Lời nhắc tiêu cực trong việc tạo hình ảnh

Diffusion models create an image by starting from random noise and removing it over many steps. At each step, the model predicts the noise twice: once conditioned on your prompt and once on an 'unconditional' input, normally an empty prompt. Classifier-free guidance, introduced by Jonathan Ho and Tim Salimans, combines the two. It starts from the unconditional prediction and moves further toward the conditional one, scaled by the guidance scale (often labeled CFG). The result follows the prompt more closely than the raw model would. A negative prompt simply replaces the empty prompt with your negative text. The formula now pushes away from the negative description as it moves toward the positive one. That is why a negative prompt is not a filter or a post-processing step: it shapes every denoising step. The technique spread through community tools such as the AUTOMATIC1111 web interface in 2022. It became especially popular with Stable Diffusion 2.x, whose results many users found improved noticeably with negatives. Negative prompts work best on concrete visual concepts the model knows well: 'blurry', 'text', 'watermark', 'monochrome', or a specific color or object. They work less well as abstract fixes. On many general-purpose models, terms like 'bad anatomy' or 'extra fingers' often do little, because those models rarely saw captions describing such failures. They can also make results worse. A long list pushes the image away from many directions at once. That can reduce variety, wash out or oversaturate colors, and produce stiff compositions, and high guidance scales amplify the effect. Some models, including guidance-distilled and few-step models, skip the unconditional pass entirely. On those, negative prompts are ignored unless the tool turns true CFG back on, which slows generation. A common misconception is that writing 'no cats' in the main prompt excludes cats. Text encoders handle negation poorly, so the phrase can add cats instead.