Black Forest Labs publishes its own prompting rules for FLUX.2, and they are more concrete than most model vendors bother to be: a named four-part structure, a stated word-order rule, three length bands, five instructions for text that has to read, and an arithmetic for how many reference images fit in one request.
Every claim below stays attached to the sentence it came from, including where the vendor stops. For the setup rather than the prompt, the FLUX.2 getting-started guide covers what this page skips.
The structure the prompting guide names
Two official pages describe how FLUX.2 reads a prompt, and they use different skeletons. The prompting guide for [pro] and [max] names four parts; Building a Good Prompt offers an eight-slot template with camera settings, lighting and colours broken out. Both belong to Black Forest Labs, and the four-part version comes first because the rest of the guide assumes it.
| Prompt part | The guide’s own definition |
|---|---|
| Subject | The main focus (person, object, character) |
| Action | What the subject is doing or their pose |
| Style | Artistic approach, medium, or aesthetic |
| Context | Setting, lighting, time, mood, or atmospheric conditions |
The guide prints those four as a sum rather than a list.
Subject + Action + Style + Context
Subject: The main focus (person, object, character)
Action: What the subject is doing or their pose
Style: Artistic approach, medium, or aesthetic
Context: Setting, lighting, time, mood, or atmospheric conditionsWord order comes before wording
One sentence governs the rest of the guide: FLUX.2 pays more attention to what comes first. The published priority order runs main subject, key action, critical style, essential context, secondary details. Building a Good Prompt says the same from the other end — name the subject, add the main action, then add mood and direction only when they improve the image. Its verdict on padding is blunt: specific detail helps, filler hurts.
How long the prompt should run
Three bands, not one limit.
| Band | Words | The guide’s stated use |
|---|---|---|
| Short | 10–30 | Quick concepts and style exploration |
| Medium | 30–80 | Usually ideal for most projects |
| Long | 80+ | Complex scenes requiring detailed specifications |
Read the middle band as the default. The long band is for scenes where the specification is the point — several subjects, a fixed layout, brand colours — not for prompts that grew long because nobody cut them. The ceiling is 32K tokens, which is a limit, not an invitation.
Text rendering, as the guide spells it out
Typography is what the model page leads with, and the guide reduces it to five moves.
- Put the words in quotation marks. The guide’s own example quotes a neon sign reading “OPEN”.
- Say where the words sit relative to everything else in frame.
- Describe the letterforms: elegant serif typography, bold industrial lettering, handwritten script.
- Name the size, from large headline text down to small body copy.
- Give the colour, and give it as a hex code.
Quotation marks are the step people skip
Marking the string is what turns a rendering request into something the model can isolate. The same rule covers brand text, so a logo lockup arrives with its letterforms and colour named in one clause.
Colour control with hex codes
Colour is a first-class input here: signal a hex value with the word “color” or “hex” immediately before the code, and the model will try to match it exactly. Gradients take two codes, and a structured JSON prompt can carry a colours array per subject.
A hex code needs an object to belong to
The sentence worth remembering: hex codes work best when clearly associated with specific objects, and a vague reference like “use #FF0000 somewhere” may produce inconsistent results. Attach the code to a noun — the vase, the pillow, the sky — or the model will place it where it likes.
Multi-reference prompting
Combining elements from several images while holding identity intact is the capability the model page leads with. The documented jobs are fashion shoots, product composites, interior layouts, and character continuity across a series.
How many references fit in one request
The count depends on the tier and the surface, and the guide gives arithmetic rather than a flat number. The input-plus-output budget for [pro] is 9MP, so a 1MP output leaves room for 8 reference images and a 2MP output for 7. The ceiling is 8 through the API and 10 in the playground for [max], [pro] and [flex], 4 for [klein], and a recommended 6 for [dev].
Labelling is the technique, not the count. The guide asks you to describe how each input should be used — subject from image one, style from image two, background from image three — because an unlabelled second image leaves the model choosing for you. The FLUX.2 API guide shows where those inputs sit in a request.
Worked example: two published prompts and their published results
Both prompts below are Black Forest Labs’ own, published with the results beside them. The first is a product advertisement, and its point is text that survives.
Samsung Galaxy S25 Ultra product advertisement, 'Ultra-strong titanium' headline, 'Shielded in a strong titanium frame, your Galaxy S25 Ultra always stays protected' subtext, close-up of phone edge showing titanium frame, dark gradient background, clean minimalist tech aesthetic, professional product photographyThe published result is a dark-gradient product shot in which both strings survive: the headline reads at poster size and the longer subtext beneath it stays legible. Two clauses do that work. The strings are quoted, so they arrive as rendering instructions rather than description, and the phone is named first, which the word-order rule says the model weights most.
The second produced the composite in the figure above.
Replace the children in the image with the animals from images 2, 3, 4, 5, and 6. Adjust them to the space and style so they sit naturally in the scene. Adjust the proportions of the animals to each other and to the space.Read it as a labelling exercise: every input is told what it contributes, the placement sentence covers the space, and the proportion sentence stops one animal arriving at the wrong scale. Six inputs also sits inside the documented ceiling for [max], [pro] and [flex].
What the two prompts have in common
Neither is long by the guide’s own bands, both open with the operation the model has to perform, and neither contains a prohibition. That last part is no coincidence: the guide states plainly that FLUX.2 does not support negative prompts, so where you would write “no extra text”, describe the frame instead.
What the official pages have not published
The documentation is specific about mechanics and quiet about outcomes.
- No per-language accuracy figure for text rendering appears in the guide, so a claim about which languages render best cannot be sourced from the vendor.
- No benchmark score and no head-to-head comparison with another vendor’s model is published on the pages referenced here.
- No free API quota is documented; the free entry point is the playground trial, which is a different thing.
- [klein] has no prompt upsampling, and no equivalent statement covers the other tiers.
The FLUX.2 topic hub collects those pages, if you want the source material rather than the reading.
Frequently asked questions
Does FLUX.2 support negative prompts? No. The guide asks you to describe what should be in frame instead — “sharp focus throughout” rather than “no blur”, an “empty scene” rather than “no people”.
Which framework should I copy? Start with Subject, Action, Style and Context; move to the eight-slot template when a scene needs camera settings, lighting and colour separated out.
Where does a hex code go? Immediately after the word “color” or “hex”, attached to a named object. A code with nothing to belong to is the documented failure case.
How many reference images am I allowed? Up to 8 through the API and 10 in the playground for [max], [pro] and [flex], 4 for [klein], a recommended 6 for [dev]. For [pro], calculate against the 9MP budget shared between input and output.
Write the four parts in order, quote the strings you want rendered, hang every colour on an object, and label every reference. The topic hub holds the rest of the family’s documentation.