VidoGridPrompt structure: the core of this FLUX prompt guide
Describe the finished image instead of issuing commands. Black Forest Labs suggests an order that suits most scenes: image type and medium, then the subject and its action, the setting, the light, and finally framing and small details. Treat it as a habit and lead with whatever matters most; the rest of this FLUX prompt guide builds on that order.
FLUX 3 Image expands a short prompt into a fuller description before rendering, so one clear sentence is a fine start; add words only for what you want to pin down, such as positions, light or text. FLUX.2 [klein] does no expansion and receives exactly what you type, so give it complete prose.
Use sentences for relationships and short phrases for style and color. Skip praise words like "masterpiece"; every phrase should name something visible.
Wide documentary photograph of a salt flat at blue hour. A lone cyclist pushes a loaded touring bike along a gravel track toward the right edge of the frame, shallow puddles mirroring a violet sky. Cool ambient light, a warm glow from the bike's front lamp, 24mm lens.
Composition, light and exact color
For photographic looks, name the camera, lens or film stock, then say what it does to the picture: grain, a warm cast, shallow focus. If results keep pulling back too wide, put the subject and shot size first and the background later.
Describe light by source, direction and quality: "soft window light from the left", not "nice lighting". Keep effects to one or two. When a color must be exact, tie a hex code to one named object; FLUX.2 [pro] and [max] accept them in text-to-image prompts, and FLUX 3 Image uses them in edits.
Studio product photograph of a glass candle jar on pale travertine. The wax is #2F5D50 and the paper label is #F3E9DC with a thin copper border. Soft key light from the upper left, one sprig of dried eucalyptus leaning on the jar, plain warm-gray backdrop, 90mm macro lens.
Rendering text in an image
Of everything in this FLUX prompt guide, text needs the most literal wording. Both generations render it, and the FLUX 3 launch examples include multi-line posters and Japanese lettering. Quote the exact words with the capitalization you want, say where they sit in the frame, and describe the type: weight, case, color, serif or sans-serif.
Write \n inside a quoted string for a line break. For FLUX.2 [pro] and [max], the docs advise keeping strings short and describing them early; FLUX.2 [flex] is the variant built for typography and small details. Proofread every string before publishing.
Screen-printed concert poster in deep navy and tangerine ink on off-white paper. A stylized lighthouse fills the center, its beam sweeping across the upper third. Across the top, tall condensed sans-serif capitals read "NIGHT SHIFT". At the bottom center, smaller tangerine lettering reads "Harbor Hall\nMarch 14". Visible paper texture.
Reference images and edits
FLUX 3 Image takes one to ten reference images in one images field, with no mask or strength setting. Refer to inputs by position ("image 1" is the first entry) and say what each supplies: subject, setting, style or product.
A good edit names the target so only one thing matches, says what it becomes, and lists what must stay. "Change the jacket from red to blue" beats "make it blue". With aspect_ratio on auto, the output follows the first image, so put the frame you want first.
The same wording works on FLUX.2, which takes fewer references: up to eight on [max], [pro] and [flex] through the API, four on [klein], and a recommended maximum of six on [dev].
Place the woman from image 1 in the kitchen from image 2. She stands at the counter on the left, kneading bread dough, sleeves rolled up. Keep her face, hairstyle and green cardigan from image 1, and keep the tiled wall and copper pans from image 2 unchanged. Morning light falls across the counter.
Settings and layout controls that change results
Wording matters more than settings, but this FLUX prompt guide would be incomplete without them. On FLUX.2, [pro] accepts an optional seed, [flex] exposes steps and guidance, and a JSON object can be sent as the prompt string for automated pipelines.
FLUX 3 Image has no negative_prompt, seed or guidance fields, and unknown fields return an error. What you can set there:
- aspect_ratio: 15 fixed ratios from 21:9 to 9:21, or auto. Without references, auto falls back to a square unless the prompt asks for a ratio, so set wide or tall ratios yourself.
- resolution: 768sq for fast drafts, 1k by default, up to 4k, which can take several minutes. Exact width and height also work.
- grounding: on by default, using web and image search to get real things right; false gives faster, prompt-only results.
- Layout boxes: append elements with an id, a bbox as [top, left, bottom, right] on a 0 to 1000 grid, and a desc.
Common mistakes and fixes
This part of the FLUX prompt guide covers problems the official docs point out, with a fix for each.
- Writing what you do not want. Describe what fills the space instead: "a deserted platform at dawn", not "a station with no passengers".
- Quality tags instead of details. Swap "highly detailed" for things a camera records: bark texture, slight grain.
- A frame that fights the prompt. An ultra-wide ratio needs content across its width.
- Judging one render. FLUX 3 Image has no seed, so compare a few results and change one detail at a time.
- Expecting a draft to scale up unchanged. A new resolution is a new generation, so composition can shift.
- Vague edit targets. If several objects match, add position, size or color, or use layout boxes.