[Feature] Automatic scene-based culling: group by time and place, keep the best shots

[Muyu5217]

I have searched the existing feature requests, both open and closed, to make sure this is not a duplicate request.

• Yes

The feature

Add an automatic culling utility that groups photos into scenes and suggests which ones to keep.

How it would work:

1. Group assets whose capture times are within a configurable window (e.g. 120s) and, when available, whose GPS coordinates are close (e.g. 200m).
2. Score each asset in a scene using signals Immich already has or can compute cheaply: preview sharpness, detected face area, exposure, resolution.
3. Suggest keeping the top K per scene (e.g. 3 for small scenes, 10% capped at 10 for large ones), and mark the rest for trash.
4. Review UI under Utilities: a per-scene grid with keep/trash suggestions, adjustable per photo, plus a one-click apply.
5. Everything goes to trash — never permanent deletion.

Why the existing duplicate detection is not enough

• It compares CLIP embeddings with maxDistance capped at 0.1. Burst frames with different expressions are typically 0.05–0.17 apart, so many never get grouped.
• It has no notion of capture time or location, so it cannot distinguish "same event" from "similar-looking events".
• It only keeps one asset per visual pair. A 30-shot birthday burst still leaves 30 photos in the timeline.

Related discussions: #24332 (manual culling mode) and #2479 (automatic stacking) are complementary — this proposal is automatic grouping plus selection.

Implementation notes

This could reuse the smart-search embeddings, asset_exif, asset_face and existing preview thumbnails; no new ML model is strictly required. Community tools such as immich-deduper and immich-stack already implement parts of this.

Would the maintainers be open to something like this in core, or would it be better kept as a companion tool?

[bo0tzz]

#10665