[Feature] Automatic "Highlights / For You" (curated Memories) view

[RWDai]

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

• Yes

The feature

Short summary Immich's Memories feature is a great start. I’d like to request an optional automatic "Highlights" (精選照片) capability that uses existing Memories/ML metadata to surface a small, curated set of high-quality photos and videos automatically (daily/weekly/monthly or on demand). The goal is to provide lightweight, privacy-respecting, self-hosted equivalents to “For You / Highlights” in other products.

Why this helps

Let users re-discover moments without manual browsing through large libraries.
Increases engagement (mobile & web) by presenting bite-sized curated content.
Leverages existing ML outputs (face groups, CLIP/text tags, OCR, quality heuristics) on the server side — no third-party sharing required.
Core proposal

Generation

A configurable server job produces “Highlights” periodically (e.g. daily/weekly) or on-demand per user.
Admins can enable/disable the feature globally; users can opt-in or opt-out per account.
Candidate selection (ranking signals)

Recency / “On this day” matches (same day previous years)
Face presence / known people prominence (prefer photos with recognizable faces)
Photo quality (sharpness, exposure, not duplicate/stacked)
Scene/CLIP tags or OCR relevance (e.g., travel, celebration, landmarks)
Low duplication / distinct shots preference
User-specified constraints (include/exclude archived assets, limit videos)
Presentation / UX

A Highlights tab (web) + Memory-style swipeable carousel (mobile).
Each highlight = a small set (3–12 assets) with a cover image and optional auto-caption (e.g., “On this day — 3 years ago”).
Per-item controls: remove from Highlights, pin another image as cover, regenerate Highlight.
Controls & privacy

Feature default: opt-in for servers / users (respect self-hosted privacy).
Option to disable ML-based selection entirely for privacy-conscious installs.
Store only references/IDs of chosen assets (no copies).
Admin UI to configure schedule, max items, whether to include videos/archived assets.
Notifications (optional)

Opt-in push notification for new Highlights (rate-limited, server-configurable).
Implementation notes (low-level hints)

Reuse existing ML pipelines (face/person metadata, CLIP tags, OCR) to rank candidates.
Store only the metadata (asset IDs / ordering) for each generated Highlight.
Expose generation as a scheduled Admin job and a manual “Generate now” action.
Provide graceful fallback when ML metadata is not available (e.g. pick recent/high-quality images).
Example user stories

“Show me highlights from last week” — user opens app, sees 6-image carousel of best shots.
“On this day” — server builds a memory from assets taken on the same date in previous years.
“Family favorites” — highlight picks prioritize images where selected People appear prominently.
Questions for maintainers

Do you prefer server-side scheduled generation (recommended for background processing) or per-user on-demand generation?
Any ranking signals you’d like to prioritize differently (faces vs scenes vs recency)?
Are there internal APIs the team prefers to reuse for selection and caching?

Platform

• Server
• Web
• Mobile

[bo0tzz]

#2836