Ph.D. Student · UT Austin ECE
Efficient generative AI for computer vision
I am a Ph.D. student in the ECE Department at UT Austin, where I am fortunate to be advised by Prof. Diana Marculescu. I received my master's and bachelor's degrees from National Taiwan University.
My research focuses on generative AI for computer vision, with a particular interest in efficiency. If you find that we have shared research interests, please feel free to reach out by email; I am always open to potential collaborations.
I am looking for a research internship. I am available anytime from late December 2026 through Summer 2027. No visa sponsorship needed for the internship.
🎉 LIPAR is accepted to NeurIPS 2026!
Extended my internship at Roblox through December 18, 2026.
Joined Roblox as a research intern (May 25 – Aug 18), working with Tianpei Gu and Xun Huang.
Gave an invited talk on Streaming LIPAR at Daydream (video).
Live demo: Streaming LIPAR editing a webcam feed in real time at ~20 FPS on a single RTX 4090 (X thread)
Proposed a training-free framework that adapts pixel-space compression techniques to latent pruning for video generation. Achieved an average 1.53× throughput boost (19.3 FPS on an RTX 4090) and a 20% memory reduction, without compromising visual quality.
We propose a pipeline that automatically injects artifacts into clean, high-quality synthetic images on a predetermined region, thereby producing pixel-level annotations without manual labeling.
4th Workshop on Image/Video/Audio Quality in Computer Vision and Generative AI, WACV 2025
We found that the presence of artifacts in images is related to the sampling process of the diffusion model, characterized by the "Similarity Trajectory".
By summing adversarial perturbations from different models, human-identifiable features emerge, showing that adversarial perturbations are in fact comprehensible to humans.
GPT-4/4o models are good at recognizing objects but struggle to identify obvious errors in images generated by diffusion models. One explanation is that they only recognize key features rather than understanding the complete structure of objects.
Streaming LIPAR: Training-Free Real-Time Video Editing
Daydream Scope Workshop, online · Video · Announcement
Ph.D. in ECE, University of Texas at Austin
Advised by Prof. Diana Marculescu
M.S. in ECE, National Taiwan University
Advised by Prof. Lin-shan Lee · GPA 4.26/4.30
B.S. in ECE, National Taiwan University
GPA 4.03/4.30