Compose V1 — one masked edit per run, scored
Pick an op in the bar at the bottom, fill in what it needs, upload a render, and run it. You'll get one result and a scorecard. Tick "dry run" to see the cost estimate without spending anything.
Pick an op in the bar at the bottom, fill in what it needs, upload a render, and run it. You'll get one result and a scorecard. Tick "dry run" to see the cost estimate without spending anything.
Upload a render (required) and, optionally, a reference/material sample. Then describe the edit you want in the chat below — the orchestrator proposes a plan (models, draws, cost) before anything runs.
Pick one of three edits below, fill in what it needs, upload a render, and Compose V2 plans a multi-model run deterministically — no chat, no orchestrator call. You'll see the cost estimate before anything runs.
This lane edits without a mask. Compose V1 (the other tab) only ever touches the pixels inside a drawn mask, so it's guaranteed not to change anything else in the image. Compose V2 has no such guarantee — it can change parts of the render you didn't ask it to touch. Every result below shows exactly how much of the frame changed, so you can judge each one for yourself. Use Compose V1 when you need certainty that the rest of the image stays untouched.