Wenyuan Zhang | 张文源

I am currently a final-year Ph.D student in School of Software, Tsinghua University, advised by Prof. Yu-Shen Liu. I got my bachelor's degree from School of Software, Tsinghua University in 2021.

My research interests lie in the area of computer vision and graphics, including multi-view 3D reconstruction and rendering, multimodal conditioned 2D video and 3D representation generation.

Email  /  Google Scholar  /  Github

(* Equal Contribution)

We introduce MaterialRefGS that significantly improves the rendering and reconstruction of reflective scenes in Gaussian Splatting by introducing PBR-based multi-view material consistency constraints and reflection strength prior.

We introduce Kling-Avatar, a novel framework that unifies multimodal instruction understanding with cascaded long-duration video generation for avatar animation synthesis.

We introduce an autoregressive human video generation framework that enables interactive multimodal control and low-latency extrapolation in a streaming manner.

We present GAP, which gaussianizes raw point clouds into high-fidelity 3D Gaussians with text guidance via depth-aware diffusion and surface-anchored optimization.

We present NeRFPrior, which adopts a neural radiance field as a prior to learn signed distance fields using volume rendering for indoor scene surface reconstruction.

we propose a general approach that explores the uncertainty of monocular depths to provide enhanced geometric priors for neural rendering and reconstruction.

We propose a method that introduces a novel prior based on inter-image matching information for indoor surface reconstruction from sparse views.

We propose a method that seamlessly merge 3DGS with the learning of neural SDFs. To this end, we dynamically align 3D Gaussians on the zero-level set of the neural SDF. Meanwhile, we update the neural SDF by pulling neighboring space to the pulled 3D Gaussians, which progressively refine the signed distance field near the surface.

We present a novel differentiable renderer to infer UDFs from multi-view images. Instead of using hand-crafted equations, our differentiable renderer is a neural network which is pre-trained in a data-driven manner, dubbed volume rendering priors.

We design a triplet network to learn VoxelPoint-to-Pixel matching via a differentiable probabilistic PnP solver.

We introduce a general strategy to speed up the learning procedure for almost all radiance fields based methods. The key idea is to reduce the redundancy by shooting much fewer rays in the multi-view volume rendering procedure.

• National Scholarship (国家奖学金, Top 2%) at Tsinghua University, 2025.
• Comprehensive Excellent Scholarship (综合优秀奖学金) at Tsinghua University, 2024,2023,2021,2019.
• Excellent undergraduate thesis (本科优秀毕业论文) at Tsinghua University, 2021.
• Excellent graduates (本科优良毕业生) at Tsinghua University, 2021.

• Program Committee Member: AAAI-26
• Conference Reviewer: ACMMM-25, NeurIPS-24/25, ICCV-25, ICML-25, CVPR-25, ICLR-25, WACV-25/26, IJCAI-24
• Journal Reviewer: IEEE TIP, IEEE TVCG, IJCV, CAD

• I currently serve as the T.A. of the undergraduate course "Fundamentals of Programming" (程序设计基础), which is taught by Prof. Yu-Shen Liu. I am also serving as an undergraduate advisor (本科生辅导员) at School of Software. My hobbies include playing the flute, running, and singing. Feel free to contact me!

template adapted from this awesome website
Last updated: October 2025