Animating a person's photo with another person's movements usually needs separate systems for face and body, and often requires slow processing steps. The researchers built FUHA, a single model that learns a compact code capturing all body motion from a driver video, then uses that code to animate a reference photo in one fast pass. It runs at 38.9 frames per second. The motion code carries almost no identity information, only 0.1% of queries matched the same person, so it works well for swapping movements between different people while preserving the original person's appearance. That matters for real-time avatars and telepresence without complex preprocessing.
AI avatars can copy face and body movements in real time
The authors demonstrate single-pass human animation that separates movement from appearance, with the strongest benchmark results on upper-body animation.
Big Tech
Seonwook Park · Hyunsoo Cha · Amrita Mazumdar · Shengze Wang · Tianye Li · Hanbyul Joo · +2 more
NVIDIA · Seoul National University
Research Digest··2 min read
In one line
A feed-forward transformer learns a compact motion embedding for real-time whole-human animation.
Before you read this
An idea this paper builds onRead it as
Park and colleagues present FUHA, a model that animates a reference photograph using another person’s movements, covering the face, hands and body.
The model compresses visual features from a driver image into a small motion representation, then combines this with a separate representation of the reference person’s appearance.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.