Sat3DGen – Street-Level 3D Scene Generation from Satellite Images

[ICLR 2026] Generate street-level 3D scenes from a single satellite image.

• 📄 Paper (OpenReview)
• 🌐 Project Page
• 💻 GitHub
• 🤗 Model
• Arxiv

Features

• 3D Mesh Generation: Upload a satellite image → get a downloadable .obj mesh with in-browser 3D preview.
• Video Rendering: Select a satellite image and sky panorama → render a walkthrough video along a trajectory.

Usage

Upload a satellite image or select one from the examples, then click "Generate". The model weights are loaded automatically from HuggingFace.

Input Requirements

The satellite image should be at zoom level 20 (same as the VIGOR dataset), resized to 512×512 pixels. You can download satellite tiles at this zoom level from any map tile API (e.g. Google Maps, Bing Maps, Mapbox).

Note: GPU hardware is recommended for reasonable inference speed.