Automation across many runs
Generating one dataset is a panel job. Generating forty variants overnight is a script, and that is what the CLI and the API are for.
Without Python
No terminal, no YAML, no scripting.
Most synthetic data tooling assumes the person generating the data is the person training the model. Often they are not: the person who knows the scene is a technical artist, and asking them to learn a command line to place some crates is a bad trade. NameFrame Studio is a panel inside the Unreal editor that does the whole job.
- Where it runs
- inside the Unreal editor
- Steps
- 5
- Code required
- none
- Same pipeline as the CLI
- yes
What the panel is actually doing
Writing a job file. The Studio is a front end over the same contract the command line consumes, which has two useful consequences: nothing is exclusive to the no-code path, and a scene authored in the editor can be handed to an engineer to run a thousand times without anyone rebuilding it.
It also means the evidence is the same. A run started from the panel produces the same manifests, the same hashed configuration snapshot and the same gates as one started from a terminal.
Five steps, in the order they appear
The panel is organised as a sequence rather than a settings tree, because the order matters and getting it wrong is the most common way a first run goes badly.
- 01SceneDraw spawn, target, exclusion and camera zones directly in the level. Assign level actors to classes, one at a time or in bulk.
- 02CaptureSet resolution, field of view, render quality, weather and time of day. Pick a lens profile if you want one.
- 03DatasetChoose the output formats, the split, and which modalities to write.
- 04GenerateValidate, smoke test, then run. Progress and failures appear in the panel rather than in a log file.
- 05ResultsBrowse the frames, switch between RGB, boxes and segmentation, read the health grade and the gates.
The Generate step is deliberately last and deliberately gated. Validate resolves every asset the spawner will need, which was 38 assets across 4 pools on the reference run. The smoke test renders a handful of frames end to end. Only then does the full run start, and by that point nothing about it is a surprise.
Where the no-code path stops
Integration into training pipelines
If dataset generation is a step in a larger workflow, it needs to be callable. The Studio cannot be called; the Python package can.
Anything genuinely novel
A distribution nobody implemented, a format nobody wrote. The extension points are in code, because that is what extension means.
None of those are reasons to start in a terminal. Author the scene in the editor, confirm it works from the panel, then automate the part that repeats.