Zhihan Zhu 朱旨函

Undergraduate Student @ SUSTech (EEE Department) 南方科技大学(SUSTech)电子系 本科生
Email: 12312326 [at] mail.sustech.edu.cn

Hi, I am a third-year undergraduate student at the Artificial Intelligence Lab of Southern University of Science and Technology, under the supervision of Professor Zhihai He (IEEE Fellow). I am interested in Generative AI, Diffusion Models, and Image Restoration. Currently, I am focusing on Rectified Flow models and their applications in high-fidelity image inversion and semantic editing, and my research interests also include exploring the capabilities of DiT.

你好,我是南方科技大学人工智能实验室的一名大三本科生, 导师是 何志海教授 (IEEE Fellow)。我的研究兴趣主要集中在 Generative AI(生成式人工智能)、Diffusion Models(扩散模型)以及 Image Restoration(图像复原)。 目前,我正专注于 Rectified Flow 模型及其在高保真图像反演(Inversion)和语义编辑中的应用。同时,我也对探索 DiT 的潜能保持着浓厚的研究兴趣。

Zhihan Zhu

Education 教育经历

Southern University of Science and Technology (SUSTech) 南方科技大学 (SUSTech)
B.S. in Electronic and Electrical Engineering (Expected June 2027) 电子与电气工程系 工学学士(预计 2027 年 6 月毕业)
GPA: 3.85/4.0 Rank: 10/46 排名: 10/46

News & Awards 新闻与奖项

2025 Gold Award, China International College Students’ Innovation Competition (Guangdong Prov.) 金奖,中国国际大学生创新大赛(广东赛区)
2025 Outstanding Student, SUSTech 优秀学生,南方科技大学
2025 Second-Class Scholarship for Academic Excellence 二等奖学金(学业优秀奖)

Papers 学术论文

Rectified Flow Inversion
Runge-Kutta Approximation and Decoupled Attention for Rectified Flow Inversion and Semantic Editing
Weiming Chen, Zhihan Zhu, Yijia Wang, Zhihai He
arXiv:2509.12888 (Sept 2025)

Proposed a high-order inversion method for rectified flow models using a Runge–Kutta solver, enabling state-of-the-art fidelity and precise semantic control via Decoupled Diffusion Transformer Attention (DDTA).

提出了一种基于 Runge–Kutta 求解器的 Rectified Flow 模型高阶反演(Inversion)方法。通过引入解耦扩散 Transformer 注意力机制(DDTA),实现了目前最优的重建保真度与精确的语义控制。

Generative Semantic Coding
Generative Semantic Coding for Ultra-Low Bitrate Visual Communication
Weiming Chen, Yijia Wang, Zhihan Zhu, Zhihai He
arXiv:2510.27324 (Oct 2025)

Developed a generative semantic coding framework integrating image generation with deep compression via rectified flow models, enabling ultra-low bitrate visual communication.

开发了一种生成式语义编码框架,通过 Rectified Flow 模型将图像生成与深度压缩相结合,实现了超低比特率的视觉通信。

Selected Projects 精选项目

one step SR
One-step Diffusion for Image Super-Resolution
May 2025 – Present 2025年5月 – 至今

Developed auxiliary plug-in SR modules for MeanFlow and SDXL-Turbo pipelines. Trained on HPC clusters (8×A100 GPUs) using a custom dataset derived from DIV2K and Flickr2K. Achieved one-step super-resolution conditioned on reference images.

为 MeanFlow 和 SDXL-Turbo 开发了辅助的外挂式超分(SR)模块。在 HPC 集群(8×A100 GPUs)上使用基于 DIV2K 和 Flickr2K 构建的自定义数据集进行训练,实现了以参考图像为条件的单步超分辨率重建。

PyTorch HPC/LSF Diffusion Models
2048 AI
Intelligent Agent for 2048 Game
Course Project (Top 5%) 课程项目 (排名前 5%)

Designed an autonomous AI agent for the 2048 game. Open-sourced implementation on GitHub.

设计并实现了一个全自动玩 2048 游戏的 AI Agent。代码已在 GitHub 开源。

Patents 专利

Zhihai He, Weiming Chen, Zhihan Zhu, et al. "Image Editing Method Based on Attention Decoupling..." "基于注意力解耦的图像编辑方法..." CN 202511281645.1.
Zhihai He, Weiming Chen, Yijia Wang, Zhihan Zhu. "Template-Replacement Image Compression & Reconstruction..." "基于模板替换的图像压缩与重建方法..." CN 202511281423.X.

Technical Skills 技术技能

Python PyTorch Hugging Face Diffusers CUDA NumPy OpenCV LaTeX Git Linux HPC (LSF, Multi-GPU Training) Stable Diffusion DiT Matplotlib FLUX STM32