Heng (Henry) Yu

I am a CS PhD student at Stanford University, advised by Prof. Ehsan Adeli in the STAI Lab. My research focuses on video generation and understanding, world modeling, and Embodied AI. I am also interested in their real-world applications, such as healthcare.

Before Stanford, I received my master's degree from the CMU Robotics Institute, where I worked on 3D vision with Prof. Laszlo Jeni. I also collaborated closely with Prof. Berkin Bilgic at Harvard Medical School on MRI reconstruction, and with Prof. Cheng Jin at Shanghai Jiao Tong University on medical vision.

I obtained my bachelor's degree from Tsinghua University, majoring in Automation with a second major in Economics and Management. My long-term goal is to build AI systems that are both technically strong and practically useful, especially in embodied perception, visual generation, and healthcare settings.

I enjoy working with motivated people across academia and industry. If you'd like to collaborate, feel free to reach out.

yuheng[at]stanford.edu

CV  /  Google Scholar  /  GitHub  /  LinkedIn

I am primarily interested in video generation and understanding, world modeling, and Embodied AI. More broadly, I want to build visual intelligence systems that can model dynamic environments, understand how the world evolves over time, and support decision-making and interaction in open-world settings.

• Nov 2025: SocialGen is accepted by 3DV 2026.
• Sep 2024: 4Real paper is accepted by NeurIPS 2024.
• Feb 2024: CoGS paper is accepted by CVPR 2024.
• Feb 2023: DyLiN paper is accepted by CVPR 2023.
• Feb 2023: SubZero abstract is accepted by ISMRM 2023 as a power pitch.
• Dec 2022: CoNFies paper is nominated as a best paper candidate.

Reviewer for CVPR, ICCV, ECCV, NeurIPS, SIGGRAPH, MICCAI, ISBI, Computer Graphics Forum, and ISMRM.

* indicates co-first author. Please see my Google Scholar for the full publication list.

SocialGen is the first unified motion-language model for multi-human interactions, enabling state-of-the-art social motion modeling with a new representation, benchmark, and dataset.

We propose 4Real, the first photorealistic text-to-4D scene generation pipeline.

CoGS enables controllable Gaussian Splatting for dynamic scenes with direct scene manipulation and real-time control.

DyLiN extends light field networks to dynamic, non-rigid scenes with strong visual fidelity and efficiency.

CoNFies is a fully automatic controllable neural representation for face self-portraits.

SubZero improves subspace-based zero-shot self-supervised MRI reconstruction with a parallel architecture and attention mechanism.

eRAKI accelerates RAKI by directly learning a coil-combined target for robust and efficient MRI reconstruction.

A multi-task deep learning framework for tumor segmentation and treatment response prediction from longitudinal medical images.

Website template adapted from Jon Barron.