Siyou Lin

Email: jsnln00@outlook.com

I am currently a Research Scientist at MeshyAI, working on generative AI with a special focus on geometry and texture generation. Prior to this, I have also spent time at ByteDance from 2025 to 2026 as a Multimedia Algorithm Intern, and Meta Zurich in 2024 as a Research Scientist Intern.

I received my PhD from the Department of Automation, Tsinghua University, supervised by Prof. Yebin Liu. My PhD study was centered around the CV & CG techniques to build realistic human avatars. I received my Bachelor's degree in Mathematics from the Department of Mathematics, Tsinghua University, as a member of the Tsinghua Xuetang Mathematics Talent Program.

During my undergraduate studies, I worked closely with Prof. Zuoqiang Shi from Yau Mathematical Sciences Center, Tsinghua University and Prof. Bin Wang from School of Software, Tsinghua University, on the problem of point cloud surface reconstruction. I also worked on monocular 3D depth estimation in a remote internship at the GRASP lab, UPenn, supervised by Prof. Jianbo Shi from 2020 to 2021.

We present Mix3R, a novel architecture that unifies TRELLIS and Pi3 using the Mixture-of-Transformers (MoT) architecture, to achieve input alignment generative reconstruction from sparse multi-view images.

We present Winding Number Normal Consistency (WNNC) and an iterative algorithm that estimates globally consistent normal vectors. Our CUDA-based treecode-accelerated algorithm handles a million of points in half a minute.

We present LayGA, a layered Gaussian avatar model based on Animatable Gaussians, with improved geometry and the ability of animatable clothing transfer.

We leverage intrinsic manifold properties and neural deformation fields and propose a coarse-to-fine two-stage method for non-rigid garment alignment, achieving wrinkle-level and texture-level alignment.

We present CaPhy, a novel method combining 3D-supervised training with unsupervised physics-based losses for reconstructing animatable human avatars. Caphy allows estimating the physical properties of garments and producing realistic dynamic clothing.

We present CloSET, a point-based human avatar modeling method that allows loose clothing modeling by explicit template decomposition. We also release a real-world capture dataset of clothed humans in various poses.

We propose a First-Implicit-Then-Explicit (FITE) framework that combines the merits of both implicit and explicit representations for modeling human avatars in clothing.

We present Parametric Gauss Reconstruction (PGR), a method that reconstructs surfaces from point clouds without normals by parametrizing the Gauss formula.

*Work done while I was an undergraduate at the Department of Mathematical Sciences, Tsinghua University.

This template is adapted from this website.