Negative prompt generator by problem type

The six categories and what they contain

Instead of one giant list copied from a forum, the tool groups terms by the problem they fix:

• Low quality and artifacts: lowres, worst quality, jpeg artifacts, pixelated, blurry, noisy.
• Bad anatomy: extra or fused fingers, malformed hands, extra limbs, distorted face, asymmetrical eyes.
• Text, logos and watermarks: text, watermark, signature, logo, caption, username, letters.
• Cropping and framing: cropped, out of frame, cut off, duplicate, cluttered background.
• Unwanted styles: cartoon, 3d render, anime, illustration, painting, cgi — useful when you want a real photo.
• Photo defects: overexposed, underexposed, grainy, motion blur, chromatic aberration, oversaturated.

Duplicates are removed automatically, and your own terms always come first so they carry the most weight.

How each model handles exclusions

Negative prompts are not a universal feature, which is why the same list is rewritten per model. Here is one result with the low quality and text categories ticked and “people” added as a custom term:

Stable Diffusion / Leonardo (negative box):
people, lowres, low quality, worst quality, jpeg artifacts, pixelated, blurry, out of focus, noisy, text, watermark, signature, logo, caption, username, letters

Midjourney (end of the prompt):
--no people, jpeg artifacts, pixelated, blurry, out of focus, noisy, text, watermark

Flux (end of the prompt):
The image is clean and sharp, without people, lowres, low quality, worst quality, jpeg artifacts, pixelated, blurry, out of focus, noisy, text.

For Midjourney, the builder drops terms starting with “bad”, “worst”, “low” or “deformed” and keeps at most eight words, because --no works on concrete nouns and long lists dilute it. For Flux and other natural-language models, the first ten terms become one sentence you append to your prompt. Stable Diffusion and Leonardo receive the full list for their dedicated field.

When a negative prompt hurts more than it helps

Bigger is not better. Pasting a hundred terms into every generation can make images flat, desaturated or oddly composed, because each term pulls the image away from something. Watch for these mistakes:

• Contradictions: excluding “illustration” while asking for a watercolor painting.
• Excluding your own subject: “hands” in the negative box when the image shows someone holding a cup.
• Using negatives to fix a vague prompt: if the image is generic, improve the positive prompt first.
• Writing “no” in the main prompt: most models see the noun and draw it anyway.

A good routine is to generate without a negative prompt, note the actual problems, and tick only those categories.

Before and after for a portrait

A portrait prompt kept producing a watermark-like scribble and six fingers. Ticking only “Bad anatomy” and “Text, logos and watermarks” for Stable Diffusion gave a short, targeted list instead of a wall of words, and the colors stayed vivid. To build the positive side, use the Stable Diffusion prompt generator, which already adds a default negative block, or the Midjourney prompt generator for the --no version inside a full prompt. Video models such as Kling and Pika also accept negative prompts; the video prompt generator has a field for them.