FLUX.2 Max

Black Forest Labs' highest-fidelity FLUX.2 tier, priced per megapixel rather than per image. It targets maximum detail and prompt adherence, and is the dearest of the FLUX.2 family to run.

Max arrived three weeks after the rest of the FLUX.2 family, and Black Forest Labs positions it as the tier where nothing is traded away: the highest editing consistency across tasks, the strongest prompt following, and the most faithful reproduction of a named style. Everything the family does - multi-reference composition, 4MP editing, hex-exact colour, legible in-image type - it does here with the fewest compromises, which is the whole argument for a tier that costs several times what pro does.

The capability that only max has is grounding search. The lab says the model can pull real-time information from the web during generation, so a prompt can ask for something the model could not know from training: today's weather in a named city, a recent news event, a current score, a moment from a specific date and place. Black Forest Labs' own worked example is a miniature isometric city scene of Freiburg with the live weather folded into the lighting and an icon and temperature laid out over it. That covers a class of prompt no other tier in the family can serve, and it is the capability Black Forest Labs puts at the top of its case for max.

Identity is the other thing the lab leans on. It says max preserves facial features, proportions, expressions and visual identity across complex edits, several references at once, and a change of environment - the failure mode that makes character-consistent campaign work expensive to produce by hand. The same claim extends to materials: retexturing an object, a product or a room while the shape, geometry and lighting stay put, rather than regenerating the scene around a new surface.

What it gives up is cost and speed. Black Forest Labs' own line on pro is that it delivers state-of-the-art quality faster and cheaper with no compromise between the two; max is the tier you move to when a specific job needs the extra consistency or the web grounding badly enough to pay for it. For high-volume generation, drafts, or anything where the prompt carries all the information the model needs, the cheaper tiers are the lab's own recommendation.