v0.4.0

What's new

Features

• YOLO animal detection: bounding boxes are computed before classification. Multi-pet photos are handled per crop, so each animal in the frame is classified separately. Improves accuracy and enables tagging photos with more than one pet.
• Visual ref search: "Find similar photos" now uses ref asset images as the CLIP query instead of text description, producing far more relevant candidates.
• Find missed photos: new "Find missed" button scores borderline candidates with the classifier and surfaces photos just below the confidence threshold. Useful for finding good refs to improve recall.
• Score-based negatives calibration: "Find not my pets" uses the classifier to exclude photos the model already considers pets, ensuring only genuinely ambiguous photos are surfaced as negatives.
• Live scan: the scan panel shows live per-category counts (Tagged, Low conf., Other, Already tagged) and the current photo date while scanning. Triggering a new scan cancels any in-progress one.
• Review low confidence: after a scan, a new "Review X low confidence" button lists photos the classifier identified as a pet but scored below threshold, sorted by score with color-coded badges.
• Open in Immich: each thumbnail in all photo grids now has a direct link icon to open the asset in Immich.
• Embedding cache persisted to disk: CLIP embeddings are saved to data/embeddings.pkl and reloaded on startup, so restarts no longer re-embed all ref and negative photos.

Fixes

• Critical: duplicate face prevention: "person": null faces (Immich-detected but unassigned) caused the existing-faces check to throw and return an empty set, bypassing deduplication and creating duplicate tags. Fixed with a null guard in all face lookup paths.
• Pagination cap removed: asset search was silently stopping at 1000 results due to a total field that Immich caps at 1000. Now paginates correctly until the page is smaller than the page size.
• Scan dedup across YOLO crops: a person was not marked as tagged when face_id retrieval failed after a successful POST, causing the same pet to be re-tagged on subsequent crops of the same photo.
• Pet delete: per-ref face deletion loop removed. Deleting the Immich person cascades face removal automatically.
• activePet stale reference: "Find missed" button stayed disabled after re-enrollment because the in-memory pet reference was not refreshed after loadPets().
• Responsive photo grid (auto-fill columns instead of fixed 4).
• Browser locale used for date formatting in tooltips and date inputs.

Docs

• README rewritten with features overview, docker-compose setup, and enrollment tutorial.
• Added guidance on picking good reference photos (skip multi-pet, blurry, or ambiguous frames).