Human-Centric Video Generation
Building controllable and interactive digital humans that can communicate naturally through audio, motion, expression, and video.
Wenzhang Sun 孙文章
I work on structured spatiotemporal modeling for generative and multimodal systems, with a focus on human-centric video generation, efficient 3D/4D world representation, and multimodal evaluation.
My recent work at Li Auto Foundation Models centers on digital humans, video generation systems, 4D/world representations, and multimodal evaluation. Before that, I received my Ph.D. from Beijing Institute of Technology, where I worked on 3D human modeling and reconstruction.
research
Research Themes
Efficient 3D/4D World Representation
Learning compact spatiotemporal representations that connect video generation, 4D scene understanding, and future-world prediction.
Multimodal Evaluation & Systems
Designing evaluation protocols and agentic systems that make multimodal models more measurable, reliable, and useful in workflows.
news
Recent updates
- RiO-DETR accepted to ECCV 2026.
- DrivingScene and PAGS accepted to ICASSP 2026.
- PREX released on arXiv with a project page.
- RiO-DETR and MUSE released as arXiv preprints.
- MoEE accepted to CVPR 2025; FaceVid-1K accepted to ICCV 2025.
selected work
Current focus
Interactive digital humans and world models
I am exploring real-time interactive digital-human generation for continuous multimodal dialogue, as well as 4D-guided video generation world models.