Flux Prompts: Writing for FLUX Instead of SDXL

The second result for flux prompts is a Reddit thread whose author admits to "a SD1.5 and SDXL history" and asks how to prompt Flux, and that's the honest shape of the query. Most people arriving here have a working vocabulary from Stable Diffusion and are finding that half of it does nothing.

Here's the short version, all of it from Black Forest Labs' own prompting guide. Write sentences, and put the subject first because word order signals priority. Drop the parenthesis weights, since they have no effect. Drop the negative prompt, because most FLUX models don't support one, and describe what fills the space instead. Put any text you want rendered inside quotation marks, then say where it goes and in what typeface. Aim for 30 to 80 words for an ordinary scene, knowing FLUX.2 will accept up to 32K tokens if you ever need them. And find out which FLUX you're on, because three of the FLUX.2 variants quietly expand a short prompt for you and one takes you completely literally.

The rest of this is the translation table, one prompt written both ways, and seven more written for the jobs people actually search for.

Which FLUX You're Prompting

The guide covers FLUX.1, FLUX.1 Kontext, and FLUX.2, and BFL's own index now says to use FLUX.2 for new text-to-image projects. Within FLUX.2 the variants differ in ways that change how you write, so here's the table I wish the copy-paste lists opened with. Prices are BFL's list prices on the day I read the page.

The klein row is the one to notice. The 4B version runs on consumer GPUs at roughly 13GB of VRAM per BFL's note, which makes it the one I'd reach for locally, and it's also the one where a lazy ten-word prompt stays a lazy ten-word prompt. On pro, max, and flex the docs say short prompts get "automatically enhanced" with detail and context, which is generous right up until you're trying to work out why your tightly written prompt and your throwaway one produced similar images. If you want to learn what your words do, klein or dev is the honest teacher.

The Translation Table

Every row is an SDXL habit on the left and what BFL's documentation says to do on FLUX on the right. Where the claim comes from a ranking page and not the vendor, I've said so.

The negative-prompt row deserves a paragraph because the results page disagrees with itself about it. One ranking how-to tells you to add "people, crowds, figures" to a negative prompt to clear a forest scene. BFL's Technical Parameters page says most FLUX models don't support negative prompts, and that even where a model can process one, "writing 'a person without glasses' causes the model to focus on 'glasses'". Their replacement table turns "no people" into "empty", "deserted", or "solitary", and "no text" into "clean surfaces" or "unmarked". So the forest fix is "an empty forest path, no other figures in sight" written into the prompt itself, or better, "a deserted forest path". The general version of why exclusion fields exist on some systems and not others is in negative prompts.

One Prompt, Written Both Ways

Same image in my head. First the way I'd have typed it into an SDXL interface, then the FLUX rewrite.

masterpiece, best quality, 8k, ultra detailed, (old fisherman:1.3) mending a net, harbour, dawn, (documentary photography:1.2), 35mm, film grain, bokeh, sharp focus
Negative prompt: blurry, deformed hands, extra fingers, text, watermark, cartoon, low quality

Close-up documentary photograph of an old fisherman mending a torn net with both hands, sitting on an upturned crate at the edge of a small harbour at dawn. Weathered face, cracked knuckles, a wool cap pulled low. Low pale sunlight from the left, cold blue shadows, thin mist over the water behind him. Shot on Kodak Portra 400, 35mm lens, shallow depth of field, natural grain. Quiet and unposed.

Six moves, each one a row from the table. The subject and the framing lead, so "close-up" isn't fighting "harbour" for the composition. "Deformed hands, extra fingers" became "mending a torn net with both hands" and "cracked knuckles", which is the replacement strategy applied to the thing SDXL users fear most. The quality block is gone and a film stock does its job. "Bokeh" became "shallow depth of field", which is the phrase the camera cheat sheet uses. The lighting got a source, a direction, and a temperature, because the style page says lighting has the greatest single impact on output quality and "good lighting" isn't a description. And it runs to about 70 words, inside the medium band.

That prompt would behave differently on klein than on pro, and I'd expect the difference to be that pro fills in atmosphere I didn't ask for. That's a prediction from the upsampling note, not a test. I run image generation locally on an M4 Pro, so testing a claim like that costs me electricity, and if you can do the same, render on both and keep the one whose result you can trace back to your own words.

Seven More, One Per Job

Each of these is written in full, with the structure that made it work explained after.

Product Shot With A Brand Colour

Studio product photograph of a matte ceramic travel mug, colour #2F5D50, standing on a pale concrete slab. Overcast, shadow-free light from above, a soft reflection under the base. 85mm lens, f/8, everything sharp. Plain warm-grey background with nothing else in frame.

The hex code is the FLUX.2 feature the copy-paste lists never use, and BFL's overview example runs a gradient between two codes on a single vase. "Overcast" is from their lighting cheat sheet, where it's listed as flat and even and "great for product shots". "Nothing else in frame" is the positive form of "no props".

Poster With Text That Has To Read

A flat poster design on a cream paper background. The headline "SATURDAY MARKET" runs across the top third in bold condensed sans-serif capitals, deep green. Beneath it, a simple line illustration of three stacked wooden crates of apples, red and yellow. At the bottom, smaller text reads "Every week, 8am to 1pm" in the same green. Screen-print texture, slight ink bleed.

Quotes, placement, typeface, in that order, which is the three-step method on the style page. Two separate strings each get their own placement sentence. The texture line is the one effect, and there's only one, because the docs say two at most and I'd rather leave room.

