AI Scientist, Roblox
I am an AI Scientist working on large-scale generative models for photorealistic gaming at Roblox Reality. Previously, I was part of the founding team at Morpheus AI under Xun Huang, which was acquired by Roblox.
I received a Ph.D. in Computer Science from the University of Michigan, advised by Justin Johnson and David F. Fouhey. I also interned at NVIDIA Deep Imagination Research with Chen-Hsuan Lin and Tsung-Yi Lin, and at Meta Reality Labs with Hung-Yu Tseng and Jia-Bin Huang.
Across these roles, I have worked on data, modeling, training and evaluation, from camera pose estimation and 3D generation to real-time video generation and world models.
Back to Pixels: Lifting Latent Video Diffusion into Pixel Space
Lifting a pretrained latent video model into pixel space enables efficient high-resolution video, precise video editing and low-latency world modeling.
My role: project lead in the final phase.
Dynamic Camera Poses and Where to Find Them
DynPose-100K is a large-scale dataset of dynamic Internet videos with camera annotations, a foundation for camera-controlled video generation and world models.
FAR: Flexible, Accurate and Robust 6DoF Relative Camera Pose Estimation
Combining the complementary strengths of correspondence-based solvers and learned models yields accurate and robust relative pose.
Scalable 3D Captioning with Pretrained Models
Pretrained models are highly effective at 3D captioning; we use them to build the large-scale 3D-text dataset Cap3D and finetune text-to-3D models.
The 8-Point Algorithm as an Inductive Bias for Relative Pose Prediction by ViTs
Small modifications to a ViT enable computations similar to the Eight-Point algorithm and are a good inductive bias for pose estimation.
PixelSynth: Generating a 3D-Consistent Experience from a Single Image
Combining 3D reasoning and autoregressive modeling enables high-quality, consistent single-image novel view synthesis.