How it works
At each denoising step the model computes two predictions: one guided by your prompt, one guided by the negative prompt. The final direction is the first minus the second, scaled by the CFG value. So the negative prompt is not a filter applied at the end: it shapes the image from the first step. That is also why a negative prompt without a CFG above 1 has no effect.
Negative prompts that work
General quality (SD 1.5, SDXL):
blurry, low quality, lowres, jpeg artifacts, watermark, text, signature, cropped, out of frame
People and portraits:
deformed, bad anatomy, extra fingers, missing fingers, extra limbs, mutated hands, poorly drawn face, asymmetric eyes
Photography:
painting, drawing, illustration, cartoon, 3d render, cgi, oversaturated, overexposed
Anime and illustration:
photorealistic, photo, realistic skin, worst quality, normal quality, monochrome
Start with the general list and add the block that matches your goal. Ten to twenty terms are plenty; long copied lists mostly add noise.
Weights in the negative prompt
Weights work exactly as in the main prompt: (blurry:1.4) pushes harder away from blur, (text:1.2) removes stray letters. Raise a weight only for a defect that keeps coming back; above 1.5 the model starts distorting other things to obey.
When it does nothing
- CFG scale at 1: the negative prompt is ignored by construction. Use 5 to 8.
- Turbo and LCM models (SDXL Turbo, LCM LoRA) run at CFG 1 to 2 and barely react to negatives.
- Stable Diffusion 3 and 3.5 use the negative prompt weakly; write the positive prompt more precisely instead.
- Concepts the model cannot separate: "no hands" in the negative removes hands from view, it does not fix hand anatomy. Inpainting does.
On this site
The inpainting tool already applies a strong negative prompt behind the scenes, so you only describe what you want. The prompt examples show the negative prompt used for each image next to the positive one; copy both to reproduce a result, then change one term at a time with a fixed seed.