Negative prompts tell the model what NOT to generate. One good negative list kills half your retries — and retries are where your credits actually go. Organized by the problem you're trying to fix.
If you only copy one thing, copy this. It covers the most common failure modes on every model:
blurry, low quality, distorted, deformed, warped, watermark, text, subtitles, logo, extra limbs, morphing, flickering
deformed hands, extra fingers, fused fingers, malformed limbs, extra limbs, disfigured face, asymmetric eyes, morphing features
face morphing, identity shift, changing facial features, inconsistent face
flickering, jittery motion, stuttering, frame skipping, unstable motion, teleporting objects
morphing objects, melting geometry, warping background, shifting proportions
watermark, text overlay, subtitles, captions, logo, timestamp, UI elements, borders
blurry, out of focus, low resolution, pixelated, compression artifacts, noise, grainy, overexposed, underexposed
oversaturated, neon colors, plastic skin, uncanny valley, cartoon, 3D render look
The same negative list behaves very differently depending on where you paste it, and knowing which model you are talking to saves a lot of wasted generations.
The universal list at the top of this page is a safe default, but the highest hit rates come from a list built for the shot in front of you. The process takes two generations, not ten.
Generate once with your positive prompt and no negatives at all. Look at exactly what went wrong — not in general, but specifically. Warped fingers on the left hand? Add the anatomy block. The background architecture melting in the last second? Add the melt block. A logo hallucinated onto a t-shirt? Add overlays. Then regenerate with only those blocks. Two or three targeted groups almost always beat the full forty-term list, because every term you add dilutes the ones that matter.
Keep a small personal library of these shot-type lists. Close-up portraits, wide landscape shots, product spins and text-heavy scenes each develop their own characteristic failures, and after a few sessions you will know which block to reach for before the first generation finishes.
Negatives are a scalpel, not a shield, and there are three common ways they backfire. Over-long lists dilute each other until none has any effect, which is why a 40-term paste often performs worse than a 10-term one. Style-word negatives can suppress the look you actually wanted — "cartoon" in the negative field flattens an intentionally illustrated scene, and "dark" can wash out a moody night shot. And negatives cannot rescue a vague positive prompt: if the model does not know what the scene is, telling it what the scene is not only narrows the guesswork slightly.
The practical rule is to fix the positive prompt first, add negatives for the specific artifact you are seeing, and remove any term the moment it stops earning its place. Every artifact prevented is a retry you did not pay for, which is the entire point.
Not as a separate field — OpenAI and Google removed explicit negative inputs. Instead, fold the intent into the positive prompt: "clean composition, no on-screen text, natural skin tones". Kling, Runway, Wan and most open models still accept a dedicated negative list.
Too many terms. Long negative lists dilute each other and can suppress things you wanted (e.g. "cartoon" in negatives can flatten stylized looks). Start with the universal list, then add only the terms for artifacts you're actually seeing.