FLUX Text to Image: Official Prompt Guide

Most FLUX prompt advice on the internet is Stable Diffusion advice with the model name swapped. You can spot it instantly: it recommends negative prompts, quality-booster token stacks, and weight syntax in brackets.

FLUX.2 does not support negative prompts. That is not an opinion — it is the first line of Black Forest Labs' official prompting guide. If your prompt template has a negative field, it is doing nothing, and the habits around it are actively costing you control.

This guide follows BFL's documented rules and shows what to do instead.

Sourcing note: every rule, structure, and technique below is taken from Black Forest Labs' official FLUX.2 prompting documentation and model pages. Example prompt patterns follow the documented syntax. Where behaviour depends on the model tier, that is noted. Last verified July 28, 2026.

What this guide solves

The pain point: prompt habits transfer badly between model families, and FLUX punishes the Stable Diffusion ones specifically. People arrive with a working SD workflow, get mediocre FLUX output, and conclude the model is overrated.

The differentiator: this is written from the vendor's own documentation rather than community folklore — including the parts most guides omit entirely: hex colour prompting, JSON-structured prompts for infographics, and prompting in the target language for cultural accuracy.

Rule 1 — The structure: Subject + Action + Style + Context

BFL's documented framework, in order:

• Subject — the main focus: person, object, character
• Action — what the subject is doing, or their pose
• Style — artistic approach, medium, aesthetic
• Context — setting, lighting, time, mood, atmospheric conditions

An example straight from the official guide's pattern:

Black cat hiding behind a watermelon slice, professional studio shot, bright red and turquoise background with summer mystery vibe

Subject (black cat), action (hiding behind a watermelon slice), style (professional studio shot), context (colour scheme and mood). No boosters, no negatives, no weights.

Rule 2 — Word order is weight

The documentation is explicit: FLUX pays more attention to what comes first. The documented priority order:

Main subject → key action → critical style → essential context → secondary details

This replaces bracket-weight syntax entirely. If something is not coming through, move it earlier rather than adding emphasis markers.

Rule 3 — Length has a documented sweet spot

BFL's guidance:

Past ~80 words you are usually adding noise. Long prompts are for compositional complexity, not for insurance.

Rule 4 — No negative prompts. Describe what you want.

There is no negative field to fill. Rewrite exclusions as positive descriptions:

The general fix: the thing you were trying to exclude usually has a positive counterpart. Name it.

Rule 5 — Be specific about optics

The documentation states plainly that "Shot on Fujifilm X-T5, 35mm f/1.4" produces more authentic results than "professional photo." Documented style descriptors:

Camera-and-lens patterns from the same source: "Shot on Hasselblad X2D, 80mm lens, f/2.8, natural lighting" or "Canon 5D Mark IV, 24-70mm at 35mm, golden hour, shallow depth of field."

Test these as you read. Flux 3 AI is a browser workspace where you can run these structures immediately — no install, no API key. Open the image generator, try one prompt with your old negative-prompt habits and one following the rules above, and compare.

Rule 6 — Text rendering has its own syntax

Text is where the FLUX family is unusually strong, and where sloppy prompting fails visibly. The documented technique:

1. Use quotation marks around the exact string — "The text 'OPEN' appears in red neon letters above the door"
2. Specify placement relative to other elements
3. Describe the style — "elegant serif typography," "bold industrial lettering," "handwritten script"
4. Set the size — "large headline text," "small body copy," "medium subheading"
5. Set the colour, including hex codes for brand work

The official guide's product-advertisement example carries a headline string, a subtext string, a described product close-up, background treatment, and an aesthetic — all in one medium-length prompt. That is the shape to copy for any marketing asset.

Rule 7 — Hex colours work, if you attach them to objects

FLUX.2 supports precise colour matching via hex codes. Signal them with a keyword like "color" or "hex" followed by the code:

a vintage illustration of an apple in color #0047AB with a heart-shaped cutout in the middle, on a white background

Gradients work by naming both ends:

the color of the vase is a gradient, starting with color #02eb3c and finishing with color #edfa3c

The documented constraint matters: hex codes work best when clearly associated with a specific object. Vague references — "use #FF0000 somewhere" — produce inconsistent results. For brand work, bind every code to a named element.

