Skyfall-GS

Why we included this project

Building large-scale 3D urban scenes usually means collecting expensive real-world scans. Skyfall-GS sidesteps that by starting from satellite imagery, which supplies rough but reliable geometry, and letting a diffusion model fill in photorealistic close-up detail. The result is a pipeline that needs no 3D annotations and renders in real time, so you can walk through a reconstructed city rather than inspect a static mesh. The method uses a curriculum-driven refinement loop that progressively completes geometry and sharpens textures across passes. If you work on generating city-scale 3D content from aerial data, the repo bundles the code, trained PLY models, and paired training and evaluation datasets for direct experimentation.