HunyuanWorld 1.0：从文字或像素生成沉浸式、可探索、交互式 3D 世界
计算机视觉与模式识别
2025-08-14 v2
摘要
从文本或图像创建沉浸式可玩 3D 世界是计算机视觉与图形学中的一项基本挑战。现有方法通常分为两类：视频基方法提供丰富多样性但缺乏 3D 一致性与渲染效率；3D 基方法保证几何一致性但受限于训练数据稀缺与内存低效表示。为此，我们提出 HunyuanWorld 1.0—a一种结合两者优势的新型框架，用于从文本和图像条件生成沉浸式、可探索、交互式 3D 场景。我们的方法具有三个关键优势：1) 通过全景世界代理实现 360{\deg} 沉浸式体验；2) 支持网格导出，实现与现有计算机图形管道的无缝集成；3) 解耦对象表示，实现增强交互性。该框架的核心是语义分层 3D 网格表示，利用全景图像作为 360{\deg} 世界代理进行语义感知的世界分解与重建，从而实现多样化 3D 世界的生成。大量实验表明，我们的方法在生成连贯、可探索、交互式 3D 世界方面实现了最先进性能，同时支持虚拟现实、物理仿真、游戏开发和交互内容创建等多种应用。
引用
@article{arxiv.2507.21809,
title = {HunyuanWorld 1.0: Generating Immersive, Explorable, and Interactive 3D Worlds from Words or Pixels},
author = {HunyuanWorld Team and Zhenwei Wang and Yuhao Liu and Junta Wu and Zixiao Gu and Haoyuan Wang and Xuhui Zuo and Tianyu Huang and Wenhuan Li and Sheng Zhang and Yihang Lian and Yulin Tsai and Lifu Wang and Sicong Liu and Puhua Jiang and Xianghui Yang and Dongyuan Guo and Yixuan Tang and Xinyue Mao and Jiaao Yu and Junlin Yu and Jihong Zhang and Meng Chen and Liang Dong and Yiwen Jia and Chao Zhang and Yonghao Tan and Hao Zhang and Zheng Ye and Peng He and Runzhou Wu and Minghui Chen and Zhan Li and Wangchen Qin and Lei Wang and Yifu Sun and Lin Niu and Xiang Yuan and Xiaofeng Yang and Yingping He and Jie Xiao and Yangyu Tao and Jianchen Zhu and Jinbao Xue and Kai Liu and Chongqing Zhao and Xinming Wu and Tian Liu and Peng Chen and Di Wang and Yuhong Liu and Linus and Jie Jiang and Tengfei Wang and Chunchao Guo},
journal= {arXiv preprint arXiv:2507.21809},
year = {2025}
}
备注
Technical Report; Project Page: https://3d-models.hunyuan.tencent.com/world/