AI Video Generator With Real People: What Actually Works

Search for an AI video generator with real people and the autocomplete tells you what everyone is actually asking: with people talking, with famous people, real person, free. Those are four different requests with four different answers, and exactly one of them is refused everywhere no matter which platform you pay.

The useful split is not by tool. It is by whose face it is. A photograph of a consenting subject, a synthetic character that happens to look photoreal, and a named celebrity are three legally and technically distinct things, and most of the frustration in this space comes from platforms treating all three as one blocked category. This walks through what each one actually does on a modern video API, what talking requires on top, and where the line genuinely does not move.

TL;DR

• Reference photos and clips of real people are supported. Seedance 2.5's parameter reference states plainly that "reference images / videos may contain real people."^[1]
• Rejected reference material refunds in full. Material passes automated review before generation, and anything refused fails with an error and a full refund of the reserve.^[2]
• Talking is a separate capability. Seedance 2.5 generates "synced speech, sound effects and background music" with generate_audio, on by default.^[1] Kling 3.0 documents lip sync explicitly.^[3]
• Named celebrities are refused on every route. That is the model's own line, not a platform policy setting, and no parameter reaches it.
• Consent is still yours to obtain. An API that accepts an upload has not acquired permission on your behalf.

Four different asks wearing one search term

Your own footage, or a consenting subject. This is the largest legitimate case: founders in their own product demos, UGC creators, actors who signed a release, a client whose brand video you are making. It is supported, and the rest of this article is mostly about it.

A synthetic person who looks photoreal. Your character does not exist, but your image generator got good enough that face classifiers score it as a photograph. This is the case that generates the most complaints, because the rejection is a false positive rather than a policy.

A named public figure. Refused. Not by reAPI as a matter of policy, but by the model, on every host that resells it. Any platform advertising otherwise is either wrong or selling you a likeness-rights problem.

People talking. A capability question rather than a permission question, and it is answered by which model you pick rather than by which platform.

Conflating these is why the advice you find online is contradictory. Someone whose consented-subject workflow runs fine and someone whose synthetic character keeps bouncing are both reporting accurately about different asks.

Using a consented person as a reference

The supported path is reference material, not prompting a description. Seedance 2.5 takes up to 30 reference images, 10 reference clips and 10 audio tracks in one generation.^[1]

curl https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer rk_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2.5-face",
    "prompt": "@Image1 walks through a sunlit workshop, handheld follow",
    "resolution": "720p",
    "duration": 8,
    "image_urls": ["https://your-cdn.example.com/subject-01.jpg"]
  }'

The constraints worth knowing before you assemble a reference set:^[1]

Two rules catch people out. Every reference must be a public HTTP(S) URL; base64 and data: payloads are rejected across the whole platform, not just this model.^[1] And image_urls cannot be combined with image_with_roles, so character references and first/last-frame chaining are separate modes.^[1]

Reference material goes through automated review before generation runs. If it is refused, the task fails with an error and the reserve is refunded in full,^[2] which means a false positive costs you a retry rather than money.

Making them talk

This is where model choice matters more than platform choice, and the two documented options behave differently.

Seedance 2.5 generates audio natively. generate_audio defaults to true and produces "synced speech, sound effects and background music (mono)."^[1] The speech is generated with the video rather than dubbed onto it, so mouth movement and audio come from the same pass. You direct it through the prompt.

Kling 3.0 documents the capability more specifically, describing synchronized native audio as "dialogue, lip sync, ambient sound."^[3] If lip sync accuracy on dialogue is the thing you are being judged on, that is the explicit claim to test against.

What neither offers is a dubbing mode where you supply a finished voice track and get precise mouth-shape matching to it. Seedance 2.5's audio_urls are documented as reference audio tracks,^[1] which is a different job from lip-sync driving. If your pipeline already has recorded VO, test that assumption on a short clip before you build around it.

One practical note for talking-head work: audio does not survive segmentation. Each generation synthesises its own track, so a sequence assembled from several clips gives you several independent audio beds. For dialogue longer than one generation, record or synthesise the voice separately and treat the video as picture only.

The line that does not move

Named living people are refused. Every honest route enforces this, and it sits with the model rather than with any platform's settings. If a host claims otherwise, the correct response is suspicion, not a signup.

