VectorSynth: Fine-Grained Satellite Image Synthesis with Structured Semantics

<-- -->

Daniel Cher*, Brian Wei*, Srikumar Sastry, Nathan Jacobs

(*Corresponding Author)

This repository is the official implementation of VectorSynth. VectorSynth is a suite of models for synthesizing satellite images with global style and text-driven layout control.

⏭️ Next

• Release PyTorch ckpt files for all models
• Release dataset

🧑‍💻 Setup and Training

Create a conda environment:

conda env create -f environment.yaml
conda activate vectorsynth

See train.md for training details.

📊 Dataset Generation

See dataset.md for detailed instructions on generating the dataset(s) from OpenStreetMap data.

🌏 Inference

See inference.md for a complete inference example.

Requirements:

• Trained ControlNet checkpoint (diffusers format)
• Render encoder checkpoint

We use 🤗 Diffusers for inference. VectorSynth requires special preprocessing through our render encoder to convert vector map embeddings into control signals for the diffusion model.

📑 Citation

@inproceedings{cher2025vectorsynth,
  title={VectorSynth: Fine-Grained Satellite Image Synthesis with Structured Semantics},
  author={Cher, Daniel and Wei, Brian and Sastry, Srikumar and Jacobs, Nathan},
  year={2025},
  eprint={arXiv:2511.07744},
  note={arXiv preprint}
}

🔍 Additional Links

Check out our lab website for other interesting works on geospatial understanding and mapping:

• Multi-Modal Vision Research Lab (MVRL) - Link
• Related Works from MVRL - Link
• See our previous work - Link