Negative Prompts: What They Do and When to Use Them
A negative prompt is a list of things you do not want in the result. Many image models and some video models have a separate field for it. Used well, it removes recurring problems. Used badly, it does nothing, or makes results worse. This post explains how negative prompts work and when they are worth using.
How a negative prompt works
In models that support it, the main prompt pushes the result toward what you describe, and the negative prompt pushes it away from what you list there. The two fields are read separately, which is why an exclusion belongs in the negative field and not in the main prompt.
Writing “no cars” in the main prompt puts the word “cars” in front of the model. Some models then add cars. In a negative field, “cars” is read as something to avoid.
What to put in it
Negative prompts work best for specific, visible problems you have actually seen in your results:
- Artifacts:
blurry, distorted hands, extra fingers, warped face - Unwanted additions:
text, watermark, logo, subtitles - Wrong style:
cartoon, illustration, 3d renderwhen you want photo-realism - Motion problems in video:
jitter, flicker, morphing
What not to put in it
- Long copied lists. A list of fifty words from the internet dilutes the few terms that matter. Start empty and add a term only when a problem appears.
- Things that are not in the image anyway. Excluding “dragons” from a kitchen scene does nothing useful.
- Abstract quality words. “Bad”, “ugly” and “low quality” are vague. Name the defect you see.
When the model has no negative field
Many video models only take one prompt box. In that case, rewrite the exclusion as a positive description:
| Instead of | Write |
|---|---|
| no people | an empty street at dawn |
| no camera movement | static camera, locked-off tripod shot |
| not cartoonish | photorealistic, shot on 35mm film |
| no clutter | a minimal room with bare white walls |
Negative prompts in JSON
Structured prompts usually give exclusions their own field, so they never mix with the scene description:
{
"scene": "A quiet library reading room in the early morning.",
"lighting": "soft window light",
"negative": "people, text, watermark, blurry"
}
Check your model’s documentation for the exact field name it expects. The JSON output of the Prompt Builder keeps things to avoid in a separate field for this reason.