Highlights:
- WorldGrow introduces a hierarchical framework for infinite 3D scene generation.
- Overcomes limitations of current 2D-lifting and object-centric 3D models.
- Combines data curation, 3D inpainting, and coarse-to-fine rendering for coherent environments.
- Achieves state-of-the-art performance on the large-scale 3D-FRONT dataset.
TLDR:
WorldGrow is a new AI-based framework that generates infinitely extendable 3D worlds with realistic textures and coherent geometry, setting a new benchmark in virtual scene synthesis and computer vision research.
A revolutionary step in computer vision research has been unveiled with the introduction of **WorldGrow**, a hierarchical framework designed to generate infinitely extendable 3D worlds. Developed by Sikuang Li, Chen Yang, Jiemin Fang, Taoran Yi, Jia Lu, Jiazhong Cen, Lingxi Xie, Wei Shen, and Qi Tian, this research advances the boundaries of virtual world creation by tackling one of the most significant challenges in computer graphics — generating continuous and coherent 3D environments at scale. Published on arXiv under the identifier arXiv:2510.21682, this study reflects a remarkable leap in both realism and scalability within synthetic 3D scene generation.
Traditional 3D world generation approaches, such as 2D-lifting models and implicit neural representations, often suffer from discontinuities, inconsistent geometry, or scalability constraints. Object-centered 3D foundation models also struggle to handle scene-level complexity required for large-scale world creation. WorldGrow resolves these issues through a structured, hierarchical approach that leverages pre-trained 3D generative priors. It uses a **data curation pipeline** to compile high-quality scene segments, transforming them into structured latent representations ideal for large-scale scene training. Additionally, the **3D block inpainting mechanism** allows the model to intelligently extend environments with contextual awareness, ensuring smooth transitions and spatial coherence.
The framework also introduces a **coarse-to-fine generation strategy**, ensuring that global layout plausibility and local detail fidelity are both maintained throughout the generative process. This method excels in maintaining photorealistic consistency and accurate geometry even across infinitely large expansions — a feat especially demonstrated through its performance on the **3D-FRONT dataset**, where it achieves state-of-the-art results in geometry reconstruction. Its ability to continuously generate seamless virtual spaces opens new possibilities for immersive simulations, digital twins, open-world gaming, and metaverse-scale virtual experiences. As researchers and developers continue exploring generative AI capabilities, WorldGrow stands as a pioneering effort towards scalable virtual realism and structured world modeling.
In summary, WorldGrow sets a new benchmark for AI-driven scene synthesis, efficiently bridging the gap between object-centric modeling and environment-scale world building. It highlights how advanced 3D priors and structured generation frameworks can be harnessed to create continuous virtual worlds, suggesting a new direction for future research in computer vision, augmented reality, and autonomous simulation.
Source:
Source:
Original research paper: WorldGrow: Generating Infinite 3D World by Sikuang Li et al., arXiv:2510.21682v1 [cs.CV], published on 24 Oct 2025. Available at https://arxiv.org/abs/2510.21682