Yuxuan Mu (Matthew)

I’m a Ph.D. student at the GrUVi Lab at Simon Fraser University, where I work on 3D computer vision and animation. My research mainly focuses on 3D character motion modeling, and I’m advised by Professor Xue bin (Jason) Peng.

Before that, I completed my master’s at the University of Alberta with Professor Li Cheng. I also spent time as a research intern at Huawei Canada (Noah’s Ark Lab), working on 3D generation and reconstruction with Juwei Lu and Xinxin Zuo. .

Email  /  GitHub  /  LinkedIn  /  X

Research

My work centers around 3D human motion—spanning estimation, generation, physics simulation, and control. I’m particularly interested in bridging the gap between the physical and virtual worlds to enable richer perception, understanding, and interaction. To that end, I design generative models and reinforcement learning systems that leverage physics simulations to better capture and replicate human behavior and interactions.

'*' indicates equal contribution.

MotionDreamer: One-to-More Motion Synthesis with Localized Generative Masked Transformer

Localized masked modeling paradigm to learn motion internal patterns from one motion with arbitrary topology.

GSD: View-Guided Gaussian Splatting Diffusion for 3D Reconstruction

The first 3D diffusion model directly upon Gaussian Splatting for real‐world object reconstruction, with fine-grained view-guided conditioning.

MoMask: Generative Masked Modeling of 3D Human Motions

We introduce MoMask, a novel masked modeling frame work for text‐driven 3D human motion generation with a hierarchical quantization scheme.

Generative Human Motion Stylization in Latent Space

We propose a flexible motion style extraction and injection method from a generative perspective to solve the motion stylization task with probabilistic style space.

RACon: Retrieval-Augmented Simulated Character Locomotion Control

We introduce an end-to-end hierarchical reinforcement learning method utilizes a task-oriented learnable retriever, a motion controller and a retrieval-augmented discriminator.

Event‐based Human Pose Tracking by Spiking Spatiotemporal Transformer

Our SNNs approach uses at most 19.1% of the computation and 3.6% of the energy costs consumed by the existing methods while achieves superior performance.

Design and source code from Leonid Keselman's website