An Illustration That Isn't A Photo

Children's book illustration of a fox reading a newspaper on a park bench, drawn in a loose ink line with flat watercolour fills. Autumn leaves in ochre and rust scattered on the path. Soft morning light, no outlines on the background shapes. Gentle, slightly comic.

Art form first, then style, then palette, then mood, following the order the Building page recommends. "No outlines on the background shapes" is a rare place I've kept a negative, because it describes a drawing convention rather than an absent object, and the positive alternative is clumsier. If it fails, the replacement is "background shapes as soft washes with no line work".

An Edit Instruction For A Reference Image

Keep the sneaker, its colours, laces and logo exactly as they are. Replace the white studio sweep with wet black asphalt at night, magenta and cyan neon reflected in the puddles around it. Keep the camera angle and the shoe's position in frame unchanged.

Editing on FLUX.2 takes up to ten reference images and plain-language instructions. What I'd underline is the ratio. Most of the words say what to keep. The docs' own single-reference example is seven words long, "the butterfly is now made of shiny silver", and everything else in the image is preserved by not mentioning it. When an edit drifts, the fix is usually one more "keep" clause. More description of the change rarely helps.

A Klein Prompt That Says Everything

Wide landscape photograph of a single white lighthouse on a grey basalt headland, late afternoon, low sun from the right throwing a long shadow across the grass. Overcast sky breaking into pale gold at the horizon. Calm dark sea, no boats. 24mm lens, f/11, deep focus, slight cool colour cast. 16:9.

Written for the variant that adds nothing. Every slot in BFL's nine-component template is filled except "additional elements", which I left out on purpose since the docs call those refinements rather than foundation. The aspect ratio is in the prompt as a reminder to set it in the interface, where their advice is that a landscape prompt gets 16:9 and a mismatched ratio forces a crop or padding.

The Same Scene For Pro, Shortened

A white lighthouse on a grey basalt headland at late afternoon, wide shot, cool overcast light.

Sixteen words, inside the short band the docs describe as good for "quick concepts, fast iteration, style exploration". On pro, flex, or max the upsampling fills the rest, and the point of showing both is that you can't compare the long prompt and the short one across variants and learn anything, because one variant rewrote the short one before rendering it.

A Structured Prompt For Automation

{
  "subject": "a brass pocket watch, open, face up",
  "background": "dark green felt, nothing else on the surface",
  "lighting": "single soft key light from the upper left, gentle falloff",
  "style": "still life photograph, muted and precise",
  "camera_angle": "45 degrees from above",
  "composition": "watch centred, square crop, generous margin"
}

FLUX.2 accepts structured prompts, and BFL's overview shows a JSON example with exactly these keys. I wouldn't write one by hand for a single image. Where it earns its place is a pipeline, where a script fills the subject key from a product list and everything else stays fixed, so the lighting and the angle can't drift between the first render and the fortieth. The keys are the nine-component template wearing a different coat.

Two Habits The Lists Skip

Iterate one thing at a time. That's the whole habit. BFL's basics page puts it as a three-step loop, start with a simple version, check what FLUX got right and wrong, adjust one important detail. The reason it matters more here than it did on SDXL is that a FLUX prompt is prose, and prose invites rewriting the whole thing when a render disappoints. Resist that. Change the lighting sentence and nothing else, render, and now you know what the lighting sentence did. Change three sentences and you know nothing.

Set the aspect ratio before you write. The Technical Parameters page lists six ratios with their uses and says a mismatched ratio forces the model to crop or pad the composition. A prompt that describes a wide headland and renders at 1:1 will lose the headland, and no amount of prompt editing fixes a frame the model was never given.

Realistic Photos, Since That's The Question People Ask

The People Also Ask box wants good Flux prompts for realistic photos, and BFL's style page answers it with four looks and their descriptors, which I'd copy before writing anything of my own.

Their one-line rule is that naming a camera, lens, and film stock "produces more authentic results than just 'professional photo'", and the fisherman prompt above is that rule applied. The lens table on the reference page is worth memorising in miniature. 35mm reads as documentary, 85mm as portrait, f/1.4 to f/2.8 blurs the background, f/8 to f/16 keeps everything sharp, ISO 1600 and up buys grain. Those six facts cover most realism prompts I'd write. The subtractive side of realism, meaning what to leave out so the image stops looking generated, is in realistic ai prompts.

The other two questions in the box, cool photo prompts and trending photo prompts, are the same question at different speeds. Cool effects are style references that wear off once you've seen them twice, and trends are the ones wearing off this month. I wrote up why in cool ai prompts, and the durable part of any of them is the lighting and lens vocabulary above, which doesn't trend.

What I'd Do First

Find out which variant you're on. If it's klein or dev, write the full lighthouse prompt and read the image against your words. If it's pro, max, or flex, write it anyway, then write the twenty-word version, and notice how much of the long one the model would have invented for you.

Then take one old SDXL prompt you liked and translate it row by row. The weights come out first, the negative block gets rewritten as positives, the quality tags become a film stock, and the subject moves to the front. The vocabulary underneath, the lenses and the light and the materials, is the part that was never SDXL-specific, and it's the part that carries to every model in ai image prompt. The syntax was the disposable layer. On FLUX there's barely any syntax left, which is either a relief or a loss depending on how much of your old prompt was parentheses. For what the other family still does with those parentheses, stable diffusion prompts has the tokens, weights, and guidance settings side.