Overview of our contributions. 1. Native 3D world model with a latent point representation and pure 3D architectures, which naturally enables 3D consistency, flexible rendering from any viewpoints, and free exploration. 2. Point-to-Gaussian VAE which transforms input RGB point cloud into a compact set of point latents and decode them into 3D Gaussian primitives. 3. Sparse point flow matching for efficient point-based generative modeling and general conditioning mechanism.
Reconstruction
Unconditional Generation
Image-Conditioned Generation
Exploration
Method Overview
Bibtex
@misc{huang2025terra,
title={Terra: Explorable Native 3D World Model With Point Latents},
author={Yuanhui Huang and Weiliang Chen and Wenzhao Zheng and Xin Tao and Pengfei Wan and Jie Zhou and Jiwen Lu},
year={2025},
eprint={2510.14977},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.14977},
}