Why refusals happen at all
Every developer (Google, ByteDance, Kuaishou and others) puts a moderation filter in front of its model. It checks the prompt and uploaded images before generation, and sometimes the result afterwards. The filter’s job is to prevent deepfakes, content with children, violence and violations of other people’s rights. Filters play it safe, so they sometimes trigger on harmless requests.
Aggregator services pass your request to the model through the developer’s API, so the developer’s filter works there too.
10 reasons for refusals
1. A real face in the photo
What happens: you uploaded a photo of a person and ask to animate or change it. Why: protection against deepfakes. What to do: for video, try Kling v3, it is more relaxed about photos of people. Seedance is the strictest. On KADR, Seedance 2.5 runs with a soft filter that refuses ordinary people less often.
2. A child in the frame
Why: all developers have especially strict rules for images of children. What to do: for family photos, try another model and the most neutral description possible. Some requests with children will not pass in any model, and that is right.
3. A celebrity
Why: the image and voice of famous people are protected, and filters recognise names and faces. What to do: remove the name, make a fictional character. Using a celebrity’s face without consent is not allowed: deepfakes and the law.
4. A well-known brand or character
Why: rights holders sue AI companies. What to do: describe the style instead of the character, invent your own name instead of the brand. More: someone else’s brand, logo or character in AI.
5. Violence and weapons
What to do: if it is a scene from a story, describe the consequences and emotions instead of the action itself.
6. Explicit content
All major models prohibit it. Beach and sports scenes sometimes touch the filter: specify the clothing («in sportswear», «in a one-piece swimsuit»).
7. Trigger words
Why: the filter looks at words without context. Examples: «shot» (a photo or a gunshot), «an explosion of emotions», «blood and milk» (a Russian idiom for a rosy complexion), «a killer look». What to do: replace them with synonyms: «frame», «vivid emotions», «a striking look». On KADR the «Improve» button rewrites the prompt neutrally.
8. Politics and real events
Models often refuse requests with politicians, flags and real tragic events. If the task is legitimate (for example, a historical reconstruction), describe the era without specific people.
9. Medicine and «before and after results»
Images of surgery, wounds and promises of a healing effect may be rejected. For advertising this is also a risk of breaking advertising law.
10. Technical reasons
Not every refusal is moderation. It happens: a file that is too large, an unsupported format, a wrong aspect ratio, an overload on the developer’s side. If the error message is not about the content, try again later or reduce the file.
Who is stricter
| Model | Strictness about real faces | Comment |
|---|---|---|
| Seedance 2.0 | high | often rejects realistic photos of people |
| Seedance 2.5 on KADR | medium | soft filter, refuses ordinary people less often |
| Veo 3.1 | medium | strict about celebrities and children |
| Kling v3 | below average | the most relaxed about photos of people |
| Nano Banana Pro | below average on ordinary portraits | strict about celebrities |
These are observations from user requests, not official data from the developers. Policies change without warning.
What to do after a refusal
- Read the message. Is it moderation or a technical error?
- Remove the obvious triggers: names, brands, words from point 7.
- Switch the model. This is the fastest way for false positives.
- Rephrase neutrally or press «Improve».
- If the request breaks the rules in essence (someone else’s face without consent, children in an unacceptable context), do not try to get around it.
A comparison of models by task: which AI to choose for video.
In short
The AI refuses when the moderation filter is triggered: real faces, children, celebrities, brands, violence, explicit content or trigger words. Models have different policies, so after a false refusal the fastest fix is to switch the model: for photos of people Kling v3 is more relaxed than Seedance. If the request breaks the rules in essence, there is no need to bypass the filter.
