Subscribe

OmniHuman-1

OmniHuman-1 animates a single image of a person from audio (talking, singing, gesturing) with a one-stage, mixed-conditioning diffusion-transformer.

Scales human animation by mixing text, audio and pose conditions during training so weakly-conditioned data is usable. Became the basis for 'talking-head' video in ByteDance's Dreamina and CapCut.

Date
Monday 3 February 2025
Lab
ByteDance
Kind
model
Access
app only

The date is the paper's first arXiv version, 2025-02-03. ByteDance's use of the model in Dreamina and CapCut is unconfirmed.

Sources

  1. arxiv.org/abs/2502.01061

This record was checked against its sources on 6 October 2026. How we check