What Is FLUX.1? The model behind a lot of current image quality

FLUX is not one model. The family splits into a fast distilled variant built to produce a usable image in very few steps, a middle variant that is openly downloadable but carries licence conditions that restrict commercial use, and a top variant available through an API rather than as a file.

The important consequence is legal rather than technical. Picking the openly downloadable variant because it is convenient, then using its output commercially, is the mistake that gets made most often, and the licence rather than the model is what decides whether that is permitted.

Check which variant a hosted service is actually running before assuming what you may do with the result. Services do not always say, and the answer changes what you own.

The practical experience people report is that it argues less. Ask for three specific objects arranged in a particular way and you are more likely to get three objects arranged that way, rather than a beautiful image with two of them and a different arrangement.

That trades against something real. A model that adheres tightly to instructions is less inclined to improve on a mediocre prompt, so a vague request returns a vague picture rather than a striking one. Models with a stronger aesthetic opinion flatter lazy prompts and fight specific ones; this is the other end of that trade.

The working implication is that detail is rewarded here in a way it is not everywhere. Naming the light, the focal length and the medium changes the output measurably, which is why the prompt guidance on this site is written in those terms.

Legible lettering was for years the reliable way to spot a generated image. Signage came out as plausible-looking gibberish because a diffusion model was assembling letter shapes from visual statistics rather than spelling anything.

FLUX is substantially better at this, and short strings — a word on a sign, a label on a package, a title on a poster — come out readable often enough to design around. It is not a typesetting engine: long passages still degrade, and the longer the string the worse the odds.

The practical technique is to ask for less text than you want and add the rest properly afterwards. A generated background with one legible word on it, finished in a real design tool, beats a generated image with a paragraph of near-miss lettering baked into the pixels.

Against Stable Diffusion, it is the newer work of overlapping people: generally stronger output and better instruction-following, against a far smaller inventory of community checkpoints and style adapters.

Against Midjourney, it is the less opinionated of the two. Midjourney will make something with a point of view from a short prompt; FLUX will make what you asked for, which is better when you know exactly what you want and worse when you are hoping to be surprised.

Against the closed assistant-attached models, it is the one you can actually run or host yourself, which matters when throughput, privacy or reproducibility are constraints rather than preferences.

The photographs on this site are generated with the FLUX family — the gallery states the model on every card alongside the prompt and the seed, so the claim is checkable rather than asserted. We have no commercial relationship with Black Forest Labs; it is a tool we pay for like anyone else.

The reason we say which model made an image is that a site arguing for prompt craft, and illustrating that argument with images whose provenance it will not disclose, has undermined its own case. If the prompts on the gallery page are real, they can be run, and a model name is part of being able to run them.