One month, three tiers
July 2025 of the demo library, cut once on each tier: Basic, GPU and Full. The library is Wikimedia Commons photographs (CC0 or CC BY) with an invented household on top: Kit's birthday, the grandparents' visit, a road trip to Nordkapp. It ran on a development build of 1.0.0 (between rc.10 and rc.11), on an M5 Max Mac with 128 GB, against a real Immich seeded with the library. No personal library was used.
Every tier got the same request: a 60-second monthly highlight, photos on, family sharing, the
app's own title, and the same bundled track (calm_acoustic_1.opus, MIT licensed like the rest of
the bundled music).
Each tier had its own home, so nothing one tier prepared reached the next.
The film
All three tiers picked the same 15 shots, in the same order, and rendered frame-identical films. This is that film:

On this month the people, the favourites and the stories settle every pick before a caption gets a say: three shots of Kit's birthday, the only picture of Pat, the visit, then the road north. A library with more lookalike pictures per moment gives the extra evidence more to decide.
What changed
| Basic | GPU | Full | |
|---|---|---|---|
| Shots in the cut | 15 | 15 | 15 |
| Pictures checked by the local classifiers | 1,131 | 1,131 | 1,131 |
| Captions written (local SmolVLM server) | none | 15 | 15 |
| Calls to the text reader (a local model) | none | none | 6 |
| Summary of the film | none | none | none |
| Time from request to finished film | 54 s | 76 s | 95 s |
No tier wrote a summary because none was asked for. Model-written titles and the summary are
experimental and off by default (title_screens.llm_titles, or --llm-title for one run). Turn
them on and the summary is what runs story prints and what goes into the
Immich description on upload. Full still asks its reader with them off: 26 model calls in that
run, 20 of them the captions and other preparation, 6 to the reader refining the rules draft.
The time is one run per tier on one machine: read the order, not the numbers. Rendering took 28 s on Basic, the first run, and 19 s on GPU and Full; the rest of the gap is preparation (24 s on Basic, 56 s on GPU, 74 s on Full).
This is one month of a small, curated library, cut once per tier. It shows what each tier adds and what it cost here, not that one tier always makes a better or a different film.
The provenance records the code
revision, the software and model versions, the exact request, and for each tier the pictures it
chose (by fixture file name), its summary, its timings and the decode check. The film passed
ffprobe stream inspection and a complete ffmpeg decode. scripts/record_tier_example.py
makes all of it again.
See Measure your setup for timing and cost numbers on bigger libraries, and Choose your setup for what each tier adds.