Loading
Research · Thread 09
Splatting without photographs
Training 3D Gaussian Splats on planetary bodies nobody has fully photographed.
3D Gaussian Splatting assumes what every other reconstruction method assumes: full camera coverage of a real object, trained from photographs taken from many known or estimated angles. A planet breaks that assumption twice over, partial imagery only, and no photographs at all of its dark side. Splatlas answers it by applying 3DGS past its usual capture-based use case: known-pose Blender renders of open NASA and USGS elevation and imagery data stand in for missing photography, skipping structure-from-motion entirely, and the splats are emission-baked, lit per body rather than depending on captured lighting, so the reconstruction reads correctly from any angle, including ones with no source photo behind them.
Paper to shipped
Implementation
Known-pose renders instead of structure-from-motionOrdinary 3DGS training solves for camera pose from a set of photos with unknown positions. Satellite imagery is not a photoset like that, so the pipeline displaces a UV sphere by the real elevation model where one exists and renders 150 to 300 views from cameras whose position is already known by construction, skipping structure-from-motion and making training tractable on one GPU instead of a cluster.
Shipped
Emission-baked splats for bodies with no photographThe gas giants have no solid surface to survey and no ground-truth photo the way Earth has. Rather than fall back to a textured sphere for them, they are built as emission-baked splats from open colour maps, rendered through the same real-time pipeline as the captured worlds, one renderer, one performance profile, for every body.
Shipped
Live at splatlas.spaceThe Sun to Neptune as navigable splats with real orbits and a date scrubber, all real-time in the browser: positions move against real orbital mechanics rather than a looping animation.