Centralize execution provider selection in providers.py with
auto-detection and graceful fallback. All ONNX sessions (embedder,
detector, face processor, recognizer) now use the configured providers.
- New VISION_EXECUTION_PROVIDERS env var: "auto" for GPU auto-detect,
or explicit "CUDAExecutionProvider,CPUExecutionProvider"
- Provider priority: CUDA > ROCm > OpenVINO > CPU (when set to "auto")
- docker-compose.yml includes commented-out NVIDIA GPU deploy section
- Supports onnxruntime-gpu as a drop-in replacement for onnxruntime
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace OpenCLIP ViT-B/32 (512-d, ~78% recall) with SigLIP2 ViT-B/16
(768-d, ~84% recall) as the default embedding model for significantly
better image-text retrieval quality.
- New SigLIP2Embedder class with 384px input and SigLIP normalization
- ONNX export pipeline for SigLIP2 visual + textual encoders
- Migration 0010: resize embeddings.vector from 512 to 768 dimensions
- Config-driven model selection: "siglip2_vitb16" (default) or
"openclip_vitb32" (legacy) — both models can coexist
- Content classifier follows the configured embedder family
- Existing embeddings cleared on migration; vision backfill regenerates
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace O(N²) pHash-only duplicate detection with a hybrid approach:
- pHash Hamming distance for exact/near-exact copies
- CLIP embedding cosine similarity via pgvector HNSW for visually
similar photos (crops, format changes, screenshots)
Post-scan now uses incremental mode: only newly added photos are
compared against the full library — O(new × log N) via HNSW index
instead of O(N²). Full regroup remains available from Settings.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Introduce username/password authentication with admin and user roles.
Each user gets their own media directory under /photos/{username}/ with
isolated photos, folders, heaps, and tags. Admins manage users and
observe the full library from a dedicated Settings page.
Backend:
- User model with bcrypt passwords and JWT access/refresh tokens
- Auth router (login, refresh, setup, change-password, status)
- Admin router (user CRUD with last-admin protection)
- user_id FK added to photos, folders, source_roots, heaps, tags
- All data routers scoped by authenticated user
- Scanner inherits user_id from source root owner
- Thumbnails stored under user-prefixed paths for isolation
- Library endpoints accept ?scope=global for admin cross-user view
- Alembic migration 0009 with data migration for existing installs
- Defensive bootstrap.py handles fresh vs existing DB startup
Frontend:
- AuthContext with token lifecycle, auto-refresh, login/logout
- Login page, first-run setup page, auth gate in App.tsx
- Bearer token interceptor on all API requests
- User identity + logout in left sidebar
- Admin-only Settings page with Library Management and Users tabs
- UserManagement panel (add, edit role, reset password, deactivate)
- Settings shows global stats across all users for admin
- Filter bar, right sidebar, keyboard hints hidden on settings page
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Lets operators fix corrupted capture dates at scale. Adds an editable
Date Taken field with a folder/filename-derived suggestion hint, a bulk
Date Taken section in the multi-select sidebar that either applies one
date to the whole selection or infers a per-photo date from each path,
a warning badge on thumbnails whose stored date disagrees with the
path, and a "Date issues" filter pill so suspicious photos can be
surfaced and fixed as a group. Edits are written back to EXIF on disk
so rescans don't clobber the fix.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a per-folder "hide from views" toggle so noisy subtrees
(screenshots, WhatsApp dumps, work archives) can be excluded from
cross-cutting views without losing indexing. Photos under a hidden
folder are still scanned, thumbnailed, embedded, OCR'd, face-
extracted — they just stop appearing in All Photos, Rated, Colors,
Map, Tags, People, Search, Duplicates, and the sidebar counts.
Navigating directly into the folder still shows every photo.
Schema (migration 0007_folder_hidden):
- folders.is_hidden user-set toggle, default false
- photos.is_hidden denormalized effective flag (true iff any
ancestor folder is hidden), indexed so cross-
cutting queries stay on the existing planner
paths
The denorm is maintained by two paths:
- The scanner walks the ancestry chain on insert, with a per-scan
memoized cache so each folder is resolved once per scan.
- POST /api/v1/folders/{id}/hide flips folders.is_hidden and runs a
WITH RECURSIVE CTE to recompute every folder's effective state in
one query, then bulk-updates photos WHERE IS DISTINCT FROM. Runs
in ~10 ms on a 13k-photo library.
