negative prompting

Negative prompting specifies what NOT to generate, helping avoid unwanted elements in image generation. **How it works**: Classifier-free guidance steers away from negative concepts. During generation, model moves toward positive prompt AND away from negative prompt. **Common negatives**: "blurry, low quality, bad anatomy, extra limbs, watermark, text, ugly, deformed, disfigured, out of frame, cropping". **Use cases**: Fix recurring issues (bad hands, extra fingers), avoid styles, remove artifacts, improve quality. **Implementation**: Negative embeddings computed, combined with unconditional and positive during CFG sampling. **Per-model negatives**: Different models have different failure modes, community-developed negative prompts per checkpoint. **Negative embeddings**: Textual inversions trained on bad outputs, e.g. "EasyNegative", "bad-hands-5". **Best practices**: Start with standard quality negatives, add specific negatives for observed problems, avoid over-negating (can distort output). **Tools**: All major diffusion UIs support negative prompts, AUTOMATIC1111, ComfyUI, InvokeAI. Essential technique for quality control.