Why Flux ignores your negative prompt
Flux-family models run without a classifier-free guidance channel, so the negative prompt field is decorative. What to write in the positive prompt instead.
The negative prompt box on a Flux request is, on a growing share of the family, a decoration. You can type into it. The request succeeds. The weights never see a signal derived from what you typed. Black Forest Labs' own FLUX.2 prompting guide is explicit: FLUX.2 does not support negative prompts. Focus on describing what you want, not what you do not want.
That is a different failure from a weak negative. A weak negative still moves pixels; you just over-specified. An ignored negative leaves the image identical to the empty-field run. People spend a session lengthening the exclusion list, conclude the terms are not strong enough, and add more. Nothing in the output changes, because nothing in the sampler changed.
There is no second prediction to push against
Classic negative prompting rides on classifier-free guidance. Each step, the model produces two predictions: one conditioned on your prompt, one not. CFG scale is the weight on the difference. The negative prompt is what you load into that second slot, so the sampler can push away from it. No second slot, no push.
Flux-family image models are trained with flow matching. On Flux.1 [dev], the usual sampler setting is CFG at 1, which zeros the unconditioned term; prompt adherence is a separate distilled-guidance scale, commonly around 3.5. That is not a negative-prompt channel. Black Forest Labs' FLUX.2 prompting guide is blunt: FLUX.2 does not support negative prompts. FLUX.2 [flex] still exposes a guidance slider (BFL documents a 4.5 default), and that slider is still not a second slot you can load an exclusion list into. Distilled siblings such as Flux Schnell go further: guidance is baked into the weights so the sampler can skip the second pass entirely, which is why they are fast, and which is also why there is nowhere for an exclusion list to plug in.
This is not a bug that a later UI toggle is going to patch. Treating the negative box as a quality lever is using a tool the architecture retired.
Two practical consequences follow.
The hosted schema often has no negative_prompt parameter at all. On Versely, a model's prompting page is generated from the published input schema. If that schema has no negative field, exclusions typed in some other UI are not being forwarded. Check the page for Flux 2 Max before you spend a session debugging a field the endpoint does not accept.
Conflicting terms are not subtraction. If a UI still accepts a negative string and you send "red dress" in the prompt plus "red" in the negative, you have not cancelled red. You have added a second string the sampler was not built to push against. Drop the negative. Rewrite the positive so the dress is the colour you actually want.
Flux 1, Flux 2, and the leftover box
The family is not one model, but the negative-prompt story is the same shape across it.
- FLUX.2 (Pro, Max, Flex, Klein). Black Forest Labs documents the gap directly: no negative prompts. Official replacements are positive restatements. "Sharp focus throughout" instead of "no blur". "Empty scene" instead of "no people".
- Flux 1.1 Pro and the 1.x line. Flow matching, and BFL's family-level line is that FLUX models do not support negative prompts. Community UIs often still render a negative box inherited from Stable Diffusion. Inheritance is not support. Confirm with the seed-locked test below if a particular stack still moves pixels.
- Schnell, Flash, Turbo, and other distilled paths. Faster because they skip the pass a negative would have ridden on. Stuffing the leftover box, or raising a leftover guidance slider, is how you get the burnt look distilled models are famous for, not how you suppress a watermark.
A reliable empirical test, when the endpoint actually lets you send the field: lock the seed, run once with the negative empty, run again with an aggressive list. If the two images are the same, the field is inert. If they differ, that stack is still doing CFG-style guidance. Do not take a visible text input as the answer. Take the pair of images.
The broader replacement habit (positive re-specification, constrain by reference, constrain by repair) lives in what to write instead of a negative prompt. The rest of this page is Flux-specific: what the positive prompt has to carry once you have accepted that the second box is not a tool.
What to put in the positive prompt instead
Flux rewards sentences, not keyword piles. That is the same advice as the Flux image prompting guide, and it becomes load-bearing once negatives are gone, because every exclusion you used to outsource to the second box now has to be a fact about the scene.
