FLUX.2 Prompting Cheat Sheet

Black Forest Labs' own formula, the 30–80 word rule, and why a negative prompt does nothing here.

TL;DR — THE SHORT VERSION

Formula: Subject + Action + Style + Context.
Word order carries weight. BFL state the model "pays more attention to what comes first". Put the subject in the first clause.
30–80 words is the documented sweet spot. 10–30 for quick concepts, 80+ only for genuinely complex scenes.
No negative prompt. BFL: "FLUX.2 does not support negative prompts." Not weak — absent.
Hex colours must attach to an object, or they will not hold.

VENDOR-DOCUMENTED

This sheet quotes the vendor's own published prompting guidance.

The formula

Subject + Action + Style + Context

Template:

[main subject, described specifically] [what it is doing] [style / aesthetic] [setting, lighting, time, mood]

Worked example:

A ceramicist in her fifties centring clay on a wheel, documentary photography, north-facing studio window at dusk, flour-dusted apron, shallow depth of field

Breakdown — what each part does

Subject

The main focus. Front-load it — this is the position the model weights most heavily. Vague subject, generic image.

Action

What the subject is doing, or its pose. A static noun gives the model nothing to compose around.

Style

The aesthetic: documentary, editorial, cinematic, painterly. One direction, not three.

Context

Setting, lighting, time of day, mood. The supporting detail that makes it specific rather than stock.

Does it take a negative prompt?

No negative prompt field of any kind.NO

Every exclusion has to be rewritten as something present. Not "no plastic skin" — write "visible pores, subsurface scattering, uneven natural skin tone". You are not blocking the default, you are crowding it out with something more specific.

What not to do

Do not bury the subject at the end of a long sentence — you are working against the attention weighting.✕

Do not append a negative prompt. It is not read.✕

Do not write a stray hex code with no object attached; "use #FF0000 somewhere" produces inconsistent results.✕

Do not pad to 150 words believing longer is stronger. 30–80 is their stated range for most work.✕

JSON PROMPTING

BFL also support a structured JSON prompt schema for production and automation work, with fields for scene, subjects, style, palette, lighting, mood, background, composition and camera.

DISTILLED WEIGHTS

The open dev and schnell weights are guidance-distilled and run at CFG 1 — the mechanism that acts on a negative prompt is not present. This is the mechanical reason behind the rule above.

The checklist

Before you send it:

1 — Subject in the first clause?□

2 — Between 30 and 80 words?□

3 — One style direction, not several competing?□

4 — Every exclusion rewritten as something present?□

5 — Any hex code attached to a specific object?□

SOURCE

Black Forest Labs, FLUX.2 [pro] & [max] prompting guide, docs.bfl.ai. Checked 25 Aug 2026.

Model versions and vendor documentation both move. Re-read the source before relying on a specific number. Errors are logged at corrections.

Every other sheet in this set: all model cheat sheets.