OSMGen: Highly Controllable Generation of Satellite Images from Structured Data

October 22, 2025 · View on GitHub

Code for “OSMGen: Highly Controllable Generation of Satellite Images from Structured Data.” This repo currently includes the core image generation components. Additional modules and full running instructions will be added.

Overview

OSMGen generates controllable satellite imagery from structured OpenStreetMap (OSM) JSON. It extends ControlNet to accept multiple conditioning signals for precise, localized control over the output.

Key features

  • Multi-modal control: semantic masks, geographic coordinates, and temporal cues
  • Extended ControlNet: custom dual-image conditioning pathway
  • Location encoding: SatCLIP-based geographic embeddings
  • Date encoding: structured temporal representations
  • Editing support: DDIM inversion for targeted image modification workflows

Reference

If you use OSMGen, please cite:

@inproceedings{
ziashahabi2025osmgen,
  title={{OSMG}en: Highly Controllable Satellite Image Synthesis using OpenStreetMap Data},
  author={Amir Ziashahabi and Narges Ghasemi and Sajjad Shahabi and John Krumm and Salman Avestimehr and Cyrus Shahabi},
  booktitle={UrbanAI: Harnessing Artificial Intelligence for Smart Cities},
  year={2025}
}

Contact

Questions or collaborations: