What a negative prompt really does

Not a prohibition, a second anchor

The negative prompt box invites a wrong model of what is happening. People type no extra fingers, no watermark, no blurry and imagine a filter that removes those things.

Go back to the guidance arithmetic. Each step runs the model twice — once with your prompt, once with an empty one — and pushes the result along the difference:

guided = negative_prediction + scale * (positive_prediction - negative_prediction)

The negative prompt simply fills in that second run. Instead of "where would the model go with no instruction", the second prediction becomes "where would the model go if asked for blurry, watermark, low quality". The image is then pushed away from that point and toward your prompt, by the guidance factor.

Three consequences follow, and each explains a common confusion.

Negative prompts have no effect at guidance 1. Look at the formula: at scale 1 the negative term cancels out entirely. On distilled models designed to run at guidance 1, or on models with no classifier-free guidance at all, the negative box does nothing whatsoever. Interfaces still show it, which is unhelpful.

A negative prompt costs you the same as a positive one. Both runs happen regardless, so a negative prompt is free in compute terms. But it is not free in effect — every token in it moves the anchor, including tokens you added carelessly.

Negation is not understood. Writing no hands in the negative box moves the anchor toward images that a caption-trained encoder associates with the word "hands". You are pushing away from hands, which is nearer to "hide the hands" than to "draw the hands correctly". This is why the fix for bad hands was never a negative prompt.

What actually belongs in there

Negative prompts work best when they name a visual register you do not want, because that is what encoders represent well.

Useful:

• watermark, signature, text overlay, stock photo — genuine visual patterns the model learned as a group.
• cartoon, illustration, 3d render when you want a photograph and the model keeps drifting toward illustration.
• oversaturated, high contrast as a partial counterweight to guidance burn.

Not useful:

• Counts. two heads in the negative does not enforce one head.
• Relations. hand behind back cannot be negated meaningfully.
• Anything specific to your image that has no visual signature of its own.

The enormous inherited negative prompts you see shared online — sixty terms including deformed, mutated, poorly drawn, extra limbs, disfigured — were tuned for one 2022 checkpoint and copied ever since. On most current models they measurably reduce quality by dragging the anchor into strange territory. Test yours by emptying it on a fixed seed. A surprising number of people find their images improve.

The related setting nobody explains

Some interfaces expose a separate control for how long the negative prompt applies — as a percentage of steps. This is more useful than it sounds. Composition is settled early; texture is settled late. A negative prompt aimed at texture problems, such as plastic, waxy skin, only needs to act in the last third of the run, and confining it there stops it from disturbing the composition.

If your tool offers it, the pattern is: composition-level negatives for the first 30% of steps, texture-level negatives for the last 40%, nothing in the middle. If it does not offer it, this is one of the reasons people move to node-based tools like ComfyUI, which expose scheduling for almost every parameter and cost nothing.

The honest limitation to hold onto: a negative prompt is a blunt instrument acting on the whole image at once. It cannot fix a local problem. When something is wrong in one region, the answer is a mask, which is the next module.

One last thing about the shared negative prompts you will meet. Many of them contain terms that are not visual at all — bad anatomy, worst quality, ugly. These entered circulation because one popular fine-tune was trained on a booru-style dataset where those exact tags appeared as human quality ratings on the images. On that checkpoint they genuinely worked, since the model had learned the tag as a visual category. On a checkpoint trained on ordinary alt text they refer to nothing, and the anchor they move is essentially random. This is the general rule under all of it: a negative term only works if the model was trained on captions containing that term. When you inherit a prompt, the first question is what it was written for, and the second is whether that is what you are running.