A negative prompt lists what you don't want in the picture: low quality, bad hands, a watermark, or a colour or object that keeps appearing. Each model reads it differently. Illustrious and other SDXL models do best with a list of quality and anatomy tags, SD 3.5 with a few words, and some newer models ignore it unless you raise the Guidance Scale. This page gives a negative prompt to start from for each kind of model in MonAI, taken from MonAI's own styles.
In MonAI the negative prompt has its own field, with suggestions you can tap to add. Choosing a style fills in that style's negative prompt for you.
¶ Which models use a negative prompt?
| Model | Does the negative prompt work? | Start with |
|---|---|---|
| Illustrious models: Nova Anime, WAI-illustrious, Illustrious XL | Yes | The model's MonAI style (below) |
| Animagine XL 4 | Yes | The Animagine 4 style (below) |
| Other SDXL models: Firenze XL v2.1, RealVisXL v5, Juggernaut XL, epiCRealism XL, CyberRealistic | Yes | A short list of weighted quality words (below) |
| SD 3.5 Large and Medium | Yes | A few words, or none (below) |
| SD 1.5 models: Anything V5, DreamShaper 8, Deliberate | Yes, and negative embeddings work too | Weighted quality words plus an embedding (below) |
| RedCraft Z-Image, Qwen Lightning, Qwen Image Edit | Only when the Guidance Scale is above 1 | Leave it empty (below) |
| Flux Schnell | No | Leave it empty (below) |
¶ Fewer negative prompts, better pictures
A shorter negative prompt often gives a better picture. Every word you add pulls the picture away from something, not only from the problem you meant to fix. Some artists paste a long negative prompt into every picture by default, but this compromises image quality.
- Start with the style. Choose the model's MonAI style. It fills in MonAI's recommended negative prompt for that model.
- Look at the picture. If nothing is wrong, leave the negative prompt alone.
- Add only what fixes this picture. If you see a specific problem, such as extra fingers, a watermark or a colour that keeps appearing, add a word for that problem.
If you weight a word, keep it light. Most of MonAI's styles weight one group at 1.2 to 1.3, as in (worst quality, low quality:1.3).
¶ Illustrious negative prompt
Illustrious models learned the Danbooru quality tags, so their negative prompts are built from tags such as worst quality, bad quality, lowres and bad hands. Each model's MonAI style uses its own:
| Model | Negative prompt of its MonAI style |
|---|---|
| Nova Anime v18 (Nova Anime v18 style) | photorealistic, realistic, 3d, glitch, deformed, ugly, disfigured, lowres, bad anatomy, bad hands, fewer digits, extra digit, missing fingers, displeasing, (bad quality, worst quality:1.2), signature, watermark, bad AI generated |
| WAI-illustrious v17 (WAI Anime v17 style) | bad quality, worst quality, worst detail, sketch |
| Illustrious XL (Illustrious style) | lowres, (low quality, worst quality:1.3), bad anatomy, displeasing, ugly, fewer digits, extra digit, missing fingers, bad hands, blurry |
Nova Anime's list starts with photorealistic, realistic and 3d to keep pictures drawn rather than photographic. WAI-illustrious needs only four tags. Add nsfw to steer away from mature content, and watermark and signature if text appears. The Illustrious Prompt Guide covers the quality tags that go in the prompt itself.
¶ Animagine XL 4 negative prompt
Animagine XL 4 uses its own score tags as well as quality tags. The Animagine 4 style uses:
lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry
¶ SDXL negative prompt (realistic and general models)
For SDXL, start with Firenze XL v2.1 and RealVisXL v5, MonAI's newest and best SDXL models. Choose their MonAI styles to apply the matching negative prompts.
