Flux2 Klein Prompting

FLUX.2 [klein] Prompting

FLUX.2 [klein] is Black Forest Labs' open-weight, size-distilled member of the FLUX.2 family. It is a rectified-flow diffusion transformer steered by a large language-model text encoder (Qwen3), which is what gives it strong prompt adherence: it reads long, structured, natural-language descriptions and follows compositional and spatial instructions far better than a tag-based model. It also renders legible text in the image and supports multi-reference image editing. Write the prompt and negative you pass to the generate_image tool per the rules below to drive FLUX.2 well on the first call.

The single most important rule: FLUX.2 wants natural language, not tags. Write the prompt as you would describe a photograph or illustration to a person - complete clauses, concrete nouns, explicit spatial relationships. Comma-spammed booru tags (1girl, solo, long hair, masterpiece) underuse the model and produce flat, generic output. There is no quality-tag prefix, no score tags, and no @artist syntax; describe the style in words instead (shot on 35mm film, a watercolour illustration, in the style of a 1950s travel poster).

Safe default when unsure: use the flux2-t2i-fast workflow, write prompt as 2-4 descriptive sentences (subject, setting, lighting, style), and pass nothing else - leave width / height / steps / cfg / seed unset and do not pass negative (it does nothing on the fast graph). Only depart from this default when the request needs editing (flux2-edit*), a specific size, or negative-prompt control (flux2-t2i). The rest of this guide is how to do better than the default when the request calls for it.

The generate_image call

You set the entire prompt string; it replaces the workflow's baked text. Omitted args keep the workflow's baked-in values, so you do not need to pass steps / cfg - the graph already ships the right pair for its variant.

Which workflow

Reach for a *-fast variant by default - it is several times quicker per render. Step up to the base (flux2-t2i / flux2-edit) only when you actually need the negative prompt or higher-CFG adherence, because that costs ~5x the steps.

CFG and the negative prompt

FLUX.2's negative prompt is classifier-free guidance, so it only influences the image when cfg > 1:

• On the base graphs (flux2-t2i, flux2-edit, flux2-edit-multi, CFG ~5) the negative arg works normally - it steers the image away from what you list.
• On the distilled *-fast graphs (CFG 1) classifier-free guidance is off, so the negative arg is effectively ignored. Do not lean on it there; bake exclusions into the positive prompt instead ("a clean, empty desk" rather than relying on negative: clutter).

When you genuinely need to suppress something, either switch to a base graph and use negative, or phrase the absence positively in the prompt.

Positive prompt recipe

Write a coherent description, not a tag list. A strong FLUX.2 prompt usually moves through these beats (in roughly this order, as prose - not as labelled fields):

1. Subject + action - who/what and what they are doing. A weathered lighthouse keeper climbs a spiral staircase, lantern in hand.
2. Appearance detail - concrete, depictable attributes: clothing, materials, age, expression, colours.
3. Setting + composition - where it is, and where things sit in the frame. FLUX.2 honours explicit spatial language: in the foreground, centred, to the left, seen from below, a wide establishing shot.
4. Lighting + atmosphere - golden-hour backlight, soft overcast light, volumetric fog, harsh noon sun.
5. Style + medium - 35mm film photograph, oil painting, cinematic still, flat vector illustration, architectural render. Name an era, film stock, or art movement rather than an artist tag.
6. Any text to render - put the literal characters in quotes (see Rendering text).

Be specific and generous with detail - FLUX.2's LLM encoder rewards a paragraph far more than a plain SDXL model does. Two to five sentences is a healthy range for a rich scene. You do not need filler quality words (masterpiece, best quality, 8k); spend those tokens on concrete description instead.

Expanding a terse user request

When the user asks for something short like "draw a cat in a hat" or "make me a cyberpunk city", do not pass that string through. Expand it into prose before calling generate_image:

1. Anchor the subject in one clause, then add 3-5 concrete, depictable details (materials, colour, age, expression).
2. Place it - name the setting and the composition/camera (a low-angle wide shot, centred portrait, overhead flat-lay). Pick one clear framing.
3. Light it - one phrase on lighting and mood.
4. Style it - one phrase on medium / era / look. Match the user's vibe; ask only if a load-bearing detail is genuinely ambiguous (e.g. "a dragon - Western or Eastern?").
5. Assemble as prose, not commas. Aim for 2-4 sentences.

Sketch: a cyberpunk city becomes

A rain-slicked cyberpunk street at night, crowded with neon signage in pink and cyan reflecting off the wet asphalt.
Steam rises from a noodle stall in the foreground where a lone figure in a translucent raincoat waits. Shot as a
cinematic wide-angle still, shallow depth of field, moody volumetric haze.

If the user wants to iterate, pass the prior seed back and tweak the wording rather than restarting from scratch.

When the user supplies their own prompt

If the user hands you a finished descriptive prompt, preserve their wording and intent - do not paraphrase it into tags or strip detail. You may append a short clarifying clause for composition or lighting if something load-bearing is missing, but keep their voice. If they hand you a tag list and want FLUX.2 output, translate it into prose (the model reads prose far better than commas) while keeping every concept they listed.

Rendering text in the image

FLUX.2 can render short, legible text. Put the exact string in quotes and say where it goes:

A vintage enamel shop sign reading "FRESH COFFEE" in bold cream lettering on a dark green background, mounted above a
café door.

Keep rendered text short (a word, a phrase, a short sign) - long paragraphs degrade. State the wording, the style of the lettering, and its placement.

Editing and reference images

The edit workflows take their output size and base content from the reference image, so they expose neither width/height nor denoise - you control the result entirely through the prompt and the reference(s).

