FLUX 3 Image places objects by coordinates, leaves the rest intact
Black Forest Labs' new image model turns layout into structured JSON, keeps untouched pixels bit-identical, renders native 4K, and is 50% off through Oct 8.
Copy markdown
Layout as structured input, not a prompt guess
Define each element with an ID, a description, and a bounding box on a normalized 0–1000 grid as [y_min, x_min, y_max, x_max]. You get deterministic placement instead of coaxing a prompt to put the logo top-left — image gen you can actually script.
Edits that leave the rest bit-identical
Edit one region and every pixel outside it stays bit-identical, so iterative refinement never silently rewrites a background or a face. Up to 10 reference images cite inline (ref_image_0–9) for compositing from assets you already have.
Native 4K, no upscaler step
Output reaches ~16.8 MP (around 5,456×3,072) at full resolution with no separate enlarge pass. One model covers text-to-image, image-to-image, text rendering, and photorealism.
Live now, on a 50%-off clock
It's in the BFL Playground and API today, with API pricing 50% off through Oct 8. A commercial weights license lets you self-host, and an open-weights edition is promised within weeks — prototype now before the discount lapses.