": 7, "type": "IMAGE", "link": null}], "outputs": [{"label": "R", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": [26]}, {"label": "G", "localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": [27]}, {"label": "B", "localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": [28]}, {"label": "A", "localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": [29]}], "properties": {"Node name for S&R": "GLSLShader"}, "widgets_values": ["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\nlayout(location = 1) out vec4 fragColor1;\nlayout(location = 2) out vec4 fragColor2;\nlayout(location = 3) out vec4 fragColor3;\n\nvoid main() {\n  vec4 color = texture(u_image0, v_texCoord);\n  // Output each channel as grayscale to separate render targets\n  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel\n  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel\n  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel\n  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel\n}\n", "from_input"]}], "groups": [], "links": [{"id": 39, "origin_id": -10, "origin_slot": 0, "target_id": 23, "target_slot": 0, "type": "IMAGE"}, {"id": 26, "origin_id": 23, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "IMAGE"}, {"id": 27, "origin_id": 23, "origin_slot": 1, "target_id": -20, "target_slot": 1, "type": "IMAGE"}, {"id": 28, "origin_id": 23, "origin_slot": 2, "target_id": -20, "target_slot": 2, "type": "IMAGE"}, {"id": 29, "origin_id": 23, "origin_slot": 3, "target_id": -20, "target_slot": 3, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Image Tools/Color adjust"}]}}

blueprints/Prompt Enhance.json

@@ -1 +0,0 @@
-{"revision": 0, "last_node_id": 15, "last_link_id": 0, "nodes": [{"id": 15, "type": "24d8bbfd-39d4-4774-bff0-3de40cc7a471", "pos": [-1490, 2040], "size": [400, 260], "flags": {}, "order": 0, "mode": 0, "inputs": [{"name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": null}, {"label": "reference images", "name": "images", "type": "IMAGE", "link": null}], "outputs": [{"name": "STRING", "type": "STRING", "links": null}], "title": "Prompt Enhance", "properties": {"proxyWidgets": [["-1", "prompt"]], "cnr_id": "comfy-core", "ver": "0.14.1"}, "widgets_values": [""]}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "24d8bbfd-39d4-4774-bff0-3de40cc7a471", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 15, "lastLinkId": 14, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Prompt Enhance", "inputNode": {"id": -10, "bounding": [-2170, 2110, 138.876953125, 80]}, "outputNode": {"id": -20, "bounding": [-640, 2110, 120, 60]}, "inputs": [{"id": "aeab7216-00e0-4528-a09b-bba50845c5a6", "name": "prompt", "type": "STRING", "linkIds": [11], "pos": [-2051.123046875, 2130]}, {"id": "7b73fd36-aa31-4771-9066-f6c83879994b", "name": "images", "type": "IMAGE", "linkIds": [14], "label": "reference images", "pos": [-2051.123046875, 2150]}], "outputs": [{"id": "c7b0d930-68a1-48d1-b496-0519e5837064", "name": "STRING", "type": "STRING", "linkIds": [13], "pos": [-620, 2130]}], "widgets": [], "nodes": [{"id": 11, "type": "GeminiNode", "pos": [-1560, 1990], "size": [470, 470], "flags": {}, "order": 0, "mode": 0, "inputs": [{"localized_name": "images", "name": "images", "shape": 7, "type": "IMAGE", "link": 14}, {"localized_name": "audio", "name": "audio", "shape": 7, "type": "AUDIO", "link": null}, {"localized_name": "video", "name": "video", "shape": 7, "type": "VIDEO", "link": null}, {"localized_name": "files", "name": "files", "shape": 7, "type": "GEMINI_INPUT_FILES", "link": null}, {"localized_name": "prompt", "name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": 11}, {"localized_name": "model", "name": "model", "type": "COMBO", "widget": {"name": "model"}, "link": null}, {"localized_name": "seed", "name": "seed", "type": "INT", "widget": {"name": "seed"}, "link": null}, {"localized_name": "system_prompt", "name": "system_prompt", "shape": 7, "type": "STRING", "widget": {"name": "system_prompt"}, "link": null}], "outputs": [{"localized_name": "STRING", "name": "STRING", "type": "STRING", "links": [13]}], "properties": {"cnr_id": "comfy-core", "ver": "0.14.1", "Node name for S&R": "GeminiNode"}, "widgets_values": ["", "gemini-3-pro-preview", 42, "randomize", "You are an expert in prompt writing.\nBased on the input, rewrite the user's input into a detailed prompt.\nincluding camera settings, lighting, composition, and style.\nReturn the prompt only"], "color": "#432", "bgcolor": "#653"}], "groups": [], "links": [{"id": 11, "origin_id": -10, "origin_slot": 0, "target_id": 11, "target_slot": 4, "type": "STRING"}, {"id": 13, "origin_id": 11, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "STRING"}, {"id": 14, "origin_id": -10, "origin_slot": 1, "target_id": 11, "target_slot": 0, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Text generation/Prompt enhance"}]}, "extra": {}}

