WorldGrow tackles a tough problem in procedural generation: creating infinite, explorable 3D worlds that maintain geometric and visual consistency across independently generated blocks. The core idea is to start from a single seed block and grow large environments incrementally, using a hierarchical generative framework that produces both coarse global layouts and detailed local geometry. This approach addresses the challenge of scale and coherence in 3D world synthesis, particularly for indoor environments with complex structure.
hierarchical generative framework for infinite 3d world synthesis
WorldGrow is a Python-based framework designed to synthesize large-scale 3D worlds using a hierarchical and block-wise approach. The system incrementally expands the world by generating new blocks adjacent to existing ones, starting from a single seed block. This growth process is guided by a coarse-to-fine refinement strategy that first produces a rough global layout and then refines details locally within each block.
Technically, the pipeline outputs both 3D Gaussian Splatting (3DGS) point clouds and reconstructed meshes, enabling the generated scenes to be walkable and usable for navigation or planning evaluation. The repo builds on the TRELLIS codebase, extending it with custom modifications to support the hierarchical growth and synthesis.
The architecture relies on modified versions of the cumm and spconv packages, which are included as submodules because the original packages are not suitable for WorldGrow’s specific 3D models. These dependencies are optimized and installed via a setup.sh script.
A key configuration parameter is world_size, which controls the scale of the generated output, allowing users to balance detail and size according to their needs. Additionally, the system supports selective output formats, including a gaussian-only mode that speeds up inference by omitting mesh reconstruction.
block-wise infinite growth with coarse-to-fine refinement ensures seamless transitions
What sets WorldGrow apart is its block-wise growth mechanism combined with a hierarchical coarse-to-fine refinement process. Each block is synthesized independently but with contextual information to maintain consistency along shared boundaries.
The coarse-level generation ensures that the global layout forms a coherent structure without abrupt changes or discontinuities. Following this, finer-level synthesis adds geometric and appearance details, refining the block’s content to integrate smoothly with its neighbors.
This approach solves a common issue in procedural world generation where independently generated sections often produce visible seams or mismatched geometry. By explicitly modeling the generation hierarchy and refining from coarse global layout to fine local detail, WorldGrow enables infinite environment expansion while preserving walkability and visual coherence.
The output formats support both 3D Gaussian Splatting point clouds, which are efficient for rendering and storage, and mesh reconstructions, which are essential for physical simulation or navigation tasks.
In terms of code quality, the project maintains a modular design, separating the hierarchical synthesis logic, data handling, and rendering utilities. The use of submodules for critical dependencies ensures the necessary custom functionality without polluting the global Python environment. The tradeoff is the increased setup complexity and dependency management, but this is well documented.
explore the project: structure, key modules, and documentation
The repo is structured around the core synthesis modules, dependency submodules, and example scripts:
submodules/: contains modified versions ofcummandspconvlibraries necessary for the custom 3D models.- Core synthesis code: implements the hierarchical block-wise generation and refinement logic.
example_world_grow.py: a script demonstrating how to generate large 3D worlds using pretrained models.
The README provides detailed instructions on cloning the repository with submodules and refers users to the original TRELLIS repository for environment setup steps.
Warnings about dependency incompatibility and instructions to use the provided setup.sh script are clearly stated, which improves developer experience by reducing common pitfalls.
Users interested in generating worlds can start by running the example script, which serves as a practical entry point for understanding the synthesis pipeline and output options.
verdict: suited for researchers and practitioners in 3d generative modeling with a focus on scalable world expansion
WorldGrow fills a niche in 3D generative modeling by enabling infinite expansion of coherent, explorable 3D worlds with a hierarchical and block-wise approach. Its strength lies in the careful architectural design that ensures seamless transitions between blocks through coarse-to-fine refinement.
The repo is best suited for researchers and developers working on procedural content generation, simulation environments, or robotics navigation where large-scale 3D worlds are needed. The dependency on modified submodules and the reliance on the TRELLIS environment setup can pose a barrier for casual users but is manageable for those accustomed to research codebases.
Limitations include the complexity of installation and the learning curve associated with understanding the hierarchical synthesis pipeline. However, the example scripts and comprehensive documentation help mitigate these.
Overall, the project offers a well-engineered framework for infinite 3D world synthesis that balances scalability, detail, and walkability, making it a useful resource for advanced 3D environment generation tasks.
→ GitHub Repo: world-grow/WorldGrow ⭐ 455 · Python