README.md

---
title: Extend3D
emoji: 🏙️
colorFrom: blue
colorTo: gray
sdk: docker
app_port: 7860
suggested_hardware: l4x1
models:
  - microsoft/TRELLIS-image-large
---
<div align="center">
    <h1>Extend3D: Town-scale 3D Generation</h1>
    <h3>CVPR 2026</h3>
    <p align="center">
      <a href="https://seungwoo-yoon.github.io">Seungwoo Yoon</a>,
      <a href="https://www.jinmo.kim">Jinmo Kim</a>,
      <a href="https://jaesik.info">Jaesik Park</a>
      <br />
      Seoul National University
    </p>
    <a href="https://seungwoo-yoon.github.io/extend3d-page">
        <img src="https://img.shields.io/badge/Project_Page-website-green?logo=GoogleChrome&logoColor=white"/>
    </a>
    <a href="#">
        <img src="https://img.shields.io/badge/arXiv-paper-red?logo=arxiv" />
    </a>
    <a href="#">
        <img src="https://img.shields.io/badge/Huggingface-demo-blue?logo=huggingface&logoColor=white" />
    </a>
</div>

![teaser](assets/images/teaser.png)


## 🛠 Preparation

### Environment
- Linux x86-64 system
- NVIDIA GPU with 24GB VRAM ($a=b=2$)
- CUDA version ≥ 12.4

Larger scene generation may require more VRAM.

### Install
```bash
conda create -n extend3d python=3.10
conda activate extend3d
pip install -r requirements.txt
```

If your GPU does not support pytorch-2.4.0, follow instructions in [SETUP.md](./SETUP.md).


## 🚀 Usage

### Quick Start
```python
from extend3d import Extend3D
from PIL import Image
import imageio

from trellis.utils import render_utils, postprocessing_utils

pipeline = Extend3D.from_pretrained("microsoft/TRELLIS-image-large").cuda()
image = Image.open("assets/examples/0.png")

output = pipeline.run(image)

video = render_utils.render_video(output['gaussian'][0], r=1.6, resolution=1024)['color']

imageio.mimsave('sample_gs.mp4', video, fps=30)

glb = postprocessing_utils.to_glb(
    output['gaussian'][0],
    output['mesh'][0],
    simplify=0.9,
    texture_size=1024
)
glb.export(os.path.join(args.output_dir, 'sample.glb'))
```
You may follow [example.py](./example.py) for detailed hyper-parameters.

### Gradio Demo
```bash
python app.py
```

## 📚 Citation
```bibtex
@inproceedings{yoon2026extend3d,
    title     = {Extend3D: Town-scale 3D Generation},
    author    = {Yoon, Seungwoo, and Kim, Jinmo, and Park, Jaesik},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference},
    year      = {2026}
}
```

##  Acknowledgement
This repository is based on the implementation from [Trellis](https://github.com/microsoft/TRELLIS/tree/442aa1e1afb9014e80681d3bf604e8d728a86ee7).
We sincerely thank the authors for releasing their code.
We also thank the anonymous reviewers for their insightful and constructive feedback.