Repolode
Submit
world-grow

world-grow/WorldGrow

Research code for WorldGrow, a method that generates open-ended, walkable 3D worlds block by block from a seed block, outputting Gaussian splats and meshes.

View on GitHub
Stars
500
Forks
28
Open issues
7
Last commit
11 months ago

What is it?

WorldGrow is a research implementation of a hierarchical framework for infinite, open-ended 3D world generation. It starts from a single seed block and grows a larger environment through block-wise synthesis and coarse-to-fine refinement, aiming for a coherent global layout with detailed local geometry and appearance.

The generated scenes are described as walkable and suitable for navigation and planning evaluation. The provided example script loads a pretrained pipeline from Hugging Face, generates a world of a chosen size in blocks, and exports a rendered video, a 3DGS point cloud (.ply), and a mesh (.glb).

The code builds on the TRELLIS environment and includes modified cumm and spconv submodules, which the project says are required for its models. Pretrained weights are published, while the full training pipeline and license are still pending.

Key features

  • 01Generates explicit 3D worlds by growing from a single seed block.
  • 02Uses block-wise synthesis combined with coarse-to-fine refinement.
  • 03Exports rendered video, 3DGS point cloud, and mesh outputs.
  • 04Loads pretrained weights from Hugging Face with a simple Python API.
  • 05Produces walkable scenes for navigation and planning evaluation.

Use cases

  • Generating large indoor or synthetic environments for robotics simulation.
  • Evaluating navigation and path planning algorithms in procedurally grown scenes.
  • Researching hierarchical or block-based 3D scene generation.

Quick start

  1. 1

    Clone the repository with submodules

    git clone --recurse-submodules https://github.com/world-grow/WorldGrow.git
  2. 2

    Set up the environment following the TRELLIS installation instructions

  3. 3

    Run the example script example_world_grow.py to generate a world

Things to know

  • No license is specified; the README says the license is TBD.
  • Inference relies on CUDA (the example calls pipeline.cuda()), so a GPU is needed.
  • Setup requires custom cumm and spconv builds from the submodules folder instead of the standard packages.

Topics

#3d-generation#gaussian-splatting#procedural-generation#deep-learning#pytorch#robotics-simulation

FAQ

What is WorldGrow?
Research code for WorldGrow, a method that generates open-ended, walkable 3D worlds block by block from a seed block, outputting Gaussian splats and meshes. WorldGrow is a research implementation of a hierarchical framework for infinite, open-ended 3D world generation. It starts from a single seed block and grows a larger environment through block-wise synthesis and coarse-to-fine refinement, aiming for a coherent global layout with detailed local geometry and appearance.
Is WorldGrow free to use? What license does it have?
WorldGrow has no declared license. Its code is public, but without a license you have no explicit permission to reuse it; ask the author before relying on it.
Is WorldGrow still maintained?
No commits for 6+ months. The last commit to WorldGrow was on Dec 3, 2025.
What can I use WorldGrow for?
Generating large indoor or synthetic environments for robotics simulation. Evaluating navigation and path planning algorithms in procedurally grown scenes. Researching hierarchical or block-based 3D scene generation.
How do I install WorldGrow?
Clone the repository with submodules: git clone --recurse-submodules https://github.com/world-grow/WorldGrow.git Set up the environment following the TRELLIS installation instructions Run the example script example_world_grow.py to generate a world
Where can I get WorldGrow?
The source code and installation instructions are on GitHub: https://github.com/world-grow/WorldGrow

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects