中文 / English

Zhihan Zhu 朱旨函 (Zhihan Zhu)

I am an undergraduate researcher in the Artificial Intelligence Lab at SUSTech, advised by Prof. Zhihai He (IEEE Fellow). My research focuses on Generative AI, particularly Diffusion Models, Flow Matching, and Rectified Flow Inversion and Semantic Editing. I am also broadly interested in Multimodal Large Language Models, LLM Agents, and Long-Video Understanding and Visual Memory.

My long-term goal is to develop unified visual models that integrate visual understanding and generation within a single framework.

我是南方科技大学电子与电气工程系信息工程专业本科生,现为南方科技大学人工智能实验室本科成员,导师为 何志海教授(IEEE Fellow)。我的研究方向聚焦于 生成式人工智能,尤其是 Diffusion ModelsFlow MatchingRectified Flow 反演与语义编辑。我同时对 多模态大语言模型 (MLLMs)LLM Agents长视频理解与视觉记忆 等方向具有浓厚兴趣。

我的长期研究目标是构建兼具视觉理解与生成能力的统一视觉模型。

Zhihan Zhu

Education 教育经历

Publications 论文

HyGRAIL Thumbnail

HyGRAIL: Cost-Aware and Evidence-Grounded Scientific Hypothesis Discovery over Knowledge Graphs

Yihang Sun, Zhihan Zhu, Zhiyuan Jiang, Jingyi Ge, Zixuan Li, Jiaxuan You
Under Review · EMNLP 2026
A cost-aware GNN–LLM framework for scientific hypothesis discovery over knowledge graphs, integrating heterogeneous graph triage, knowledge-graph evidence retrieval, and LLM-based hypothesis review.
提出了一种面向科学假设发现的成本感知 GNN–LLM 框架,融合异质图分流、知识图谱证据检索与基于 LLM 的假设审查,实现对稀疏与模糊候选关系的可靠验证。
Rectified Flow Inversion Thumbnail

Runge-Kutta Approximation and Decoupled Attention for Rectified Flow Inversion and Semantic Editing

Weiming Chen, Zhihan Zhu, Yijia Wang, Zhihai He
Under Review · IEEE Transactions on Image Processing, 2025 · arXiv:2509.12888
We propose 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 模型高阶反演方法。通过引入解耦扩散 Transformer 注意力机制(DDTA),实现了极高的重建保真度与精确的语义控制。
Generative Semantic Coding Thumbnail

Generative Semantic Coding for Ultra-Low Bitrate Visual Communication and Analysis

Weiming Chen, Yijia Wang, Zhihan Zhu, Zhihai He
Under Review · IEEE Transactions on Image Processing, 2025 · arXiv:2510.27324
A generative semantic coding framework that combines deep compression with rectified-flow generation, supporting both ultra-low-bitrate visual communication and downstream visual analysis.
提出了一种生成式语义编码框架,将深度压缩与 Rectified Flow 生成相结合,支持超低比特率的视觉通信以及下游视觉分析任务。
Latent Bias Alignment Thumbnail

Latent Bias Alignment for High-Fidelity Diffusion Inversion in Real-World Image Reconstruction and Manipulation

Weiming Chen, Qifan Liu, Siyi Liu, Yijia Wang, Zhihan Zhu, Zhihai He
Under Review · IEEE Transactions on Circuits and Systems for Video Technology, 2026 · arXiv:2603.23903

Patents 发明专利

Honors & Awards 荣誉与奖项

Technical Skills 技术能力

Programming: Python, C/C++, Java, MATLAB
Deep Learning: PyTorch, Hugging Face Diffusers, Stable Diffusion, FLUX, SDXL, DiT, ControlNet, VAE, GNNs, vLLM, verl, LLM/MLLM pipelines
Systems: Linux, Git, CUDA, HPC/LSF, Multi-GPU Training, LaTeX
Languages: Chinese (native), English (IELTS 7.0, CET-6 545)