Negative Prompts for AI Video: What Actually Works

I spent about two weeks convinced negative prompts were broken. I'd type no extra fingers, no warped hands, no text into every generation, and the clips came back with warped hands and text anyway. Then I checked which models I was actually running, and the answer was embarrassing: most of them have no negative prompt field at all. I'd been typing wishes into a positive prompt.

That's the thing nobody says plainly. "Negative prompt" is not one feature. On some models it's a real, separate input the sampler uses. On others it's a phrase you tack onto the end of your prompt and hope. On the rest it does literally nothing. Knowing which of the three you're in front of is most of the skill.

What a negative prompt actually is

In an image or video diffusion model, a true negative prompt is a second conditioning vector. The model steers toward your positive prompt and away from the negative one on every denoising step. It's not a filter applied afterwards, and it's not a rule — it's a direction in the same space, pulling the other way.

That mechanism explains the single most useful fact about them: a negative prompt can only push away from something the model already understands as a concept. "Blurry" works because blur is a strong, well-represented visual concept. "Six fingers" barely works because the model has no clean internal concept of finger-counting — it has a concept of "hand", and hands come out however they come out. This is why my two weeks of no extra fingers did nothing even on models that do support the field.

Which models actually have the field

This is the part worth checking before you write a single word of negative prompt. As of September 2026:

If you're on an open-source local setup, you have the real thing. Everywhere else you're mostly writing English at a model that reads it as one blob. I go into the local path in more detail in my guide to the open-source AI video generators.

The five terms that earn their place

When I do have a real negative field, I've stopped writing paragraphs. Five terms, reused everywhere, and they're all about rendering quality rather than content:

• blurry, out of focus — the highest-value term by a wide margin. Cheap generations default soft.
• low resolution, compression artifacts — cleans up block noise in fast-moving areas.
• oversaturated, overexposed — kills the plasticky glow that makes a clip read as AI.
• watermark, text, caption, subtitle — this one genuinely works, because text is a strong concept. Worth keeping even though most models now suppress it by default.
• static, frozen frame — for image-to-video, where the model sometimes just barely moves the picture.

Everything else I've tried — anatomy terms, "no CGI", "not cartoon", naming things that shouldn't be in the scene — ranged from neutral to actively harmful. The last category is the dangerous one, and it deserves its own section.

Why naming a thing can summon it

This is the failure I still make. Put no dog in a negative prompt and you'll sometimes get a dog. Not always, but often enough that I stopped doing it.

The reason is that conditioning isn't a clean subtraction. The token for "dog" is in the computation either way, and how strongly the model steers away depends on the guidance scale, the seed and where in the sampling it lands. At low guidance the negative barely bites, and you've just handed the model a dog-shaped hint.

So the rule I follow now: negative prompts are for rendering defects, positive prompts are for content. If I don't want a dog in the shot, I describe what is in the shot, completely enough that there's no room left. "An empty gravel courtyard, nothing else in frame" beats "no dog" every time.

Same logic applies to the hands problem, which I wrote up separately in why AI videos look fake — the fix is almost never a negative term.

Length: shorter wins, and I can show why

I ran the same Wan 2.2 prompt three ways on a fixed seed — no negative, my five-term list, and a 40-term kitchen-sink list I copied off a forum. Twelve clips each, 5 seconds, same positive prompt.

The five-term version was clearly the best: sharper, less blown-out, no captions. The 40-term version was worse than no negative at all — flat, grey, oddly stiff motion. That surprised me until it didn't. Every term you add spreads the conditioning thinner and drags the whole thing toward a bland average. You're not adding forty rules, you're diluting five good ones.

If you take one thing from this: a long negative prompt is not a thorough negative prompt. It's a weak one.

Where negative prompts can't help you

Being honest about the limits, because this is where people waste days:

• Character consistency across shots. No negative term holds a face together. That's a reference-image job — see keeping characters consistent.
• Counting anything. Fingers, people, objects. The models don't count.
• Physics. Cloth that behaves, water that falls right. You get that from model choice, not from negating "unrealistic physics".
• Text rendering. You can suppress accidental text. You cannot negate your way to correct text — add it in post.
• Anything the model has never seen. Negation needs a concept to push against.

What I actually do now

Concretely, my working process: check whether the model has a real negative field. If yes, paste the same five rendering-quality terms and leave them alone. If no, spend that effort on the positive prompt instead — more specific scene, explicit camera move, named lighting — which is what my prompt guide is mostly about.

And then change one thing at a time. The reason it took me two weeks to notice I was typing into the void is that I was changing the negative prompt, the positive prompt and the model together, then judging the result. One variable per test. Boring, and it's the only thing that works.

Looking for something else? Browse all 142 AI video guides in one list.

If you want the specific prompt sets I run, the seed-locked test clips from the length experiment above, and the running list of which models have gained or lost a negative field, that's all in the community — come and grab it: join the AI Video Generator community on Skool. And if you'd rather just have the finished clip, I make them to order.