Filters added (cross-cutting queries):
- /library/stats — every sidebar badge via a shared `visible` filter
- /photos (list) — only when neither folder_id nor heap_id is set;
folder browse and heap browse always show everything
- /photos/map
- /library/duplicates/groups
- /folders/tree photo_count subquery
- /tags count_subq (drives Tags + People sidebar counts)
- services/duplicates.regroup_duplicates (so hidden dupes never
contaminate the Duplicates view)
- services/search.hybrid_search — both semantic (pgvector) and FTS
legs join photos so rankings don't include hidden results
Intentionally NOT filtered:
- /photos?folder_id=X and /photos?heap_id=X (user-intentional browse)
- /library/maintenance/pipeline-stats (tracks real worker state)
- cleanup service (disk-level ops, not views)
Frontend:
- sourceFolders.setHidden(id, hidden) API client method
- FolderTreeNode.is_hidden carried through the tree into TreeItem
- LeftSidebar kebab menu: "Hide from views" / "Show in views" with a
mutation that invalidates folders, photos, stats, and tags caches
- Hidden folder rows swap the Folder icon for EyeOff and render the
label italic/muted so the state is visible at a glance
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three overlapping fixes so the ingestion pipeline actually runs and the
user can see what it's doing:
Pipeline recovery
- app/database.py: use NullPool when MULITA_CELERY_WORKER=1 so each
Celery task opens a fresh asyncpg connection on its own event loop.
Fixes "another operation in progress" and "Future attached to a
different loop" errors that were dropping ~every thumbnail +
extract_metadata task on the floor.
- app/tasks/thumbs.py: initialize photo=None before the try and rollback
on error so a transport failure in the initial SELECT doesn't raise
UnboundLocalError in the except block and leak rows stuck in 'pending'.
- app/services/vision/bootstrap_models.py: on missing model files,
invoke export_models automatically instead of just warning. First
boot of a fresh install now self-heals.
- app/services/vision/export_models.py: shutil.move instead of
Path.rename so the YOLO export survives the /app → /data/models
cross-volume hop.
- requirements.txt: add ultralytics so export works in a stock image.
Worker topology
- docker-compose.yml: replace the single worker with worker-light
(default/high/low queues, c=2, IO-bound) and worker-vision (vision
queue, c=5, OMP_NUM_THREADS=1 to avoid oversubscription on 6 cores).
Vision is pinned to ≤5 parallel inferences so ONNX doesn't each
spawn an all-cores intra-op pool.
- .env / .env.example: CELERYD_CONCURRENCY replaced with
CELERY_LIGHT_CONCURRENCY + CELERY_VISION_CONCURRENCY.
- Backfill queries in thumbs / scan / vision now ORDER BY taken_at
DESC NULLS LAST so newest photos finish first — the library fills
in top-down in the UI instead of arbitrary insertion order.
Settings visibility
- routers/library.py: new GET /maintenance/pipeline-stats returning
done/total per stage (thumbnails, exif, gps, phash, embeddings,
tags, ocr, faces, face clusters, duplicate groups). Worker-status
now also reports the `vision` queue depth, which was missing.
- services/api.ts: PipelineStats / PipelineStage / ScanStatus types
and the matching client call.
