Publications

2026

1. 

   Learning Convex Decomposition via Feature Fields.

   Yuezhi Yang ,  Qixing Huang ,  Mikaela Angelina Uy ,  Nicholas Sharp

   Conference on Computer Vision and Pattern Recognition (CVPR) oral , 2026

   TLDR PDF Video Website

   A feed-forward model for convex decomposition of 3D shapes by learning a continuous feature field where distances between point embeddings indicate whether they belong to the same convex component, trained by a geometric contrastive loss.

2025

1. 

   ART-DECO: Arbitrary Text Guidance for 3D Detailizer Construction.

   Qimin Chen ,  Yuezhi Yang ,  Yifan Wang ,  Vladimir G. Kim ,  Siddhartha Chaudhuri ,  Hao Zhang ,  Zhiqin Chen

   SIGGRAPH Asia(Conference) , 2025

   TLDR PDF Website

   A reusable neural 3D detailizer that converts a coarse shape proxy into a detailed 3D asset within seconds, by distilling knowledge from a pretrained multi-view diffusion model via score distillation sampling.

2. 

   An Efficient Global-to-Local Rotation Optimization Approach via Spherical Harmonics.

   Zihang He ,  Yuezhi Yang ,  Congyue Deng ,  Jiaxin Lu ,  Leonidas Guibas ,  Qixing Huang

   Symposium on Geometry Processing (SGP) , 2025

   TLDR PDF

   A global-to-local optimization method for estimating the relative rotation between two 3D shapes, by aligning spherical harmonic coefficients of spherical distance functions with a robust weighted objective.

3. 

   GenAnalysis: Joint Shape Analysis by Learning Man-Made Shape Generators with Deformation Regularizations.

   Yuezhi Yang ,  Haitao Yang ,  George Kiyohiro Nakayama ,  Xiangru Huang ,  Leonidas Guibas ,  Qixing Huang

   SIGGRAPH (Journal) , 2025

   TLDR PDF Website

   We study the variation of shapes in the tangential latent space of a learned generative model, by enforcing piece-wise affine deformation regularization and analyzing the resulting tangent-space vector fields. This structure enables downstream 3D understanding tasks such as joint shape segmentation and cross-shape correspondence.

4. 

   GenVDM: Generating Vector Displacement Maps From a Single Image.

   Yuezhi Yang ,  Qimin Chen ,  Vladimir G. Kim ,  Siddhartha Chaudhuri ,  Qixing Huang ,  Zhiqin Chen

   Conference on Computer Vision and Pattern Recognition (CVPR) highlight , 2025

   TLDR PDF Video Code Website

   A generative model that produces reusable geometric detail stamps (Vector Displacement Maps) from a single image using multi-view diffusion model and neural deformation field.

2022

1. 

   Discovering Design Concepts for CAD Sketches

   Yuezhi Yang ,  Hao Pan

   Advances in Neural Information Processing Systems (NeurIPS) spotlight , 2022

   TLDR PDF Video Code Website

   We study the discovery of modular design concepts in CAD sketches, by representing sketches as constraint graphs in a declarative DSL and learning reusable concept libraries through self-supervised program library induction. The learned concepts capture reusable geometric primitives and constraint patterns, enabling structured sketch interpretation and more accurate sketch auto-completion.

2021

1. 

   ToothInpaintor: Tooth Inpainting from Partial 3D Dental Model and 2D Panoramic Image

   Yuezhi Yang ,  Zhiming Cui ,  Changjian Li ,  Wenping Wang

   Arxiv , 2021

   TLDR PDF

   We study tooth inpainting from a partial 3D dental model and a 2D panoramic image, by learning a latent space of full tooth shapes using SDF representation and projecting the inputs into this space via neural optimization. An adversarial constraint on the latent code helps ensure globally plausible tooth geometry while reconstructing the missing roots.

2. 

   Domain Generalization for Mammography Detection via Multi-style and Multi-view Contrastive Learning

   Zheren Li ,  Zhiming Cui ,  Sheng Wang ,  Yuji Qi ,  Xi Ouyang ,  Qitian Chen ,  Yuezhi Yang ,  Zhong Xue , and 2 more authors

   Medical Image Computing and Computer Assisted Intervention Conference(MICCAI) , 2021

   TLDR PDF

   We study domain generalization for mammography lesion detection, by learning domain-invariant features through multi-style and multi-view contrastive learning that pairs style-transferred images and complementary CC/MLO views as positives.