A negative prompt tells the AI what to avoid. In Stable Diffusion, Flux, and DALL·E, this is as important as the positive prompt — sometimes more so.
The Universal Base Negative Prompt
Start every generation with this, then customise for your specific use case:
text
blurry, out of focus, low quality, pixelated, watermark, signature, text, logo, compression artifacts, noise, grain, overexposed, underexposed, bad composition, amateurFor Portrait Photography
text
deformed face, bad anatomy, extra limbs, missing fingers, fused fingers, poorly drawn hands, ugly, disfigured, cross-eyed, asymmetric eyes, bad teeth, double chin (unless intentional)💡
In Flux, negative prompts have less influence than in SD. Compensate by being more specific in your positive prompt instead.
For Architecture & Interiors
text
unrealistic proportions, impossible geometry, floating objects, wrong perspective, warped lines, bad lighting, cluttered, messyThe Weight System (Stable Diffusion)
In SD-based models, you can weight negative terms to increase their influence. Syntax: (term:1.4) means 40% stronger emphasis.
text
(blurry:1.3), (bad anatomy:1.4), (extra fingers:1.5), watermarkWhat Not to Put in Negative Prompts
- Don't negate the subject itself ("no people" in a portrait prompt confuses the model)
- Don't overload with 50+ terms — models have attention limits
- Don't use vague antonyms like "not beautiful" — be specific about what you don't want
