OmniHuman-1

OmniHuman-1, from ByteDance Intelligent Creation (paper 'OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models', February 2025), is an end-to-end human animation framework built on a diffusion transformer. Its signature idea is 'omni-conditions' training: audio, video, and pose signals are mixed during training so the model can learn from far more data than the small pool of high-quality paired samples, then infer from weak inputs like a single image plus an audio track. It handles portraits, half-body, and full-body images at any aspect ratio, generates gestures and lip-sync that match speech or singing, and even animates cartoons and objects.

The official project page states plainly that ByteDance offers no downloads or services for the research model itself, warning users against fraudulent lookalikes, so there are no public weights or API. The technology did reach the public indirectly: it powers the 'Master Mode' of ByteDance's Jimeng/Dreamina digital-human module, and a successor, OmniHuman-1.5, added dual-system cognitive simulation for minute-long videos. Trained on 18,700 hours of human video, it set a new bar for expressive full-body animation, with the authors reporting an image quality score of 3.875 on the CelebV-HQ test set, above prior specialized models.

Author-reported image quality score of 3.875 on the CelebV-HQ test set, claimed above prior specialized models; no independent published benchmarks