Author 👨💻: Ziyang Xie @ Manifolds-AI
Contact Email 📧: ziyangxie01@gmail.com
Feel free to contact me for any questions or collaborations!
🌏 WorldGen can generate 3D scenes in seconds from text prompts and images. It is a powerful tool for creating 3D environments and scenes for games, simulations, robotics, and virtual reality applications.
- Instant 3D Generation ⚡️ : Create full 3D scenes from input data in seconds
- 360° Free Exploration 🧭 : WorldGen supports free 360° consistent exploration of the generated 3D scene with loop closure.
- Diverse Scenes Support 🌈 : WorldGen supports both indoor and outdoor scenes, both realistic and unrealistic scenes in any style.
- Flexible Rendering 📸 : WorldGen supports rendering at any resolution with any camera setting and trajectory in real-time.
Two lines of code to generate a 3D scene in seconds
# Use our API to generate a 3D scene
worldgen = WorldGen()
worldgen.generate_world("<TEXT PROMPT to describe the scene>")-
05.10.2025🔥 Add support for low-vram generation (Only use ~10GB VRAM for generation). -
04.26.2025📄 New Relase a project page for WorldGen -
04.22.2025🏡 Add support for mesh scene generation (Should give better results than splat) -
04.21.2025🎉 Opensource the WorldGen codebase -
04.19.2025🖼️ Add support for image-to-scene generation -
04.17.2025📝 Add support for text-to-scene generation - 🏃🏻♀️ Release a better lora checkpoint.
- Release technical report and video
- Support higher resolution generation (e.g., 1080P, currently limited by computational resources)
- Release huggingface demo.
- Support better background inpainting (Invisible region inpainting)
Getting started with WorldGen is simple!
# Clone the repository
git clone https://github.com/ZiYang-xie/WorldGen.git
cd WorldGen
# Create a new conda environment
conda create -n worldgen python=3.11
conda activate worldgen
# Install torch and torchvision (with GPU support)
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu128
# Install worldgen
pip install .Quick start with WorldGen (mode in t2s or i2s) and generate your first 3D scene in seconds:
- 📝 Text to Scene: Generate a 3D scene from a text prompt
# Example using the Python API
from worldgen import WorldGen
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
worldgen = WorldGen(mode="t2s", device=device, low_vram=False) # Set low_vram to True if your GPU VRAM is less than 24GB.
splat = worldgen.generate_world("<TEXT PROMPT to describe the scene>")
splat.save("path/to/your/output.ply") # Save splat file as a .ply file, which can be loaded and visualized using a standard gaussian splatting viewer- 🖼️ Image to Scene: Generate a 3D scene from an image
worldgen = WorldGen(mode="i2s", device=device, low_vram=False) # Set low_vram to True if your GPU VRAM is less than 24GB.
image = Image.open("path/to/your/image.jpg")
splat = worldgen.generate_world(
image=image,
prompt="<Optional: TEXT PROMPT to describe the image and the scene>",
)- 🏡 Generate a 3D scene in mesh mode
mesh = worldgen.generate_world("<TEXT PROMPT to describe the scene>", return_mesh=True)
o3d.io.write_triangle_mesh("path/to/your/output.ply", mesh) # Save mesh as a .ply fileTip
We also support background inpainting for better scene generation, but it's currently an experimental feature, which may not work for all scenes.
It can be enabled by setting WorldGen(inpaint_bg=True).
# If want to use background inpainting feature, install iopaint
pip install iopaint --no-dependenciesWe provide a demo script to help you quickly get started and visualize the 3D scene in a web browser. The script is powered by Viser.
# Generate a 3D scene from a text prompt
python demo.py -p "A beautiful landscape with a river and mountains"
# Generate a 3D scene from an image
python demo.py -i "path/to/your/image.jpg" -p "<Optional: TEXT PROMPT to describe the scene>"
# 🔥 New feature: Generate a 3D scene in mesh mode
# Make sure you installed my customized viser to correctly visualize the mesh without backface culling
# pip install git+https://github.com/ZiYang-xie/viser.git
python demo.py -p "A beautiful landscape with a river and mountains" --return_meshAfter running the demo script, A local viser server will be launched at http://localhost:8080, where you can explore the generated 3D scene in real-time.
Note
WorldGen internally support generating a 3D scene from a 360° panorama image 📸, which related to how WorldGen works: You can try it out if you happen to have a 360° panorama (equirectangular) image. Aspect ratio of the panorama image should be 2:1.
pano_image = Image.open("path/to/your/pano_image.jpg")
splat = worldgen._generate_world(pano_image=pano_image)Give a star to WorldGen if you like it!
If you find this project useful, please consider citing it as follows:
@misc{worldgen2025ziyangxie,
author = {Ziyang Xie},
title = {WorldGen: Generate Any 3D Scene in Seconds},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/ZiYang-xie/WorldGen}},
}This project is built on top of the follows, please consider citing them if you find them useful:
Some of the core methods and ideas in this project are inspired by the following projects, special thanks to them:
- WonderWorld [RGBD to GS conversion]
- WorldSheet [Mesh Generation]