GeoGen I: Towards General Geospatial Point Data Generation from Text
Generating realistic geospatial vector data is important for evaluating algorithms, index structures, and systems under diverse conditions. Existing synthetic data generators typically rely on simple statistical or procedural models that fail to capture the complexity of real-world spatial patterns. This paper introduces GeoGen I, a generative framework that produces geospatial point distributions from natural language prompts. The system combines contrastive learning, region context, and a diffusion-based generator to create plausible datasets. In the experiments, we test variations of the model and provide both qualitative and quantitative evaluations. Our experiments show that it can generate spatial patterns aligned with different prompts. While the results are promising, many challenges still remain, including in dataset curation and quality, and the model’s ability to capture subtle geospatial constraints.