Jinyuan Mao
Portrait of Jinyuan Mao

Undergraduate Researcher · Zhejiang University

Jinyuan Mao 茅晋源

I am a senior undergraduate at Zhejiang University, pursuing a dual degree in Statistics and Information Management & Information Systems. My research focuses on making image and video generation high-quality, efficient, and interactive. I am applying to PhD programs for Fall 2027.

Most recently, I was a research intern at PixAI, working with Chenxi Liu (University of Victoria) on interactive image generation. Before that, I spent eight months at the ZIP Lab advised by Prof. Bohan Zhuang, working on sparse attention and KV-cache design for interactive, minute-long video generation.

I also enjoy high-performance computing: our team placed 1st globally at the SC25 IndySCC, and I look after the GPU infrastructure at ZJU X-Lab.

  • Image & Video Generation
  • Interactive Generation
  • Efficient AI
jymao@zju.edu.cn CV GitHub Notes
1/20Major rank · GPA 92.5
3×National Scholarship
3First / co-first author papers
#1SC25 IndySCC, global

News

Publications

* equal contribution

SIGGRAPH Asia 2026

BIFE: Better Interaction, Fewer Errors for Minute-Long Video Generation

Zeyu Zhang*, Jinyuan Mao*, Shuning Chang, Yuanyu He, Yizeng Han, Jiasheng Tang, Fan Wang, Bohan Zhuang

SIGGRAPH Asia 2026, Technical Communications

A semantic sparse KV cache retrieves the long-range context that minute-long generation needs, cutting error accumulation and improving VDE-Subject by 22.2%.

arXiv PDF
NeurIPS 2026 · Reviewed

WorldAttention: An Efficient Attention Architecture for Interactive Video World Models

Zeyu Zhang*, Jinyuan Mao*, Dakai An*, Wangbo Zhao, Hanfeng Lu, Jiasheng Tang, Yinghao Yu, Wei Wang, Bohan Zhuang

NeurIPS 2026 Reviewed 5/5/5 · Accept / Accept / Accept

Attention heads in video diffusion transformers are sparse in very different ways. Hybrid Sparse Attention pairs a linear global branch with head-adaptive block-sparse attention for a 2.21× end-to-end speedup at matched quality.

arXiv PDF
Eurographics 2027

HimmBO: High-Dimensional Interactive Model-Merging Bayesian Optimization for Image Generation

Jinyuan Mao, Zhihao Shen, Chenxi Liu

Under review at Eurographics 2027

Two-stage preferential Bayesian optimization scales interactive LoRA merging from a few dozen to 1,080 adapters. A style-aware kernel shares preferences across similar adapters, so it needs under 1/10 of the inference used by prior work.

Paper coming soon
arXiv 2025 · Preprint

IntuiTF: MLLM-Guided Transfer Function Optimization for Direct Volume Rendering

Yiyao Wang, Bo Pan, Ke Wang, Han Liu, Jinyuan Mao, Yuxin Liu, Minfeng Zhu, Xiuqi Huang, Weifeng Chen, Bo Zhang, Wei Chen

arXiv preprint, 2025

An evolution-driven explorer searches the transfer function space while an MLLM evaluates renderings, forming a trial–insight–replanning loop that turns user intent into volume visualizations.

arXiv PDF

Research Experience

  1. PixAI Research Intern

    Jun 2026 – Oct 2026

    Interactive image generation · Advised by Chenxi Liu (University of Victoria)

    • As first author and sole student researcher, proposed HimmBO, a two-stage preferential Bayesian optimization framework that scales interactive LoRA merging from a few dozen to 1,080 adapters. Its level-set screening stage needs no image generation during interaction.
    • Designed a style-aware kernel that transfers preference information across adapters. It reaches a given similarity with under 1/10 of the inference prior work needs, and stays stable as the library grows while baselines degrade.

  2. ZIP Lab, Zhejiang University Research Intern

    Oct 2025 – Jun 2026

    Efficient video generation & world models · Advised by Prof. Bohan Zhuang

    • WorldAttention: found that attention heads in video diffusion transformers have highly heterogeneous sparsity, and designed Hybrid Sparse Attention, pairing a linear global branch with head-adaptive block-sparse attention. Implemented it end to end for a 2.21× inference speedup at matched quality.
    • BIFE: co-designed a semantic sparse KV cache that retrieves relevant history instead of keeping a sliding window, preserving long-range interactions in minute-long video generation. Improved VDE-Subject by 22.2% and VDE-Clarity by 19.4%.

  3. Provincial SRTP Project Project Leader

    Apr 2024 – Jun 2025

    Database agent framework · Zhejiang University

    • Proposed a two-stage trial-and-error learning framework that improves how database agents interact. Filed one national invention patent.

  4. ZJU Supercomputing Team Member

    2024 – Present

    Student cluster competitions & HPC application optimization

    • SC25 IndySCC (1st place globally): built, parallelized and tuned the SST architecture simulator, and searched for the best-performing simulation configurations.
    • ICT Competition (1st Place, National Finals): ported a multimodal LLM (InternVL2.5) to an NPU-based distributed training stack for a 1.15× training speedup.
    • PAC 2025 (National 2nd Prize): sped up floating-point matrix kernels 4× with multi-threading, SME vectorization and block prefetching.

Education

Zhejiang University

Sep 2023 – Jun 2027

B.S. in Statistics & B.Mgt. in Information Management and Information Systems (dual degree) · Hangzhou, China

GPA 92.5/100Rank 1/20

Honors & Awards

Other Experience

  1. ZJU X-Lab Operations & Infrastructure Lead

    Jun 2025 – Present

    Student-run innovation organization

    • Run the GPU clusters with Docker-isolated environments for 10+ projects. Built the private Overleaf, NAS storage, VPN and the organization's website.

  2. Nodesk AI AI Engineering Intern

    Jun 2025 – Aug 2025

    Algorithm development · Hangzhou

    • Built a smart-reply cross-validation module and fine-tuned it for a 15% accuracy gain. Co-authored a patent on a guardrail that combines reranking models with scenario recognition.

Skills

Languages
PythonC/C++SQLLaTeX
ML & Systems
PyTorchDiffusersMPIOpenMPDockerLinux
HPC & Profiling
CUDASlurmNsight SystemsIntel VTuneperf