OmniHuman-1
ByteDance's single-image human animator: one photo plus audio becomes a full-body talking video with near-flawless lip sync — the research demo that reset expectations for AI dubbing.
About
OmniHuman-1 (February 2025, arXiv 2502.09630) is ByteDance's multimodal video generation framework for turning a single image into a realistic human video. Where most talking-head models animate just the face, OmniHuman goes full-body — lip-synced speech plus hand gestures, head tilts, and body language matched to the audio — trained on more than 18,000 hours of human video with a diffusion-transformer backbone. The demos were the story: near-flawless lip sync, minimal glitches, and it works on cartoons, animals, and objects too, not just photoreal people. Third-party writeups credit it with 37-language audio handling and ~94% phoneme accuracy, but those figures come from fan and vendor-adjacent sites, not ByteDance — treat them as unverified. ByteDance never open-sourced it; the only public access is OmniHuman v1.5 served through third-party API hosts. For dubbing workflows it's the reference for what audio-driven re-animation can look like, even if you can't download the model itself.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Third-party sites claim 1.2mm lip-sync error and 94% phoneme accuracy (unverified, not from ByteDance)\n- No peer-reviewed benchmark numbers published
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 4/10
Scored 4/10 (TIER metrics, metrics-v1). Evidence: ByteDance research paper (arXiv 2502.09630); wide media coverage of the demos; third-party API availability. Strengths: landmark quality bar for audio-driven full-body animation, directly relevant to lip-sync dubbing. Docked for: (1) no open weights and no official API — nothing reproducible; (2) the most impressive numbers (37 languages, phoneme accuracy) trace to third-party sites, not the vendor; (3) commercial use terms unclear. License: proprietary, not released.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 1.0/10
- Manual review · sanity read + red-flag check
- 5.0/10
- · Paper arXiv 2502.09630; trained on 18,700+ hours; full-body audio-driven animation
- · Never open-sourced; only v1.5 via third-party API hosts — nothing reproducible
- · The flashiest numbers (37 languages, 94% phoneme accuracy) come from third-party sites, not ByteDance
- · officialUrl is a community site; no official ByteDance product page verified