- components/dialogs/SettingsDialog.tsx:
- new Pipeline Progress card with one progress bar per stage
- inline scan banner (processed/total/current folder) inside the
Library section while a scan is running
- Tasks/min throughput computed by diffing worker processed counters
between polls
- Workers section calls out the vision queue and documents the
CELERY_LIGHT/VISION_CONCURRENCY + docker compose up -d scale path
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix Photo.created_at → Photo.added_at (column doesn't exist)
- Fix Postgres DISTINCT + ORDER BY conflict by using a subquery for
tag_id filtering instead of JOIN + DISTINCT on the outer query
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rewrite faces.py to use cv2.FaceDetectorYN instead of raw ONNX
(handles multi-scale anchor decoding and NMS internally)
- Load original photo files at up to 4000px for face detection instead
of 240px thumbnails — faces were too small to detect at thumbnail res
- Falls back to thumbnail if original is unavailable
Tested: 33 faces extracted from 13 photos, clustered into 1 person.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rewrite all vision tasks to use sync psycopg2 sessions instead of
asyncpg — fixes 'another operation in progress' and event loop errors
when Celery forks workers sharing the async connection pool
- Letterbox-pad images to exactly 640x640 for YuNet face detector
(was crashing on non-square thumbnails)
- Deduplicate object detections per label per photo — keep highest
confidence only to avoid photo_tags PK violation on multiple
detections of the same class
- Add all queues (-Q default,high,low,vision) to worker command
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add export_models.py for OpenCLIP ViT-B/32 and YOLOv8n ONNX export
- Fix ArgMax(13) ORT ARM64 incompatibility by passing eot_indices as a
separate ONNX input (computed outside the graph in embed.py)
- Use legacy TorchScript exporter (dynamo=False) for IR version 9 compat
- Upgrade onnxruntime to 1.18.1
- Rewrite bootstrap_models.py with clear separation of auto-downloadable
models (YuNet, SFace) vs manually-exported ones (OpenCLIP, YOLOv8n)
- Wire bootstrap into worker CMD (runs before Celery)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Create ocr_text table for storing per-region OCR results
- Add tsvector search_vector column to photos with GIN index and
auto-update trigger on filename/user_title/user_notes
- Implement ocr_photo Celery task using rapidocr-onnxruntime
- Add FTS leg to hybrid search: queries photos.search_vector and
ocr_text via UNION, fused with semantic results via RRF (k=60)
Migration 0004 backfills search_vector for existing rows.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Wire the full embedding flow:
- Rewrite Embedding model to use pgvector Vector(512) with HNSW index
- Add embed_photo, vision_fanout, backfill_vision Celery tasks on
dedicated `vision` queue
- Hook vision_fanout into generate_thumbnails completion
- Add POST /api/v1/photos/search with hybrid RRF ranking (semantic-only
for now; FTS leg added in PR5)
- Stub ocr_photo, detect_objects, extract_faces tasks for later PRs
Migration 0003 drops/recreates the embeddings table (was never populated).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a new Map sidebar entry that plots photos by their EXIF GPS
coordinates on a clustered Leaflet map. While wiring this up, the
metadata extractor was reading unprefixed GPS keys that never exist
in `exiftool -G -j` output AND assumed coordinates were already
floats — every photo silently lost its GPS. The new extract_gps
helper handles Composite/EXIF group prefixes and parses DMS strings,
and lat/lon are stored as first-class indexed columns so the map
can query them without parsing exif_json on every request.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Duplicates section was useless: SHA-256-only detection only caught
byte-identical files, not the actual duplicates a real library
accumulates (re-encoded JPEGs, screenshots, resized exports), and the
view was a flat date-sorted list with no grouping or actions. This
replaces the whole flow.
Detection
- New phash + duplicate_group_id columns on Photo, added via an
idempotent ALTER TABLE pass in init_db (the project has no Alembic).
- Thumbs worker computes a 64-bit pHash from the original-resolution
decoded frame just before the destructive thumbnail loop. Falls back
silently — phash is nice-to-have, not a blocker for thumbnails.
- backfill_phashes Celery task fills in phashes for photos that
predated the column, reading the existing thumb_large rather than
re-decoding the original.
- regroup_duplicates service runs union-find over Hamming distance
(threshold 6), persists duplicate_group_id, and maintains is_duplicate
as derived state so existing badges/counts keep working. Chained
after scan_all_source_roots with a 60s countdown.
API
- GET /library/duplicates/groups returns all groups with members,
bucketed in Python from one query. Each group has a reason ("exact"
iff every member shares a SHA-256, "similar" otherwise).
- POST /library/maintenance/{regroup-duplicates,backfill-phashes}.
Frontend
- New DuplicatesView (sectioned grid, one section per cluster) replaces
the timeline when the user is in the duplicates section. Each section
shows a "Keep best, discard N" button that picks the highest-pixel
copy and reuses the existing undoable bulk-discard so Cmd+Z works.
- Manual best override: hover any non-best thumbnail and click "Keep
this" (Crown icon, top-right) to override the auto-pick. The header
annotates "(manual)" so it's obvious which copy will be kept.
- Keyboard nav within the duplicates view walks the flat member list,
with ↑/↓ jumping by the measured column count and scrollIntoView on
every move. Timeline's keyboard handler now early-returns in the
duplicates section so the two don't fight.
