cv

Education, research experience, internships, publications, and selected honors.

Basics

Name Yi LIU
Label Master's Student
Email liuyi61@tongji.edu.cn
Phone (+86) 183-2641-8258
Url https://ly403.github.io/
Summary Yi Liu is a master's student at ADMIS Lab, Tongji University, advised by Associate Professor Wengen Li. His research focuses on efficient and controllable diffusion and flow matching models for image restoration, image and video generation, and digital-human motion generation.

Education

  • 2024.09 - 2027.03

    Shanghai, China

    M.Eng.
    Tongji University
    Computer Science and Technology, School of Computer Science and Technology
    • Average score: 93.88/100
    • Advisor: Associate Professor Wengen Li
    • Expected graduation: March 2027
    • National Scholarship for Graduate Students
  • 2021.09 - 2023.03

    Shanghai, China

    Minor
    Fudan University
    Finance
  • 2020.09 - 2024.07

    Shanghai, China

    B.Eng.
    Tongji University
    Computer Science and Technology, College of Electronics and Information Engineering
    • Cumulative GPA: 4.65/5.0 (91.49/100)
    • Admitted to the master's program by recommendation (examination waived)

Work

  • 2026.05 - Present

    Shanghai, China

    Research Intern
    ByteDance, GenAI & Ads Infra
    Efficient generative models and inference acceleration.
    • Research on generation efficiency and hardware-aware acceleration.
  • 2025.06 - 2026.05

    Hangzhou, China

    Research Intern
    Alibaba Cloud, Tongyi Lab (Multimodal Interaction)
    Long-horizon co-speech motion generation for 3D digital humans.
    • Proposed PASPA, a blockwise flow matching model that summarizes history into a persistent state and models current motion with blockwise attention.
    • Developed long-sequence generation with a fixed-size history state; first-author paper at NeurIPS 2026.

Publications

Projects

  • 2025.07 - 2026.03
    Controllable Image Restoration via Disentangled Stochastic Interpolants
    Project Leader, ADMIS Lab, Tongji University; advised by Associate Professor Wengen Li.
    • Proposed DiSI, which separates regression and generation components in stochastic interpolants to reconcile efficient reconstruction with realistic generation.
    • Developed a single model with inference-time control over the reconstruction–generation trade-off across deraining, denoising, super-resolution, and inpainting.
  • 2023.10 - 2024.11
    Effective Cloud Removal by an Improved Mean-Reverting Denoising Model
    Project Leader, ADMIS Lab, Tongji University; advised by Associate Professor Wengen Li.
    • Proposed EMRDM, a modular mean-reverting diffusion framework that directly transforms cloudy images into cloud-free images.
    • Designed a unified network for mono-temporal and multi-temporal inputs; first-author paper at CVPR 2025, with a released implementation.
  • 2024.02 - 2024.06
    Efficient Reconstruction for Snapshot Compressive Imaging
    Visiting Student, SCI Lab, Westlake University; mentored by Professor Xin Yuan.
    • Adapted efficient architectures, including linear attention and state space models, for lower-cost snapshot compressive imaging reconstruction.
    • Reviewed more than 50 papers and presented a technical report at group meetings.
  • 2023.03 - 2023.09
    SenseAir: AI-Driven Compressed Air System Management Platform
    Core Team Member, Tongji University.
    • Developed a digital-twin platform for compressed-air system operation, maintenance, and optimization.
    • First inventor of a granted patent, with a registered software copyright.

Awards

Skills

Programming and Tools
Python
PyTorch / PyTorch Lightning
C/C++
Linux / Git
LaTeX

Languages

Chinese
Native speaker
English
IELTS 7.0 (Reading 8.5, Listening 7.5, Speaking 6.0, Writing 6.5); CET-6 557

Interests

Efficient and Controllable Generative Models
Diffusion models
Flow matching
Few-step sampling
Conditional control
Visual Generation and Restoration
Image restoration
Image and video generation
Digital-human motion generation
Temporally coherent long-sequence modeling
Efficient Architectures and Implementation
Lightweight architectures
Linear attention and state space models
Hardware-aware acceleration

Volunteer

  • 2021.09 - 2023.01

    Shanghai, China

    Volunteer
    Volunteer Work Department of Youth League Committee in Tongji University
    1) Served as a dedicated long-term volunteer in the Volunteer Work Department, facilitating package deliveries across various campuses within Tongji University by following the department head's instructions. 2) Engaged in short-term volunteering opportunities during school events, with a specific focus on charity activities for education. 3) Assisted school staff during school events to provide valuable support.
  • 2020.09 - 2022.01

    Shanghai, China

    Volunteer
    Tongji Association Center
    1) Assumed the role of meeting secretary, diligently documenting meeting proceedings weekly following departmental requirements at the Tongji Association Center. Collaborated closely with co-workers to ensure accurate recording of meetings. 2) Actively contributed to the organization of various club activities, including the Club Gala and Club Carnival at Tongji University. Collaborated effectively to assist clubs in planning and executing successful events.

References

Associate Professor Wengen Li
Professor Wengen Li is the supervisor of Yi Liu. He is an associate professor at the College of Computer Science and Technology, Tongji University. His research interests include Multimodal Intelligence and Spatial Temporal Data Analysis.