Preprint · 2026
Taming Outlier Tokens in Diffusion Transformers
Dual-Stage Registers address outlier tokens in both the encoder and diffusion transformer, improving ImageNet-256 FID from 5.89 to 4.58 at 80 epochs.
Hi, I'm Yifei Wang.
Second-year PhD student · Rice University
I work on generative models, including diffusion, flow matching, and the building blocks that make them more efficient and controllable.
At Rice, I am advised by Prof. Chen Wei. My research connects the theory of generative modeling with the practical challenges of training and inference.
Selected work
All publications
NeurIPS 2025
Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction
A unified f-divergence framework connecting more than 10 one-step diffusion distillation methods, with a one-step FID of 1.02 on ImageNet 64 × 64.
* Equal contribution.
News
- New blogpost: Struggling Toward Generative Model Evaluation.
- New post: EMA: A Quiet Hyperparameter That Moves Diffusion Leaderboards.
- Visited Prof. Alan Yuille's lab at Johns Hopkins University.
- Released DSR in collaboration with Apple.
Earlier updates
- Uni-Instruct, with Xiaohongshu Inc., accepted to NeurIPS 2025.
- Started my PhD at Rice University with Prof. Chen Wei.
- EM-Diffusion accepted to NeurIPS 2024.
Background
Before Rice, I received my B.S. from Peking University, advised by Prof. He Sun and Dr. Weijian Luo. In May 2026, I visited Prof. Alan Yuille's lab at Johns Hopkins University.