World models capture the structure, appearance, and dynamics of environments to support reconstruction and simulation. Mundo builds such a world from images of a specific place. It represents the scene’s layout and objects in a 3D proxy and defines their physical properties in a scene program. A simulator computes how the objects move and interact, while a video model renders the results using the input images as visual context. You can explore this environment from new viewpoints, modify its objects, and simulate robot tasks within it.
- Project page
- Code and weights (end of November)
@misc{luo2026mundo,
author = {Yihang Luo and Zunnan Xu and Shangchen Zhou and Yushi Lan and
Kang Liao and Zerui Gong and Xiaoming Wu and Qingyi Tao and
Zhonghua Wu and Bixuan Jin and Fei Wang and Zhi Hou and
Xiaogang Wang and Xingang Pan and Chen Change Loy},
title = {{Mundo}: Customized World Model},
year = {2026},
url = {https://mundo-world-model.github.io/}
}