Negative Prompts: What They Are, Which Models Use Them, and the Lists That Actually Work

What is a negative prompt?

A negative prompt is a second text field that some image models read alongside the main prompt. Where the main prompt pulls the image toward certain concepts, the negative prompt pushes it away from others. Technically the model computes how the image would change if it moved toward the negative text, and steers in the opposite direction at every denoising step.

That is why negative prompts work best for things a model drifts toward on its own: extra fingers, blurry backgrounds, watermarks, the generic glossy look. They are weak for things the model was never going to draw anyway. Writing “no elephants” into a landscape negative prompt does nothing useful; writing “oversaturated, HDR, plastic skin” into a portrait negative prompt changes the result.

Which models support negative prompts

Stable Diffusion 1.5, SDXL and most community checkpoints on Automatic1111, ComfyUI, Forge and Fooocus have a dedicated negative prompt field. This is where negative prompts came from and where they matter most.

Midjourney has no negative field. It uses the --no parameter: --no text, watermark, blur. Midjourney does not understand prose negation, so “no text” written inside the prompt is read as a request for text.

Flux, ChatGPT Images (GPT Image), Gemini and Nano Banana do not use negative prompts. Flux’s T5 text encoder reads the whole prompt as a description, so negatives inside it are ignored or inverted. ChatGPT and Gemini follow natural-language constraints instead: “keep the background plain, no visible text” works because they are instruction-following models, not because they have a negative channel.

• Stable Diffusion, SDXL, community checkpoints: dedicated negative prompt field, high impact.
• Midjourney: --no parameter, comma-separated, moderate impact.
• Flux: no negative prompt; describe what you want instead.
• ChatGPT Images, Gemini, Nano Banana: no negative prompt; use short positive constraints.

How to write a negative prompt that helps

Keep it short and specific to the failure you are seeing. A 60-term negative prompt copied from a forum is mostly noise: every term takes a small share of the guidance budget, and terms that do not apply to your image pull the result in random directions.

Start empty. Generate once. Look at what went wrong, add two or three terms for exactly that, and generate again. Most images need fewer than ten negative terms. Put the most important ones first, because CLIP reads the text in 77-token chunks and weights the front of each.

Use weights sparingly. (deformed hands:1.3) in SDXL raises the pressure on one term without inflating the whole list. Weighting everything to 1.4 is the same as weighting nothing.

Negative prompt lists by use

These lists are starting points for Stable Diffusion and SDXL. Take the block that matches your problem, not all of them. For Midjourney, move the same words into --no.

Photorealistic portraits: deformed hands, extra fingers, fused fingers, extra limbs, asymmetrical eyes, plastic skin, waxy skin, oversharpened, HDR, oversaturated, airbrushed, doll-like, uncanny.

Hands and anatomy: extra fingers, missing fingers, fused fingers, six fingers, bad hands, malformed limbs, extra arms, disconnected limbs, mutated anatomy.

Text, logos and artifacts: text, watermark, signature, logo, caption, letters, jpeg artifacts, compression noise, frame, border, split screen, collage.

Anime and illustration: bad anatomy, extra digits, lowres, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry, cropped, realistic skin texture (when a flat cel look is wanted).

Product and interior renders: warped perspective, floating objects, duplicated objects, uneven reflections, clutter, text, watermark, fisheye, over-exposed highlights.

Style drift toward the default look: digital painting, 3D render, smooth, glossy, cinematic lighting, dramatic, bokeh (when the reference is flat, documentary or hand-made).

• Photorealistic portrait: plastic skin, oversaturated, HDR, deformed hands, extra fingers, waxy, airbrushed
• Hands: extra fingers, fused fingers, missing fingers, bad hands, malformed limbs
• Text and artifacts: text, watermark, signature, logo, jpeg artifacts, border, collage
• Anime: bad anatomy, extra digits, lowres, worst quality, signature, blurry, cropped
• Product render: warped perspective, floating objects, duplicated objects, clutter, fisheye

SDXL negative prompts are different from SD 1.5

SD 1.5 checkpoints were trained on captions full of “lowres, bad anatomy, worst quality” tags, so those quality tags have real effect there. SDXL was trained differently: the classic quality-tag block does much less, and a short, concrete negative prompt performs better. In SDXL, describe the specific defect and the specific unwanted style rather than stacking quality words.

Many SDXL fine-tunes, including Pony and Illustrious derivatives, come back with their own recommended negative block in the model card. Use that as the base and add your two or three specific terms on top.

When the answer is not a negative prompt

If the model keeps producing a polished, generic version of your reference, the fix is usually in the positive prompt: name the medium and the imperfections you want kept, and lead with them. Image to Prompt does this automatically when you choose Stable Diffusion as the target model: it writes comma-separated tags with the style first and emits a matching negative prompt line built from the specific default-aesthetic drift that style needs to avoid.

For ChatGPT Images and Gemini, replace every negative with a positive constraint. “No text” becomes “leave the sign blank.” “Not blurry” becomes “sharp focus across the whole frame.” The models follow the second form far more reliably.