Almost every AI video service refuses the same four categories. If you know where the lines are, you can predict a refusal before you spend a render — and more usefully, you can tell whether your idea is genuinely over the line or just phrased badly.
The four lines
1. Real people who haven't agreed. A recognisable face generated into a situation that person never consented to. This is the line most strictly enforced, and it's the one that turns a fun project into a legal problem.
2. Minors. Universal, absolute, and not a matter of degree. Any depiction of a child in a sexualised or violent context is refused everywhere, and most platforms extend this to any ambiguous case.
3. Sexual content involving real or realistic people. Deepfake pornography is the specific target. Platforms generally permit stylised, non-explicit romantic content; they don't permit anything that reads as a real person in a sexual scenario.
4. Gratuitous violence and gore. This is the fuzzy one, and it's the one that matters for a zombie video.
Where a zombie couple video sits
Comfortably inside the lines, in the ordinary case. Two consenting adults, a stylised undead look, ten seconds, no injury depicted.
The look the genre implies — grey-green skin, dark around the eyes, a torn collar — is makeup vocabulary, not violence. Platforms treat it as costume. The refusal risk rises when a request drifts toward injury: wounds, blood pooling, visible damage to a body.
The practical distinction is between undead and wounded. One is a costume; the other is graphic harm, and filters are tuned to notice the difference.
The gore question, concretely
Things that usually pass: pallid skin, sunken eyes, dirt, torn clothing, a blank stare.
Things that raise risk: open wounds, blood on skin or clothing, missing limbs, anything depicting the moment of death.
If you want a harsher look, the reliable way is lighting and framing — hard shadows, a cold colour cast, a tight crop — rather than explicit injury. It reads as more frightening anyway. Suggesting violence is more effective than depicting it, in horror generally.
Using someone else's photo
This is where people get caught, and not by a filter.
Uploading a photo of an ex-partner, a friend, a colleague, or a public figure to generate content they haven't agreed to isn't a policy edge case — it's the core prohibition, and it's the one with real consequences outside the platform.
For a couple video the rule is simple: both people in the frame know the video exists and agreed to it being made. If that's not true, nothing else about the prompt matters.
What a refusal looks like, and why rewording won't help
Refusals usually come as a generic failure rather than a specific accusation — partly to avoid teaching people how to get around the filter.
Rewording to dodge a refusal is the wrong response, for two reasons. It doesn't work: the decision is a classification made upstream of sampling, so a euphemism is still classified. And it's a bad idea even when it does work.
The right response is to change the request, not the phrasing. If the scene needs blood to function, the scene needs redesigning.
What to do instead of arguing with a filter
Three moves that solve most "refused" situations:
Restage the moment. Instead of the bite, the reaction to the bite. Instead of the wound, the hand over it.
Change the register. Stylised and theatrical passes where photorealistic fails. A mask reads as costume; realistic torn skin doesn't.
Cut duration. A shorter clip gives less time for a borderline element to accumulate.
The thing worth internalising
Content policy isn't an obstacle between you and the fun version. It's the reason the tool exists in a form you can use without verifying anyone's identity. The two-consenting-adults, costume-not-injury zone is wide, and it's where the good ideas mostly live anyway.
Try it at the generator; price at pricing.
Frequently asked questions
Will a zombie couple video get refused for violence?
Not in the ordinary case. Stylised undead makeup reads as costume. Risk rises when the request depicts injury rather than undeath.
Can I make a video using my partner's photo as a surprise?
Only if they'd be fine with it. Consent is the line, not the filter — and a filter passing doesn't make it acceptable.
Can I make a video of a celebrity?
No. That's the clearest prohibition on every platform and in a growing number of jurisdictions.
Why did my prompt fail with no explanation?
Refusals are usually deliberately vague so the filter can't be reverse-engineered. Change the request rather than the wording.
Does adding "fictional" or "for a movie" help?
Not meaningfully. The classification looks at what's depicted, not at the framing sentence around it.