What a negative prompt actually does
A negative prompt tells the model what to steer away from — it’s a second conditioning input, not magic. The model pushes the image away from those concepts, so short and specific beats long and generic.
The short list that earns its place
- worst quality, low quality — a broad quality steer that genuinely helps on SD 1.5 and SDXL.
- blurry, out of focus — when you want sharpness.
- extra fingers, extra limbs, deformed hands — the classic anatomy fixes.
- watermark, signature, text — when stray marks or text show up.
- jpeg artifacts — for a clean, uncompressed look.
What’s usually cargo-culted
Those 200-token mega-negatives copied from a model page rarely help, and can suppress detail. Every token you add dilutes the rest. If you can’t explain why a token is there, cut it and compare — you’ll usually keep the cut.
It depends on your model
- SD 1.5 / SDXL: negatives help; keep them focused.
- Pony: put low score tags (score_4, score_5, score_6) in the negative to push quality up.
- Illustrious: booru-tag negatives like bad anatomy, jpeg artifacts.
- Flux: you usually don’t need a negative prompt at all — the base pipeline ignores it.
A reusable baseline
Save your baseline once and stamp it into every build, then add model-specific tweaks per project instead of retyping. In Prompt Builder a negative block is just a button you press.