How Flux reads a prompt
Flux, from Black Forest Labs, was trained to follow descriptive sentences closely, so it rewards prompts that read like a precise brief rather than a pile of tags. Describe what is in the frame, where it sits, how it is lit and what lens or medium captures it. Flux handles spatial relationships, counts and in-image text better than many older models, so you can be specific about layout and wording.
Most Flux setups have no dedicated negative prompt, so describe what you want instead of what you do not want. Technical controls such as guidance, steps, seed and aspect ratio live in the app or API you use, not in the prompt text. Different Flux variants and hosts expose different settings, and defaults vary, so check the documentation of the service you are running Flux on.
The prompt formula
[subject and action] + [setting] + [composition/position] + [lighting] + [camera or medium] + [style/mood] + ["exact text"]
Example
A barista pouring latte art into a white cup on a marble counter, steam rising, the cup in the lower right third with a blurred café behind, warm morning window light, shot on a 50mm lens at f/2, natural editorial style, a chalkboard in the background reading "Fresh Roast Today".
Flux parameters and controls
| Control | Syntax | What it does |
|---|---|---|
| Natural-language prompt | Full sentences | Flux follows complete descriptive sentences well. Order details from most to least important.e.g. A red fox sitting in fresh snow, looking toward the camera, soft overcast light. |
| Quoted text | "your text" | Text inside quotation marks is treated as words to render in the image. Keep it short and say where it appears.e.g. a neon sign above the door that says "OPEN LATE" |
| Guidance | guidance value in the app or API | Controls how strictly the image follows the prompt. Moderate values usually look most natural; very high values can look over-processed. Defaults differ by variant.e.g. guidance 3.5 |
| Steps | steps / num_inference_steps | Number of generation steps on variants that expose it. Faster variants are tuned for few steps.e.g. 28 steps |
| Seed | seed | Fixes randomness so you can reproduce a result or test single changes.e.g. seed 90210 |
| Aspect ratio / size | aspect_ratio or width × height | Sets image dimensions through the host's controls rather than inside the prompt.e.g. 16:9 or 1344×768 |
Do this
- Write in full sentences and describe positions explicitly: "on the left", "in the foreground", "behind her".
- Put any text you need rendered in quotation marks and keep it to a few words for the most reliable spelling.
- Describe the camera like a photographer would: lens length, aperture, angle and distance.
- State what you want instead of what you do not want, since most Flux setups have no negative prompt.
- Specify realistic lighting sources ("a single desk lamp", "late afternoon sun through blinds") for convincing photorealism.
- Keep a seed fixed while refining wording, then randomize to explore final variations.
Avoid this
- Pasting tag-style prompts full of weights like (word:1.4), which Flux does not use the way Stable Diffusion does.
- Writing "no people, no text" in the prompt, which can introduce the very things you named.
- Asking for long paragraphs of in-image text, which increases spelling errors.
- Including Midjourney flags such as --ar in the prompt instead of setting size in the app.
Flux prompt examples
Poster with text
A vintage travel poster of a lighthouse on a rocky coast at sunset, bold flat shapes and grainy print texture, deep orange sky and navy sea, large title at the top reading "VISIT CAPE NORTH" in cream serif letters.
Street photo
A cyclist waiting at a crosswalk on a rainy Tokyo street at night, reflections of red and blue signs on the wet asphalt, shot from across the street on a 35mm lens, shallow depth of field, cinematic color grading.
Product flat lay
Top-down flat lay of a leather notebook, a brass fountain pen and a cup of black coffee on a walnut desk, soft diffused daylight from the upper left, neat grid composition with generous empty space on the right.
Character portrait
Close-up portrait of an old fisherman with a white beard and a knitted navy cap, deep wrinkles and weathered skin, overcast harbor light, 85mm lens at f/2, muted blue and gray tones, honest documentary style.
Flux prompt FAQ
Does Flux support negative prompts?
Generally no. Standard Flux pipelines do not use a negative prompt, although a few third-party interfaces add workarounds. The more reliable approach is to describe the image you want precisely: say "clean empty background" rather than "no clutter". If a stubborn element keeps appearing, rephrase the scene so it has no reason to be there.
How do I get readable text in Flux images?
Put the exact words in quotation marks and say where they appear and how they look, for example a sign that says "OPEN" in red block letters. Short phrases render far more reliably than sentences. If spelling slips, regenerate with a new seed or simplify the surrounding scene so the text is the clear focal point.
What guidance value works best for Flux?
It depends on the variant and host. Many setups default to a guidance value around 3 to 4, which balances prompt adherence and natural look. Raise it slightly if the image ignores details, and lower it if colors look harsh or plastic. Distilled fast variants may not expose guidance at all, so check your provider's documentation.
How long should a Flux prompt be?
Flux handles long prompts well, so two to four detailed sentences are usually fine. What matters more is clarity: one subject, a defined setting, specific lighting and a camera description. Avoid contradicting yourself and avoid padding with generic quality words, which add length without adding information the model can act on.