Ziyi Wang

I am a researcher at Tencent Hunyuan, where my work centers on large multimodal models and foundation models for physical AI.

I obtained my Ph.D. from the Department of Automation at Tsinghua University, advised by Prof. Jiwen Lu . In 2020, I received my B.Eng. from the Department of Electronic Engineering, Tsinghua University, along with a dual B.Admin. degree from the School of Economics and Management, Tsinghua University.

I am broadly interested in computer vision and deep learning. My research focuses on vision-language models, physical AI foundation models, world models, 3D/4D generation, and 3D vision.

Email  /  Google Scholar  /  Github

Publications

* indicates equal contribution

HY-Embodied-0.5 is a family of embodied foundation models built for real-world agents, achieving state-of-the-art performance across 22 benchmarks in visual perception, spatial reasoning, and embodied understanding, with effective downstream robot control.

MoRe4D synthesizes high-quality 4D scenes from a single image by jointly performing motion generation and 3D geometry reconstruction.

We conduct a comprehensive survey on vision generalist models that support multimodal inputs and can handle various downstream tasks.

OGGSplat is designed to expand the field-of-view of the Gaussian-based 3D scene reconstructed from sparse views and feedforward / generalizable models.

UniPre3D is a unified pre-training method that can be applied to both object-level and scene-level point clouds. It is supported by cross-modal Gaussian splatting technique.

XMask3D is a framework that propose mask-level reasoning techniques to empower 3D segmentation model with open vocabulary capacity under the assistance of the pre-trained 2D mask generator.

P2P++ is the extended journal version of P2P. We further propose Pixel-to-Point Distillation to make P2P applicable in scene-level perception tasks.

DPV-RAFT is the extended journal version of PV-RAFT. We further propose Spatial Deformation and Temporal Deformation to enhance PV-RAFT.

TAP is a 3D-to-2D generative pre-training method that generate projected images of point clouds from instructed perspectives.

P2P is a framework to leverage large-scale pre-trained image models for 3D point cloud analysis.

We present Semantic-Affine Transformation that transforms decoder mid-level features of the encoder-decoder segmentation network with class-specific affine parameters.

PoinTr is a transformer-based framework that reformulates point cloud completion as a set-to-set translation problem.

We present a deep interpretable metric learning (DIML) that adopts a structural matching strategy to explicitly aligns the spatial embeddings by computing an optimal matching flow between feature maps of the two images.

We present point-voxel correlation fields for 3D scene flow estimation which migrates the high performance of RAFT and provides a solution to build structured all-pairs correlation fields for unstructured point clouds.

Website Template