Yang Li

I'm a researcher at Tencent (Hunyuan3D).

I finished my Ph.D. with Tatsuya Harada at The University of Tokyo. During my Ph.D., I did an internship at Technical University Munich with Matthias Nießner, and an internship with Bo Zheng at Huawei Japan Research Center. I did a master in bioinformatics with Tetsuo Shibuya at The University of Tokyo.

My research interests lie in the intersection of 3D computer vision, artificial intelligence, particularly focusing on registration, 3D/4D reconstruction, 3D Generative modeling, with applications in VR/AR, robotics, etc.

Email  /  Scholar  /  Github

Research Highlights

We introduce X-Part, a controllable generative model designed to decompose a holistic 3D object into semantically meaningful and structurally coherent parts with high geometric fidelity.

We propose a native 3D point-promptable part segmentation model termed P3-SAM, designed to fully automate the segmentation of any 3D objects into components.

We introduce SeamGPT, an auto-regressive model that generates cutting seams by mimicking professional workflows.

We present BAG, a Body-aligned Asset Generation method to output 3D wearable asset that can be automatically dressed on given 3D human bodies.

PhyCAGE generates physically plausible compositional 3D assets from a single image.

This paper proposes a novel approach for automatically generating interactive (i.e., playable) 3D game scenes from users' casual prompts, including hand-drawn sketches and text descriptions.

We present Frankenstein, a diffusion-based framework that can generate semantic-compositional 3D scenes in a single pass. Unlike existing methods that output a single, unified 3D shape, Frankenstein simultaneously generates multiple separated shapes, each corresponding to a semantically meaningful part.

We introduce the first 3D diffusion based approach for directly generating large unbounded 3D scene in both inddor and outdoor scenarios. At the core of this approach is a novel tri-plane diffusion and tri-plane extrapolation mechanism.

Neural Deformation Pyramid (NDP) break down non-rigid point cloud registration problem via hierarchical motion decomposition. NDP demonstrates advantage in both speed and registration accuracy.

We design Lepard, a novel partial point clouds matching method that exploits 3D positional knowledge. Lepard reaches SOTA on both rigid and deformable point cloud matching benchmarks.

We introduce 4DComplete, the first method that jointly recovers the shape and motion field from partial observations. We also provide a large-scale non-rigid 4D dataset for training and benchmaring. It consists of 1,972 animation sequences, and 122,365 frames.

We learn the tracking of non-rigid objects by differentiating through the underlying non-rigid solver. Specifically, we propose ConditionNet which learns to generate a problem-specific preconditioner using a large number of training samples from the Gauss-Newton update equation. The learned preconditioner increases PCG’s convergence speed by a significant margin.

This website is based on source code.