Skip to main content

FAQ

Does it modify my Immich library?

By default, no. It downloads copies of your videos for analysis and processing. If you enable --upload-to-immich, the generated compilation video is uploaded as a new asset (optionally into an album). Your original videos are never modified or deleted.

What video formats does it support?

Anything FFmpeg can decode, which is basically everything: MP4, MOV, AVI, MKV, WebM, you name it. Output supports mp4 and mov containers with h264, h265, or ProRes codecs.

Can I use it without face recognition?

Yes. Skip the --person flag and it'll pull everything eligible from the selected time period. Face recognition just narrows the pool to assets containing a specific person.

How long does analysis take?

Depends mostly on whether analysis runs on Apple Silicon / a GPU or on a CPU-only box. First-run analysis is roughly:

  • Apple Silicon or GPU: ~1 minute per 10 clips
  • CPU-only (4-core NAS class): ~1-2 minutes per clip

Analysis already downscales to 480p by default, so the resolution knob is not the one to reach for. What is left is --analysis-depth fast, which restricts the LLM pass to favorites, and preset: fast, the whole CPU-only profile. Results are cached, so a library only pays this once; re-runs over the same period are quick. The README resource table has RAM and encoding numbers, and the NAS-only guide has a Celeron-class table.

Can I run it headless?

Yes. The CLI works without a display. Use immich-memories generate with flags instead of immich-memories ui. Works fine over SSH, in Docker containers, and in CI pipelines.

Is it safe for production?

The codebase is AI-written (on purpose, as an experiment) with 5,600+ tests (5,000+ unit, 600+ integration/E2E) and strict quality gates. The output (music, clip selection, mood analysis) is AI-generated too, so results vary. Review what it produces before showing it at grandma's birthday party.

Can I generate for multiple people at once?

Yes. Use --person "Alice" --person "Bob" with --memory-type multi_person. By default it finds videos where both people appear together. See the generate CLI docs for all memory type options.

How much disk space does it need?

Downloads are a bounded cache, not per-run scratch, so the ceiling is set by config rather than by your library size. The defaults under cache: are 10 GB of downloaded video (evicted after 7 days), 2 GB of clip previews, 500 MB of thumbnails, and cache.db keeping 30 days of analysis results.

The output is small next to that. One measured run: 62 seconds of 1080p H.264 came out at 87 MB, or 30 MB under preset: fast. The NAS-only guide has the rest of that measurement.

Can it use iPhone Live Photos?

Yes. Live Photos are included by default (analysis.include_live_photos: true). Live Photos are ~3 second video clips captured with every iPhone photo. When you took photos in rapid succession, the tool detects the overlap and merges them into one continuous moment, cutting at the midpoint between consecutive shutter presses. Two real examples from the Live Photos page: three Live Photos merged to 4.5 seconds, six merged to 8.4 seconds.

Tested on iPhones. Samsung and Google Pixel motion photos should work (Immich normalizes them to the same field), but I only use iOS so it hasn't been tested firsthand. PRs from Android users welcome.

How big should my PRs be?

About 300 lines of diff, excluding generated and lock files, one concern per PR. Smaller PRs get reviewed faster and catch bugs earlier. If your change is bigger, split it into focused chunks. See CONTRIBUTING.md for the full guidelines.

Does it work on Apple Silicon?

Yes. VideoToolbox hardware acceleration is auto-detected. For music generation, ACE-Step works via MLX on Apple Silicon. For mood detection LLM, mlx-vlm is the recommended server.