Black Forest Labs launches FLUX 3 Image with grid controls

Black Forest Labs has launched FLUX 3 Image, the image generation and editing component of its multimodal FLUX 3 family. Unlike traditional text-to-image systems that infer layouts from prose, this model allows developers to define scenes using structured coordinates. By mapping a canvas to a normalized 0-to-1000 grid, users can specify the exact location of elements using bounding boxes formatted as y-min, x-min, y-max, and x-max. This structured approach aims to eliminate layout drift, making it highly useful for complex designs like magazine covers, product composites, and collages.

The model also introduces robust multi-turn editing capabilities and reference image integration. During iterative edits, pixels outside the designated bounding boxes remain entirely unchanged, preserving the surrounding image with bit-identical accuracy. Additionally, developers can incorporate up to 10 reference images in a single generation, citing them inline from ref_image_0 to ref_image_9. The model automatically determines the scale and placement of these references before rendering the final scene, which is ideal for programmatic workflows like fashion lookbooks and e-commerce composites.

FLUX 3 Image supports native high-resolution outputs up to approximately 16.8 megapixels, or roughly 5456 by 3072 pixels. By rendering directly at these resolutions, the model avoids separate upscaling passes that often degrade text and fine textures. The Freiburg-based startup, founded in August 2024 by former Stability AI researchers, is offering the model through its API and Playground. To mark the launch, the company is offering a 50 percent discount on API usage through October 8. While commercial weights are currently available for licensing, Black Forest Labs plans to release an open-weights version of the model in the coming weeks.