- BEST pill / Keep-this button live at top-right with a ring outline so
they don't collide visually with the cyan selection ring around a
selected cell. Dimensions chip moved to bottom-left to free both
right corners for the keep affordances.
- New "Duplicates" section in SettingsDialog: shows group/member counts
and exposes both backfill + re-detect actions, sharing a query cache
with DuplicatesView via DUPLICATE_GROUPS_QUERY_KEY.
- PhotoInfoPanel "Basic Info" section now shows the photo's full file
path in monospace below the size/dimensions/date grid.
- New imagehash==4.3.1 dep in requirements.txt.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
prune_missing_photos() previously only deleted Photo rows whose files
were gone, leaving every folder row from the old library in the DB —
which made the sidebar tree wildly out of sync with the on-disk
structure (still showing /photos/2024/, /photos/2026/03/, etc. that
no longer exist).
Now also drops Folder rows whose path doesn't resolve under a mounted
source root, with the same defensive "skip if source root unmounted"
guard. The Settings orphan card surfaces both counts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds /api/v1/library/maintenance/{missing-stats,prune-missing} backed
by a new cleanup helper that deletes Photo rows whose files no longer
exist on disk under a *mounted* source root. Skips photos under
unmounted roots so a temporarily-disconnected drive doesn't get
silently nuked.
Settings panel surfaces the orphan count with a destructive Prune
button, plus a "Kick pending" action that re-queues photos stuck in
processing_status='pending' (typically left behind when the scanner
created the row but the worker never picked up the thumbnail task).
Common trigger: PHOTO_DIRS in .env was repointed at a different
library root, leaving every old row dangling.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Diagnosis: every backend restart was dispatching watch_folders.delay()
unconditionally. watch_folders is an infinite-loop celery task
(for changes in watch(*paths)). With CELERYD_CONCURRENCY=4 and several
restarts during dev, all four worker slots ended up pinned by stale
watch_folders instances, leaving zero workers free for scan_folder.
The result: clicking "Scan all folders" successfully queued a task
that then sat in the queue forever, the new /photos/sub folder was
never walked, and the user's newly added photo never appeared.
The watcher was only opportunistically useful and the user already
triggers scans manually. Disabling it removes the foot-gun. Re-
enabling needs:
- a Redis lock so only one watcher runs at a time
- or a dedicated long-running container with concurrency=1
- or a celery beat schedule with a singleton flag
Until then, manual scans work. Cleared the backlog by wiping the
redis broker volume so the stale watch_folders tasks are gone.
Verified: post-fix, scan_folder runs in 0.12s and reports
"Processed 7/7 files. Errors: 0", picking up the previously missing
/photos/sub/Samuel_Colman... file.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The frontend AddSourceFolderDialog let users register source roots
from inside the app, but with the bootstrap auto-creating one for
the /photos mount on first boot, the dialog was redundant in the
common case and confusing in every other (users had to know which
container path corresponded to their host directory). Going
config-driven matches Plex/Photoprism/Immich and matches the
mental model "the docker mount IS the library".
Frontend
- Deleted components/dialogs/AddSourceFolderDialog.tsx entirely.
- LeftSidebar drops the "+ Add Source Folder" button + bottom-bar
layout, the addFolderMutation, the dead Plus action button on
the (no-longer-existing) folders/heaps tree headers, and the
Plus icon import.
- api.ts: removed sourceFolders.add(), library.browse(), and the
BrowseChild / BrowseResponse types. The remaining sourceFolders
surface is read-only (list + manual scan).
- LeftSidebar bottom strip is now just the "Scan all folders"
button when there's at least one source root.
Backend
- Dropped POST /folders (no consumers) along with FolderCreate /
FolderResponse pydantic models. The folders router header now
documents the config-driven approach.
- Dropped GET /library/browse (no consumers). Removed the unused
os/HTTPException/SourceRoot imports it brought in.
- cleanup_data_integrity now also walks the source roots and logs
a warning for any whose path is missing on disk. Doesn't auto-
delete (a missing path could be a temporarily unmounted drive)
but surfaces enough hint to fix it. Returns the count in the
summary dict alongside merged-duplicates.
Docs
- README "How libraries are managed" section rewritten to spell
out that mounts ARE source roots, edit .env + restart, no UI for
managing source roots. New "Changing or adding libraries"
section walks through the typical edit-restart loop including
the optional volume-nuke for a clean slate.
