Seonglae ChoIt reconstructs rough geometry using 3D Gaussian Splatting, then leverages pre-trained text-to-image diffusion models to progressively fill in occluded regions (building facades, etc.) through curriculum-based iterative refinement. Without additional 3D data or street-view images, it significantly improves geometric accuracy, texture realism, and view consistency, outperforming existing methods like Sat-NeRF, CityDreamer, and GaussianCity in both quantitative and qualitative evaluations, while enabling real-time exploration. (Satellite → Rough 3D → Probabilistic correction → Back to 3D absorption)
arxiv.org
https://arxiv.org/pdf/2510.15869
Skyfall-GS: Synthesizing Immersive 3D Urban Scenes from Satellite Imagery
Skyfall-GS converts satellite images to explorable 3D urban scenes using diffusion models, with real-time rendering performance.
https://skyfall-gs.jayinnn.dev/
