Negative Prompts in AI Image Generation Are Just Redirected Guidance, Not Filters

Negative prompts in diffusion models do not function as filters or constraints but instead replace the empty string in the unconditional branch of classifier-free guidance. The model runs two denoising passes per step — one with the user's prompt and one with the negative text — then extrapolates away from the negative prediction rather than a generic empty one. The negative prompt's effective weight equals one minus the guidance scale, meaning at a guidance scale of 1.0 it has zero influence and is often skipped entirely by the model. Raising or lowering the guidance scale simultaneously affects both the positive and negative prompt strengths, as they share the same dial. On guidance-distilled or dropout-free models, negative prompts have no effect at all, even if the UI still displays the input field.