WorldGen Turns a Text Prompt Into a Walkable 3D World in Seconds
WorldGen turns a text prompt or image into a fully explorable 3D Gaussian Splat scene in seconds, with just two lines of Python
ProTakeaways
- WorldGen is an open-source Python library that generates fully explorable 3D Gaussian Splat scenes from a text prompt or image in seconds.
- The pipeline works in two stages: generate a 360-degree panorama with FLUX.1-dev, then lift it into 3DGS or mesh using DA-2 depth estimation.
- Supports text-to-scene, image-to-scene, mesh output, and 360° panorama input; runs on as little as ~10GB VRAM with the low-VRAM flag.
- Free and Apache 2.0 licensed; model weights available on Hugging Face, code on GitHub.
- Key limitation: single-viewpoint panorama anchor means large open worlds and multi-floor environments are not natively supported.
- Cited by Tencent's HunyuanWorld 1.0 paper; technical report not yet published but the codebase is actively maintained.
WorldGen is an open-source Python library that converts a text prompt or a reference image into a fully navigable 3D scene in seconds. Built by researcher Ziyang Xie, it has quietly accumulated over 2,000 GitHub stars and nearly 200 forks since its release, making it one of the more practically usable open-source entries in the fast-moving 3D generation space.
The problem it solves
Creating 3D environments has always been a specialist skill. Game studios, robotics labs, and VR developers spend enormous time hand-crafting scenes or painstakingly reconstructing them from multi-view captures. WorldGen is designed to create interactive environments suitable for games, simulations, robotics research, and virtual reality, rather than just static 3D assets. The core idea is that you describe a world in natural language and WorldGen produces a navigable 3D scene that you can freely explore in 360 degrees, with loop closure so that the space remains consistent as you move around.
Two stages under the hood
The pipeline is deliberately straightforward. First, WorldGen generates a high-resolution, 360-degree panorama image based on the input text prompt or image. This panorama captures the overall scene layout and appearance. Next, the generated panorama is converted into a full 3D scene, involving converting the panorama image into 3DGS or mesh representations, allowing for interactive exploration and consistent views from any angle.
3DGS, or 3D Gaussian Splatting, is a rendering technique that represents a scene as millions of tiny, soft, ellipsoid-shaped blobs in 3D space. Unlike traditional meshes, these blobs can be rendered in real-time at high quality from any viewpoint, which is why it has become the go-to output format for generative 3D systems. WorldGen uses