Junhan Zhu 诸俊涵

News

Publications

• First author / co-first (listed first)
• Co-first (listed second)

A summary of my research is available in these slides. ^† Equal contribution; ^* corresponding author.

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

Junhan Zhu, Hesong Wang, Mingluo Su, Zefang Wang, Huan Wang^*

ICLR 2026 · First author

project page / code / paper / SDXL / SD3.5-Large

Propose the first training-free, one-shot pruning framework for diffusion models, demonstrating broad applicability across diverse architectures and pruning granularities.

Cross-Resolution Diffusion Models via Network Pruning

Jiaxuan Ren^†, Junhan Zhu^†, Huan Wang^*

CVPR 2026 Findings · Co-first author (second listed)

project page / paper

Introduce CR-Diff, a framework that repurposes network pruning to enhance generalizability by removing "adverse weights" that cause degradation at non-default resolutions.

MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond Code

Xueyi Chen^†, Shiyu Liu^†, Xin Jin, Yuhua Zheng, Xin Li, Haolei Bai, Junhan Zhu, Huan Wang^*

arXiv, 2026

project page / code / paper

A benchmark of 74 paired GPU kernel problems that evaluates whether experience distilled into natural-language skills transfers between CuTe DSL and TIRx.

LVOmniBench: Pioneering Long Audio-Video Understanding Evaluation for Omnimodal LLMs

Keda Tao^†, Yuhua Zheng^†, Jia Xu, Wenjie Du, Kele Shao, Hesong Wang, Xueyi Chen, Xin Jin, Junhan Zhu, Bohan Yu, Weiqiang Wang, Jian Liu, Can Qin, Yulun Zhang, Ming-Hsuan Yang, Huan Wang^*

Under Review · Adopted by Qwen3.8-Omni for evaluation.

project page / code / paper / dataset

Launch LVOmniBench, a pioneering, comprehensive benchmark for evaluating OmniLLMs in Long Audio-Video Understanding.

Selected Projects

SparAlloc: A Simple and Modular Framework for Decoupled Sparsity Allocation in Layerwise Pruning for LLM

Junhan Zhu

Open Source Project

code

A modular benchmark toolkit for decoupled sparsity allocation in LLM pruning. Enables fair comparison and flexible combination of pruning algorithms.

Paper2Demo: Turn Any Academic PDF into a Stunning Demo Video — Fully Automatically

Junhan Zhu

Open Source Project

code

An automated pipeline that converts research papers into ~30s animated video presentations. Extracts text and figures via PyMuPDF, structures content with LLM APIs, and renders polished videos with domain-aware theming using Remotion.

Writing

Lonely Zone

I share my research and personal thoughts on my blog.

Education

The Chinese University of Hong Kong

Incoming Ph.D. Student

Starting 2027

Advisor: Prof. Weiyang Liu
Research interests: Efficient AI

Nanyang Technological University

Exchange Student

Jan 2026–Apr 2026

Research interests: Efficient AI

Westlake University

B.Eng. in Electronic and Information Engineering

Jul 2023–present (expected 2027)

Rank: 1 / 23. Admitted by skipping the final year of high school.
Research interests: Efficient AI

Selected Honors & Awards

• 2026
  Zhejiang Provincial Government Scholarship, Zhejiang Provincial Government (6,000 RMB).
• 2026
  SenseTime Scholarship, SenseTime (20,000 RMB) — one of only 30 recipients nationwide.
• 2025
  Outstanding Bachelor's Student, Westlake University, for outstanding academic performance.
• 2024, 2025
  Outstanding Undergraduate Scholarship, Westlake University (2,000 RMB each year).
• 2024, 2025
  Innovation Award, Westlake University, for innovative research.
• 2024
  Hongyi Scholarship, Westlake University, for meaningful social activities.

Academic Services

Conference reviewer: NeurIPS 2026, ICLR 2027.
Journal reviewer: IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).