blueprints/Sharpen.json

@@ -1 +0,0 @@
-{"revision": 0, "last_node_id": 25, "last_link_id": 0, "nodes": [{"id": 25, "type": "621ba4e2-22a8-482d-a369-023753198b7b", "pos": [4610, -790], "size": [230, 58], "flags": {}, "order": 4, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": null}], "outputs": [{"label": "IMAGE", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": []}], "title": "Sharpen", "properties": {"proxyWidgets": [["24", "value"]]}, "widgets_values": []}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "621ba4e2-22a8-482d-a369-023753198b7b", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 24, "lastLinkId": 36, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Sharpen", "inputNode": {"id": -10, "bounding": [4090, -825, 120, 60]}, "outputNode": {"id": -20, "bounding": [5150, -825, 120, 60]}, "inputs": [{"id": "37011fb7-14b7-4e0e-b1a0-6a02e8da1fd7", "name": "images.image0", "type": "IMAGE", "linkIds": [34], "localized_name": "images.image0", "label": "image", "pos": [4190, -805]}], "outputs": [{"id": "e9182b3f-635c-4cd4-a152-4b4be17ae4b9", "name": "IMAGE0", "type": "IMAGE", "linkIds": [35], "localized_name": "IMAGE0", "label": "IMAGE", "pos": [5170, -805]}], "widgets": [], "nodes": [{"id": 24, "type": "PrimitiveFloat", "pos": [4280, -1240], "size": [270, 58], "flags": {}, "order": 0, "mode": 0, "inputs": [{"label": "strength", "localized_name": "value", "name": "value", "type": "FLOAT", "widget": {"name": "value"}, "link": null}], "outputs": [{"localized_name": "FLOAT", "name": "FLOAT", "type": "FLOAT", "links": [36]}], "properties": {"Node name for S&R": "PrimitiveFloat", "min": 0, "max": 3, "precision": 2, "step": 0.05}, "widgets_values": [0.5]}, {"id": 23, "type": "GLSLShader", "pos": [4570, -1240], "size": [370, 192], "flags": {}, "order": 1, "mode": 0, "inputs": [{"label": "image0", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": 34}, {"label": "image1", "localized_name": "images.image1", "name": "images.image1", "shape": 7, "type": "IMAGE", "link": null}, {"label": "u_float0", "localized_name": "floats.u_float0", "name": "floats.u_float0", "shape": 7, "type": "FLOAT", "link": 36}, {"label": "u_float1", "localized_name": "floats.u_float1", "name": "floats.u_float1", "shape": 7, "type": "FLOAT", "link": null}, {"label": "u_int0", "localized_name": "ints.u_int0", "name": "ints.u_int0", "shape": 7, "type": "INT", "link": null}, {"localized_name": "fragment_shader", "name": "fragment_shader", "type": "STRING", "widget": {"name": "fragment_shader"}, "link": null}, {"localized_name": "size_mode", "name": "size_mode", "type": "COMFY_DYNAMICCOMBO_V3", "widget": {"name": "size_mode"}, "link": null}], "outputs": [{"localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": [35]}, {"localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": null}, {"localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": null}, {"localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": null}], "properties": {"Node name for S&R": "GLSLShader"}, "widgets_values": ["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / u_resolution;\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}", "from_input"]}], "groups": [], "links": [{"id": 36, "origin_id": 24, "origin_slot": 0, "target_id": 23, "target_slot": 2, "type": "FLOAT"}, {"id": 34, "origin_id": -10, "origin_slot": 0, "target_id": 23, "target_slot": 0, "type": "IMAGE"}, {"id": 35, "origin_id": 23, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Image Tools/Sharpen"}]}}

blueprints/Video Upscale(GAN x4).json

@@ -1 +0,0 @@
-{"revision": 0, "last_node_id": 13, "last_link_id": 0, "nodes": [{"id": 13, "type": "cf95b747-3e17-46cb-8097-cac60ff9b2e1", "pos": [1120, 330], "size": [240, 58], "flags": {}, "order": 3, "mode": 0, "inputs": [{"localized_name": "video", "name": "video", "type": "VIDEO", "link": null}, {"name": "model_name", "type": "COMBO", "widget": {"name": "model_name"}, "link": null}], "outputs": [{"localized_name": "VIDEO", "name": "VIDEO", "type": "VIDEO", "links": []}], "title": "Video Upscale(GAN x4)", "properties": {"proxyWidgets": [["-1", "model_name"]], "cnr_id": "comfy-core", "ver": "0.14.1"}, "widgets_values": ["RealESRGAN_x4plus.safetensors"]}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "cf95b747-3e17-46cb-8097-cac60ff9b2e1", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 13, "lastLinkId": 19, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Video Upscale(GAN x4)", "inputNode": {"id": -10, "bounding": [550, 460, 120, 80]}, "outputNode": {"id": -20, "bounding": [1490, 460, 120, 60]}, "inputs": [{"id": "666d633e-93e7-42dc-8d11-2b7b99b0f2a6", "name": "video", "type": "VIDEO", "linkIds": [10], "localized_name": "video", "pos": [650, 480]}, {"id": "2e23a087-caa8-4d65-99e6-662761aa905a", "name": "model_name", "type": "COMBO", "linkIds": [19], "pos": [650, 500]}], "outputs": [{"id": "0c1768ea-3ec2-412f-9af6-8e0fa36dae70", "name": "VIDEO", "type": "VIDEO", "linkIds": [15], "localized_name": "VIDEO", "pos": [1510, 480]}], "widgets": [], "nodes": [{"id": 2, "type": "ImageUpscaleWithModel", "pos": [1110, 450], "size": [320, 46], "flags": {}, "order": 1, "mode": 0, "inputs": [{"localized_name": "upscale_model", "name": "upscale_model", "type": "UPSCALE_MODEL", "link": 1}, {"localized_name": "image", "name": "image", "type": "IMAGE", "link": 14}], "outputs": [{"localized_name": "IMAGE", "name": "IMAGE", "type": "IMAGE", "links": [13]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "ImageUpscaleWithModel"}}, {"id": 11, "type": "CreateVideo", "pos": [1110, 550], "size": [320, 78], "flags": {}, "order": 3, "mode": 0, "inputs": [{"localized_name": "images", "name": "images", "type": "IMAGE", "link": 13}, {"localized_name": "audio", "name": "audio", "shape": 7, "type": "AUDIO", "link": 16}, {"localized_name": "fps", "name": "fps", "type": "FLOAT", "widget": {"name": "fps"}, "link": 12}], "outputs": [{"localized_name": "VIDEO", "name": "VIDEO", "type": "VIDEO", "links": [15]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "CreateVideo"}, "widgets_values": [30]}, {"id": 10, "type": "GetVideoComponents", "pos": [1110, 330], "size": [320, 70], "flags": {}, "order": 2, "mode": 0, "inputs": [{"localized_name": "video", "name": "video", "type": "VIDEO", "link": 10}], "outputs": [{"localized_name": "images", "name": "images", "type": "IMAGE", "links": [14]}, {"localized_name": "audio", "name": "audio", "type": "AUDIO", "links": [16]}, {"localized_name": "fps", "name": "fps", "type": "FLOAT", "links": [12]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "GetVideoComponents"}}, {"id": 1, "type": "UpscaleModelLoader", "pos": [750, 450], "size": [280, 60], "flags": {}, "order": 0, "mode": 0, "inputs": [{"localized_name": "model_name", "name": "model_name", "type": "COMBO", "widget": {"name": "model_name"}, "link": 19}], "outputs": [{"localized_name": "UPSCALE_MODEL", "name": "UPSCALE_MODEL", "type": "UPSCALE_MODEL", "links": [1]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "UpscaleModelLoader", "models": [{"name": "RealESRGAN_x4plus.safetensors", "url": "https://huggingface.co/Comfy-Org/Real-ESRGAN_repackaged/resolve/main/RealESRGAN_x4plus.safetensors", "directory": "upscale_models"}]}, "widgets_values": ["RealESRGAN_x4plus.safetensors"]}], "groups": [], "links": [{"id": 1, "origin_id": 1, "origin_slot": 0, "target_id": 2, "target_slot": 0, "type": "UPSCALE_MODEL"}, {"id": 14, "origin_id": 10, "origin_slot": 0, "target_id": 2, "target_slot": 1, "type": "IMAGE"}, {"id": 13, "origin_id": 2, "origin_slot": 0, "target_id": 11, "target_slot": 0, "type": "IMAGE"}, {"id": 16, "origin_id": 10, "origin_slot": 1, "target_id": 11, "target_slot": 1, "type": "AUDIO"}, {"id": 12, "origin_id": 10, "origin_slot": 2, "target_id": 11, "target_slot": 2, "type": "FLOAT"}, {"id": 10, "origin_id": -10, "origin_slot": 0, "target_id": 10, "target_slot": 0, "type": "VIDEO"}, {"id": 15, "origin_id": 11, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "VIDEO"}, {"id": 19, "origin_id": -10, "origin_slot": 1, "target_id": 1, "target_slot": 0, "type": "COMBO"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Video generation and editing/Enhance video"}]}, "extra": {}}

