Skip to main content

One month, three tiers

These are finished films from the same June 2024 fixture library: 133 assets and three synthetic people in a disposable Immich instance. The pictures are credited CC0 stock photographs; the dates, names and household story are invented. No personal library was used. The bundled soundtrack, calm_acoustic_1.opus, is MIT licensed.

Every request asked for a 60-second monthly highlight with photos, family sharing and template titles. Upload was disabled.

Basic and GPU​

Asking for the same June 2024 month, Basic and GPU each kept 14 shots and all four favourites. Twelve sources overlapped; GPU selected two different videos.

BasicGPU
Basic at 37 seconds: a swimmer underwater.GPU at 37 seconds: a mountain lake and tents.
IMG_2501.mp4IMG_2491.mp4
Basic at 48 seconds: a lake below a mountain and forested shore.GPU at 48 seconds: a waterfall and green rock face.
IMG_2524.mp4IMG_2521.mp4

Basic excluded one screen-like source; GPU excluded three (a screenshot, a scatter plot and a logo), none of which made Basic's cut either. This shows extra source filtering, not Basic shipping unsafe footage. Both films passed complete audio/video decoding.

Frame and film provenance records source filenames and sampling times.

Basic and Full​

Four actual frames from the Basic film: a playground swing, misty forest, market stall and mountain scenery.

Four actual frames from the Full film: a playground swing, birthday cake, forest fern and mountain scenery.

Full used real SmolVLM captions, detectors, the cached Laya checkpoint and a local text reader. There was no reader or caption fallback. Each cut selects 14 shots and keeps four favourites. 13 of 14 shots overlap: Basic includes IMG_2483.jpg; Full instead includes IMG_2432.jpg. Full rejects three screen-like assets at eligibility; Basic rejects one.

Recorded resultBasicFull
Finished duration56.0 s56.5 s
Selection wall time11.5 s42.7 s
Rendering wall time35.3 s35.5 s
File size36.9 MB36.1 MB
Video and audio1280×720 H.264, AAC1280×720 H.264, AAC
Complete decode checkPassedPassed

The duration is a target, not an exact output length. This is one paired example, not a quality guarantee or a speed benchmark across every library. A changed cut is something you can watch, not proof that one tier always makes a better film.

Limits of these runs​

Seven clips had no motion evidence in the Full run, so Full used plain clip facts for them. The store and cache started fresh and were shared between tiers; Full reused Basic's facts and acquired its own missing model facts. Both machines ran the same software environment, so these numbers are not a cross-platform speed comparison. Both finished files passed ffprobe stream inspection and complete video/audio decoding with ffmpeg.

The sanitized provenance records the sources used and selected filenames. The fixture credit manifest records each photograph's source and licence. No credentials or execution logs are included.

See Measure your setup for timing and cost numbers from the project's own hardware, and Choose your setup for what each tier adds.