Publications

HyperBones — realtime neural garment simulation conditioned via hypernetworks. SIGGRAPH Asia 2026.

Large-Scale Codec Avatars (LCA) — high-fidelity full-body 3D avatars that generalize via pretraining on 1M in-the-wild videos then post-training on curated studio data. Meta Reality Labs / Human AI. CVPR 2026.

PhySkin — physics-based bone-driven neural garment simulation. Meta Reality Labs human clothing work. CVPR 2026.

MHR — Momentum Human Rig, collaborative Meta effort on human rigging.

FRESA — feedforward personalized skinned avatar reconstruction from few images. CVPR 2025.

FaceMap — perceptual facial saliency maps driven by visual distortion, for better avatar faces.

RCAD — retrieval from counterfactually augmented data to test video-text models, with Feint6K dataset.

A declarative spec for unstructured mesh editing: we show a system that lets you describe mesh editing algorithms at high level while the runtime handles concurrency and robustness. Presented at SIGGRAPH Asia 2022.

Coarse yet accurate geometry with high order elements. We introduce a robust and automatic approach to convert triangle meshes to coarse high-order tetmeshes. We bound geometric error, ensure valid elements, and preserve features precisely. We can also transfer attributes and solutions between input and output boundary.

Adjusting the common used orthogonal projection between discrete surfaces a reliable tool, by constructing a shell to restrict the domain. The bijective association between meshes in the shell provide a robust way to construct high quality computational proxy.

Microstructure family defined on rhombic meshes, with better approximation on (planar) irregular geometric domains. A smoother material-to-geometry map and fabricated results.

A numerically stable, and provably correct replacement for Tutte’s embedding. Avoid exponential area change and serve as a robust building block for locally injective maps with hard constraints, in quad meshing.

One million CAD models in B-Rep form, with rich information to learn from.

An adapation of graph neural networks on triangle meshes, taking advantage of discrete differential operators like Laplacians and Diracs.

A robust and efficient way optimize the distortion of bijective maps. Applicable to 2D (bijective surface parameterization) as well as 3D (intersection free deformation) problems.