Negative Prompt Generator for Stable Diffusion
Choose your model and the problems you actually see, then copy a focused negative prompt. Anatomy terms are skipped for landscapes, products and architecture.
FreeNo signupRuns in your browser
Optional, comma separated. Add a term after you see the defect.
Working on the words around your images? Browse free, copy-ready text prompts for marketing and SEO, sales, careers and coding.
Browse free copy-ready prompts →Match negative terms to visible problems
A useful negative prompt starts with a defect you can point to. If the image has a watermark, try watermark. If the composition is already correct, leave cropping terms off. This makes your next comparison easier to interpret and keeps unrelated instructions out of the negative field.
| Problem | Starting terms | When to try them |
|---|---|---|
| Hands and anatomy | bad anatomy, bad hands | Human figures with visible anatomical defects |
| Blurry or low detail | blurry, lowres | Unwanted softness in the main subject |
| Text and watermarks | text, watermark | Unwanted writing or a mark in the image |
| Duplicate subjects | duplicate subjects | An extra copy of the intended subject |
| Cut-off framing | cropped | A subject cut off by the frame |
| Color and compression defects | oversaturated, jpeg artifacts | Excessive color or visible compression-like blocks |
| Photographic style guard | cartoon, 3d render | A photograph drifting toward an illustrated look |
| Illustration style guard | photorealistic, photograph | An illustration drifting toward a photographic look |
These terms are starting points, not repair commands. Describe the intended pose and framing positively as well. Use the image prompt generator to keep the subject, action and composition explicit before you tune exclusions.
Negative prompts by model family
| Family | Uses negatives? | Starting point |
|---|---|---|
| SD 1.5 | Yes | low quality, worst quality |
| SDXL | Yes | low quality, worst quality, then only relevant defects |
| Pony Diffusion XL | Yes | score_6, score_5, score_4 |
| Illustrious / NoobAI | Yes | worst quality, low quality, lowres |
| SD 3.5 | Yes | A short list of observed defects; no preset here |
| FLUX.1 dev / schnell | No effect at CFG 1 | Describe what you want in the positive prompt |
FLUX.1 dev and schnell are distilled models that normally run at CFG 1. Selecting Flux therefore leaves the preset output empty. Extra terms still copy through if you enter them, but that does not make them effective at CFG 1. The guide to negative prompts in Stable Diffusion provides further context for exclusions in models that use them.
How to use the negative prompt generator
- Select the actual model family. Its base terms appear first, ahead of your selected problems.
- Choose the image type. Anatomy terms are skipped for landscapes, products and architecture, even if that checkbox remains selected.
- Uncheck problems you have not observed. Start with Short and add your own comma-separated terms only when needed.
- Copy the result into the negative prompt field of your image app. Keep the positive prompt and generation settings steady while comparing changes.
Terms keep a stable order: family, selected problems, then extras. Repeats are removed case-insensitively, so adding Watermark does not duplicate an existing watermark term. If the result needs a different weighting syntax for your interface, the prompt weight converter can translate an explicitly weighted prompt.
Keep the list short until a defect appears
On SDXL, start with a short, targeted list instead of the very long stock lists that circulated for SD 1.5. Every extra term is one more thing you cannot evaluate. A landscape does not benefit from a stock list of hand defects. A deliberately soft painting does not need every softness-related term selected just because the preset offers it.
Use Thorough when you want to try more descriptions of an observed problem, then remove terms that do not help. Extra terms are never truncated to meet a length target. Save a useful comparison by recording both prompts and the settings, so you know which change produced the result you preferred.
Frequently asked questions
What is a negative prompt in Stable Diffusion?
A negative prompt describes things you want the model to avoid. SD 1.5, SDXL and SD 3.5 support negative prompts. Put the desired scene in the positive prompt and unwanted defects in the negative field.
What is a good SDXL negative prompt?
Start with a short list that matches defects you can see, such as low quality, blurry or watermark. Add a term only when a defect actually shows up. Remove terms that do not help your particular image.
Do negative prompts work with Flux?
FLUX.1 dev and schnell normally run at CFG 1, where negative prompts have no effect. Describe the desired image positively instead. The Flux option only copies your extra terms so you can keep your own notes without adding an irrelevant preset.
Which negative tags should I use for Pony?
Start with score_6, score_5, score_4 in the negative prompt. Keep score_9, score_8_up, score_7_up at the beginning of the positive prompt. The Pony option adds the lower score tags before any selected problem terms.
Can a negative prompt fix bad hands?
You can try targeted terms such as bad hands and bad anatomy, but they are not a guaranteed repair. Compare the result before adding more terms. For landscapes, products and architecture, this tool leaves anatomy terms out because they add noise.
How do I add my own negative prompt terms?
Enter them in Extra terms separated by commas. They appear after the model and problem terms. Repeated terms are removed without changing the order of the first occurrence.
