CityGaussian

This repo contains official implementations of our series of work in large-scale scene reconstruction with Gaussian Splatting, Star ⭐ us if you like it!

• CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes (ICLR 2025)
• CityGaussian: Real-time High-quality Large-Scale Scene Rendering with Gaussians (ECCV 2024)

The links above point to the papers. The main branch now has been rebased to Gaussian Lightning v0.10.1. If you need to visit the original V1 code, please turn to the V1-Original branch. If you need support for VGGT-X, please visit guidance here. Feel free to explore the repository!

👏 Features

• CityGaussian-style multi-gpu reconstruction with controllable memory cost and no limit on GPU amount
• Analysis of model partition and data assignment
• 2DGS-style mesh extraction & Large-scale scene geometric performance evaluation
• Trajectory aligned rendering & mesh video generation with floater removement
• Joint pose and 3DGS optimization for imperfect COLMAP results like from 3D foundation models.
• Features of Gaussian Lightning, including web viewer, MipSplatting, AbsGS, StopThePop, etc.

Scene	SSIM↑	PSNR↑	LPIPS↓	Precision↑	Recall↑	F1-Score↑	#GS(M)
LFLS	0.744	23.44	0.246	0.556	0.400	0.466	8.19
SMBU	0.794	24.00	0.185	0.559	0.523	0.541	5.33
Upper Campus	0.779	25.78	0.186	0.654	0.394	0.491	7.87
MatrixCity Aerial	0.859	27.26	0.175	0.432	0.790	0.559	8.57
MatrixCity Street	0.791	22.32	0.344	0.325	0.797	0.461	7.40

Note for street view, the F1-Score is lower than that reported in paper, since we sacrifice precision for a better recall and more complete road surface. If unbroken road is prefered, you can adjust depth_ratio to 0.0, but surface reconstruction performance will be worse. The checkpoints of CityGSV2 can be found here:

• Baidu Netdisk: https://pan.baidu.com/s/1tRKiJzMLk2-zoyvoa9bkqA?pwd=1b4r
• Hugging Face: https://huggingface.co/TeslaYang123/CityGaussianV2

📰 News

[2025.10.18] Adds support and guidance for joint pose and 3DGS optimization of VGGT-X!

[2025.04.04] A Discord is now available. Come for a deeper and more instant communication about city reconstruction!

[2025.03.26] The repo now belongs to the research organization Linketic.

[2025.01.31] Checkpoints of CityGaussian V2 has been released!

[2025.01.22] CityGaussian V2 has been accepted by ICLR 2025!

[2025.01.22] Code of CityGaussian V2 is now released. Welcome to try it out!

[2024.11.04] Announcement of our CityGaussianV2!

[2024.10.12] Checkpoints of V1 on main datasets have been released!

[2024.08.05] Code of CityGaussian V1 is available!

🛠 Getting Started

• Installation
• Data Preparation
• Run and Eval
• Video Rendering on GS and Mesh
• VGGT-X

📝 TODO List

• [ ] Official Implementation of Appearance Embedding.
• [ ] Support of V1 style LoD.
• [x] Support of VGGT-X.
• [x] Release the checkpoint of CityGaussian V2.
• [x] Release the V2 of CityGaussian.
• [x] Release code and checkpoints of CityGaussian.
• [x] Release ColMap results of main datasets.

Star History

📄 License

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

🤗 Citation

If you find this repository useful, please use the following BibTeX entry for citation.

@inproceedings{
    liu2024citygaussianv2,
    title={CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes},
    author={Yang Liu and Chuanchen Luo and Zhongkai Mao and Junran Peng and Zhaoxiang Zhang},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=a3ptUbuzbW}
}

@inproceedings{liu2025citygaussian,
  title={Citygaussian: Real-time high-quality large-scale scene rendering with gaussians},
  author={Liu, Yang and Luo, Chuanchen and Fan, Lue and Wang, Naiyan and Peng, Junran and Zhang, Zhaoxiang},
  booktitle={European Conference on Computer Vision},
  pages={265--282},
  year={2025},
  organization={Springer}
}

👏 Acknowledgements

This repo benefits from 3DGS, 2DGS, TrimGS, LightGaussian, Gaussian Lightning. Thanks for their great work!

❓ FAQ

• Out of memory occurs in training. To finish training with limited VRAM, downsampling images or adjusting max_cache_num (we used a rather large 1024) in train_large.py can be a useful practice. Besides, you can increase prune_ratio in parallel tuning to further reduce memory cost.

• Generation of COLMAP results. We use the ground-truth poses offered by datasets and separately match the train and test sets. And this will be faster and more robust than match from scratch. But indeed it still costs a lot of time.

• Most blocks are not trained. The main reason here is the data assigned to most blocks are too few (<50), and to prevent overfitting these blocks won't get trained. This can be attributed to unreasonable aabb setting, please try to adjust it and see if things work.