WorldGen
🌍 WorldGen - Generate Any 3D Scene in Seconds
What is WorldGen?
🌍 WorldGen - Generate Any 3D Scene in Seconds
Overview of WorldGen
What it is
WorldGen is a tool that generates 3D scenes from text prompts or images in seconds. It is designed for creating 3D environments for games, simulations, robotics, and virtual reality applications. The project is authored by Ziyang Xie and is open-sourced under the Apache-2.0 license.
Key features
- Instant 3D generation from text prompts or images in seconds.
- Supports 360° free exploration of the generated 3D scene with loop closure.
- Supports both indoor and outdoor scenes, realistic and unrealistic styles.
- Supports rendering at any resolution with any camera setting and trajectory in real-time.
- Offers both text-to-scene and image-to-scene generation modes.
- Provides a mesh generation mode as an alternative to splat-based output.
- Includes an experimental ml-sharp feature for potentially better Gaussian Splatting results.
Use cases
WorldGen can be used to quickly prototype 3D scenes for game environments, simulation training data, robotics navigation scenarios, and virtual reality experiences. It is also suitable for generating background scenes for creative projects or research in 3D scene understanding.
Getting started
bash# Clone the repositorygit clone --recursive https://github.com/ZiYang-xie/WorldGen.gitcd WorldGen# Create a new conda environmentconda create -n worldgen python=3.11conda activate worldgen# Install torch and torchvision (with GPU support)pip3 install torch torchvision# Install worldgenpip install .# Install DA-2 (360 depth estimation) -- use --no-deps to avoid version conflictspip install git+https://github.com/EnVision-Research/DA-2.git#subdirectory=src --no-deps# Install pytorch3d dependenciespip install git+https://github.com/facebookresearch/pytorch3d.git --no-build-isolation
License and pricing
WorldGen is licensed under Apache-2.0. The pricing type is listed as open source.
Based on the project's public README and repository metadata.