Three types of point cloud generative models: (a) diffusion-based methods that iteratively denoise shapes starting from Gaussian noise; (b) vanilla autoregressive (AR) methods that predict the next point by flattening the 3D shape into a sequence; and (c) our proposed PointNSP, which predicts next-scale level-of-detail in a coarse-to-fine manner.
Overview of PoinNSP: Illustration of training a multi-scale VQ-VAE for 3D point cloud reconstruction across scales \( s_{1} \) to \( s_{4} \), resulting in a multi-scale token sequence \( Q = (q_{1}, \dots, q_{4}) \).
Illustration of training a multi-scale 3D point cloud causal transformer with intermediate decoding, upsampling, position-aware soft masks, and block-wise causal masks.
Our generation results (right) compared to baseline models (left). PointNSP generates high-quality and diverse 3D point clouds.
Visualization of multi-scale point clouds during the PointNSP generation process as the scale K increases.
Visualizations of point cloud completion results.
Visualizations of point cloud upsampling results.
To cite the paper, please use the below:
@misc{meng2025pointnspautoregressive3dpoint,
title={PointNSP: Autoregressive 3D Point Cloud Generation with Next-Scale Level-of-Detail Prediction},
author={Ziqiao Meng and Qichao Wang and Zhiyang Dou and Zixing Song and Zhipeng Zhou and Irwin King and Peilin Zhao},
year={2025},
eprint={2510.05613},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.05613},
}