OmniHuman-1 — A Single Photo Sings and Speaks into a Full Video

ByteDance's OmniHuman-1 generates full video with lip-sync, gestures and expression from a single still image plus audio. Einstein demo went viral. Research-stage only — no public access yet.

Developer: ByteDance (TikTok parent company)

Modality: Single image + audio → lip-synced video with gesture

Status: Research-stage (no public product)

Hot Demo

A single still photo (Einstein) + audio track → full video with natural lip-sync + hand gestures + facial expression changes. Synchronization accuracy at movie lip-sync level + motion estimation from a single image.

Technical Highlights

- Multimodal Input — Any combination of text/audio/video reference can be used as a driver

- Single Image Source — Dynamic character synthesis from a single photo, not a video

- Lip-sync Accuracy — Facial expressions, jaw movements, and phoneme alignment are at movie dubbing level

Risks

- Politician fake speech / resurrection videos of the deceased / unauthorized deepfakes. The reason ByteDance intentionally did not release a public API.

Synopsis

OmniHuman-1 is an AI video generation model developed by ByteDance, covering the technology of transforming a single image into a realistic human video. This falls under informational, industry trend, or tech meta-content, including introductions, demonstrations, and explanations of technology (AI/software). It is not a video explaining the plot (protagonist, conflict, world-building) of narrative works like movies, dramas, or animations. Therefore, the output [for a narrative synopsis] should be exactly 'NONE'.

Maru