Haiyang Liu
I'm a researcher working on multimodal AI models for video, 3D animation, and robotics. My work primarily focuses on generating human videos/3D animation from multimodal conditions such as text, audio, and image.
At Hedra, I led research on the real-time talking heads model Livatar-1 and the multimodal video generation model Omnia. I worked hands-on across the data, model architecture, and training infrastructure.
Before that, I completed my Ph.D. at 
The University of Tokyo, with internships at 
Hedra Research, 
Adobe Research, 
CyberAgent AI Lab, and 
Huawei Research Tokyo. I advised 3D animation research in Alaya Lab.
Recent Work
Livatar-1
2025Real-Time Talking Heads Generation with Tailored Flow Matching
FloodDiffusion
2025Tailored Diffusion Forcing for Streaming Motion Generation
Video Motion Graphs
2025Controllable Human Motion Video Generation from Music and Motion Tags
TANGO
2024Co-Speech Gesture Video Reenactment with Hierarchical Audio-Motion Embedding and Diffusion Interpolation
EMAGE
2023Towards Unified Holistic Co-Speech Gesture Generation via Expressive Mask Audio Gesture Modeling
Academic Services
- Reviewer: SIGGRAPH, SIGGRAPH Asia, Eurographics, ICLR, CVPR, ICCV, ECCV, ACMMM, WACV, IJCV, TVCG