FLUX 3 Image is BFL's new text-to-image/editing model with spatial composition as a first-class input. The canvas is normalized to a 0–1000 grid regardless of aspect ratio; you specify elements as bounding boxes in [y_min, x_min, y_max, x_max] form, each with a semantic id and description. The full prompt is a global caption plus a JSON element table, with the caption citing each element by id. Key operational details: untouched boxes are guaranteed to survive edits verbatim — same id, same coordinates — so you can do multiple targeted edits per round without composition drift. A prompt upsampler expands terse requests but never rewrites your boxes; its added elements only persist if the caption references them. An LLM agent can plan the entire layout from one line, so you don't hand-draw boxes unless you want to. Up to 10 reference images, native 4K rendering, commercial weights license for self-hosting. In 1987 I did this exact thing on my Amiga with DPaint and a grid overlay, except the model was me and the guaranteed coordinate fidelity was "hope." Black Forest Labs has finally replaced hoping with a JSON array, which is roughly the entire history of computer graphics in one sentence. Sam Altman is reportedly furious that someone shipped a box-based product without first scanning an eyeball into it. Open question: does the edit-preservation guarantee hold under fine-tuning, or is "untouched boxes stay put" a property of the base weights only?