About

Hi, I’m Kevin, currently a Research Scientist at Meta Superintelligence Labs (MSL). My research interests are in 3D perception and world models, multimodal foundation models, and a mechanistic understanding of how neural networks learn, forget, and break.

I previously spent 4 years as a researcher in Fundamental AI Research (FAIR) at Meta, as well as a brief stint in Facebook AI Applied Research (FAIAR). Prior to that, I studied at Duke University, of which I am a triple alumnus: I completed my PhD (+MS) while advised by Lawrence Carin, and before that, earned an ECE/BME double-major BSE. As an undergraduate, I was able to work on image processing with Guillermo Sapiro and brain-computer interfaces with Leslie Collins.

Outside of research, I was a member of Sabrosura, Duke’s premier Latin dance troupe, and enjoyed playing Ultimate frisbee and tennis (though these days I quite enjoy pickleball). I also enjoy mentoring: I’ve mentored and advised many students over the years through Engineering a Community, Engineering Alumni Council, and E-team. I’ve also served as an interviewer for Duke undergraduate admissions.

Email: kevinjliang [at] meta [dot] com

Selected Projects

SAM 3D: 3Dfy Anything in Images

SAM 3D

Reconstruct the full 3D shape, texture, and layout of objects from a single image. CVPR 2026 Best Paper Honorable Mention. [website][paper][demo][code]

Ego-Exo4D

Ego-Exo4D

A massive multimodal, multiview dataset capturing skilled human activity from synchronized first- and third-person perspectives. CVPR 2024 Oral. [website][paper][video]

Fast3R

Fast3R

3D reconstruction from 1000+ images in a single forward pass, no global alignment required. CVPR 2025. [paper][code]