Write for Flux the way you’d brief a photographer
Flux models use a large language-based text encoder, so they understand full sentences, spatial relationships and fairly long descriptions. Where tag-based models get confused by grammar, Flux benefits from it: “a ceramic mug on the left of a linen table, steam rising towards a bright window” gives it a scene it can lay out correctly.
Our Flux format produces three to five sentences that cover the foreground, midground and background, the materials and textures, the direction and quality of the light, and the camera or lens. That structure gives the model enough to work with without drifting into a wall of adjectives.
Getting text in the image right
One of Flux’s strengths is rendering legible text. When your reference contains a sign, a label or a headline that matters, the generator quotes it exactly, for example a sign that reads "OPEN". Keep the quotes and the exact spelling when you edit; the model uses them as a cue to draw those letters.
Short phrases render more reliably than long sentences. If a long headline comes out garbled, shorten it, or describe where it should appear (“in bold white letters across the top”).
No weights, no negative prompts
Syntax such as (tag:1.3) or a separate negative prompt comes from Stable Diffusion interfaces. Most Flux setups ignore weight brackets and don’t use negative prompts in the usual way. To emphasise something, describe it earlier and in more detail. To exclude something, say what the scene contains instead, for example “an empty street” rather than “no people”.
If you switch between models, use the matching format: the same image produces a tag list for Stable Diffusion and flowing prose for Flux, because each model responds best to its own style.
Refining your Flux prompt
Change one sentence at a time and keep the rest stable. Adjusting only the lighting sentence, for example, shows you exactly what that change does. Your ten most recent prompts are saved in this browser, so earlier versions are always one tap away.
The structure of a Flux result
A typical result opens with the medium and the subject (“A photograph of a lone lighthouse on a basalt cliff”), follows with the surroundings and spatial layout, then describes the light (“low sun from the left, long shadows across wet rocks”), the camera (“shot from a low angle with a wide lens”) and finally the palette and mood. Reading it back should feel like a short, precise scene description in a novel.
Guides from our blog
- Why your AI images look generic, and seven prompt habits that fix it
AI image prompts that say how a picture should look, not just what is in it: seven habits, with before-and-after examples, for sharper results.
- How to reverse-engineer any image into a prompt, and when it works
A practical reverse prompting workflow: describe an image, format it for your generator, and refine one element at a time. What works, and what never will.