Complete SDK workflow¶

1. Prepare¶

prepared = studio.prepare("sketch.jpg")

Preparation corrects EXIF orientation, converts to RGB, selects a model-safe size, extracts the Canny control, and returns diagnostics. It does not modify the source file. Review low contrast, edge density, fragmentation, and crop-risk signals before generation.

See Prepare.

2. Explore¶

study = studio.explore(
    prepared,
    intent=intent,
    outputs=4,
    seed_plan=SeedPlan.scout(4),
)

explore() resolves intent against the preset and backend, generates each candidate with an independent generator, and preserves creation order. The resolved recipe and all actual seeds belong to the study.

3. Pick¶

selected = study.pick(1)

Picking adds a selection record; it does not remove the other candidates. Use a stable candidate ID when integrating with another interface, and treat the index as a display convenience.

4. Vary¶

variations = studio.vary(
    selected,
    outputs=4,
    strength="subtle",
    locks=("structure",),
)

The new study points back to the selected candidate. A lock expresses intent, but the capability report remains the source of truth for how a backend applied it.

See Explore, pick, and vary.

5. Export¶

variations.export("design-study")

The export is a directory with normalized inputs, controls, candidate images, a contact sheet, and manifest.json. File hashes make later mutation visible.

6. Replay¶

report = studio.replay("design-study/manifest.json", mode="strict")

Inspect the returned report even on success. It describes resolved resources, runtime differences, and whether the generated artifact hashes match.

See Export and replay.