How to Write Flux Prompts (2026)

If you learned prompting on Midjourney, your instincts will actively work against you on Flux. The comma-separated keyword stacks, the “8k, ultra detailed, masterpiece” quality tags, the --ar and --stylize flags at the end — Flux ignores all of it, or worse, treats it as noise that dilutes the words that actually matter. Prompts that would win on Midjourney routinely produce flat, confused images on Flux, and beginners conclude the model is worse when the real problem is a dialect mismatch.

The consensus is total: 100% of our 37 tested Flux prompts run guidance at 3.5, with a median of 28 steps — the tightest settings cluster of any model we test (full corpus data).

Flux speaks a different language: plain, descriptive English. It was trained with a text encoder that reads full sentences the way a person would, so the prompts that work best look less like a tag list and more like a paragraph from a novel or a photographer’s shot notes. Once you make that one mental switch, Flux becomes the easiest of the major models to get precise results from — and the only one where readable text inside the image is genuinely reliable. This guide covers the writing style, the two settings worth knowing, and the traps that catch people arriving from Midjourney-style prompting.

Which Flux is which

“Flux” is actually a family of three variants, and knowing which one you’re using explains a lot of result differences:

• Flux.1 dev — the open-weights model most hosted playgrounds and local UIs run. Quality is close to the top tier, it’s free to run yourself, and it’s the default for anyone who isn’t paying per image. Every Flux prompt in our library is tested on Flux.1 dev, so results should transfer directly.
• Flux.1 schnell — the speed-optimized variant, producing an image in as few as four steps. Great for rapid drafting and composition checks; noticeably softer on fine detail. Draft on schnell, finish on dev.
• Flux.1 pro — the strongest variant, available only through hosted APIs and paid platforms. If you’re on a commercial image tool built on Flux, this is often what’s under the hood.

The prompting advice below applies to all three — the writing style is the same; only speed and polish differ.

Write scenes, not keyword lists

The core Flux skill is scene-writing: its text encoder (a T5 language model) parses grammar — subjects, verbs, spatial relationships, which adjective belongs to which noun. A keyword list throws all that away, leaving Flux to guess how the fragments connect. A sentence tells it exactly.

Compare a keyword-stacked prompt with a scene-written one:

lighthouse, storm, cliff, dramatic waves, moody lighting, fog, 8k, ultra detailed

A weathered white lighthouse on a storm-battered cliff at dusk, waves
exploding against the rocks below it, its beam cutting through low fog.
One warm lit window in the keeper's cottage. Overcast blue-grey light,
photographed from the beach with a telephoto lens.

The first version gets you a lighthouse and some storm, arranged however the model feels like. The second controls the relationships: the waves are below, the beam goes through the fog, the one warm window plays against the cold palette. That relational control is what keyword lists can’t buy at any length.

The same principle turns a generic interior into a specific place:

bakery, fresh bread, morning light, cozy, rustic, warm tones, photorealistic

Morning light through a bakery window falls across a flour-dusted wooden
counter, where a tray of golden sourdough loaves cools beside a
handwritten card reading "FRESH TODAY". Soft warm window light, shallow
depth of field.

Notice the second prompt spends its words on things you could point at — flour dust, a specific card with specific words — instead of mood adjectives. Flux converts concrete nouns into detail far more reliably than it converts “cozy” into anything.

The two settings that matter

Flux exposes far fewer dials than Stable Diffusion, and only two deserve your attention as a beginner:

• Guidance (CFG). The sweet spot is around 3.5, and it’s a genuinely narrow window. Push it to 6 or 7 — normal territory for Stable Diffusion — and Flux images turn harsh: burned highlights, oversaturated color, plasticky skin. Drop below 2 and things go dreamy and loose, drifting away from your prompt. When a Flux image looks “fried”, the fix is almost never in the prompt; it’s the guidance.
• Steps. Around 28 is the reliable zone for dev (20–30 works; more buys little). Schnell is built for about 4 steps — running it longer doesn’t make it dev.

Just as important is a setting Flux doesn’t have: there is no true negative prompt channel. You can’t hand it a list of things to avoid the way Stable Diffusion’s negative prompts work. The workaround is to phrase positively: instead of “no clutter”, write “a clean, empty counter”; instead of “no people”, write “a deserted street at dawn”. Describe the world where the unwanted thing doesn’t exist, and Flux builds that world.

Flux’s superpowers

Three things Flux does better than any model in its class, worth designing prompts around:

1. Photorealism. Skin texture, natural lighting falloff, and material rendering — brushed metal, wet stone, fabric weave — are where Flux earns its reputation. Naming materials and light sources explicitly plays directly to this strength.
2. Legible text. Flux is the one model where short readable text inside an image is a viable request, not a gamble. Signs, labels, a word on a mug — put the exact wording in quotes in your prompt. Keep it to a few words; whole sentences still degrade.
3. Hands. The classic AI tell is far less of a problem here. Prompts involving hands holding, pouring, or cooking — long a lost cause elsewhere — are safe territory.

See it in practice: three tested prompts

Everything above shows up in the Flux entries from our library. First, photorealism applied to a portrait — the genre where Flux’s skin rendering matters most.

Why it works: “Natural skin texture with freckles” is the anti-plastic clause — without it, Flux airbrushes faces into rendered perfection. And “still from an indie film” quietly sets the entire color grade in four words, more effectively than any list of color adjectives. Note the guidance at 3.5: raising it is exactly what makes AI portraits look uncanny.

Next, material rendering and light control in a product shot.

Why it works: The load-bearing sentence is “one long controlled specular highlight along the bezel” — naming the highlight explicitly steers Flux’s lighting the way a studio photographer would place a softbox. The “no logos or brand text on the dial” line matters too: leave it out and Flux, good at text as it is, will invent plausible-looking brand marks you can’t legally use.

Finally, a complex scene that shows how sentence-style prompting handles density.

Why it works: The prompt spends its words on documentary details — “handwritten menu cards”, “mismatched stools” — that make the stall read as a real place instead of a matte painting, and the billowing steam gives all that neon something physical to light. Keep guidance at 3.5 here as well; higher values oversharpen the rain and turn the puddle reflections glassy and fake.

Mistakes that waste your generations

1. Pasting Midjourney flags. --ar 16:9 and --stylize 400 are just words to Flux — they do nothing except pollute the prompt. Aspect ratio, guidance, and steps are set in your UI or API call, not in the text.
2. Keyword soup. Ten mood adjectives cost you the token budget that two concrete details would have spent better. If you can’t point at it in the imagined photo, cut it.
3. Cranking guidance to force detail. High CFG doesn’t add detail on Flux; it adds contrast and artifacts. Detail comes from describing more specific things, at guidance 3.5.
4. Expecting LoRA-style control without LoRAs. A recurring character, a locked art style, your own product — base Flux prompting won’t hold identity across images. That’s what LoRA fine-tunes are for; no phrasing substitutes for one.

Is Flux better than Midjourney?

For photorealism, legible in-image text, and anything you want to run on your own hardware with open weights, yes — Flux is the stronger choice. For heavily stylized art and fast, opinionated iteration where the model’s own aesthetic does half the work, Midjourney still wins. They’re not competitors so much as different tools: pick Flux when the image needs to look like a photograph or contain words, pick Midjourney when it needs to look like art, and let the job decide.

Where to go next

Browse the full Flux prompt library for tested starting points — every entry lists the exact variant, guidance, and steps we used. If product shots are your goal, the product photography use-case hub collects the prompts that trade on Flux’s material rendering. And even though Flux skips negative prompts, understanding how negative prompts work in Stable Diffusion will sharpen your instinct for what to phrase positively here.