MeshPack: Towards High-Resolution Artistic Mesh Generation via Part-wise Autoregressive Modeling

Qinglin Yang1,2  ·  Xianze Fang2  ·  Jingnan Gao3,2  ·  Jiangjing Lyu2  ·  Yichao Yan3  ·  Xiaowei Zhou1  ·  Jiazhi Xia4  ·  Sida Peng1

1State Key Lab of CAD & CG, Zhejiang University    2Alibaba Group    3Shanghai Jiao Tong University    4Central South University

teaser

Abstract

TL;DR: We propose MeshPack, a part-wise autoregressive framework with a part-condition fusion module and FurniSet3D dataset, achieving superior high-resolution furniture mesh generation over existing methods.
MeshPack Overview

This paper addresses the task of generating artistic meshes with high resolution and part-level decomposition. Recent autoregressive methods generate meshes by learning the sequential distribution of mesh vertices and faces, inherently capturing strong mesh priors for high-quality topology. However, their scalability is limited by the quadratic complexity of the attention mechanism, which restricts the output resolution. To address this challenge, we introduce MeshPack, a part-wise autoregressive modeling approach that decomposes meshes into multiple part-level sequences and iteratively generates each sequence, significantly reducing computational complexity and naturally enhancing the mesh resolution. In addition, an information passing module is designed to ensure consistency across part-level sequences by fusing previously generated parts with current part, imposing global context awareness during generation. Furthermore, we introduce FurniSet3D, a high-quality furniture dataset with detailed geometry, clean topology, and semantic part decompositions for training our model. Extensive experiments demonstrate that MeshPack generates fine-resolution meshes with high topological quality and reasonable part decomposition, substantially outperforming existing methods on furniture generation.

Qualitative Comparison


comparison

More Results


more results

Citation