comfy/checkpoint_pickle.py

@@ -0,0 +1,13 @@
+import pickle
+
+load = pickle.load
+
+class Empty:
+    pass
+
+class Unpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        #TODO: safe unpickle
+        if module.startswith("pytorch_lightning"):
+            return Empty
+        return super().find_class(module, name)

comfy/cli_args.py

@@ -146,7 +146,6 @@ 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("--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.")
 
@@ -160,6 +159,7 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
+    DynamicVRAM = "dynamic_vram"
 
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
@@ -260,4 +260,4 @@ else:
     args.fast = set(args.fast)
 
 def enables_dynamic_vram():
-    return not args.disable_dynamic_vram and not args.highvram and not args.gpu_only and not args.novram and not args.cpu
+    return PerformanceFeature.DynamicVRAM in args.fast and not args.highvram and not args.gpu_only

comfy/comfy_types/node_typing.py

@@ -176,8 +176,6 @@ class InputTypeOptions(TypedDict):
     """COMBO type only. Specifies the configuration for a multi-select widget.
     Available after ComfyUI frontend v1.13.4
     https://github.com/Comfy-Org/ComfyUI_frontend/pull/2987"""
-    gradient_stops: NotRequired[list[list[float]]]
-    """Gradient color stops for gradientslider display mode. Each stop is [offset, r, g, b] (``FLOAT``)."""
 
 
 class HiddenInputTypeDict(TypedDict):

comfy/conds.py

@@ -4,25 +4,6 @@ import comfy.utils
 import logging
 
 
-def is_equal(x, y):
-    if torch.is_tensor(x) and torch.is_tensor(y):
-        return torch.equal(x, y)
-    elif isinstance(x, dict) and isinstance(y, dict):
-        if x.keys() != y.keys():
-            return False
-        return all(is_equal(x[k], y[k]) for k in x)
-    elif isinstance(x, (list, tuple)) and isinstance(y, (list, tuple)):
-        if type(x) is not type(y) or len(x) != len(y):
-            return False
-        return all(is_equal(a, b) for a, b in zip(x, y))
-    else:
-        try:
-            return x == y
-        except Exception:
-            logging.warning("comparison issue with COND")
-            return False
-
-
 class CONDRegular:
     def __init__(self, cond):
         self.cond = cond
@@ -103,7 +84,7 @@ class CONDConstant(CONDRegular):
         return self._copy_with(self.cond)
 
     def can_concat(self, other):
-        if not is_equal(self.cond, other.cond):
+        if self.cond != other.cond:
             return False
         return True
 

comfy/context_windows.py

@@ -214,7 +214,7 @@ class IndexListContextHandler(ContextHandlerABC):
         mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep[0], rtol=0.0001)
         matches = torch.nonzero(mask)
         if torch.numel(matches) == 0:
-            return  # substep from multi-step sampler: keep self._step from the last full step
+            raise Exception("No sample_sigmas matched current timestep; something went wrong.")
         self._step = int(matches[0].item())
 
     def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:

comfy/controlnet.py

@@ -297,30 +297,6 @@ class ControlNet(ControlBase):
         self.model_sampling_current = None
         super().cleanup()
 
-
-class QwenFunControlNet(ControlNet):
-    def get_control(self, x_noisy, t, cond, batched_number, transformer_options):
-        # Fun checkpoints are more sensitive to high strengths in the generic
-        # ControlNet merge path. Use a soft response curve so strength=1.0 stays
-        # unchanged while >1 grows more gently.
-        original_strength = self.strength
-        self.strength = math.sqrt(max(self.strength, 0.0))
-        try:
-            return super().get_control(x_noisy, t, cond, batched_number, transformer_options)
-        finally:
-            self.strength = original_strength
-
-    def pre_run(self, model, percent_to_timestep_function):
-        super().pre_run(model, percent_to_timestep_function)
-        self.set_extra_arg("base_model", model.diffusion_model)
-
-    def copy(self):
-        c = QwenFunControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
-        c.control_model = self.control_model
-        c.control_model_wrapped = self.control_model_wrapped
-        self.copy_to(c)
-        return c
-
 class ControlLoraOps:
     class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp):
         def __init__(self, in_features: int, out_features: int, bias: bool = True,
@@ -584,7 +560,6 @@ def load_controlnet_hunyuandit(controlnet_data, model_options={}):
 def load_controlnet_flux_xlabs_mistoline(sd, mistoline=False, model_options={}):
     model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options)
     control_model = comfy.ldm.flux.controlnet.ControlNetFlux(mistoline=mistoline, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
-    sd = model_config.process_unet_state_dict(sd)
     control_model = controlnet_load_state_dict(control_model, sd)
     extra_conds = ['y', 'guidance']
     control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
@@ -630,53 +605,6 @@ def load_controlnet_qwen_instantx(sd, model_options={}):
     control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
     return control
 
-
-def load_controlnet_qwen_fun(sd, model_options={}):
-    load_device = comfy.model_management.get_torch_device()
-    weight_dtype = comfy.utils.weight_dtype(sd)
-    unet_dtype = model_options.get("dtype", weight_dtype)
-    manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
-
-    operations = model_options.get("custom_operations", None)
-    if operations is None:
-        operations = comfy.ops.pick_operations(unet_dtype, manual_cast_dtype, disable_fast_fp8=True)
-
-    in_features = sd["control_img_in.weight"].shape[1]
-    inner_dim = sd["control_img_in.weight"].shape[0]
-
-    block_weight = sd["control_blocks.0.attn.to_q.weight"]
-    attention_head_dim = sd["control_blocks.0.attn.norm_q.weight"].shape[0]
-    num_attention_heads = max(1, block_weight.shape[0] // max(1, attention_head_dim))
-
-    model = comfy.ldm.qwen_image.controlnet.QwenImageFunControlNetModel(
-        control_in_features=in_features,
-        inner_dim=inner_dim,
-        num_attention_heads=num_attention_heads,
-        attention_head_dim=attention_head_dim,
-        num_control_blocks=5,
-        main_model_double=60,
-        injection_layers=(0, 12, 24, 36, 48),
-        operations=operations,
-        device=comfy.model_management.unet_offload_device(),
-        dtype=unet_dtype,
-    )
-    model = controlnet_load_state_dict(model, sd)
-
-    latent_format = comfy.latent_formats.Wan21()
-    control = QwenFunControlNet(
-        model,
-        compression_ratio=1,
-        latent_format=latent_format,
-        # Fun checkpoints already expect their own 33-channel context handling.
-        # Enabling generic concat_mask injects an extra mask channel at apply-time
-        # and breaks the intended fallback packing path.
-        concat_mask=False,
-        load_device=load_device,
-        manual_cast_dtype=manual_cast_dtype,
-        extra_conds=[],
-    )
-    return control
-
 def convert_mistoline(sd):
     return comfy.utils.state_dict_prefix_replace(sd, {"single_controlnet_blocks.": "controlnet_single_blocks."})
 
@@ -754,8 +682,6 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
             return load_controlnet_qwen_instantx(controlnet_data, model_options=model_options)
         elif "controlnet_x_embedder.weight" in controlnet_data:
             return load_controlnet_flux_instantx(controlnet_data, model_options=model_options)
-    elif "control_blocks.0.after_proj.weight" in controlnet_data and "control_img_in.weight" in controlnet_data:
-        return load_controlnet_qwen_fun(controlnet_data, model_options=model_options)
 
     elif "controlnet_blocks.0.linear.weight" in controlnet_data: #mistoline flux
         return load_controlnet_flux_xlabs_mistoline(convert_mistoline(controlnet_data), mistoline=True, model_options=model_options)

comfy/k_diffusion/sampling.py

@@ -1,11 +1,12 @@
 import math
+import time
 from functools import partial
 
 from scipy import integrate
 import torch
 from torch import nn
 import torchsde
-from tqdm.auto import tqdm
+from tqdm.auto import trange as trange_, tqdm
 
 from . import utils
 from . import deis
@@ -14,7 +15,34 @@ import comfy.model_patcher
 import comfy.model_sampling
 
 import comfy.memory_management
-from comfy.utils import model_trange as trange
+
+
+def trange(*args, **kwargs):
+    if comfy.memory_management.aimdo_allocator is None:
+        return trange_(*args, **kwargs)
+
+    pbar = trange_(*args, **kwargs, smoothing=1.0)
+    pbar._i = 0
+    pbar.set_postfix_str("  Model Initializing ...  ")
+
+    _update = pbar.update
+
+    def warmup_update(n=1):
+        pbar._i += 1
+        if pbar._i == 1:
+            pbar.i1_time = time.time()
+            pbar.set_postfix_str(" Model Initialization complete!  ")
+        elif pbar._i == 2:
+            #bring forward the effective start time based the the diff between first and second iteration
+            #to attempt to remove load overhead from the final step rate estimate.
+            pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time)
+            pbar.set_postfix_str("")
+
+        _update(n)
+
+    pbar.update = warmup_update
+    return pbar
+
 
