P2Voxel: Pyramid Pivot Voxelization for 3D Mesh Tokenization

---

News

[TBA] The code is being organized and will be gradually released. Please stay tuned.

Overview

P2Voxel retargets mesh tokenization as local surface evidence sampling. Instead of storing dense volumetric fields or edge-level contouring constraints, it represents each active voxel with a compact pivot point and orientation sign that can induce the corner values required for deterministic Sparse Marching Cubes reconstruction.

The framework contains three main components:

• Pivot Voxelization, which stores minimal local surface evidence in each active voxel.
• Pyramid Pivot Voxelization, which adaptively allocates finer tokens to geometrically complex regions while keeping smooth regions coarse.
• Pyramid VAE, which learns compact multi-resolution latent codes over locally reconstructable pivot blocks.

Code Release

This repository currently contains the project page and paper assets.

The core implementation of P2Voxel is currently being organized. We will gradually open-source the code, scripts, and documentation in future updates. Please stay tuned.

Repository Structure

code/
├── README.md
└── docs/
    ├── index.html
    ├── styles.css
    ├── script.js
    └── assets/

Contact

For questions about the paper or implementation release, please contact the authors listed on the project page.

BibTeX

@article{sun2026p2voxel,
  title={P2Voxel: Pyramid Pivot Voxelization for 3D Mesh Tokenization},
  author={Sun, Zhenhong and Liu, Haozhe and Wang, Yifu and Song, Xibin and Wang, Senbo and Mo, Huadong and Dong, Daoyi and Li, Hongdong and Ji, Pan},
  journal={arXiv preprint},
  year={2026}
}