Third-party characters are refused too. The same check catches recognisable copyrighted characters, which is a recurring frustration for anime and fan work and has no workaround worth publishing.

Consent is not transferred by an upload form. The API accepting your reference photo is not permission from the person in it. Releases, likeness rights and, in several jurisdictions, biometric-data rules are your responsibility and do not become the platform's because a request returned 200.

That last one is not legal boilerplate. It is the reason the supported path exists at all: the check is designed to stop the non-consensual case, and it works better when the legitimate case has somewhere to go.

When a person who does not exist gets refused

The most common failure in this space is a fully synthetic character being blocked as a real person, because the classifier scores photorealism rather than provenance. It has no way to know your character was generated. Users have reported the same class of rejection with errors like "Input image may contain real person" on other platforms.^[4]

The mechanics, the exact error strings, and how to tell a platform-level face detector from a model-level check are covered in reapi.ai/blog/seedance-face-detected-real-person-error. The short version: time the rejection. A refusal that lands in under a second never reached a GPU.

What it costs

Reference-driven work is priced the same as any other generation on Seedance 2.5, with one wrinkle. Reference images and audio do not change the rate; a reference video moves you to a lower per-second rate that is billed over input plus output seconds.^[2]

Rates from the published per-second band.^[2] A failed moderation check refunds in full, so exploratory reference sets cost time rather than credits.

FAQ

Can I use an AI video generator with real people at all?

Yes, with reference material of a consenting subject. Seedance 2.5's parameter reference states that reference images and videos may contain real people.^[1]

Can I generate a famous person?

No. Named public figures are refused at the model level on every route, and no parameter changes that.

Which model handles people talking?

Seedance 2.5 generates synced speech alongside the video with generate_audio.^[1] Kling 3.0 documents dialogue and lip sync explicitly.^[3]

Can I supply my own voice recording and get lip sync to it?

Not as a documented mode. Seedance 2.5's audio_urls are reference tracks rather than a lip-sync driver.^[1] Test it on one short clip before designing a pipeline around it.

Why was my AI-generated character rejected as a real person?

Because the check scores photorealism, not origin. See the face-detected explainer for how to identify which layer refused you.

Do I get charged when a reference is rejected?

Not on reAPI. A task that fails review refunds the reserve in full.^[2]

Can I upload the reference file directly?

No. Every reference must be a public HTTP(S) URL. Base64 and data: payloads are rejected platform-wide.^[1]

How many reference images can one generation take?

Up to 30, plus 10 clips and 10 audio tracks.^[1]

Working with faces instead of around them

Decide which of the four asks you actually have before you pick a tool, because three of them are ordinary product decisions and one of them is not available anywhere. A consenting subject and a good reference set is a supported workflow with a published parameter surface and a refund when review says no. A photoreal synthetic character is the same workflow plus a false-positive problem you can diagnose in thirty seconds. Dialogue is a model-selection question with two documented answers.

Only the fourth ask, a named person who did not agree, has no route, and that is worth defending rather than routing around. An AI video generator with real people is a normal tool when the people said yes, and the reason the supported path stays open is that the unsupported one stays closed.

Full parameter reference at reapi.ai/docs/seedance-2-5; current rates on the Seedance 2.5 model page.

References

1. reAPI. doubao-seedance-2.5-face — request body, reference limits, generate_audio and platform media rules. Retrieved August 2026 from reapi.ai/docs/seedance-2-5
2. reAPI. Seedance 2.5 — model page, real-person reference support, published rates and refund behaviour. Retrieved 9 August 2026 from reapi.ai/models/seedance-2-5
3. reAPI. Kling 3.0 — native audio with dialogue and lip sync. Retrieved August 2026 from reapi.ai/docs/kling-3-0
4. r/generativeAI. Seedance keeps rejecting my AI images as "real person"; any fix? Posted 22 May 2026, retrieved August 2026 from reddit.com/r/generativeAI/comments/1tkkxjq

Further reading

• reAPI. Seedance Face Detected: Real Person Errors, Explained. reapi.ai/blog/seedance-face-detected-real-person-error
• reAPI. AI Video Generator Long Videos: Chaining Past the Cap. reapi.ai/blog/ai-video-generator-long-videos