Negative Prompt
Specifying what you dont want in a generated image - telling the model to avoid certain concepts or styles.
Think of it like
A restaurant order: "Ill have a burger, extra cheese. No pickles, no mustard." The "no" parts guide the chef.
Example
"A dog running, negative: blurry, low quality, deformed." "A sunset, negative: people, buildings, text."
How it actually works
Implemented via classifier-free guidance by steering away from the negative condition. If you want to avoid "blurry," compute pred_avoid_blurry and pred_blurry, then interpolate away: final = pred_avoid - scale x (pred_blurry - pred_avoid). Doesnt require training on negative prompts; just reframes guidance. Effective but sometimes over-corrects (oversaturation of details). Combined with positive prompts: balance both. Empirically improves coherence and reduces artifacts.
For product teams
Lets users refine outputs without iterating - powerful for creative control.
For engineers
Uses guidance formula with reversed interpolation. Works best with concrete descriptions (blurry, ugly) vs. abstract (bad).
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome