Geometry-Controlled Satellite Image Synthesis via Diffusion Models

Researchers propose a method for geometry-controlled high-resolution satellite image synthesis using pre-trained diffusion models. The approach leverages skip connection features with windowed cross-attention modules to control the generation process. Comparisons with existing techniques show comparable performance but better alignment with geometry control maps. The study highlights limitations in current evaluation methods for alignment assessment. This work addresses the scarcity of high-resolution satellite images for machine learning applications in land-cover classification, change detection, and disaster monitoring.