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])

comfy/latent_formats.py

@@ -776,10 +776,3 @@ class ChromaRadiance(LatentFormat):
 
     def process_out(self, latent):
         return latent
-
-
-class ZImagePixelSpace(ChromaRadiance):
-    """Pixel-space latent format for ZImage DCT variant.
-    No VAE encoding/decoding — the model operates directly on RGB pixels.
-    """
-    pass

comfy/ldm/ace/ace_step15.py

@@ -7,67 +7,6 @@ from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
 from comfy.ldm.flux.layers import timestep_embedding
 
-def get_silence_latent(length, device):
-    head = torch.tensor([[[ 0.5707,  0.0982,  0.6909, -0.5658,  0.6266,  0.6996, -0.1365, -0.1291,
-                        -0.0776, -0.1171, -0.2743, -0.8422, -0.1168,  1.5539, -4.6936,  0.7436,
-                        -1.1846, -0.2637,  0.6933, -6.7266,  0.0966, -0.1187, -0.3501, -1.1736,
-                        0.0587, -2.0517, -1.3651,  0.7508, -0.2490, -1.3548, -0.1290, -0.7261,
-                        1.1132, -0.3249,  0.2337,  0.3004,  0.6605, -0.0298, -0.1989, -0.4041,
-                        0.2843, -1.0963, -0.5519,  0.2639, -1.0436, -0.1183,  0.0640,  0.4460,
-                        -1.1001, -0.6172, -1.3241,  1.1379,  0.5623, -0.1507, -0.1963, -0.4742,
-                        -2.4697,  0.5302,  0.5381,  0.4636, -0.1782, -0.0687,  1.0333,  0.4202],
-                        [ 0.3040, -0.1367,  0.6200,  0.0665, -0.0642,  0.4655, -0.1187, -0.0440,
-                        0.2941, -0.2753,  0.0173, -0.2421, -0.0147,  1.5603, -2.7025,  0.7907,
-                        -0.9736, -0.0682,  0.1294, -5.0707, -0.2167,  0.3302, -0.1513, -0.8100,
-                        -0.3894, -0.2884, -0.3149,  0.8660, -0.3817, -1.7061,  0.5824, -0.4840,
-                        0.6938,  0.1859,  0.1753,  0.3081,  0.0195,  0.1403, -0.0754, -0.2091,
-                        0.1251, -0.1578, -0.4968, -0.1052, -0.4554, -0.0320,  0.1284,  0.4974,
-                        -1.1889, -0.0344, -0.8313,  0.2953,  0.5445, -0.6249, -0.1595, -0.0682,
-                        -3.1412,  0.0484,  0.4153,  0.8260, -0.1526, -0.0625,  0.5366,  0.8473],
-                        [ 5.3524e-02, -1.7534e-01,  5.4443e-01, -4.3501e-01, -2.1317e-03,
-                        3.7200e-01, -4.0143e-03, -1.5516e-01, -1.2968e-01, -1.5375e-01,
-                        -7.7107e-02, -2.0593e-01, -3.2780e-01,  1.5142e+00, -2.6101e+00,
-                        5.8698e-01, -1.2716e+00, -2.4773e-01, -2.7933e-02, -5.0799e+00,
-                        1.1601e-01,  4.0987e-01, -2.2030e-02, -6.6495e-01, -2.0995e-01,
-                        -6.3474e-01, -1.5893e-01,  8.2745e-01, -2.2992e-01, -1.6816e+00,
-                        5.4440e-01, -4.9579e-01,  5.5128e-01,  3.0477e-01,  8.3052e-02,
-                        -6.1782e-02,  5.9036e-03,  2.9553e-01, -8.0645e-02, -1.0060e-01,
-                        1.9144e-01, -3.8124e-01, -7.2949e-01,  2.4520e-02, -5.0814e-01,
-                        2.3977e-01,  9.2943e-02,  3.9256e-01, -1.1993e+00, -3.2752e-01,
-                        -7.2707e-01,  2.9476e-01,  4.3542e-01, -8.8597e-01, -4.1686e-01,
-                        -8.5390e-02, -2.9018e+00,  6.4988e-02,  5.3945e-01,  9.1988e-01,
-                        5.8762e-02, -7.0098e-02,  6.4772e-01,  8.9118e-01],
-                        [-3.2225e-02, -1.3195e-01,  5.6411e-01, -5.4766e-01, -5.2170e-03,
-                        3.1425e-01, -5.4367e-02, -1.9419e-01, -1.3059e-01, -1.3660e-01,
-                        -9.0984e-02, -1.9540e-01, -2.5590e-01,  1.5440e+00, -2.6349e+00,
-                        6.8273e-01, -1.2532e+00, -1.9810e-01, -2.2793e-02, -5.0506e+00,
-                        1.8818e-01,  5.0109e-01,  7.3546e-03, -6.8771e-01, -3.0676e-01,
-                        -7.3257e-01, -1.6687e-01,  9.2232e-01, -1.8987e-01, -1.7267e+00,
-                        5.3355e-01, -5.3179e-01,  4.4953e-01,  2.8820e-01,  1.3012e-01,
-                        -2.0943e-01, -1.1348e-01,  3.3929e-01, -1.5069e-01, -1.2919e-01,
-                        1.8929e-01, -3.6166e-01, -8.0756e-01,  6.6387e-02, -5.8867e-01,
-                        1.6978e-01,  1.0134e-01,  3.3877e-01, -1.2133e+00, -3.2492e-01,
-                        -8.1237e-01,  3.8101e-01,  4.3765e-01, -8.0596e-01, -4.4531e-01,
-                        -4.7513e-02, -2.9266e+00,  1.1741e-03,  4.5123e-01,  9.3075e-01,
-                        5.3688e-02, -1.9621e-01,  6.4530e-01,  9.3870e-01]]], device=device).movedim(-1, 1)
-
-    silence_latent = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
-                                2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
-                                -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
-                                7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
-                                2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
-                                -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
-                                5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
-                                -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
-                                2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
-                                1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
-                                -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
-                                -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
-                                7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, length)
-    silence_latent[:, :, :head.shape[-1]] = head
-    return silence_latent
-
-
 def get_layer_class(operations, layer_name):
     if operations is not None and hasattr(operations, layer_name):
         return getattr(operations, layer_name)
@@ -244,7 +183,7 @@ class AceStepAttention(nn.Module):
             else:
                 attn_bias = window_bias
 
