# RANGE

The official implementation of a CVPR 2025 paper on retrieval-augmented neural fields for embedding geographic coordinates, with a pip-installable package that returns high-resolution visual-feature embeddings for any latitude and longitude.

RANGE (Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings) is a retrieval-augmented framework that directly estimates visual features for a geographic location, so the resulting embeddings capture high-resolution information. It enforces a spatial smoothness constraint that lets embeddings be generated at different desired frequencies, and the authors report it outperforms prior location-embedding methods such as SatCLIP, GeoCLIP and CSP on a range of downstream tasks.

The rangegeo package on PyPI provides inference only: it downloads the SatCLIP backbone and the RANGE retrieval database from Hugging Face on first use, and model.encode takes an array of longitude and latitude pairs and returns embeddings, batching automatically on GPU. The model is a regular torch.nn.Module and can be swapped for a fine-tuned backbone or a regenerated database.

Built with Python and PyTorch; the full research code for training, evaluation and baselines is separate from the inference package and requires cloning the repository and downloading precomputed databases from Hugging Face. MIT licensed.

Website: https://github.com/mvrl/RANGE