SDXL & Stable Diffusion Negative Prompts: Copy-Paste Lists (2026)

A negative prompt lists what Stable Diffusion should steer away from. Paste it into the Negative prompt field (or the negative_prompt API parameter), not into the main prompt.

Every block below has a copy button. Start with the short list for your model, generate, then add terms only for problems you actually see. To combine lists in a few clicks, use the negative prompt generator.

Quick answer: a starter SDXL negative prompt

lowres, blurry, jpeg artifacts, watermark, signature, text, deformed, bad anatomy, extra fingers, cropped, out of frame

SDXL needs far fewer quality tags than SD 1.5 did. If an image already looks clean, an empty negative field is a valid choice.

Essential Negative Prompt Templates

SD 1.5 general list

SD 1.5 checkpoints are more prone to anatomy and quality problems, so longer lists are common:

lowres, worst quality, low quality, normal quality, jpeg artifacts, blurry, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, extra limbs, deformed, disfigured, cropped, out of frame, watermark, signature, text, username

Some SD 1.5 users also add negative embeddings such as EasyNegative. Those only work if the embedding file is installed in your interface; typed as plain text they do nothing.

Realistic photos

For photoreal SDXL or SD 1.5 checkpoints, exclude other media first, then the "AI look":

cartoon, anime, illustration, painting, drawing, 3d render, cgi, plastic skin, airbrushed, oversaturated, lowres, blurry, jpeg artifacts, watermark, text

Do not exclude blurry if you asked for shallow depth of field or motion blur. The two instructions will fight.

Anime and illustration

For anime-style checkpoints trained on Danbooru-style tags:

worst quality, low quality, lowres, bad anatomy, bad hands, extra digits, fewer digits, jpeg artifacts, signature, watermark, username, blurry, photorealistic, 3d

Illustrious XL

The Illustrious XL model card gives this example negative prompt:

worst quality, comic, multiple views, bad quality, low quality, lowres, displeasing, very displeasing, bad anatomy, bad hands, scan artifacts, monochrome, greyscale, signature, twitter username, jpeg artifacts, 2koma, 4koma, guro, extra digits, fewer digits

The same card suggests Euler a, 20–28 steps and CFG 5–7.5. Remove monochrome, greyscale when you want a black-and-white image.

Pony Diffusion V6 XL

Pony works differently. Its model card says it is designed to not need negative prompts in most cases. The quality control goes at the start of the positive prompt:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,

If you still want a negative field, a common community habit is to put the low score tags there. The model card does not recommend this, so treat it as optional:

score_4, score_5, score_6

Text, watermarks and logos

text, watermark, signature, logo, caption, username, artist name, letters, words

Skip this list when the image should contain text, such as a poster, sign or product label.

SFW (safe-for-work) list

nsfw, nude, naked, explicit, suggestive, cleavage, lingerie, underwear

This lowers the chance of unwanted content, but it is not a safety filter. If a workflow must never show explicit images, use the safety checker or moderation step your platform provides as well.

Category-Specific Negative Prompts

Portraits and hands

asymmetrical eyes, cross-eyed, deformed iris, bad teeth, extra fingers, fused fingers, malformed hands, bad anatomy, double face, plastic skin, airbrushed

Hand terms reduce mistakes but do not reliably fix them. If hands are still wrong, change the pose, crop closer, try a new seed, or use inpainting on the hand.

Landscapes and architecture

people, person, crowd, cars, power lines, signs, text, tilted horizon, distorted perspective, warped lines, oversaturated, lowres

Excluding people and cars empties the scene. Leave them out of the list if you want a sense of scale.

Product photography

cluttered background, dust, scratches, fingerprints, harsh reflections, distorted label, misspelled text, watermark, lowres, blurry

Logos and flat graphics

photograph, photorealistic, 3d render, gradient, shadow, texture, noise, cluttered, extra text, misspelled text, watermark

Which models support negative prompts?

Two details from the Midjourney docs: --no is the same as giving that word a −0.5 weight, and each word after --no is read separately, so --no modern clothing excludes "modern" and "clothing". For Leonardo, see the Leonardo negative prompt guide.

FLUX.1 [dev] is guidance-distilled. In Diffusers, negative_prompt is ignored until you raise true_cfg_scale above 1, which also makes each image slower. For most FLUX work, write the positive prompt more precisely instead ("clean white background" rather than a negative "clutter").

Making a term stronger

AUTOMATIC1111 and compatible interfaces support attention weights in both fields:

(watermark:1.3), (extra fingers:1.2), blurry

Start without weights. Add a modest weight (1.1–1.3) only to a term that still appears after a few seeds. Very high weights can distort the whole image. This syntax does not work in APIs such as Stability's, which take plain text.

How to tell if a term helps

1. Fix the seed, sampler, steps, CFG and size.
2. Generate once with your current negative prompt.
3. Add or remove one or two terms.
4. Generate again with the same seed and compare.

A long copied list changes many things at once. Testing in small steps shows which terms actually matter for your checkpoint.

Common mistakes

• Contradicting the prompt. Asking for a foggy night street while excluding dark, fog cancels your own request.
• Using one list everywhere. Danbooru tags such as 2koma mean nothing to a photoreal checkpoint, and SD 1.5 lists are overkill on SDXL.
• Writing "no" in the main prompt. "A beach with no people" can add people, because the model still sees the word "people". Move it to the negative field.
• Expecting a fix for everything. If a problem survives a good negative prompt, change the seed, pose, resolution or model.

Build your prompt

The negative prompt generator combines the lists on this page for your model without duplicates. The prompt generator keeps the positive prompt and the negative field together so you can copy both at once. It runs in your browser and does not generate images.

Primary sources

• Stability AI Developer Platform: Stable Image API
• AUTOMATIC1111 Web UI: negative prompt
• AUTOMATIC1111 Web UI features (attention syntax)
• Hugging Face Diffusers: Stable Diffusion 3 pipeline
• Hugging Face Diffusers: FLUX pipeline
• Pony Diffusion V6 XL model card
• Illustrious XL model card
• Midjourney docs: No parameter