Choose your setup
Every tier makes a complete film with titles and music. Pick the one your hardware can run. Start with Basic if you are unsure: you can switch tiers later and keep your work.
| Tier | What changes in the film | Install |
|---|---|---|
| Basic | Dates, favourites, people and places guide the edit; small CPU classifiers read the pictures. Output is up to 1080p. | Docker Compose, NAS or native Mac/Linux |
| GPU | Adds picture descriptions, document/sensitive-content checks and a family-viewing pre-screen. | Choose GPU in Quick start or the setup builder; on Apple Silicon use native Mac |
| Full | Adds a text reader that refines the draft. | Choose Full in Quick start or the setup builder; on Apple Silicon use native Mac |
For an existing cluster, Kubernetes covers all three tiers. Compare the same source material to see what the tiers can change.
Basic: start with the film
Basic runs on the app's CPU without a separate model server. Allow two cores, 4 GiB for the app, and about 25 GB of app data plus 0.35 GB per 1,000 pictures, plus the images and finished films. A plain NAS does the whole job.
The first cut prepares the pictures it needs; later cuts reuse that work. Rendering still runs each time. Start with a small album to get a film before preparing a large period.
GPU: understand more of the pictures
GPU adds inference, a caption service and the Laya family-viewing check. On NVIDIA, the Compose route starts the inference and caption containers together. On Apple Silicon, the native app uses Metal and a local caption server. A NAS can use services on a separate NVIDIA machine.
Allow room for the models as well as the app: the GPU Compose app limit is 8 GiB, with separate limits for its model services. Check hardware and memory requirements. Review the cut before sharing; model checks can miss things.
Full: refine the draft
Full needs the GPU setup and an explicitly enabled reader with a 32k context window. Native installs can run the app-owned reader; Docker and Kubernetes connect to a reader server. Reader setup covers local and hosted choices.
The reader receives annotation text, including people and places. Each proposed edit must pass the selection checks. If refinement fails, the app keeps the rules draft and reports why.
What each tier includes
| Feature | Basic | GPU | Full |
|---|---|---|---|
| Dates, favourites, people, trips, titles and bundled music | Yes | Yes | Yes |
| Captions and additional sharing checks | No | Yes | Yes |
| Text-reader refinement | No | No | Yes |
| Maximum resolution | 1080p | 4K when the renderer supports it | 4K when the renderer supports it |
| Generated music | Optional | Optional | Optional |
Hardware video encoding speeds up rendering and does not select the GPU tier. Encoding, remote rendering and generated music can be configured separately. A reader on Basic picks the music mood (and names special days when model titles are on); it does not turn on Full selection by itself.
Change services later
Use Settings for inference and caption URLs and the reader's URL, model and enabled switch. A locked field names the configuration source that controls it. Compatible facts and review decisions survive a tier change.
After changing tiers, use Download models in Settings if it appears, then check the new configuration. The resolved tier and its required checks must agree with what you requested. Then make a film.