PointNSP: Autoregressive 3D Point Cloud Generation with Next-Scale Level-of-Detail Prediction
2510.05613v1
cs.CV, cs.AI, cs.LG
2025-10-09
Авторы:
Ziqiao Meng, Qichao Wang, Zhiyang Dou, Zixing Song, Zhipeng Zhou, Irwin King, Peilin Zhao
Abstract
Autoregressive point cloud generation has long lagged behind diffusion-based
approaches in quality. The performance gap stems from the fact that
autoregressive models impose an artificial ordering on inherently unordered
point sets, forcing shape generation to proceed as a sequence of local
predictions. This sequential bias emphasizes short-range continuity but
undermines the model's capacity to capture long-range dependencies, hindering
its ability to enforce global structural properties such as symmetry,
consistent topology, and large-scale geometric regularities. Inspired by the
level-of-detail (LOD) principle in shape modeling, we propose PointNSP, a
coarse-to-fine generative framework that preserves global shape structure at
low resolutions and progressively refines fine-grained geometry at higher
scales through a next-scale prediction paradigm. This multi-scale factorization
aligns the autoregressive objective with the permutation-invariant nature of
point sets, enabling rich intra-scale interactions while avoiding brittle fixed
orderings. Experiments on ShapeNet show that PointNSP establishes
state-of-the-art (SOTA) generation quality for the first time within the
autoregressive paradigm. In addition, it surpasses strong diffusion-based
baselines in parameter, training, and inference efficiency. Finally, in dense
generation with 8,192 points, PointNSP's advantages become even more
pronounced, underscoring its scalability potential.
Ссылки и действия
Дополнительные ресурсы: