When a model declines your shot
Every model Breme runs has a content filter on the provider’s side. Filters misread film work. Here is how to tell what happened.
Joey Brass, Head of Communications
A moderation refusal means a content filter classified your request, not that anyone reviewed your intent. Read which filter spoke first: a Breme policy refusal is final, a provider refusal naming an input means you must change that input, and a provider refusal naming the output is worth one retry because the output pass sees a different video each run. Refusals are never charged.
Every model Breme runs has a content filter on the provider’s side. Filters misread film work. This page explains what a refusal means, how to tell which one you hit, and what to do next.
A refusal never costs you anything. The reservation releases and you are not charged.
Read the error first
Breme’s refusals name their author, and the author tells you what kind of problem you have.
Breme does not generate this. House policy. Our own screen ran before anything left the platform. Rewording will not change it and no other model on Breme will accept it, because the standard is the same across the catalog.
A reference appears to show a real, identifiable person. The provider’s likeness check, on the way in. It scores your reference against a set of known faces and blocks on a close match, so an ordinary face you hold the rights to can trip it. Likeness checks differ genuinely between vendors. This is the one case where trying another model is the right move.
The model’s own content filter declined this request. The provider’s general content filter. It fires on a category, not on your intent. Change the framing, the wording, or the reference. Carrying the same input around the catalog until something accepts it is not the remedy and is not something we will point you at.
The generated audio may match copyrighted material. The output pass heard something close to a real recording. Describe the music you want instead of naming a song or artist, or turn native audio off and lay your own track.
Why it passed yesterday and fails today
Most video models moderate twice: once over what you send, once over what comes back. The input pass is stable, so a refused reference stays refused. The output pass looks at a fresh video every run, which is why an identical request sometimes goes through on the second attempt. If the error names an input, stop re-running and change the input. If it names the output, one retry is reasonable before you change anything.
Classes that commonly false-positive
These are the categories where legitimate film work gets caught, drawn from provider documentation and from our own refusal log.
Children and anyone who reads as under eighteen. The hardest class in the industry, and the one where the filters are deliberately over-inclusive. Nothing about the innocence of the scene enters into it. Age-signaling words in the prompt raise scrutiny across the whole request, including parts of it that have nothing to do with the child.
Real faces on reference images. Photographic faces are checked harder than illustrated or rendered ones. A face repeated across a composition, on screens, posters, or mirrors, gets scored once per instance.
Signage, logos, and lettering near a face. Trademark checks and privacy checks compound. The same street scene without readable text is quieter.
Weapons and violence without cinematic framing. A sparse line reads as an instruction. A shot description with a lens, a light, and a location reads as a scene, and passes more often.
Named characters, brands, and living artists’ styles. Describe the look instead of naming the source.
Heavily compressed reference footage. Blocking and banding can imitate the artifacts filters are trained to flag.
Relative strictness by model family
Directional only, and it moves whenever a provider retunes. Use it to pick a starting model, not to plan a production around.
| Model family | Real faces on references | Children as subjects | Brands, logos, signage | Cinematic violence |
|---|---|---|---|---|
| Seedance | Strictest. Photographic faces are refused on the way in, with a dedicated error | Strictest. Minor safety is a named filter category at the lowest threshold | Strict, on both the input and the output pass | Passes with production framing, refuses on bare instructions |
| Veo | Restricted by design. Person generation is a documented API control | Will not generate children. The permissive setting is allowlist-gated, so this is a hard no rather than a false positive | Moderate | Moderate |
| Kling | Moderate to strict, inconsistent between runs | Strict | Moderate | Strict, with reported false positives on ordinary words |
| Hailuo | Moderate. Input and output declines carry separate codes | Strict | Moderate | Moderate |
| Luma Ray | Moderate, with a word list on the prompt and a frame check on the output | Strict | Moderate | Moderate |
| Grok Imagine | Looser on style and framing, output still reviewed against policy | Strict | Moderate | Looser, lower confidence |
One thing the table will not give you: a permissive route for children on screen. There is not one. Every general-audience provider guards that category hardest, and the two that document their behavior guard it in opposite ways, one by refusing ambiguous cases and one by declining to generate children at all. If your project lives there, budget for it. Stylized and animated rendering clears more often than photoreal, invented character references clear more often than photographic ones, and describing a character by build and bearing clears more often than describing them by age.
What to do, in order
- Read which filter spoke. House policy stops here.
- If an input was named, change that input. Re-running will not help.
- If the output was named, retry once before changing anything.
- Add production framing to the prompt: shot size, lens, light, location, mood.
- Replace photographic references with generated or illustrated ones where the shot allows it.
- For a likeness refusal on a reference you hold the rights to, switch models from the refusal and run it again.
If a shot you believe is clean keeps getting refused across attempts, send it to us. Refusal patterns are how we learn which model to put in front of which kind of project, and that work goes back into the model picker. The reasoning behind all of this is in whose filter said no.
Common questions
- What does a content moderation refusal actually mean?
- That a classifier sorted your request into a category it is configured to block. No person reviewed it and it is not a judgment about your intent. Filters misread legitimate film work regularly, particularly around children, real faces, and signage.
- Should I retry a refused generation?
- Only when the error names the output. The input pass is deterministic, so a refused reference or prompt stays refused however many times you send it. The output pass sees a different video on every run, so one retry there is reasonable.
- Which AI video model is least strict?
- None of them are permissive about children on screen, and that is the class that most often blocks legitimate work. For rights-cleared real faces, likeness checks differ genuinely between vendors, so switching models is worth trying. For other sensitive-content refusals, changing the framing beats changing the model.
Keep reading
- Whose filter said no
Why the same frame passes on one platform and gets refused on another, and what Breme does about it.