Negative Prompts Explained (And Why Yours Is Too Long)
Copy-pasted 200-word negative prompts are making your images worse. Here is what a negative prompt is really doing and how to write a short, targeted one.
A negative prompt pushes the generation away from a region of the model's latent space. Every word you add moves the result, including the words you did not need.
Why long lists hurt
If you negate "dark" on a night scene, you fight your own prompt. If you negate "painting" on a stylised illustration, you flatten the style you asked for. Generic mega-lists contain dozens of terms that contradict the positive prompt, and the model splits the difference.
Write for the failure mode
Generate once. Look at what actually went wrong. Then negate that. A useful negative prompt is usually eight to fifteen terms, all specific to the shot:
- Portraits: plastic waxy skin, extra fingers, deformed eyes, asymmetrical face, heavy retouch.
- Products: visible support stand, cluttered background, colour fringing, warped edges, dust.
- Landscapes: tilted horizon, cloned repeating patterns, heavy haze, oversaturated sky.
- Video: jump cuts, flickering, morphing hands, sudden zoom, frame stutter.
Some models barely use it
Diffusion models with classifier-free guidance respond strongly to negatives. Several newer models treat them as a weak hint, and conversational image tools often ignore them entirely — with those, phrase it positively instead: "clean uncluttered background" rather than negating clutter.
Rule of thumb
If you cannot say why a term is in your negative prompt, delete it. Every prompt in this library ships with a negative prompt written for that specific image, not a copy-pasted block.
