Stable Diffusion Negative Prompts: What They Do and What to Use

How negative prompts work

Stable Diffusion builds an image over many small steps. At each step it compares where your prompt pulls the image with an unconditioned baseline, and pushes further in the prompt's direction. A negative prompt takes the place of that baseline, so the model is actively pushed away from whatever you list.

That's why a negative prompt can clean up defects no amount of positive wording fixes, like watermarks or mangled hands. It's also why a long, unfocused negative prompt can flatten an image: every term pulls on the result.

A starter negative prompt

For most images: lowres, blurry, jpeg artifacts, watermark, signature, text, cropped, out of frame, worst quality, low quality.

For people, add: bad anatomy, extra fingers, missing fingers, deformed hands, extra limbs, disfigured face, cross-eyed.

For photorealism, add: cartoon, illustration, 3d render, painting, oversaturated. For illustration, do the opposite and add photo, photorealistic.

Tips that actually help

Target the problem you see. If hands come out wrong, add hand terms. If there's text in the corner, add text and watermark. Adding everything at once makes it hard to tell what works.

Use weights for stubborn defects. In AUTOMATIC1111, Forge and ComfyUI, (extra fingers:1.3) pushes harder than the plain phrase. Stay under about 1.5, or the image starts to degrade.

Don't contradict your prompt. If the prompt says 'night city' and the negative prompt says 'dark', the two fight each other.

Negative embeddings, small files like EasyNegative, bundle a whole list into one word. They're handy but tied to a model family (SD 1.5 or SDXL), so check they match your checkpoint.

Which models use negative prompts

Stable Diffusion 1.5, SDXL and SD 3.5 all support negative prompts, and so do most community checkpoints built on them.

Some newer models, such as Flux, are trained without that baseline comparison and ignore negative prompts in their standard setup. With those, describe what you want instead: 'clean background' rather than a negative 'cluttered'.

Get both prompts from an image

1. 1Open savelark.com/image-to-prompt and upload the image you want to match.
2. 2Choose the Stable Diffusion style and tap Generate prompt.
3. 3You get comma-separated tags ordered by importance, then a line starting with Negative prompt: listing defects to avoid.
4. 4Paste the tags into the prompt box and the rest into the negative prompt box of your Stable Diffusion interface.