סמינר: Machine Learning Seminar
Diffusion Models for Human Motion Synthesis
Lecturer:
Guy Tevet
Research Areas:
| Realistic character motion synthesis remains a fundamental challenge in computer animation, requiring both natural expressiveness and precise control. The introduction of diffusion models has transformed the field, significantly improving synthesis quality while enabling intuitive control mechanisms such as text and music. One of the earliest and most widely used approaches is the Motion Diffusion Model (MDM) [ICLR 2023], which integrates domain expertise into the diffusion process to enhance motion generation and editing capabilities. While MDM sets a strong foundation, recent advancements push its capabilities even further, enabling novel forms of control and synthesis beyond traditional pipelines.Multi-view Ancestral Sampling (MAS) [CVPR 2024] is an inference-time algorithm that generates 3D animations from 2D keypoint diffusion models. This method facilitates the synthesis of character animations for scenarios that are challenging to capture in motion capture systems yet are prevalent in in-the-wild videos, such as horse racing and professional rhythmic gymnastics. Closing the Loop between Simulation and Diffusion (CLoSD) [ICLR 2025, Spotlight] combines motion diffusion with reinforcement learning for physics-based character control. Using a diffusion model as a real-time motion planner enables interactive, text-driven behaviors that respond dynamically to the environment while ensuring physically plausible motion. To conclude, I will share my perspective on key challenges in the field and highlight ongoing research in our lab aimed at addressing them. |
| Guy Tevet is a prospective Postdoctoral Fellow at Stanford University, where he will join the Digital Athlete Moonshot team under the supervision of Prof. Karen Liu and Prof. Scott L. Delp. He recently completed his Ph.D. in the Computer Graphics Lab at Tel Aviv University, advised by Prof. Amit Bermano and Prof. Daniel Cohen-Or. His research focuses on human motion generation. In the summer of 2023, he was a visiting research student at the University of British Columbia, advised by Prof. Michiel van de Panne and Dr. Xue Bin (Jason) Peng (Simon Fraser University). Previously, he was a research intern at Google (2020–2021) in natural language processing and a computer vision researcher at Apple (2018–2020).
|