Juze Zhang

I am currently a postdoc at Stanford University working with Prof. Ehsan Adeli, affiliated with the Stanford Vision and Learning Lab (SVL) and the Stanford Translational AI Lab (STAI). Before joining Stanford, I worked on digital humans at ShanghaiTech University, where I was fortunate to work closely with Jingyi Yu, Jingya Wang, and Lan Xu. I received my Ph.D. from the University of Chinese Academy of Sciences.

I consider myself a multidisciplinary researcher who enjoys building across the full stack — hardware, systems, algorithms, and frontier multimodal frameworks. On the systems side, I maintained a multi-view dome with over 100 cameras at ShanghaiTech, built the DailyCap Studio from scratch as the first student on the project, and assembled the HOI-X Studio (12 Z-CAM + 8 OptiTrack + RealSense). I also led a small team to design wireless IMUs from scratch, embedding them into objects, clothing, and even flexible PCBs worn on the face (see Projects).

On the algorithm side, my current research centers on multimodal alignment — in particular world/video models, speech LLMs, and the hardware stack for robotics. At Stanford, I built ViBES, an early exploration of the mixture-of-modality-experts (MoME) design — self-attention with modality-specific parameters and fractional embeddings on top of a speech LLM — which became a 3D conversational agent in early 2025. Encouraged by the success of this mixture-of-transformers (MoT) recipe, I then initiated OpenWAM, applying the same design to video generation in place of the speech LLM. I view this line of work as essentially sequential-transformer-to-sequential-transformer modeling.

Email  /  Scholar  /  Twitter  /  Github

Research

Selected papers are listed below (* denotes equal contribution). For the full list, please see my Google Scholar.

Projects

Beyond papers, I love building the full stack with my hands — capture studios, sensors, and the systems that power the datasets above.