.ai

nejlepší open source AI modely

Hledat

Negative prompts: what they do, when they help and why Flux does not need them

Q: Image got blander after adding the negative

A long generic negative list averaging out the style. Cut to the three words that matter.

Q: Doubling generation time on Flux

A UI forcing CFG above 1 so the negative is evaluated. Set CFG to 1 and guidance to 3.5; leave the negative empty.

Téma: Promptování

Autor: Lookout

Publikováno 2025-11-05

The negative prompt explained through how classifier-free guidance works: a short list that helps on SD 1.5 and SDXL, the negative embeddings people actually use, why long copy-pasted negatives hurt, and what to do instead on Flux, Qwen and other guidance-distilled models that have no negative at all.

Tato stránka zatím nebyla přeložena, proto je zobrazena v angličtině.

Přehled

SD 1.5

and

SDXL

every step runs the model twice: once with your prompt and once with the negative prompt (empty by default). The

sampler

then moves the image toward the first and away from the second, scaled by CFG. The negative prompt is therefore not a filter applied afterwards; it is a second prompt that the image is pushed away from. That explains both its power (a few words remove a recurring flaw) and its limits (it cannot remove what the positive prompt keeps asking for).

Flux dev

FLUX.2

Z-Image Turbo

, SDXL Lightning and other distilled models run once per step with guidance baked in. They have no second pass to push away from, so the negative box is ignored or, in UIs that force a CFG above 1, it doubles generation time for little effect. On those models you fix problems in the positive prompt, with a

LoRA

, or by

inpainting

The copy-pasted 200-word negative prompt of 2023 ("deformed, mutated, ugly, disgusting, bad anatomy...") is mostly harmless on SDXL and mostly useless. Short, specific negatives work better: name the thing that actually appeared in your last four images.

Reference

Název	Typ	Co to je
Negative prompt	text (SD 1.5, SDXL)	Comma-separated tags the image is pushed away from. 5 to 15 words. Specific beats generic.
CFG scale	number	How hard the push is in both directions. A negative prompt does nothing at CFG 1.
Negative embedding	file	A trained token summarising a bad look: easynegative, bad-hands-5, negative_hand, FastNegativeV2 (SD 1.5); ac_neg1, unaestheticXL (SDXL). Typed by file name in the negative box.
[[double brackets]]	PirateDiffusion	The negative prompt inside a /render command: /render prompt [[blurry, watermark]].
Negative weight	syntax	(word:1.4) in the negative box pushes harder away from that word.
Flux / distilled models	no negative	Guidance-distilled; the negative box has no effect at CFG 1. Rephrase the positive prompt instead.

Test a negative in seconds

In BitVector Prism the negative box sits under the prompt; run the same seed with and without it and compare side by side. SDXL and Flux models from this site are preloaded.

Vyzkoušet BitVector

Krok za krokem

Generate four images with an empty negative prompt first. Look for the recurring flaw: extra fingers, text, a watermark, a blurry background, a second person.
Name that flaw in the negative box in two to four words: "extra fingers, watermark, text".
For anime models add the standard short set the model page recommends (lowres, bad anatomy, bad hands, extra digits, signature); for photo models usually nothing more.
If hands or faces are the problem on SD 1.5, add a negative
embedding
(bad-hands-5, easynegative) instead of a word list.
Keep CFG where the model likes it (5 to 7 on SDXL). Raising CFG to make the negative "work harder" burns the image.
On
Flux
,
Qwen
,
Z-Image
and Lightning models leave the negative empty; put the fix in the positive prompt ("hands in pockets", "plain background") or inpaint.
Save the short negative with the prompt. A negative that was right for one model is often wrong for the next.

Příklady

Short, specific negatives for an SDXL photo

Positive: a woman reading on a tram at night, window reflections, 50mm photo

Negative: text, watermark, extra fingers, second person

Illustrious standard set

Negative: lowres, bad anatomy, bad hands, extra digits, fewer digits, worst quality, low quality, signature, watermark, username

Negative in a PirateDiffusion command

/render <juggernaut-xl> a woman reading on a tram at night, window reflections, 50mm photo [[text, watermark, extra fingers]] /size:832x1216

Tipy

Negatives are good at removing styles and objects ("cartoon", "glasses", "car"), weak at fixing anatomy. Anatomy is fixed by a better model, resolution or inpainting.
A negative word that also appears in the positive prompt cancels out. "No blur" in the positive prompt is read as "blur".
Long negative lists shift the whole image toward a bland average look; if your images look generic, shorten the negative.
Negative embeddings are family specific. An SD 1.5 embedding in an SDXL negative box does nothing.
Pony
needs score_6, score_5, score_4 in the negative to match its score tags in the positive. Illustrious does not use scores.
PirateDiffusion
and
BitVector
accept negatives on SDXL models and quietly ignore them on Flux models, so one habit works across both.

Install the embedding in models/embeddings and refresh, or remove the word.

Negatives in [[double brackets]]

PirateDiffusion puts the negative prompt inside double brackets in the same /render command, remembers it for your session and ignores it on Flux models automatically. Unlimited renders, fixed price.

Vyzkoušet PirateDiffusion

Otázky

Is there a universal negative prompt?

No. "text, watermark" is the closest thing. Everything else depends on the model and the picture.

Why do anime prompts always carry a long negative?

Because Illustrious, NoobAI and Pony were trained with quality tags; the negative set tells them which end of the quality scale to avoid. Photo models were not trained that way.