// the find
jayin92/Skyfall-GS
[ECCV 2026] Skyfall-GS: Synthesizing Immersive 3D Urban Scenes from Satellite Imagery
Skyfall-GS is the official implementation for an ECCV 2026 paper on generating large, explorable 3D Gaussian-splat scenes of cities from satellite imagery, using a diffusion model to fill in close-up texture detail that satellite data can't provide. It's for computer vision researchers working on 3D reconstruction or novel-view synthesis from overhead imagery, not for anyone wanting a drop-in tool.
The two-stage pipeline (coarse reconstruction, then iterative dataset update with diffusion-based refinement) is documented with the actual hyperparameters used for each stage, not just hand-wavy descriptions. It ships pretrained PLY models, preprocessed datasets, and eval data all on Hugging Face so you can skip the expensive parts and go straight to inspecting results. Evaluation covers five real metrics (PSNR, SSIM, LPIPS, CLIP-FID, CMMD) against six baseline methods, which is more rigorous than most paper repos bother with.
Install is the usual 3DGS research nightmare: a pinned CUDA 12.8 toolkit, three custom CUDA submodules (diff-gaussian-rasterization-depth, simple-knn, fused-ssim) that have to compile from source, and a force-reinstall of torch after the fact — any version drift and you're debugging nvcc errors instead of running the model. Custom dataset support just points you at a separate SatelliteSfM repo for preprocessing, so this isn't self-contained for anyone without JAX/NYC-style satellite data already in hand. No GPU/VRAM requirements or training time estimates anywhere, and there's no test suite or CI, so you're trusting that the scripts in scripts/ still match the current code.