
In our runs, Seedance 2.0 on fal refused an AI-generated character for real-person likeness on 3 of 6 reference-to-video and 2 of 6 image-to-video attempts, while Kling v3 Pro rendered all 6, and a faceless version of the scene passed 3 of 3. This article lays out every refusal we've logged by provider and shot type; six attempts per engine is not a refusal rate.
Method. Our own tests: a six-aspect-ratio run on 25 September 2026 (one fictional character, 4 video models) plus a faceless re-run (3 aspect ratios × 2 video models), a 12-beat film run on 27 July 2026, and single observations from September 2026. Vinc makes an AI film studio, and these are our own tests of our own pipeline.
What this means for you
- Have a second video model ready for close-ups and two-shots. In our runs, Seedance 2.0 refused face-forward shots of fictional characters, and Kling v3 rendered the same frames.
- Run a cheap refusal test before you commit to one model. Generate your main character at the aspect ratios you need, on two models, at the lowest quality tier, and note which ones refuse.
- If a shot is refused, try it as a wide, an insert or an empty set. Those shot types went through in every run we logged.
- Expect a refusal to take about a minute and a half, and read the error text. Ours took 86 to 105 seconds, and the message names the likeness filter, so you can tell it apart from a crash.
- Don't treat one pass as a guarantee. The same image passed one Seedance endpoint and failed the other, so test on the exact model and mode you plan to use.
You generate a fictional character, ask a video model to animate them, and the request is refused because the image "may contain likenesses of real people". This article collects every time we've logged that refusal, by provider and shot type. All were AI-generated or fictional characters: no test used a photo of a real person, and we didn't alter any image to get past a filter.
The cleanest test: one character, six aspect ratios, every engine
On 25 September 2026 we checked every film format Vinc supports against the real providers. The test was about aspect ratios, not filters, but the refusals showed up anyway, and the identical setup makes it our cleanest comparison.





