世界生长：生成无限3D世界

世界生长：生成无限3D世界
WorldGrow: Generating Infinite 3D World

原始链接: https://github.com/world-grow/WorldGrow

## WorldGrow：无限3D世界生成 WorldGrow是一种新的生成方法，用于创建广阔、明确的3D世界，为传统世界模拟器提供替代方案。该方法由上海交通大学、华为和华中科技大学的研究人员开发，利用分层框架从单个种子生长环境。该系统采用分块合成和粗到细的优化，生成具有详细局部几何和外观的连贯全局布局。重要的是，这些生成的世界被设计为可导航的，并且适合于测试导航和规划算法。目前该项目为预印本（arXiv:2510.21682），展示了大规模环境的创建——包括一个19x39的室内世界（~1,800平方米），具有重建的网格和纹理渲染。代码正在积极开发中，并计划公开发布，包括预训练权重和完整的训练流程。

## WorldGrow：无限3D世界生成 - 摘要 WorldGrow是一个新的框架，旨在利用人工智能生成无限延伸、连贯的3D世界。该项目在Hacker News上分享，利用预训练的3D模型和分层场景合成方法，专注于生成逼真且结构一致的环境。讨论的中心在于，这种方法是否比现有的程序化生成技术（如波函数坍塌（WFC）——一些人认为它具有类似目的）或传统PCG方法（在游戏中使用了数十年）具有显著优势。人们对生成空间的“趣味性”表示担忧，许多人指出，创造*引人入胜*的虚拟环境才是真正的挑战，而不仅仅是大型环境。一些评论员指出Minecraft、No Man’s Sky和Dwarf Fortress等游戏是成功的大世界生成的例子，而另一些人则强调了潜在的应用，例如训练人工智能机器人或为恐怖游戏（如Backrooms）创建环境。一个关键问题围绕资源消耗和控制——人工智能驱动的生成是否比成熟的技术提供优势，或者引入了新的限制。代码目前正在准备公开发布。

原文

Sikuang Li^1*, Chen Yang^2*, Jiemin Fang^2✉, Taoran Yi³, Jia Lu³,
Jiazhong Cen¹, Lingxi Xie², Wei Shen¹, Qi Tian^2✉

¹Shanghai Jiao Tong University ²Huawei ³Huazhong University of Science and Technology
^*Equal contribution ^✉Corresponding author

We propose WorldGrow — a generative method which creates infinite EXPLICIT 3D worlds, an alternative to the extensible, realistic, interactive world simulator.

WorldGrow is a hierarchical framework for infinite (open-ended) 3D world generation. Starting from a single seed block, the system grows large environments via block-wise synthesis and coarse-to-fine refinement, producing coherent global layouts and detailed local geometry/appearance. The generated scenes are walkable and suitable for navigation/planning evaluation.

If you use any part of this repository, please consider starring ⭐ the project and citing our paper.

2025-10-27 — 🚧 Paper released and repository initialized. The code is being prepared for public release; pretrained weights and full training/inference pipelines are planned.

⚠️ The repo is under active development. Interfaces may change. Placeholders below will be updated as components are released.

Requirements (placeholder)

Installation (placeholder)

Visualization (placeholder)

Gallery: diverse generated scenes at multiple scales.
Large-scale example: a 19x39 indoor world (~1,800 m²) with reconstructed mesh and textured rendering.

Please visit the project page for more figures, videos, and metrics.

TBD (to be finalized before full code release).

@article{worldgrow2025,
  title   = {WorldGrow: Generating Infinite 3D World},
  author  = {Li, Sikuang and Yang, Chen and Fang, Jiemin and Yi, Taoran and Lu, Jia and Cen, Jiazhong and Xie, Lingxi and Shen, Wei and Tian, Qi},
  journal = {arXiv preprint arXiv:2510.21682},
  year    = {2025}
}