Benteng Sun bio photo

Benteng Sun

Undergraduate Student

πŸ“ Shenzhen, China

🏫 Harbin Institute of Technology

Email

Google Scholar

GitHub

LinkedIn

Research Interests

Machine Learning
Computational Imaging
Generative Models

About Me πŸ‘‹

Benteng Sun

Hi there, I’m Benteng Sun 🎨

I am Benteng Sun (ε­™ε₯”θ…Ύ), an undergraduate student in Data Science (Mathematics Category) at Harbin Institute of Technology, Shenzhen (graduating in 2026). My research centers on machine learning, sparse representation, and computational imaging, advised by Prof. Yongyong Chen.

I have published at ICLR and participated in top-tier competitions. Currently, I am exploring efficient and controllable generation paradigms, aiming to build scalable frameworks for images, videos, and 3D content.

πŸ“§ Contact: SMARK2019@outlook.com

Research Interests

  • Multimodal Understanding & Generation β€” cross-modal synthesis, language-guided visual generation, modality fusion

  • Generative Modeling & Controllable Generation β€” structure-guided diffusion, efficient sampling, scalable 2D/3D/video frameworks

  • Computational Imaging & Low-Level Vision β€” spectral imaging, MRI acceleration, physics-driven restoration, compressed sensing

Selected Publications

Deep LoRA-Unfolding Networks for Image Restoration
Xiangming Wang, Haijin Zeng, Benteng Sun, Jiezhang Cao, Kai Zhang, Qiangqiang Shen
IEEE TIP 2026 Β· IF 10.6
LoRun Framework
A novel deep unfolding framework that introduces stage-specific LoRA adapters to dynamically modulate denoising behavior across unfolding stages. Achieves NΓ— parameter reduction for N-stage networks while maintaining on-par or better performance across multiple image restoration tasks.
Keywords: deep unfolding, low-rank adaptation, image restoration
πŸ“„ Paper
Spectral Compressive Imaging via Unmixing-driven Subspace Diffusion Refinement
Haijin Zeng*, Benteng Sun*, Yongyong Chen, Jingyong Su, Yong Xu
ICLR 2025 Spotlight Β· Top 4.79%
PSR-SCI Pipeline
A novel subspace diffusion refinement framework for spectral compressive imaging, recovering high-frequency details from minimal measurements. Achieved 38.14 dB PSNR on KAIST dataset with 10Γ— speedup over existing diffusion models.
Keywords: spectral imaging, sparsity modelling, diffusion refinement
πŸ“„ Paper πŸ’» Code 🌐 Project

Competitions & Projects

  • The 19th β€œChallenge Cup” National Competition First Prize
    Achieved over 95% accuracy in radar-infrared fusion object detection under limited-sample, noise-intensive conditions.

  • NTIRE Challenge (CVPR 2025)
    Developed a diffusion-based image restoration system, compressing model parameters from 1.77B to 2.2M while retaining high visual quality.

  • SMARK Media Tools Open Source
    GPU-accelerated photography management toolbox with intelligent grouping and aesthetic scoring. πŸ’» GitHub

  • MetaMusic: Cross-Modal Music-Driven Visual Generation Top 1/83
    Cross-modal generation pipeline using CLIP embeddings for music-to-image translation. πŸ’» Project

Honors & Awards

  • First Prize (National), "Challenge Cup" β€” MoE, PRC, 2024
  • Meritorious Winner (Top 6%), Mathematical Contest in Modeling β€” COMAP, USA, 2024
  • First Prize, 15th Chinese Mathematics Competition, 2023
  • Champion (1/83), Freshman Annual Research Project β€” HITSZ, 2023

Scholarships

  • National Scholarship (Top 0.2%) β€” MoE, PRC, 2024–2025
  • High-Level Innovation Award (Top 0.1%) β€” HIT, 2024 (Awarded 50,000 RMB)