Skyfall-GS: Synthesizing Immersive 3D Urban Scenes from Satellite Imagery

This repository contains the fused 3D Gaussian Splatting (3DGS) .ply files for Skyfall-GS, a framework for synthesizing large-scale, immersive 3D urban scenes.

Project Page | Paper | GitHub

Description

Skyfall-GS is a hybrid framework that synthesizes city-block scale 3D urban scenes by combining satellite reconstruction with diffusion refinement. It leverages readily available satellite imagery to provide realistic coarse geometry and uses open-domain diffusion models to synthesize high-quality close-up appearances. This approach eliminates the need for costly 3D annotations and allows for real-time, immersive 3D exploration.

Visualization

The .ply files in this repository are intended for use with Gaussian Splatting viewers. For the best experience, use a fused PLY file with the following tools:

1. Mip-Splatting Viewer: Use the online viewer. Recommended settings:
   • Up vector: 0,0,1
   • SH degree: 1
   • Camera origin: 0,0,200
2. SuperSplat: You can also use the web-based SuperSplat Editor.

Sample Usage

To render a video from a .ply file using the scripts provided in the official repository:

python render_video_from_ply.py \
    --ply_path <path_to_ply_file> \
    --camera_path <path_to_camera.json>

Citation

If you find this work useful, please consider citing:

@article{lee2025SkyfallGS,
  title = {{Skyfall-GS}: Synthesizing Immersive {3D} Urban Scenes from Satellite Imagery},
  author = {Jie-Ying Lee and Yi-Ruei Liu and Shr-Ruei Tsai and Wei-Cheng Chang and Chung-Ho Wu and Jiewen Chan and Zhenjun Zhao and Chieh Hubert Lin and Yu-Lun Liu},
  journal = {arXiv preprint},
  year = {2025},
  eprint = {2510.15869},
  archivePrefix = {arXiv}
}

Acknowledgement

This codebase is built upon several open-source projects: Mip-Splatting, WildGuassians, FlowEdit, MoGe, and SatelliteSfM.