-        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True, low_precision_attention=False)
+        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True)
         attn_output = self.o_proj(attn_output)
 
         return attn_output
@@ -738,7 +677,7 @@ class AttentionPooler(nn.Module):
     def forward(self, x):
         B, T, P, D = x.shape
         x = self.embed_tokens(x)
-        special = comfy.model_management.cast_to(self.special_token, device=x.device, dtype=x.dtype).expand(B, T, 1, -1)
+        special = self.special_token.expand(B, T, 1, -1)
         x = torch.cat([special, x], dim=2)
         x = x.view(B * T, P + 1, D)
 
@@ -789,7 +728,7 @@ class FSQ(nn.Module):
         self.register_buffer('implicit_codebook', implicit_codebook, persistent=False)
 
     def bound(self, z):
-        levels_minus_1 = (comfy.model_management.cast_to(self._levels, device=z.device, dtype=z.dtype) - 1)
+        levels_minus_1 = (self._levels - 1).to(z.dtype)
         scale = 2. / levels_minus_1
         bracket = (levels_minus_1 * (torch.tanh(z) + 1) / 2.) + 0.5
 
@@ -804,8 +743,8 @@ class FSQ(nn.Module):
         return codes_non_centered.float() * (2. / (self._levels.float() - 1)) - 1.
 
     def codes_to_indices(self, zhat):
-        zhat_normalized = (zhat + 1.) / (2. / (comfy.model_management.cast_to(self._levels, device=zhat.device, dtype=zhat.dtype) - 1))
-        return (zhat_normalized * comfy.model_management.cast_to(self._basis, device=zhat.device, dtype=zhat.dtype)).sum(dim=-1).round().to(torch.int32)
+        zhat_normalized = (zhat + 1.) / (2. / (self._levels.to(zhat.dtype) - 1))
+        return (zhat_normalized * self._basis.to(zhat.dtype)).sum(dim=-1).round().to(torch.int32)
 
     def forward(self, z):
         orig_dtype = z.dtype
@@ -887,7 +826,7 @@ class ResidualFSQ(nn.Module):
         x = self.project_in(x)
 
         if hasattr(self, 'soft_clamp_input_value'):
-            sc_val = comfy.model_management.cast_to(self.soft_clamp_input_value, device=x.device, dtype=x.dtype)
+            sc_val = self.soft_clamp_input_value.to(x.dtype)
             x = (x / sc_val).tanh() * sc_val
 
         quantized_out = torch.tensor(0., device=x.device, dtype=x.dtype)
@@ -895,7 +834,7 @@ class ResidualFSQ(nn.Module):
         all_indices = []
 
         for layer, scale in zip(self.layers, self.scales):
-            scale = comfy.model_management.cast_to(scale, device=x.device, dtype=x.dtype)
+            scale = scale.to(residual.dtype)
 
             quantized, indices = layer(residual / scale)
             quantized = quantized * scale
@@ -1096,26 +1035,28 @@ class AceStepConditionGenerationModel(nn.Module):
                     audio_codes = torch.nn.functional.pad(audio_codes, (0, math.ceil(src_latents.shape[1] / 5) - audio_codes.shape[1]), "constant", 35847)
                 lm_hints_5Hz = self.tokenizer.quantizer.get_output_from_indices(audio_codes, dtype=text_hidden_states.dtype)
             else:
-                lm_hints_5Hz, indices = self.tokenizer.tokenize(refer_audio_acoustic_hidden_states_packed)
+                assert False
+                # TODO ?
 
             lm_hints = self.detokenizer(lm_hints_5Hz)
 
         lm_hints = lm_hints[:, :src_latents.shape[1], :]
-        if is_covers is None or is_covers is True:
+        if is_covers is None:
             src_latents = lm_hints
-        elif is_covers is False:
-            src_latents = refer_audio_acoustic_hidden_states_packed
+        else:
+            src_latents = torch.where(is_covers.unsqueeze(-1).unsqueeze(-1) > 0, lm_hints, src_latents)
 
         context_latents = torch.cat([src_latents, chunk_masks.to(src_latents.dtype)], dim=-1)
 
         return encoder_hidden, encoder_mask, context_latents
 
-    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, replace_with_null_embeds=False, **kwargs):
+    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, **kwargs):
         text_attention_mask = None
         lyric_attention_mask = None
         refer_audio_order_mask = None
         attention_mask = None
         chunk_masks = None
+        is_covers = None
         src_latents = None
         precomputed_lm_hints_25Hz = None
         lyric_hidden_states = lyric_embed
@@ -1127,7 +1068,7 @@ class AceStepConditionGenerationModel(nn.Module):
         if refer_audio_order_mask is None:
             refer_audio_order_mask = torch.zeros((x.shape[0],), device=x.device, dtype=torch.long)
 
