[ICLR 2026] Sat3DGen: Comprehensive Street-Level 3D Scene Generation from Single Satellite Image
Code Page, Project Page
Author: Ming Qian, Zimin Xia, Changkun Liu, Shuailei Ma, Wen Wang, Zeran Ke, Bin Tan, Hang Zhang, Gui-Song Xia
Upload a satellite image to generate a 3D mesh or render a walkthrough video.
📌 Input requirements: The satellite image should be at zoom level 20 (same as the VIGOR dataset), then will be resized to the input size. You can download satellite tiles at this zoom level from any map tile API (e.g. Google Maps, Bing Maps, Mapbox).
⏳ After generation completes, the 3D preview may take ~10-200 seconds to load. Please wait patiently.
Sample Images — click to load
⚠️ Note: The 3D mesh preview may show slight color distortion. The cause is currently under investigation.
Select a demo image or upload your own.
Default demo panorama: default_panorama.jpg
If you do not select another demo panorama, this one will be used.
⏳ Running on CPU — video rendering is slow (~5 min for 80 frames). Please be patient.
🛰️ Sample Satellite Images — click to load
🌤️ Demo Panorama Street Images — the first one is the default
The panorama image and its corresponding sky mask are used to extract a sky region color histogram, which serves as a lighting condition hint during street-view rendering. This only affects the appearance (illumination/color tone) of the rendered views — it does not alter the underlying 3D NeRF geometry.