ctrl + enter to run · fine controls below the canvas
Write a prompt or shuffle for one, then press Generate.
More steps settle fine detail — and cost more GPU seconds.
How literally the model obeys the sentence.
Strength of the second, detail-recovery pass.
How far the returned frames drift from each other.
Lock it and the same prompt returns the same frame.
Recent prompts
Prompts you run above land here, with the style, ratio and seed that produced them.
Saved recipes
Prompt library →Save a frame above and its prompt, style family, ratio and seed are kept together — enough to reproduce that exact frame later.
What Generate actually triggers
The progress bar above is standing in for four real stages. In production each one runs on the pooled GPU fleet, and the timings are what the home page reports.
Your sentence becomes conditioning: subject, style cues and composition, read by hybrid multimodal encoders.
The latent core denoises toward that target across a series of steps, on a GPU worker pulled from the warm pool.
A second pass recovers edges and texture the base pass leaves soft — the difference between a preview and a deliverable.
Moderation screens the frame, it lands in object storage, and the edge cache streams the preview to you before the full render finishes.