-        if src_latents is None:
+        if src_latents is None and is_covers is None:
             src_latents = x
 
         if chunk_masks is None:
@@ -1140,9 +1081,6 @@ class AceStepConditionGenerationModel(nn.Module):
             src_latents, chunk_masks, is_covers, precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz, audio_codes=audio_codes
         )
 
-        if replace_with_null_embeds:
-            enc_hidden[:] = self.null_condition_emb.to(enc_hidden)
-
         out = self.decoder(hidden_states=x,
                            timestep=timestep,
                            timestep_r=timestep,

comfy/ldm/anima/model.py

@@ -179,8 +179,8 @@ class LLMAdapter(nn.Module):
             if source_attention_mask.ndim == 2:
                 source_attention_mask = source_attention_mask.unsqueeze(1).unsqueeze(1)
 
+        x = self.in_proj(self.embed(target_input_ids))
         context = source_hidden_states
-        x = self.in_proj(self.embed(target_input_ids, out_dtype=context.dtype))
         position_ids = torch.arange(x.shape[1], device=x.device).unsqueeze(0)
         position_ids_context = torch.arange(context.shape[1], device=x.device).unsqueeze(0)
         position_embeddings = self.rotary_emb(x, position_ids)
@@ -195,20 +195,8 @@ class Anima(MiniTrainDIT):
         super().__init__(*args, **kwargs)
         self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations"))
 
-    def preprocess_text_embeds(self, text_embeds, text_ids, t5xxl_weights=None):
+    def preprocess_text_embeds(self, text_embeds, text_ids):
         if text_ids is not None:
-            out = self.llm_adapter(text_embeds, text_ids)
-            if t5xxl_weights is not None:
-                out = out * t5xxl_weights
-
-            if out.shape[1] < 512:
-                out = torch.nn.functional.pad(out, (0, 0, 0, 512 - out.shape[1]))
-            return out
+            return self.llm_adapter(text_embeds, text_ids)
         else:
             return text_embeds
-
-    def forward(self, x, timesteps, context, **kwargs):
-        t5xxl_ids = kwargs.pop("t5xxl_ids", None)
-        if t5xxl_ids is not None:
-            context = self.preprocess_text_embeds(context, t5xxl_ids, t5xxl_weights=kwargs.pop("t5xxl_weights", None))
-        return super().forward(x, timesteps, context, **kwargs)

comfy/ldm/chroma/layers.py

@@ -3,6 +3,7 @@ from torch import Tensor, nn
 
 from comfy.ldm.flux.layers import (
     MLPEmbedder,
+    RMSNorm,
     ModulationOut,
 )
 
@@ -28,7 +29,7 @@ class Approximator(nn.Module):
         super().__init__()
         self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
         self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
-        self.norms = nn.ModuleList([operations.RMSNorm(hidden_dim, dtype=dtype, device=device) for x in range( n_layers)])
+        self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
         self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
 
     @property

comfy/ldm/chroma/model.py

@@ -152,7 +152,6 @@ class Chroma(nn.Module):
         transformer_options={},
         attn_mask: Tensor = None,
     ) -> Tensor:
-        transformer_options = transformer_options.copy()
         patches_replace = transformer_options.get("patches_replace", {})
 
         # running on sequences img
@@ -229,7 +228,6 @@ class Chroma(nn.Module):
 
         transformer_options["total_blocks"] = len(self.single_blocks)
         transformer_options["block_type"] = "single"
-        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
         for i, block in enumerate(self.single_blocks):
             transformer_options["block_index"] = i
             if i not in self.skip_dit:

comfy/ldm/chroma_radiance/layers.py

@@ -4,6 +4,8 @@ from functools import lru_cache
 import torch
 from torch import nn
 
+from comfy.ldm.flux.layers import RMSNorm
+
 
 class NerfEmbedder(nn.Module):
     """
@@ -143,7 +145,7 @@ class NerfGLUBlock(nn.Module):
         # We now need to generate parameters for 3 matrices.
         total_params = 3 * hidden_size_x**2 * mlp_ratio
         self.param_generator = operations.Linear(hidden_size_s, total_params, dtype=dtype, device=device)
-        self.norm = operations.RMSNorm(hidden_size_x, dtype=dtype, device=device)
+        self.norm = RMSNorm(hidden_size_x, dtype=dtype, device=device, operations=operations)
         self.mlp_ratio = mlp_ratio
 
 
@@ -176,7 +178,7 @@ class NerfGLUBlock(nn.Module):
 class NerfFinalLayer(nn.Module):
     def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None):
         super().__init__()
-        self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device)
+        self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
         self.linear = operations.Linear(hidden_size, out_channels, dtype=dtype, device=device)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -188,7 +190,7 @@ class NerfFinalLayer(nn.Module):
 class NerfFinalLayerConv(nn.Module):
     def __init__(self, hidden_size: int, out_channels: int, dtype=None, device=None, operations=None):
         super().__init__()
-        self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device)
+        self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
         self.conv = operations.Conv2d(
             in_channels=hidden_size,
             out_channels=out_channels,

comfy/ldm/cosmos/predict2.py

@@ -335,7 +335,7 @@ class FinalLayer(nn.Module):
         device=None, dtype=None, operations=None
     ):
         super().__init__()
-        self.layer_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
         self.linear = operations.Linear(
             hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
         )
@@ -463,8 +463,6 @@ class Block(nn.Module):
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
         transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
-        residual_dtype = x_B_T_H_W_D.dtype
-        compute_dtype = emb_B_T_D.dtype
         if extra_per_block_pos_emb is not None:
             x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
 
