The negative prompt
Saying what you do not want, and when it does nothing at all.
In plain words
The negative prompt is the text that goes into the sampler's negative input. At every step the model makes two predictions: one from your prompt and one from the negative prompt. Then it moves away from the second towards the first. How far is set by CFG.
So the negative prompt names the very things to move away from: blurry, extra fingers, watermark, text. There is no need to write "no".
Where you meet it in a graph
A second CLIP Text Encode (Prompt) node, fed by the same CLIP, with its output in negative. Some pack nodes keep both texts in one node (WanVideoWrapper's text node has positive_prompt and negative_prompt).
Values and what they change
- A short list of defects that really appear in your pictures works better than a long generic one.
- The ComfyUI templates for Wan carry a long negative prompt in Chinese (overexposure, still frame, blurred detail, subtitles, extra fingers, deformed hands); it can be kept as it is.
- Some newer models, and every model at CFG 1, ignore it.
When it does nothing
At CFG 1 the second prediction is not made at all, so the negative prompt has no effect, whatever it says. This is the case with Flux dev and schnell, Z-Image Turbo, Flux.2 Klein and other distilled models, and with speed-up LoRAs (lightx2v for Wan). For them, everything that matters must be in the main prompt.
Typical mistakes
- Negations in the main prompt (
no hat): the wordhatpulls towards hats. Move it here. - The same word in both prompts: two opposite instructions.
- Expecting an effect at CFG 1.
- Empty negative prompt left unconnected: the
negativeinput is required. Connect an encode node with empty text (or, for some models, a ConditioningZeroOut node).
How the app helps
Check says when a sampler has cfg 1 so the negative prompt does nothing, and when the same thing stands in both prompts, and when the main prompt holds a negation.