Bing He

I'm a PhD student at Shanghai Jiao Tong University, where I am advised by Li Song. My research interests lie at the intersection of 3D reconstruction, generative modeling, and rigging, with a specific focus on advancing state-of-the-art 3D vision tasks.
My work has been recognized at premier venues including ICCV, NeurIPS, CVPR, ECCV and ICLR. Most recently, I developed SurfSplat, which was accepted at ICLR 2026.
I am currently doing a research internship at Tencent, Shenzhen, and I warmly welcome academic exchange and discussion—feel free to reach out.
I am currently seeking research internship opportunities in the United States for 2026 and 2027 (with a preference for California). If you're interested in collaborating or discussing my work, please feel free to reach out via email or DM.

Email  /  CV  /  Scholar  /  LinkedIn  /  Github

Research

I'm interested in computer vision, generative AI, and image processing. Most of my research is about inferring the physical world (shape, motion, color, light, etc) from images, usually with gaussian splatting. Some papers are highlighted.

Reconstructing 3D scenes with feed-forward 2DGS! 🏄🏻 By combining neural networks with graphics priors, we’ve solved common convergence problems and persistent visual artifacts. The result? Crystal-clear reconstructions that hold up even when you zoom in.

We propose GLIC, a content-adaptive learned image compression framework that leverages dual-scale Graph Neural Networks with adaptive connectivity to flexibly capture both local and global redundancy, achieving state-of-the-art rate-distortion performance.

We reinterpret ultra-low-bitrate image compression as a next-frame prediction task, decoding from a compact, semantically faithful anchor frame to the full image via a pretrained video diffusion model—achieving over 50% bitrate savings and up to 5× faster decoding than DiffC on CLIC2020.

Content-Aware Mamba meets Learned Image Compression! 🐍 While standard SSMs are often restricted by rigid, content-agnostic scans, our latest work, CMIC, introduces Content-Aware Mamba (CAM). By replacing fixed raster scans with a content-adaptive token permutation strategy and injecting global priors to break sequential dependency, we’ve pushed the boundaries of efficiency and performance. The result? A state-of-the-art LIC model that surpasses VTM-21.0 by up to 21.34% in BD-rate across standard benchmarks.

Exploring heterogeneous 3D motion representation for deformable 3D Gaussian Splatting! ⚡ By bridging neural networks with graphics priors, we’ve significantly accelerated the training process and achieved robust results—even on datasets with extreme, high-frequency motion. H3D-DGS makes deforming scenes faster and more stable than ever.

Knowledge distillation meets learned image compression! We’ve developed a streamlined framework that delivers a smaller, faster model with significantly reduced computational overhead, all while maintaining state-of-the-art performance.

Multi-view video compression is a major bottleneck for the future of VR. Our new framework introduces an implicit-explicit integrated representation that tackles this head-on.

Miscellanea