Train a style model

Train a custom image model (a LoRA) on a brand's visual identity from 10 to 50 reference images, then generate new imagery in that exact look from any prompt. This is the durable, at-scale answer when a team needs the same hand across endless asset variations. For a few consistent images right now, prefer character-consistency (references) instead. Training takes about two hours and is submitted once, then polled.

Inputs to collect

• The training dataset. A single ZIP of 10 to 50 images that share one visual style across deliberately varied subjects. (Ask for this if not provided. It is the whole job.)
• The target AIR to register the trained model under (your org namespace, e.g. yourorg:brand-illustration@1). Suggest a stable, versioned one if the user has none.
• Model name + short description for the registered model, and whether it should be private (default true while iterating).
• Optional: what subjects they will generate after, so you can sanity-check the dataset has the right subject variety.

Models

• Default: Exactly Illustrative Training (exactly:illustrative@training, live) - trains a style LoRA from a ZIP dataset and registers it for imageInference. The only live op:train model today.
• Confirm the live trainer and its schema via runware-models (capability op:train) + runware-run before calling. Never hardcode a stale choice. The catalog moves.

Workflow