Rule 8 — JSON prompts for structured design work

For infographics and layout-driven design, the documentation shows structured prompts. The pattern:

• a type and title
• a subtitle for supporting context
• sections, each with a heading, content, and a visual
• a color_scheme array of hex values
• a style descriptor such as "modern, clean, corporate"

The same structure works for product flat-lays: a scene, a subjects array where each subject has a description and a colours array, plus style and lighting fields.

Use this when the output has structure — infographics, data visualisation, layouts with defined regions. For a single photographic subject, plain natural language is better.

Rule 9 — Prompt in the language of the content

Documented behaviour: FLUX.2 understands multiple languages, and prompting in the native language of the content you are creating often produces more culturally authentic results — local markets, architecture, and atmosphere rendered with greater accuracy.

Writing a French village scene? Write the prompt in French. This is a documented model behaviour, not a superstition.

Rule 10 — Character consistency comes from references, not adjectives

For sequential work — comic panels, campaign series, storyboards — the documented approach is to define your character in detail and maintain that description across panels, generating each panel separately.

At the model level, FLUX.2 supports up to 10 reference images simultaneously, which BFL positions as the best character, product and style consistency currently available. If identity has to hold across a set, references do the work that prompt adjectives cannot.

A working template

[Subject], [action/pose], [style/medium], [setting],
[lighting], [camera + lens if photographic],
[text in quotation marks with placement + size if needed],
[hex colours bound to named objects if brand work]

Keep it in the 30–80 word band, put the subject first, and never write a negative.

FAQ

Do negative prompts work in FLUX? No. FLUX.2 does not support them. Describe what you want instead.

How long should a FLUX prompt be? 30–80 words for most projects, per BFL's documentation.

Can FLUX render readable text? Yes, and it is a documented strength. Put the string in quotation marks and specify placement, style, size and colour.

Can I control exact brand colours? Yes, via hex codes bound to specific objects. Loose colour references produce inconsistent results.

Should I prompt in English? Not necessarily — the docs recommend prompting in the language of the content for culturally accurate results.

Do these rules apply to FLUX 3? FLUX 3-specific prompting documentation has not been published. These are the current documented FLUX family rules, and structural behaviour of this kind generally carries across a family. Expect updates when BFL publishes FLUX 3 guidance.

Do weight brackets like (word:1.3) work? That syntax belongs to other model families. In FLUX, use word order for emphasis.

Bottom line

FLUX prompting is less about incantations than most model families and more about clear, ordered description: subject first, medium length, positive phrasing, concrete optics, quoted strings for text, hex bound to objects.

Unlearning the negative prompt is the single highest-value change if you are coming from Stable Diffusion. Everything else is refinement.

Put it into practice in the Flux 3 AI image generator, and upscale the keepers when you have a frame worth delivering.

Sources

1. Official FLUX.2 prompting guide (BFL documentation) — prompt structure, no-negatives rule, word order, length guidance, style descriptors, camera simulation, text rendering, hex and JSON prompting, multilingual behaviour, sequential art
2. BFL documentation overview — model family and recommended usage
3. FLUX.2 pro / flex model page — multi-reference control, up to 10 reference images
4. FLUX.2 max model page — text rendering, character consistency, spatial reasoning
5. FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — text rendering and prompt-following improvements, 4MP editing
6. BFL API pricing — reference-image billing rules relevant to multi-reference prompting
7. FLUX.2 klein model page — model variants for fast iteration
8. black-forest-labs/flux on GitHub — reference inference code
9. FLUX.1-schnell model card — prompt-following characteristics of the open models
10. FLUX 3 — Real World Models (BFL, July 23, 2026) — status of FLUX 3 and its improved complex-prompt and text handling

Scope note: documented behaviour applies to the FLUX.2 family; FLUX.1 differs in some respects and FLUX 3 guidance has not been published. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.