¶ Firenze XL v2.1 negative prompt
The Firenze v2 Photo style uses:
(deformed, distorted, disfigured), ugly, mutation, blurry, (worst quality, low quality, bad quality:1.2)
¶ RealVisXL v5 negative prompt
The RealVis5 Photo style uses:
(deformed, distorted, disfigured:1.3), ugly, blurry, glitch, lowres, mutation, (worst quality, low quality, bad quality:1.3)
Other photo styles use variations of the same list:
- Juggernaut 8 Photo (Juggernaut XL v8):
(deformed, distorted, disfigured:1.3), ugly, blurry, (worst quality, low quality, bad quality:1.3) - Epic Photo (epiCRealism XL):
(deformed, distorted, disfigured:1.2), ugly, blurry, mutation, worst quality, bad quality - Cyber v12 Photo (CyberRealistic CyberIllustrious v12):
lowres, worst quality, bad quality, bad anatomy, bad hands, extra fingers, ugly, deformed
CyberRealistic is based on Illustrious XL, so its list mixes the Illustrious tags with the photo words. To keep a picture from looking like a painting or a cartoon, add those words to the negative prompt; to keep a drawing from looking like a photo, add photo. Prompt Weighting explains the bracket syntax.
¶ SD 3.5 negative prompt
SD 3.5 needs far less. MonAI's SD 3.5 styles use two to four words, such as ugly, deformed or deformed, ugly, photo, photography, and many examples use none at all. In the SD 3.5 prompting tests, removing a long negative prompt made the text more accurate and fixed the eyes. Start empty, then add a word only for a problem you can see.
¶ SD 1.5 negative prompt and negative embeddings
SD 1.5 models such as Anything V5, DreamShaper 8 and Deliberate can also use negative embeddings such as easynegative, BadDream and ng_deepnegative_v1_75t. MonAI's styles pair one with a weighted quality phrase:
- DreamShaper 8:
(deformed, distorted, disfigured:1.3), BadDream - Anything 5 Anime:
(deformed, distorted, disfigured:1.3), Unspeakable-Horrors-32v
Negative embeddings only work with SD 1.5 models in MonAI. With Illustrious, SDXL and the newer models, a name such as easynegative is read as ordinary words, so leave it out.
¶ Qwen Image Edit, Qwen Lightning and Z-Image negative prompt
Qwen Image Edit, Qwen Lightning and RedCraft Z-Image start at a Guidance Scale of 1. At exactly 1 they skip the negative prompt, so anything in that field has no effect. Above 1 it is used.
To use a negative prompt with these models, raise the Guidance Scale. Use it for things you don't want added, not to undo something your prompt asks for.
None of MonAI's 31 Qwen Lightning and Z-Image styles uses a negative prompt, and the wiki's Qwen Image Edit examples leave it empty too. These models are prompted in plain sentences, so say what you want in the prompt instead: "keep her face and hair the same" or "a plain white background".
¶ Flux Schnell negative prompt
Flux Schnell doesn't use a negative prompt. Its Guidance Scale is fixed at 1, so whatever is in the field is ignored. Describe what you want in the prompt.
¶ Questions
¶ How long should a negative prompt be?
Usually shorter than you think. Start with the negative prompt of the model's MonAI style, and add to it only to fix something specific in a picture. A long negative prompt used by default can lower image quality.
¶ What is a good negative prompt for Illustrious?
Use the negative prompt of the model's MonAI style. For WAI-illustrious v17 that is bad quality, worst quality, worst detail, sketch. For Illustrious XL it is lowres, (low quality, worst quality:1.3), bad anatomy, displeasing, ugly, fewer digits, extra digit, missing fingers, bad hands, blurry. Nova Anime v18 uses a longer list that starts with photorealistic, realistic, 3d.
¶ Does Qwen Image Edit use a negative prompt?
Only when the Guidance Scale is above 1. At 1, where it starts in MonAI, the negative prompt is skipped. Most Qwen Image Edit examples leave it empty and describe the change in the prompt instead.
¶ Does EasyNegative work with Illustrious or SDXL?
No. In MonAI, negative embeddings such as easynegative only work with SD 1.5 models. With Illustrious and other SDXL models the name is read as ordinary words.
¶ What negative prompt should I use for SDXL?
Start with MonAI's Firenze v2 Photo style for Firenze XL v2.1 or RealVis5 Photo for RealVisXL v5. Each style applies its own negative prompt. Add words only for problems you can see.
¶ Does Flux Schnell use a negative prompt?
No. Its Guidance Scale is fixed at 1, so the negative prompt is ignored. Describe what you want in the prompt instead.
Source: the MonAI wiki · Updated · First published