Models

Prompting FLUX Models

Practical, model-aware guidance for structuring FLUX prompts, controlling composition and text, avoiding conflicts, and iterating systematically.

Updated Sep 2026 image-generationfluxpromptingworkflow

Use this guide with an AI agent

AI-readable: use the clean Markdown version without navigation or ads.

Markdown
Agent instructions
On this page

Prompting FLUX Models#

FLUX-family behavior varies by checkpoint, host, guidance settings, adapters, and software version. Treat the following as a durable starting method rather than a substitute for the documentation of a specific endpoint.

Use a descriptive hierarchy#

FLUX generally handles coherent natural language well. Put the most important visual facts early and arrange details by role:

  1. Image type and primary subject — photograph, illustration, diagram; identity, count, age, species, or object class.
  2. Action and pose — what is happening, body direction, gaze, and interaction.
  3. Setting — location, era, season, weather, and meaningful background elements.
  4. Composition — shot size, angle, placement, aspect-ratio-aware negative space.
  5. Optics and lighting — lens character, focus, source, direction, softness, and color.
  6. Materials and color — surfaces, palette, wardrobe, and small identifying details.
  7. Finish — medium, rendering approach, print process, or restrained quality targets.
Editorial food photograph of one handmade ceramic bowl filled with tomato soup,
placed slightly left of center on a worn oak table. A folded cream napkin and silver spoon
sit behind it. Eye-level three-quarter view, 70mm perspective, shallow but sufficient depth
of field, soft north-window light from the left, muted brick red and sage palette, realistic
steam and natural crumbs, quiet rustic magazine styling.

Prefer observable facts#

Concrete visual language gives the model something depictable.

Vague wording More controllable wording
beautiful lighting large soft window key from camera-left with gentle shadow detail
dynamic pose runner leaning into a turn, left knee raised, arms counterbalancing
cinematic 2.39:1 framing, low-key side light, restrained teal shadows, foreground occlusion
futuristic city dense elevated walkways, ceramic-white towers, autonomous trams, cyan wayfinding lights
highly detailed visible woven fibers, tiny seam stitching, brushed metal fasteners, fine skin texture
professional logo flat two-color geometric mark, uniform stroke weight, strong silhouette, no gradients

Control entities explicitly#

When count or relationships matter, make them grammatical rather than comma-separated:

Exactly three objects are on the shelf: one red radio on the left, one small green plant in
the center, and one blue book standing upright on the right. Nothing else is on the shelf.

Useful relationship language includes:

  • behind, in front of, partially occluding, inside, attached to, and reflected in;
  • camera-left and camera-right for image placement;
  • the subject's left hand for anatomy;
  • foreground, middle ground, and background for depth;
  • one, a pair of, or exactly three when quantity is important.

Avoid overloading a scene with many equally emphasized subjects. Generate the structural composition first, then add secondary details during later iterations or editing.

Composition language#

State shot scale, viewpoint, and placement independently:

full-body portrait, eye-level camera, front three-quarter view, subject on the right third,
large clean negative space on the left for a headline

Useful controls include extreme close-up, close-up, medium shot, medium full shot, full-body, wide establishing shot, top-down, low angle, over-the-shoulder, symmetrical, centered, rule of thirds, edge-to-edge, and isolated on a plain background.

If an output aspect ratio is selected outside the prompt, compose for it explicitly: describe vertical stacking for tall frames, lateral relationships for wide frames, and safe empty zones for text or cropping.

Camera and lighting#

Use photographic terms to communicate a visible consequence, not merely equipment prestige.

  • A wide lens suggests broad context, spatial depth, and possible edge stretching.
  • A normal lens suggests natural proportions.
  • A telephoto lens suggests narrow field of view and compressed background layers.
  • A wide aperture suggests shallow focus but may hide details that need to remain readable.
  • A small aperture / deep focus keeps layered scenes legible.
  • A named lighting source plus direction is stronger than dramatic lighting.
35mm environmental portrait, camera at chest height, moderate depth of field,
late-afternoon sun as a warm rear-side key, cool open-sky fill, level horizon

Do not combine mutually exclusive cues such as flat shadowless light and hard chiaroscuro, or extreme shallow focus and every distant object tack sharp, unless the contradiction is deliberately surreal.

