Yuxiang Fu

I am a PhD student in the ECE department at UBC, and a member in DSL lab, advised by Prof. Lele Wang and Prof. Renjie Liao. My PhD is supported by the Four Year Doctoral Fellowship (4YF).

Previously, I received an Honours degree in CS from UBC under the supervision of Prof. Andrew Roth. During my undergraduate studies, I am fortunate to intern at both the BCCRC Roth Lab and Vector Institute.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  Github

• [2026/10] 🏆 I am selected as a Top Reviewer for NeurIPS 2026.
• [2026/09] I am serving as an invited reviewer for AAAI 2027.
• [2026/09] I started my PhD in ECE at UBC, supported by the Four Year Doctoral Fellowship (4YF).
• [2026/05] Served as a core developer of the AI/CRV 2026 conference website.
• [2025/08] Checkout our new blogs. Explore the power of diffusion distillation and the future of generative models in motion prediction.
• [2025/02] 🎉 One paper gets accepted to CVPR 2025.
• [2025/02] 🎉 One paper gets accepted to Robotics and Automation Letters 2025.
• [2025/01] I received the GSI award from UBC.
• [2024/04] 🎉 One paper gets accepted to CRV 2024 Workshop and it is selected as Oral Presentation.

My research focuses on developing efficient and scalable distillation methods for generative models, particularly diffusion and flow-based models. I am interested in extending these techniques beyond image generation to video generation, language modeling, world modeling, and vision-language-action (VLA) models for embodied intelligence and robotics. Some papers are highlighted.

We propose Consistent Distribution Matching (CDMD), a data-free and simulation-free distillation method that unifies sample generation and score estimation in one student network, using only a frozen teacher and a trainable student with a single objective. CDMD attains a 1-NFE FID of 2.04 and a 4-NFE FID of 1.37 on ImageNet 256×256.

We present a novel Motion prediction conditional Flow matching model, termed MoFlow, to generate K-shot future trajectories for all agents in a given scene. In addition, by leveraging the Implicit Maximum Likelihood Estimation (IMLE), we propose a novel distillation method for flow models that only requires samples from the teacher model.

We introduce TGD, a diffusion-based human trajectory generation model, along with a trainable student model leveraging the IMLE scheme to align with the teacher diffusion model’s distribution at any intermediate diffusion timestamp.

Accurate single-cell segmentation is essential for spatial omics analysis, yet existing methods rely heavily on expert-driven annotations and separate statistical or biological evaluation strategies, highlighting the need for a unified assessment approach. ESQmodel alleviates this limitation.

We aim to elucidate the fundamental nature of rotational impedance and presents a comprehensive, unified framework for formulating rotational impedance using Lie algebra and Noether's theorem. In particular, we utilized quaternions and rotation matrices to represent rotational motion within our proposed framework to ensure theoretical validity.

A slight tweak of the variational autoencoder that controls the primary tissue effect of the bulk RNA sequencing data from PCAWG/ICGC dataset. PCVAE offers the ability to uncover novel connections across heterogeneous cancers in a site-effect-free environment.

This webpage is build based on Jon Barron's website and deployed on Github Pages.

Last updated: © 2026 Yuxiang (Felix) Fu