Negative prompts are the most over-recommended trick in AI video. They work, sort of, in image generation. In video they're weaker, less predictable, and occasionally they cause the exact thing they were meant to prevent.
Here's the mechanism, and what to do instead.
What a negative prompt does mechanically
In a diffusion model with guidance, a negative prompt produces a second conditioning signal and the sampling is pushed away from it. It's a steering correction, not a filter.
That means it works on the distribution of the output, not on a checklist. It makes "no text in the image" slightly less likely to sample text. It doesn't forbid text.
Video models add a second problem: temporal consistency. A correction that flips on and off between frames is worse than no correction, because it produces flicker. This is why negatives that work fine on a still image behave badly on a clip.
Why naming a thing can summon it
The classic failure: you write "no blood" and get blood.
Part of this is that the negative signal is weak relative to the positive one, and the token "blood" still passed through the text encoder attached to your scene. Part of it is that many video pipelines only loosely apply the negative branch, so the word leaks in as positive material.
The practical lesson is uncomfortable: describing what you don't want can function as describing what you want.
Positive constraints that work instead
Every useful negative has a positive phrasing. Use those.
| Instead of | Write |
|---|---|
| no blood | clean faces, intact skin |
| no text, no watermark | empty background, plain wall behind them |
| not blurry | sharp focus on both faces |
| no extra people | only two people in frame |
| not moving fast | slow, minimal movement |
| no distorted hands | hands out of frame |
The right-hand column has a property the left doesn't: it describes a state the model can generate toward, rather than a state to avoid.
When negatives do help
Two cases where I'd still use them:
Colour casts. "No green tint" sometimes helps, because colour is a stable global property and the correction is cheap.
One specific recurring artefact. If the same wrong element appears every single time — a hat that keeps materialising, a third figure in the background — a targeted negative occasionally breaks the pattern. Test it; don't assume it.
And one case where they actively help: you can't remove something, so you crop or reframe instead. That's not a prompt fix, it's a framing fix, and it's more reliable than either.
The order to try things
When a clip comes back wrong, try these in this order:
- Change the source photo. Wrong wardrobe is usually a photo problem.
- Shorten the prompt. Remove the thing by removing the attention spent on it.
- Simplify the scene. Fewer objects, less to get wrong.
- Reduce motion. Drift scales with movement.
- Add a negative. Last, and only for one specific artefact.
Most people start at step 5. That's why they think it doesn't work.
What negatives can't do
They can't fix identity drift. They can't add detail that resolution doesn't support. They can't override an unsafe-content filter — a negative prompt is not a policy argument.
If a request is refused, rewording it negatively won't help. The refusal is a classification decision upstream of sampling, and it's answered by changing the request, not by dressing it up.
The honest summary
A negative prompt is a nudge on a probability distribution. That's a real effect and a small one. Treat it as a last-mile adjustment, not as a control surface — and spend your effort on the source photo and the scene, where the leverage actually is.
Try it at the generator; price at pricing.
Frequently asked questions
Do negative prompts work in AI video at all?
Weakly and inconsistently. They're much less reliable than in still-image generation because they can introduce flicker.
Why does "no blood" sometimes produce blood?
The negative signal is weak, and the token still reaches the text encoder tied to your scene. Naming a thing can pull it toward the output.
What should I use instead?
A positive description of the state you want: "clean faces, intact skin" rather than "no blood".
Can a negative prompt get past a content filter?
No. Refusals happen upstream of generation and are answered by changing the request.
Is there a case where negatives are clearly better?
Yes — global colour corrections, and breaking one specific artefact that recurs identically every time.