Materials and physical behavior#

Name a material, finish, condition, and relevant interaction with light:

  • brushed stainless steel with fine directional grain and soft elongated reflections;
  • unglazed terracotta, porous matte surface, chipped rim, pale mineral deposits;
  • clear ribbed glass containing condensation droplets and refracted window highlights;
  • heavy ivory cotton paper, deckled edge, blind-debossed lettering.

Physical specificity often improves realism more reliably than generic quality boosters. Mention gravity, contact shadows, weight, folds, reflections, weathering, and scale clues where relevant.

Styles and references#

Describe observable style components—medium, mark-making, shape language, palette, period, and production method—instead of relying only on a creator's name.

hand-painted gouache editorial illustration, simplified angular shapes, opaque layered color,
dry-brush edge texture, limited navy, rust, and cream palette, generous unprinted paper

For design work, specify whether the result should look like a photograph, finished graphic, print artifact, interface mockup, packaging dieline, or presentation board. Mixing too many production modes (oil painting, 3D render, vector icon, analog film) usually weakens consistency.

Text in images#

Treat text as a separate, exact requirement:

A simple cream storefront sign containing exactly the words “NORTH STAR” in large uppercase
dark-blue sans-serif letters. The spelling must be N-O-R-T-H, space, S-T-A-R. No other text,
numbers, logos, labels, or watermarks appear anywhere.

For better reliability:

  1. Keep copy short.
  2. Quote the exact wording and specify capitalization.
  3. Describe location, hierarchy, font class, color, and contrast.
  4. State whether any other text is allowed.
  5. Use a straight-on or minimally distorted surface.
  6. Verify spelling visually; regenerate or edit if accuracy is essential.
  7. For long body copy or regulated content, generate the visual without text and typeset it in a layout tool.

Negative instructions and exclusions#

Some FLUX workflows support a dedicated negative-prompt field while others do not or give it limited influence. When no reliable negative channel exists, write positive structural constraints first: one centered bottle on a clean empty surface is often stronger than a long list of unwanted objects.

Keep exclusions short and observable:

No people, hands, labels, captions, border, watermark, or duplicate objects.

Do not use indiscriminate boilerplate negatives. Terms such as bad anatomy or ugly are ambiguous; identify the actual failure (extra fingers, merged utensils, unreadable lettering) only if the current system responds to exclusions.

Prompt weighting and syntax#

Parentheses, numeric weights, token breaks, and emphasis syntax are interface-dependent. A syntax recognized by one UI may be ignored or interpreted differently by another. Before using (term:1.4) or similar notation:

  • confirm that the specific pipeline documents it;
  • prefer sentence order and explicit language first;
  • change one weight at a time;
  • avoid extreme weights, which can produce artifacts or suppress context.

Iteration workflow#

  1. Write the invariant brief: subject, action, setting, and required composition.
  2. Generate a baseline: omit decorative micro-detail.
  3. Diagnose one failure: count, pose, spatial relation, lighting, text, or style.
  4. Rewrite the relevant clause: do not append endless corrective keywords.
  5. Lock successful variables: keep seed and generation settings when the platform allows it.
  6. Add secondary detail: material wear, props, atmosphere, or finishing texture.
  7. Edit locally: use inpainting or image editing for isolated defects rather than rerolling a nearly correct image.
  8. Validate at full resolution: inspect hands, edges, reflections, repeated patterns, legibility, and accidental text.

Reusable templates#

Photograph#

[genre] photograph of [subject with count and defining traits], [action], in [specific setting].
[shot size], [angle], [placement], [lens/focus behavior]. Lit by [dominant source and direction]
with [fill/accent]. [palette], [materials], [atmosphere], [realism cues].

Illustration#

[medium] illustration of [subject and action] in [setting], composed as [layout].
[shape language], [line quality], [paint/print texture], limited palette of [colors],
[background treatment], [intended editorial or decorative use].

Product image#

One [product] made of [materials and finish], shown in [orientation] on [surface/background].
[viewpoint and lens], [lighting setup], [reflection and shadow behavior], [required label text].
Clean commercial styling; no unrelated props, duplicate products, or unintended text.

Poster with copy#

Vertical [ratio] poster for [topic]. [Main visual] occupies [area]; leave [negative-space area].
Headline reads exactly “[TEXT]” in [font class and treatment]. [Secondary copy requirements].
[grid, palette, medium, print texture]. No other text or logos.

The strongest prompt is not necessarily the longest. It is the shortest description that preserves the intended subject, relationships, composition, physical cues, and visual treatment without contradiction.