BFL's own substitutions are the right template. Name the state you want.
| You were excluding | Weak (still a negative) | Flux-shaped positive |
|---|---|---|
| Blur, softness | no blur, sharp | sharp focus throughout, 50mm, subject crisp, background a smooth falloff |
| Extra people | no people, no crowd | an empty street at dawn, one figure, closed shopfronts |
| Plastic skin | no cgi, not airbrushed | visible pores, faint vellus hair, uneven specular on the forehead |
| Text, watermarks | no text, no watermark | a clean frame, picture only, type added in layout |
| Cartoon / illustration slide | no cartoon, not anime | shot on Kodak Portra 400, natural grain, a documentary still |
| Extra fingers | no extra fingers | hands in pockets / hands out of frame / a relaxed five-finger grip on the mug |
| Wrong colour | negative: red | a navy wool coat, hex-locked if the model accepts a code |
A few Flux-native habits make the rewrite stick.
Front-load the thing you used to negate. Word order matters; BFL says the model pays more attention to what comes first. If the failure was "a pedestrian walked into my empty street", start with the empty street, then the subject. Do not bury the absence-turned-presence at the tail, where truncation and attention both weaken it.
Use photographic facts, not quality pleas. "Photorealistic, 8k, masterpiece" is the old incantation and does close to nothing here. "Shot on a Canon 5D Mark IV, 35mm, late-afternoon side light, slight film grain" actually moves the output, and it also crowds out the illustration prior you used to ban with "no cartoon".
Hands, text, and object-class absences are repair jobs, not prompt jobs. "No extra fingers" is not a working instruction on this family. If a hand is wrong, inpaint the hand. If a sign is wrong, composite real type. If a second person appeared, describe a genuinely empty scene in the first clause or mask them out. The negative box was never good at these, and on Flux it is not even connected.
Worked prompt, product on a table.
Before (two fields):
- Positive:
ceramic mug on a walnut table, morning window light, photorealistic, 8k - Negative:
blurry, extra fingers, text, watermark, cartoon, ugly, low quality
After (one field):
A single ceramic mug on a walnut table, morning light from a window at camera left. Empty room, no other objects in frame. Shot on a 50mm lens, sharp on the glaze, wood grain in focus. Clean picture, type added in layout.
The after version spends its words on count ("a single", "empty room"), on light, and on where type will be handled. It does not mention blur, fingers, cartoons, or quality. Those were never Flux's job.
A working sequence on Flux
- Leave the negative field empty. If the schema does not expose one, you are already doing this.
- Write the scene as sentences in the five slots Flux actually uses: subject, action, setting, light, camera. The prompting guide above is the template.
- Generate twice before diagnosing. One ugly hand is variance. The same watermark twice is a pattern.
- For each pattern, name the desired state in the positive prompt. Put it near the front if it is the reason you rerolled.
- If the defect is structural (anatomy, spelling, a whole extra object), go to repair: inpaint, composite, or a masked second pass. Do not grow an exclusion list.
- If you are unsure whether a leftover negative field is live, run the seed-locked empty-versus-full test once. Then stop using it on that model.
On Versely, run that loop in the text-to-image tool. Swapping Flux tiers does not connect a negative channel the family does not have.
FAQ
How do I know Flux is ignoring the field, rather than my terms being weak?
Lock the seed. One run with the negative empty, one run with a long, aggressive list. Identical images mean the field is inert. Different images mean that particular stack is still doing something with it, and you can treat it as live. Visible UI is not evidence. The pair of images is.
Does Flux 1.1 Pro ignore negatives the way FLUX.2 does?
Treat it as the same class until the seed-locked test says otherwise. BFL's current line is that FLUX models do not support negative prompts, not only FLUX.2. Schnell and other distilled paths are even less likely to read the box, because the second pass was trained out.
What if I need a whole object class gone, and "empty street" still grows pedestrians?
Absence of a class is the one job positive language struggles with, because "street" statistically implies people. State the emptiness as a scene fact, up front: "an empty street at dawn, one parked bicycle, closed shutters, no pedestrians in frame" is still partly a negative, but it is a scene description rather than a standing rule in a second box. If the extra person survives two takes, mask them. That is cheaper than inventing a Flux-negative workflow the weights will not run.
Why do so many UIs still show a negative prompt box on Flux?
They were built for Stable Diffusion-class models, where CFG is real and the second slot is real. The layout survived the architecture change. If the published schema has no negative_prompt parameter, the box is not part of the request.