Zefang Wang

ENCODE Lab

Westlake University

Hangzhou, Zhejiang, China

Hi there! I am Zefang Wang (王泽芳).

I am currently a visiting student at the ENCODE Lab, Westlake University, advised by Prof. Huan Wang. I am also pursuing my Master’s degree in Control Engineering at Zhejiang University, supervised by Dr. Guanzhong Tian.

Previously, I received my B.S. degree in Rail Transit Signal and Control from North University of China, where I graduated with the top ranking (1%) in my major.

Research Interests

My research focuses on Efficient AI and Model Compression, with particular emphasis on:

Pruning techniques for visual autoregressive models and diffusion models
Network quantization and binarization
Knowledge distillation for edge deployment
Mobile and edge AI optimization

I am passionate about developing principled and effective compression methods that enable large-scale AI models to run efficiently on resource-constrained devices.

news

Nov 20, 2025	Submitted EVAR to CVPR 2026.
Mar 01, 2025	Started my visiting position at the ENCODE Lab, Westlake University, advised by Prof. Huan Wang. Focus on iOS model deployment and visual autoregressive model compression.
Nov 01, 2024	Our paper on Information Theoretic Framework for Channel Pruning has been accepted to IEEE TNNLS!

selected publications

arxiv

EVAR: Edge Visual Autoregressive Models via Principled Pruning

Zefang Wang, Yanyu Li, Mingluo Su, Simin Xu, Guanzhong Tian, and Huan Wang

In arxiv, 2026

Abs Code Website

We propose a principled OBS-based structured pruning method for visual autoregressive models. Our approach introduces a progressive scale-aware distillation method to address gradient imbalance during next-scale autoregressive fine-tuning. The method achieves 1.8× speedup with only 10% quality loss on single-image generation scenarios on edge devices.
arxiv

OBS-Diff: Accurate Pruning For Diffusion Models in One-Shot

Junhan Zhu, Hesong Wang, Mingluo Su, Zefang Wang, and Huan Wang

In arxiv, 2026

Abs arXiv Website

We propose an OBS-based training-free multi-granularity pruning method adapted for diffusion models. Our approach introduces a timestep-aware Hessian construction with logarithmically decreasing weights to mitigate error accumulation and optimize group pruning efficiency. The method achieves accelerated inference while minimizing visual quality degradation.
TNNLS

An Effective Information Theoretic Framework for Channel Pruning

Yihao Chen and Zefang Wang

IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2024

TNNLS Abs arXiv

IEEE TNNLS

We use a fusion function combining entropy and rank to obtain the information concentration of each layer, which serves as the pruning rate. We propose an explainable Shapley value-based contribution evaluation method as the intra-layer pruning criterion. The method achieves good results in image classification and object detection, demonstrating its effectiveness.