Yushi Lan

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.

Email  /  CV  /  Twitter  /  Github  /  Scholar

My interests lie in the intersection of CV, CG, and ML. I have worked on Shape Analysis, Inverse Graphics / Neural rendering, and 3D Avatar. I am now interested in Spatial AI, Embodied AI, Physical AI, 3D Generative Model,

STream3R resolves streaming 3D/4D reconstruction as a sequential registration task with causal attention.

ArtiLatent generates articulated 3D object.

WORLDMEM enables long-term memory for video world models.

GaussianAnything generates high-quality and editable surfel Gaussians through a cascaded 3D diffusion pipeline, given single-view images or texts as the conditions.

GaussianAnything generates high-quality and editable surfel Gaussians through a cascaded 3D diffusion pipeline, given single-view images or texts as the conditions.

SAR3D generates and understands 3D object in a scale-level autoregressive way.

3DEnhancer enhances low-quality 3D assets through multi-view diffusion priors.

3DTopia-XL scales high-quality 3D asset generation using Diffusion Transformer (DiT) built upon an expressive and efficient 3D representation, PrimX.

LN3Diff is a native 3D diffusion model that creates high-quality 3D object mesh from image or text within 8 seconds.

Gaussian3Diff adopts 3D Gaussians defined in UV space as the underlying 3D representation, which intrinsically support high-quality novel view synthesis, 3DMM-based animation and 3D diffusion for unconditional generation.

We propose a novel face reenactment framework, which adopts tri-planes as fundamental NeRF representation and decomposes face tri-planes into three components: canonical tri-planes, identity deformations, and motion.

We propose DeformToon3D, an 3D toonification methods that achieves high-quality geometry and texture stylization under given styles.

We propose E3DGE, an encoder-based 3D GAN inversion framework that yields high-quality shape and texture reconstruction.

EVA3D is a high-quality unconditional 3D human generative model that only requires 2D image collections for training.

We study dense correspondence, which plays a key role in 3D scene understanding but has been ignored in NeRF research. DDF presents a novel way to distill dense NeRF correspondence from pre-trained NeRF GAN unsupervisedly.

We propose MagnifierNet, a triple-branch network which accurately mines details from whole to parts in person re-identification (ReID).

Design and source code from Jon Barron's website