I am a fifth-year PhD candidate in the Department of Computer Science at the University of Rochester, advised by Prof. Chenliang Xu. Previously, I spent one wonderful year as a research assistant at the Chinese University of Hong Kong, working with Prof. Chi-Wing Fu on 3D vision. I received my B.Eng. from ESE Department, Nanjing University in 2019. In my undergrad, I worked with Prof. Zhan Ma on image compression.

I am broadly interested in developing machine learning models to understand how human perceive the surrounding scenes from multi-modal inputs and utilize the perception for action. Specifically, I am working on multimodal video understanding and generation.

Research opportunities: I am open to collaborating on research projects. Shoot me an email if you are insterested.

Email  /  CV  /  Google Scholar

No fine-tuning, no task-specific data, just latent inversion + text-conditioned denoising to isolate any sound you describe.

Where and why you should care about frequency space in diffusion models.

We learn from movies to transform audio to deliver appropriate highlighting effects guided by the accompanying video.

A survey on the recent Large Language Models for video understanding.

We introduce VidComposition, a benchmark designed to assess MLLMs' understanding of video compositions

We use pretrained text-guided diffusion models to scale up/down concepts in image/audio.

A new take on the audio-visual separation problem with the recent generative diffusion models.

We achieve joint audio-visual editing under language guidance.

Thinking of the equivalent of 3D Gaussian Splatting and volumetric primitives for the human body soundfield? Here, we introduce Acoustic Primitives.

We propose a novel method of synthesizing real-world audio-visual scenes at novel positions and directions.

We explore the problem of sound source visual localization in egocentric videos, propose a new localization method and establish a benchmark for evaluation.

A non-local attention based method for point cloud denoising in both synthetic and real scenes.

An image compression system under extreme condition, e.g., < 0.05 bits per pixel (bpp).

TrustFM: Workshop on Trustworthy Foundation Models @ ICCV 2025

CVPR (2023 - 2025), AAAI (2023 - 2025), ACM MM (2023 -2025), ICCV (2025)

TMM, TIP, SIGGRAPH

The template is based on Jon Barron's website.