Shanghai Jiao Tong University
BEng
Electrical and Computer Engineering
China
Advisor: Prof. Ming Yang
About
My research is centered on building human-centric world models that perceive, anticipate, and generate how people and their environments evolve over time. I develop predictive and generative models across human motion, dynamic 3D scenes, and long-horizon video generation, with the goal of learning general models of human-centered dynamics.
I am currently looking for research roles and collaborations in computer vision and generative AI. Please get in touch if you think there could be a good fit!
News
- 2026Excited to co-organize the Human Motion-Informed World Models Workshop at ECCV 2026.
- 2026Social Mamba was accepted at ECCV 2026. Congrats to Po-Chien!
- 2026Released EverAnimate, a framework for minute-scale human animation via latent flow restoration.
- 2026OmniTraj was accepted in Transportation Research Part C and Precognition at CVPR 2026.
- 2026Deformable Gaussian Occupancy (DeGO) was accepted at CVPR 2026.
- 2026Stable Video Infinity was accepted as an oral presentation at ICLR 2026. Congrats to Wuyang!
- 2026RAP was accepted at ICLR 2026. Congrats to Lan!
Selected Publications
* Joint first-author · Full list on Google Scholar
Showing all 9 publications
-
EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration
Generates minute-scale human animation using persistent latent memory and restorative flow matching to preserve identity, motion, and visual quality over long sequences.
-
OmniTraj: Pre-Training on Heterogeneous Data for Adaptive and Zero-Shot Human Trajectory Prediction
Pre-trains on heterogeneous trajectory data to support adaptive and zero-shot prediction across datasets, frame rates, and observation lengths.
-
Deformable Gaussian Occupancy: Decoupling Rigid and Nonrigid Motion with Factorized Distillation
† Corresponding author
Represents dynamic scenes with deformable Gaussian occupancy, separating rigid and nonrigid motion through factorized foundation-model distillation.
-
Stable Video Infinity: Infinite-Length Video Generation with Error Recycling
Extends video generation to long horizons through error-recycling fine-tuning, improving temporal consistency and reducing accumulated drift.
-
RAP: 3D Rasterization Augmented End-to-End Planning
Uses 3D rasterization to connect simulation and real-world features for end-to-end driving, enabling scalable training without photorealistic rendering.
-
Social-Pose: Enhancing Trajectory Prediction with Human Body Pose
Combines human body poses with observed trajectories to capture spatiotemporal social interactions and improve predictions of future motion.
-
Unified Human Localization and Trajectory Prediction with Monocular Vision
Unifies human localization and future trajectory prediction from a single camera by connecting 2D pose observations with 3D motion forecasting.
-
Multi-Transmotion: Pre-trained Model for Human Motion Prediction
Learns transferable representations through pre-training for human motion prediction, integrating diverse motion cues for future forecasting.
-
Social-Transmotion: Promptable Human Trajectory Prediction
Uses a promptable transformer to forecast human trajectories from varied motion cues, including trajectories and body poses.
Academic Service & Teaching
Academic Service
- Workshop organizer, ECCV 2026: Human Motion-Informed World Models and Socially Intelligent Action
- Conference reviewer: CVPR, ICCV, ECCV, NeurIPS, ICLR, ICRA, IROS
- Journal reviewer: RA-L, IEEE T-ITS, TR-C
Teaching Assistant
- CIVIL-459 · Deep Learning for Autonomous Vehicles
- CS-119 · Information, Computation, Communication (Python Programming)
Education
KTH Royal Institute of Technology
MSc
Embedded Systems
Sweden
Advisor: Prof. Hedvig Kjellström
École polytechnique fédérale de Lausanne (EPFL)
PhD Candidate
Computer Science
Switzerland
Advisor: Prof. Alexandre Alahi