Skyfall-GS – Synthesizing Immersive 3D Urban Scenes from Satellite Imagery (skyfall-gs.jayinnn.dev)

Skyfall-GS is a new two-stage framework that synthesizes large, explorable, and geometrically accurate city-block scale 3D urban scenes using only multi-view satellite imagery and open‑domain diffusion models—no costly ground-truth 3D scans or annotations required. In the Reconstruction stage the system extracts coarse but consistent geometry from satellite views; in the Synthesis stage it uses diffusion-based image priors to generate high‑quality, close-up appearance. A curriculum-driven iterative refinement strategy progressively fills geometric gaps and sharpens photorealistic textures, and the pipeline renders results with 3D Gaussian Splatting for real-time, free‑flight navigation and interactive exploration. This approach is significant because it delivers the first city-block scale 3D scene creation method that avoids expensive 3D data collection while achieving better cross‑view geometric consistency and more realistic textures than prior methods. Technical highlights include leveraging widely available satellite imagery as a geometry prior, integrating open-domain diffusion models for appearance synthesis, and the iterative refinement loop that balances completeness and photorealism. The result enables immersive, embodied applications—such as simulation, urban planning, AR/VR, and robotics—at scale, and demonstrates a practical path for scaling realistic 3D urban datasets from remote sensing sources.