Benchmarking your own setup
A service benchmark is not a whole-film benchmark. Record the release, hardware, resolved tier and cache state before comparing runs. For the project's own numbers, see Measure your setup. Current deployment choices are in Requirements and Optional upgrades.
Keep four measurements separate
| Measurement | What to record |
|---|---|
| Picture preparation | Number and type of previews, enabled producers, downloads, fresh or reused facts |
| Selection and text work | Selected shots, model, prompts, retries, input and output tokens |
| Rendering | Source media, output dimensions, frame rate, codec, encoder and title backend |
| Music | Bundled or generated track, generator model, stem separation and fallback |
Make the same cut twice to distinguish first-run acquisition from reused work:
immich-memories runs show RUN_ID
immich-memories report RUN_ID
Compare the same saved cut with the same output settings. Record failures and fallback routes rather than counting them as successful accelerated runs. The measured examples show whole-run timings and memory scopes by release and hardware.
Memory
Measure the app and each service separately. Process RAM does not establish GPU memory use, and a reachable CUDA service does not prove that every producer ran on CUDA. Use the service's health response to inspect its loaded providers; use container statistics and your GPU's monitoring tools for memory.
The unified GPU worker admits one phase at a time and releases model resources between phases. This changes residency, so measurements from separate always-resident services do not establish its memory requirement.
Cost and quality
Use the provider's reported input, cached-input and output tokens with the prices applicable to your account. Include retries, failed requests and incomplete usage reports. A local reader has hardware and electricity costs even without an API bill.
Watch the finished film before choosing an upgrade. Correct JSON, a passing API probe and a fast model do not establish that its edit is the one you prefer.