Negative prompts: when saying "no" helps an AI image and when it backfires
Why "no people" in a prompt can add people, where a real Negative prompt field exists, and how to phrase what you do not want so models listen.
Write "a quiet beach at dawn, no people, no boats" and there is a fair chance you get a figure on the sand and a dinghy near the horizon. It feels like the model is being stubborn. It is not; it is reading your words the only way it can. This post is about the gap between what you mean by "no" and what a model does with it, and the three places where an exclusion actually works.
Why "no" summons the thing
Image and video models learned from pictures paired with captions. Captions describe what is in a picture; they almost never list what is absent. So the model has a strong sense of what "boat" looks like and a weak sense of "no". When it reads "no boats", the strongest signal in that phrase is the word boats. In practice the exclusion often works as an invitation, especially with simpler or faster models.
Newer models that read whole sentences handle negation better than older ones, and some follow "without" quite well. But it varies from model to model in the catalog, so the safe habit is to avoid relying on it.
Place 1: describe the presence of the alternative
Most of the time an exclusion can be turned into a description of what is there instead. You are not asking for less of something, you are describing a different scene. This works on every model, because it uses the kind of words captions are made of.
- "no people" becomes "an empty beach, untouched sand, a deserted shoreline"
- "no text" becomes "a plain unmarked wall" or "a blank label"
- "not cartoonish" becomes "photograph, natural skin texture, real fabric"
- "no clutter" becomes "a single object on a bare table, lots of empty space"
- "not too dark" becomes "bright overcast daylight, soft even light"
A deserted beach at dawn, untouched smooth sand, calm empty sea to the horizon, pale pink sky, soft even light, wide shotThe same thinking drives the post on quality buzzwords: concrete visible words do the work, and abstract ones get averaged away.
Place 2: the Negative prompt field, when a model has one
Some models have a separate field for exclusions. In the studio, open the Advanced section under the model's settings; if the model supports it, you will see a field called Negative prompt. Text in that field is used to steer the result away from what it describes, which is exactly the mechanism the main prompt lacks. In the catalog this field is far more common on video models than on image models; most image models do not have it at all, which is one more reason to master Place 1.
A negative field reads best as a short list of nouns and adjectives, not sentences, and without the word "no" (the field already means no):
blurry, distorted face, extra fingers, flickering, text, watermark, camera shakePlace 3: keep clauses in an edit
When you edit an existing picture, the exclusion you usually need is "do not touch the rest". That one works well when phrased as a positive instruction: "Keep the face, the label and the composition exactly as they are." Edit models are trained to follow instructions, so a keep clause is read as a constraint, not as an invitation. The guide to writing edit prompts covers the pattern.
Change the sky to a clear evening blue. Keep the buildings, the street, the cars and the lighting on the facades unchanged.A quick test to run yourself
- Pick one subject and one thing you want gone, for example a city street without cars.
- Run it once with "no cars" in the main prompt.
- Run it once with "a pedestrian street, empty asphalt, closed to traffic" instead.
- If the model has a Negative prompt field, run the second version once more with "cars, vehicles" there.
Each run shows its price on the Generate button before you press it, and a draft-quality model is enough to see the difference. Three or four renders teach you how that particular model treats exclusions, which is worth more than any general rule, this one included.
What exclusions cannot do
No phrasing guarantees an absence. A negative prompt lowers the odds of something appearing; it does not forbid it, and on a busy scene a stray figure or a sign with nonsense letters still slips in sometimes. A negative field cannot remove something that is in your reference image or the first frame you upload; for that you need an edit or an object removal model. It cannot fix a problem the model simply has, such as unreadable lettering or wrong finger counts, though it can make them less frequent; the post on hands and text has the workarounds. And it is not a way around content rules: a model's own checks apply to what is generated, whatever the negative field says. When the unwanted thing keeps coming back after a few tries, change the scene so it has no reason to be there, or remove it afterwards.
Questions
Does writing "no people" in an AI image prompt work?
Often not. Models learned from captions that describe what is present, so the word people can pull people in. Describe the alternative instead, such as an empty street or a deserted beach.
Where is the negative prompt in the studio?
On models that support it, in the Advanced section of the model's settings, as a field called Negative prompt. Many image models do not have one; it is more common on video models.
What should I put in a negative prompt?
A short list of the specific things that went wrong in your last result, such as blurry or watermark, without the word no. Avoid long copied lists; they flatten the picture.
