# Negative prompts Instead of only telling the model what you want, you can explicitly tell it what to **avoid**. Negative prompts work by a technique called [[CFG|CFG (Classifier-Free Guidance)]]. The model generates two latent paths in parallel: one steered by your initial prompt (also known as positive prompt), and one steered by your negative prompt. It then interpolates *away* from the negative direction, amplifying the positive concept while suppressing unwanted ones. <br>This is why negative prompts are so effective: they actively push the generation *away* from what you don't. ^overview ### Types of Negative Prompts 1. **Quality Fixers**: These fix common model failures. - These are often "universal" and can be copy-pasted across many prompts. > [!example] `blurry, bad anatomy, missing fingers, extra limbs, low resolution, text, watermark`. 2. **Specific Exclusions**: These remove undesired elements in the scene. - These will be more context-dependent and opposite to the subject > [!example] Need an empty forest with only `forest landscape` as a prompt? > Add `people, humans, hikers` to the negative prompt. ### Common Pitfalls * **Don't Over-Negate**: Adding hundreds of negative terms can restrict the model too much, leading to "fried" or generic imagery because the latent space has nowhere left to go. * **Don't Use Negatives for Positives**: Saying `no red dress` in the negative prompt is less effective than saying `blue dress` in the positive prompt. Use negatives to remove concepts, not to flip toggles.