Black Forest Labs has launched Flux 3 Image, completing the image arm of its multimodal Flux 3 family. The model introduces a structured JSON bounding-box composition system and guarantees bit-identical pixel preservation outside edited regions, eliminating the visual drift that plagues conventional inpainting. This architecture allows LLM agents to treat image layouts as persistent, editable project files. Flux 3 Image supports native 4K output and is available via API with a 50% discount through October 8, 2026.
Structured Layout and coordinate system
- ▪Flux 3 Image accepts a structured JSON array of scene elements where each element carries a unique identifier, a natural-language description, and coordinates mapping to a normalized 0-to-1000 grid
- ▪The bounding-box grid coordinates in Flux 3 Image are decoupled from pixel resolution and aspect ratio, allowing a layout to be applied consistently across different output sizes
Region-specific editing and pixel preservation
- ▪When editing a selected bounding box, Flux 3 Image regenerates only the specified region while leaving all pixels outside that box numerically unchanged
- ▪Black Forest Labs' published pixel-difference measurements for Flux 3 Image show that a cat-removal operation left 80.7 percent of pixels identical, while a hair-recolor operation left 67.8 percent unchanged
Integration with LLMs and prompt systems
- ▪Flux 3 Image includes a prompt upsampler that expands short instructions, but stable box IDs and coordinate arrays bypass this rewrite step to reach the model unchanged
- ▪Because element IDs and coordinates are stable JSON state, an LLM agent can treat the Flux 3 Image element table as a persistent project file across conversation turns
Native resolution capabilities
- ▪Black Forest Labs demonstrated Flux 3 Image's native resolution capabilities with a soba shop scene where hand-painted Japanese characters roughly 225 pixels tall remained legible
- ▪Flux 3 Image supports native output up to approximately 5,456 × 3,072 pixels without requiring a separate upscaling pass
Debatable claims
- ▪Programmatic JSON layouts are superior to natural-language prompting for AI image editing
- ▪Staggering open-weight AI releases behind hosted API launches is a justified business strategy
- ▪Bit-identical pixel preservation is essential for professional AI image editing
Story comments
Loading comments…