WorldGrow redefines 3D world generation by enabling infinite, continuous 3D scene creation through a hierarchical block-wise synthesis and inpainting pipeline. Developed by researchers from Shanghai Jiao Tong University, Huawei Inc., and Huazhong University of Science and Technology, it achieves unbounded, photorealistic, and geometrically coherent environments paving the way for scalable virtual world modeling for games, AR/VR, and embodied AI.
Key Highlights:
- Block-Wise Infinite Generation: Builds large 3D worlds progressively from a single seed block using structured latent representations (SLAT) with coarse-to-fine refinement.
- Scene-Friendly SLAT: Introduces an occlusion-aware adaptation of TRELLIS for realistic geometry and texture continuity across 3D blocks.
- 3D Block Inpainting: Enables context-aware scene extension with consistent spatial and visual transitions during block-by-block expansion.
- Coarse-to-Fine Refinement: Ensures global layout plausibility and local detail fidelity, balancing large-scale structure with fine geometric realism.
- SOTA Performance: Outperforms BlockFusion, SynCity, and DiffInDScene in both geometry (FID↓ 3.95) and visual realism, generating massive indoor scenes spanning 1,800 m² with photorealistic textures.
Why It Matters:
WorldGrow is a breakthrough toward training-free, scalable 3D environment synthesis, allowing embodied agents to explore open-ended, coherent virtual worlds. It bridges object-centric 3D generation and large-scale world modeling, a crucial step toward creating autonomous AI-driven worlds for simulations, robotics, and metaverse applications.
Explore More:
- Paper: arXiv:2510.21682
- Project Page: https://world-grow.github.io
The post WorldGrow: Generating Infinite 3D Worlds appeared first on OpenCV.