@@ -514,7 +512,7 @@ class Block(nn.Module):
         result_B_T_H_W_D = rearrange(
             self.self_attn(
                 # normalized_x_B_T_HW_D,
-                rearrange(normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
+                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
                 None,
                 rope_emb=rope_emb_L_1_1_D,
                 transformer_options=transformer_options,
@@ -524,7 +522,7 @@ class Block(nn.Module):
             h=H,
             w=W,
         )
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
 
         def _x_fn(
             _x_B_T_H_W_D: torch.Tensor,
@@ -538,7 +536,7 @@ class Block(nn.Module):
             )
             _result_B_T_H_W_D = rearrange(
                 self.cross_attn(
-                    rearrange(_normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
+                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
                     crossattn_emb,
                     rope_emb=rope_emb_L_1_1_D,
                     transformer_options=transformer_options,
@@ -557,7 +555,7 @@ class Block(nn.Module):
             shift_cross_attn_B_T_1_1_D,
             transformer_options=transformer_options,
         )
-        x_B_T_H_W_D = result_B_T_H_W_D.to(residual_dtype) * gate_cross_attn_B_T_1_1_D.to(residual_dtype) + x_B_T_H_W_D
+        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
 
         normalized_x_B_T_H_W_D = _fn(
             x_B_T_H_W_D,
@@ -565,8 +563,8 @@ class Block(nn.Module):
             scale_mlp_B_T_1_1_D,
             shift_mlp_B_T_1_1_D,
         )
-        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D.to(compute_dtype))
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
         return x_B_T_H_W_D
 
 
@@ -878,14 +876,6 @@ class MiniTrainDIT(nn.Module):
             "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
             "transformer_options": kwargs.get("transformer_options", {}),
         }
-
-        # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream
-        # in fp32, but run attention and MLP modules in fp16.
-        # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable
-        # quality degradation and visual artifacts.
-        if x_B_T_H_W_D.dtype == torch.float16:
-            x_B_T_H_W_D = x_B_T_H_W_D.float()
-
         for block in self.blocks:
             x_B_T_H_W_D = block(
                 x_B_T_H_W_D,
@@ -894,6 +884,6 @@ class MiniTrainDIT(nn.Module):
                 **block_kwargs,
             )
 
-        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D.to(crossattn_emb.dtype), t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
         x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]]
         return x_B_C_Tt_Hp_Wp

comfy/ldm/flux/layers.py

@@ -5,9 +5,9 @@ import torch
 from torch import Tensor, nn
 
 from .math import attention, rope
+import comfy.ops
+import comfy.ldm.common_dit
 
-# Fix import for some custom nodes, TODO: delete eventually.
-RMSNorm = None
 
 class EmbedND(nn.Module):
     def __init__(self, dim: int, theta: int, axes_dim: list):
@@ -87,12 +87,20 @@ def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dt
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
 
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
+
+    def forward(self, x: Tensor):
+        return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6)
+
 
 class QKNorm(torch.nn.Module):
     def __init__(self, dim: int, dtype=None, device=None, operations=None):
         super().__init__()
-        self.query_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
-        self.key_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
+        self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+        self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
 
     def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
         q = self.query_norm(q)
@@ -161,7 +169,7 @@ class SiLUActivation(nn.Module):
 
 
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
@@ -189,6 +197,8 @@ class DoubleStreamBlock(nn.Module):
 
         self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
 
+        self.flipped_img_txt = flipped_img_txt
+
     def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
         if self.modulation:
             img_mod1, img_mod2 = self.img_mod(vec)
@@ -196,9 +206,6 @@ class DoubleStreamBlock(nn.Module):
         else:
             (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
-        transformer_patches = transformer_options.get("patches", {})
-        extra_options = transformer_options.copy()
-
         # prepare image for attention
         img_modulated = self.img_norm1(img)
         img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
@@ -217,23 +224,32 @@ class DoubleStreamBlock(nn.Module):
         del txt_qkv
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
 
-        q = torch.cat((txt_q, img_q), dim=2)
-        del txt_q, img_q
-        k = torch.cat((txt_k, img_k), dim=2)
-        del txt_k, img_k
-        v = torch.cat((txt_v, img_v), dim=2)
-        del txt_v, img_v
-        # run actual attention
-        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
-        del q, k, v
+        if self.flipped_img_txt:
+            q = torch.cat((img_q, txt_q), dim=2)
+            del img_q, txt_q
+            k = torch.cat((img_k, txt_k), dim=2)
+            del img_k, txt_k
+            v = torch.cat((img_v, txt_v), dim=2)
+            del img_v, txt_v
+            # run actual attention
+            attn = attention(q, k, v,
+                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
+            del q, k, v
 
-        if "attn1_output_patch" in transformer_patches:
-            extra_options["img_slice"] = [txt.shape[1], attn.shape[1]]
-            patch = transformer_patches["attn1_output_patch"]
-            for p in patch:
-                attn = p(attn, extra_options)
+            img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
+        else:
+            q = torch.cat((txt_q, img_q), dim=2)
+            del txt_q, img_q
+            k = torch.cat((txt_k, img_k), dim=2)
+            del txt_k, img_k
+            v = torch.cat((txt_v, img_v), dim=2)
+            del txt_v, img_v
+            # run actual attention
+            attn = attention(q, k, v,
+                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
+            del q, k, v
 
-        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
+            txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
         img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
@@ -312,9 +328,6 @@ class SingleStreamBlock(nn.Module):
         else:
             mod = vec
 
-        transformer_patches = transformer_options.get("patches", {})
-        extra_options = transformer_options.copy()
-
         qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
@@ -324,12 +337,6 @@ class SingleStreamBlock(nn.Module):
         # compute attention
         attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         del q, k, v
-
-        if "attn1_output_patch" in transformer_patches:
-            patch = transformer_patches["attn1_output_patch"]
-            for p in patch:
-                attn = p(attn, extra_options)
-
         # compute activation in mlp stream, cat again and run second linear layer
         if self.yak_mlp:
             mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]