Golden Noise and Sampling Enhancement of Diffusion Models
Organizer
Speaker
Zeke Xie
Time
Wednesday, April 9, 2025 10:00 AM - 11:00 AM
Venue
A3-4-301
Online
Zoom 230 432 7880
(BIMSA)
Abstract
In this talk, we introduce our recent research on noise intialization and sampling enhancement strategies of diffusion models, spaning from image generation to video generation. Diffusion models is a class of mainstream generative modeling method for modern AIGC. Diffusion models can gradually denoise a Gaussian noise into a clean generated result, where such random Gaussian noise can naturally lead to diversitified results. What can this motivate us? We try to understand why some noise initializations are “golden noises” that can generate better results, and leverage such “golden noise” insight to further enhance diffusion sampling quality. This talk includes three parts: 1) Golden Noise for Text-to-Image Diffusion Models; 2) [ICLR 2025] Zigzag Diffusion Sampling; 3) [ICLR 2025] Leveraging Image Diffusion Models for Enhanced Video Synthesis.
Speaker Intro
Dr. Zeke Xie is an Assistant Professor at Information Hub, Hong Kong University of Science and Technology (Guangzhou). He is leading Xie Machine Learning Foundations Lab (xLeaF Lab) that generally interested in understanding and solving fundamental issues of modern AI, particularly large models, by scientific principles and methodology. He currently focuses on optimization and inference of Large Models and Generative AI. Previously, he was a researcher at Baidu Research responsible for large models and AIGC research. He obtained Ph.D. and M.E. both from The University of Tokyo. He received multiple faculty research awards from the industry, including ByteDance and Baidu.