- "Adding more libraries" subsection covers multi-mount via
edited compose with a note that auto-registration is roadmap.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Cleans up the maze of overlapping ways folders entered the app, plus
removes the dead trash plumbing left over from the soft-discard
refactor.
Setup model (now)
- ONE env var: PHOTO_DIRS in .env, set to the host path of your
library. Compose mounts that at /photos. That's the entire setup.
- On first boot, the backend auto-creates a SourceRoot row named
"Library" pointing at /photos so the user sees their photos
immediately without configuring anything.
- Source roots and discard live in the database; mulita.yml only
carries operational settings (thumbnails, scanner, performance).
- The "Add Source Folder" dialog is now a directory browser
restricted server-side to /photos and any existing source root —
the user clicks through actual mounted directories instead of
typing container paths they can't possibly know.
Backend
- New services/scanner.bootstrap_default_source_root(): if no
SourceRoot rows exist and /photos is mounted, create one. Wired
into the lifespan handler before cleanup + initial scan.
- New GET /library/browse?path= returning the immediate child
directories of `path`, validated to live under one of the allowed
roots (default mount + every active SourceRoot). Hidden entries
are filtered. Children are tagged with is_existing_root so the UI
can show an "Added" badge. Returns parent path for up-nav, or
null when at the top of the allowed scope.
- scan_all_source_roots now reads from the DB instead of the YAML
config so DB-managed source roots are honoured by initial scan.
- Dropped the placeholder source_roots block from mulita.yml — the
paths /photos/main and /photos/iphone never existed and just
produced startup warnings.
- Dropped TrashSettings, settings.trash, settings.source_roots,
and the SourceRoot pydantic model from config.py. Soft discard
has owned this for a while; it was dead code.
Compose
- Single ${PHOTO_DIRS:-./photos}:/photos:rw mount in both backend
and worker.
- Removed the hardcoded ~/Pictures:/host/Pictures:rw mount — the
PHOTO_DIRS variable is the single source of truth now.
- Removed the trash_data named volume + mounts (no consumers).
- backend/Dockerfile no longer creates /data/trash; it now creates
/data/proxies (which the proxy endpoint actually uses).
Frontend
- AddSourceFolderDialog rewritten as a directory tree picker:
loads /library/browse on open, lets the user navigate up via a
ChevronUp button or down by clicking subfolders, shows the
current path inline, and adds whatever directory is currently
shown. Existing source roots are tagged "Added" so the user
knows what's already registered. Errors from the backend (e.g.
trying to navigate outside the allowed scope) surface inline.
- New library.browse() helper + BrowseChild / BrowseResponse types
in services/api.ts.
Docs
- README Quick Start rewritten around the single PHOTO_DIRS env
var, with macOS/Linux/Windows examples.
- New "How mounted folders and source folders relate" section that
spells out the two-layer model (mount = visibility, source root
= scanning) so the most common confusion is addressed up front.
- Added a "Read-only libraries" subsection that lists exactly which
endpoints fail under :ro.
- "Configuration" section reframed: source roots are managed by the
UI/API now, mulita.yml is operational settings only.
- .env file now has examples for the common host paths.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two related fixes:
1. Prevention — scanner now normalizes paths before lookup/insert
in get_or_create_source_root and get_or_create_folder. Trailing
slashes, redundant separators, and `.` segments all collapse to
the same row. _normalize_path uses os.path.normpath; symlinks
are intentionally NOT resolved so mount paths stay intact for
cross-machine portability.
2. Cleanup — new app/services/cleanup.py runs on backend startup
(idempotent) and merges any pre-existing duplicates left over
from older scanner versions:
- Groups source_roots by normalized path. Picks the canonical
row (preferring one with a non-empty name and the earliest
added_at), re-points child Folder rows via UPDATE, and
deletes the duplicates.
- Same for folders, with photo_count as the tiebreaker. Photos
get re-pointed to the canonical folder via UPDATE.
- Recomputes folder.photo_count from the actual non-discarded
photo membership so the sidebar count matches reality.
Wired into main.py's lifespan handler. On the dev DB this merged
the empty-name "/host/Pictures/MulitaTest/" duplicate that was
showing up alongside the canonical MulitaTest source root.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