• flux2-edit / flux2-edit-fast - one reference via inputImages: ["~/in.png"]. The prompt is an edit instruction: describe the change, not the whole scene. Change the car's colour to matte black and add rain on the windshield. Keep everything you do not mention implicitly unchanged.
• flux2-edit-multi - two references via inputImages: ["ref1.png", "ref2.png"]. Use it to combine elements (Place the character from the first image into the room from the second image, matching its lighting.). Output size follows the first reference.

You can also reuse a prior render as the edit input with the tool's refine arg (point it at a generation id and set workflow to an edit variant) instead of passing inputImages.

Generation settings

Steps/CFG live baked in each workflow graph, so the usual move is to pick the right workflow (fast vs base) rather than to override steps/cfg per call.

Prompt enhancement (enhance)

The generate_image tool has an opt-in enhance option that routes your prompt through a separate model before the render to rewrite it into the target workflow's native protocol. For FLUX.2 that means turning a terse phrase or a loose tag list into the kind of rich, descriptive paragraph this guide describes, folding in composition / lighting / style detail the bare prompt lacked.

• It is opt-in. Pass enhance: true on the call (or a workflow/config default turns it on). When off, your prompt is sent verbatim.
• Use it when the incoming prompt is thin (a few words, or a comma list) and you want a one-shot upgrade to full prose without hand-expanding it yourself. When you have already written a careful descriptive prompt, leave enhance off - it only adds latency and risks drifting from your intent.
• Scene continuity. Pass the context arg alongside enhance to hand the enhancer background to honour (character facts, ongoing scene, wardrobe) without depicting it literally. It is ignored when enhance is off.
• Negative. The enhancer builds on whatever baseline negative you pass and returns a refined one - but remember the CFG caveat: on a *-fast (CFG 1) graph the resulting negative still has no effect.

If you are enhancing a prompt rather than rendering

You may be reading this not to call generate_image yourself, but as guidance handed to the prompt-enhancement step: you were given a rough positive prompt, a baseline negative, and the target workflow's protocol, and you must return a single JSON object {"prompt", "negative"} and nothing else. In that role:

• Act only on the prompt-writing rules above - prose over tags (positive recipe), the terse-request expansion, text rendering, and the CFG-vs-negative caveat. The workflow matrix, tool-arg tables, and worked examples do not apply - you do not pick a workflow or call any tool.
• Translate and enrich, do not reinvent. Keep the incoming subject and intent; turn a terse phrase or a comma-tag list into a rich descriptive paragraph in FLUX.2's natural-language style.
• Always add detail - even to an already-rich prompt. Do not just echo back a positive that arrives as a full paragraph. FLUX.2 rewards more concrete, depictable specifics, so layer in what the prompt leaves unstated: lighting and atmosphere, materials and textures, palette, setting depth, camera framing and lens, expression and mood, medium or era. The enhanced positive should always be richer than what you were handed, as long as the additions stay faithful to the stated subject and do not contradict it.
• Mine any background context you were given (a scene/continuity note, recent conversation) for that extra detail, and fold it in as FLUX.2-native prose. Treat it as source material to pick depictable specifics from, not a checklist to dump and not a subject that overrides the explicit prompt; ignore chatter with no visual bearing.
• Build on the baseline negative rather than discarding it. You cannot know whether a base or *-fast graph will render, so return a sensible negative regardless - but do not over-invest in it, since at CFG 1 it has no effect.

Worked examples

These illustrate the shape of a finished prompt - adapt them to the actual request, do not reuse them verbatim. Pass the prompt (and negative) value as a single string.

Text-to-image (workflow: flux2-t2i-fast):

prompt: A close-up portrait of an elderly fisherman with a weathered, sun-creased face and a thick white beard, wearing a yellow oilskin coat beaded with rain. He gazes off-frame with a quiet, resolute expression. Soft overcast daylight, shallow depth of field, the blurred grey of a harbour behind him. Shot on 85mm film, muted colour palette.

Text-to-image with negative control (workflow: flux2-t2i, CFG ~5):

prompt:   A serene Japanese rock garden at dawn, raked gravel in concentric rings around three mossy stones, a single maple in autumn red at the edge. Soft directional morning light casting long shadows. Minimalist, tranquil, high detail.
negative: people, footprints, clutter, harsh shadows, oversaturated colours

Single-image edit (workflow: flux2-edit-fast, inputImages: ["~/room.png"]) - the prompt is the edit instruction:

prompt: Replace the daytime view through the window with a starry night sky and a full moon, and dim the room's lighting to a warm lamp glow. Keep the furniture and layout unchanged.

Anti-patterns

• Booru / tag spam. 1girl, solo, long hair, detailed, masterpiece, 8k underuses FLUX.2. Write a sentence: A young woman with long hair, rendered in fine detail. Translate any handed-in tag list into prose.
• Quality-word filler. masterpiece, best quality, ultra-detailed, 8k, award-winning does little here - those tokens are better spent on concrete subject, lighting, and style description.
• Leaning on negative at CFG 1. On a *-fast graph the negative prompt is ignored. Either switch to a base graph or phrase the exclusion positively in the prompt.
• Artist tags. There is no @artist / artist:name syntax. Name a medium, era, or movement in words instead (in the style of a 1960s screen-printed poster).
• Over-long rendered text. FLUX.2 renders a word or short phrase well, not a paragraph - do not promise a wall of legible text on a sign.
• Describing the whole scene in an edit prompt. For flux2-edit*, give the change as an instruction; re-describing the entire image can cause the model to regenerate parts you wanted kept.
• Cranking steps/CFG instead of switching graphs. The steps/CFG split is baked per workflow; pick fast vs base rather than hand-tuning a *-fast graph up to base-graph values.