Why Seedance Blocks Real Faces — and What Actually Works Instead
Seedance and Veo reject real human likenesses. We spent April 2026 testing where that filter lives, which bypasses work (none), and which models let you legitimately put a real person in AI video.
If you have tried to feed a photo or clip of a real person into Seedance 2.0 and received a content-policy rejection, the behaviour is not a glitch and not tier-dependent. We ran this down properly in April 2026 because our own pipeline depended on the answer.
Where the filter lives
The short version: in ByteDance's inference layer, not in the API wrapper you happen to be using.
We tested the same rejected input through fal.ai, through an independent third-party host, and through the image endpoints rather than the video ones. Same rejection, same shape, everywhere. Community claims that a particular reseller or the "first frame" path is looser did not reproduce.
Three specific findings from that round of testing:
- The image endpoints reject real faces too. A pipeline that stylises a face into an image first and feeds that forward does not slip through — the image call is filtered on the same basis.
- A grid-overlay trick that circulated in March 2026 stopped working by 20 April. Image-layer bypasses that get shared publicly have an observed lifetime of about a month.
- Stylising the source video enough to clear the filter destroys the facial dynamics that made the shot worth copying. It is a real trade-off, not a tuning problem.
We are documenting this so other people stop paying to rediscover it. There is no configuration of Seedance that accepts a real person's likeness.
Seedance 2.5 did not change it (September 2026 update)
ByteDance shipped Seedance 2.5 in late August 2026 with the most impressive spec sheet in the category: up to 50 reference inputs and a genuine 30-second single take, which nothing else offers. The reference-to-video endpoint is exactly the shape of a "put me in this" tool.
We have not re-tested it, and we should say so plainly rather than imply a September benchmark we did not run. The filter lives in ByteDance's inference layer, upstream of every reseller and of every endpoint in the family — that is what the April testing established, and a new model version does not move a filter that sits above the model. So we treat it as unchanged until the vendor says otherwise, and we are not spending on another bypass attempt to confirm it. All three 2.5 rows sit in our registry marked deprecated on that standing verdict: the capability is real, the use case is still blocked.
The one endpoint the filter cannot touch is text-to-video, and there it loses on price. Seedance 2.5 bills $0.0214 per 1,000 video tokens, which works out near $0.462 per output second at 720p24 — about $13.90 for that headline 30-second take, against $0.085 per second for the faceless model we actually route to. See the autumn model round-up for the rest of that release wave.
Which models do allow real people
| Model | Real-face input | Notes |
|---|---|---|
| Seedance 2.0 (any host) | No | filter is upstream of every wrapper |
| Seedance 2.5 (any host) | No | same upstream filter; family verdict carried over, not re-tested |
| Veo 3.1 | No | same category of restriction |
| HappyHorse 1.x | Yes | up to 9 character references, 1080p |
| Kling 3.0 | Yes, with limits | public-figure restrictions still apply |
| Wan (open weights) | Yes | different vendor, different policy |
| MiniMax H3 | Yes | released 31 July 2026 |
| Gemini Omni Flash 1.1 | Unknown | new in late August 2026; the vendor's other frontier video model refuses real likenesses |
| HeyGen | Yes, with rights | requires you to hold rights to the source video |
Two of those need a caveat. Kling's allowance is not unconditional — likeness restrictions around public figures are enforced, and a rejection there is a policy call rather than a fault. HeyGen's translation products require that you actually have rights to the video you upload, which is a licensing obligation and not a checkbox to click past.
What to do instead
If your job is "put a real person into a generated shot", the working paths in September 2026 are HappyHorse reference-to-video (up to nine reference images, native 9:16), Kling 3.0 Motion Control when the movement has to match an existing clip, or Wan 2.7 edit-video when the original background must survive. All three are priced per output second — $0.10 to $0.168 — so a ten-second clip is between $1.00 and $1.68.
If your job is "generic B-roll with no identifiable person", Seedance is genuinely good and the policy never fires. Route by whether a real identity is involved, and the whole problem disappears.
The part nobody says out loud
Policy filters are a product constraint, not an obstacle to route around. Every published bypass we tested was patched within weeks, and the ones that "worked" degraded output badly enough that the result was unusable anyway. Building on a model that permits your use case is cheaper than building on one that tolerates it briefly.
