I am a postdoctoral researcher at the Visual Geometry Group, University of Oxford, working with Prof. Andrea Vedaldi.
I received my Ph.D. from Nanyang Technological University, advised by Prof. Chen Change Loy and Asst. Prof. Xingang Pan at MMLab@NTU. Earlier, I earned my B.Eng. in Software Engineering from Yepeida Honors College, BUPT.
My current focus is Physical AI: bringing spatial intelligence (3D/4D perception, reconstruction, and generation) to embodied agents that understand and act in the physical world.
In 2025, I was selected as an Outstanding Prize recipient of the Meshy Fellowship 2025. My Ph.D. thesis, Expanding the Generality of Neural Fields [PDF], is available online.
All | World Models | Agentic 3D | Video Generation | Spatial Intelligence | 3D/4D Generation | 3D Correspondence
SceneConductor is a harness framework for agentic 3D scene reconstruction that turns an open-source model (Qwen 27B) into a comprehensive 3D scene reconstructor.
project page / arXiv / code / bibtex
Mundo turns context images into editable 3D worlds for camera exploration, simulation, and robot planning.
project page / blog / code / bibtex
Holo4D reshoots a video along new camera paths, using reconstructed persistent 3D tracks as a scaffold to tame a pretrained video diffusion model.
project page / arXiv / code / bibtex
PhysiFormer learns physically plausible 4D mesh dynamics in world coordinates for multi-object, multi-material simulation.
project page / arXiv / code / demo / bibtex
Syn4D is a multiview synthetic 4D dataset with dense annotations for dynamic scene reconstruction, tracking, pose, and view synthesis.
project page / arXiv / data / code / bibtex
4RC is a unified feed-forward framework for 4D reconstruction from monocular videos with conditional querying across time and views.
project page / arXiv / code / bibtex
VFMF performs stochastic forecasting of future worlds in the latent space of vision foundation models.
project page / arXiv / code / bibtex
Video4Spatial studies complex visuospatial reasoning directly from video generation.
project page / arXiv / bibtex
STream3R resolves streaming 3D and 4D reconstruction as a sequential registration task with causal attention.
project page / paper / code / bibtex
GaussianAnything generates high-quality editable surfel Gaussians from images or text.
project page / arXiv / code / demo / bibtex
ArtiLatent generates articulated 3D objects with fine-grained geometry, accurate articulation, and realistic appearance via structured latents.
project page / arXiv / code / bibtex
WorldMem enables long-term memory for video world models.
project page / arXiv / code / demo / bibtex
ObjCtrl-2.5D enables training-free object control in image-to-video generation using depth and camera trajectories.
project page / arXiv / code / demo / bibtex
Textured 3D Regenerative Morphing uses a 3D diffusion prior to create smooth, plausible morphing sequences between textured 3D objects.
project page / arXiv / code / bibtex
SAR3D generates and understands 3D objects in a scale-level autoregressive way.
project page / arXiv / code / video / bibtex
3DEnhancer enhances low-quality 3D assets through multi-view diffusion priors.
project page / arXiv / code / video / bibtex
3DTopia-XL scales high-quality 3D asset generation via primitive diffusion.
project page / arXiv / code / bibtex
LN3Diff creates high-quality 3D meshes from image or text within seconds.
project page / TPAMI / arXiv / code / bibtex
Gaussian3Diff uses UV-space 3D Gaussians for view synthesis, animation, and unconditional 3D diffusion for full heads.
project page / arXiv / bibtex
DeformToon3D decomposes 3D toonification into geometry deformation and texture stylization on a pre-trained 3D GAN.
project page / arXiv / code / bibtex
E3DGE is an encoder-based 3D GAN inversion framework for high-quality shape and texture reconstruction.
EVA3D is a high-quality unconditional 3D human generative model that only requires 2D image collections for training.
project page / arXiv / video / code / bibtex
DDF studies dense correspondence in NeRF-based GANs and distills it without supervision.
project page / arXiv / Springer / bibtex
MagnifierNet is a triple-branch network for mining details from whole to parts in person re-identification.
Workshop Host: PhysAI @ ECCV 2026
Conference Reviewer: CVPR'21-26, ICCV'23/25, ECCV'24/26, NeurIPS'22-25, ICLR'24-26, SIGGRAPH'23/24/25, SIGGRAPH Asia'23/24/25
Journal Reviewer: TPAMI, IJCV, TVCG, CVIU, TOG
2025 @ GAMES Webinar: 三维基础模型之重建与生成 [Video]
