WorldGrow: Generating Infinite 3D World

TL;DR

WorldGrow extends coherent 3D scenes block by block, combining structured 3D priors, context-aware inpainting, and coarse-to-fine synthesis to generate worlds without a fixed spatial boundary.

Abstract

Generating an infinitely extendable 3D world requires large continuous environments with coherent geometry and realistic appearance. Prior 2D-lifting methods often accumulate cross-view inconsistencies, implicit 3D representations are difficult to scale, and most 3D foundation models focus on isolated objects. WorldGrow addresses scene-level generation by reusing pretrained 3D priors for structured scene blocks. The hierarchical framework contains a data curation pipeline for high-quality training blocks, a 3D block-inpainting mechanism for context-aware spatial extension, and a coarse-to-fine strategy that balances plausible global layout with local geometric and textural detail. On the large-scale 3D-FRONT dataset, WorldGrow achieves strong geometry reconstruction while uniquely supporting open-ended scene generation with photorealistic, structurally consistent outputs. The approach provides a path toward scalable virtual environments and future generative world models.

Key contributions

• Introduces a hierarchical framework for unbounded, block-based 3D scene generation.
• Uses context-aware 3D inpainting to extend scenes while preserving local continuity.
• Combines coarse global planning with fine geometry and texture synthesis.

Citation

@inproceedings{li2026worldgrow,
  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},
  booktitle={AAAI},
  year={2026},
  doi={10.1609/aaai.v40i8.37571}
}