Can multimodal prompts make change detection work across any landscape?

Satellite change detection—spotting urban growth, disaster damage, or environmental shifts—fails when trained on one region and tested on another. OmniCD combines image and text prompts (descriptions, semantic maps, metadata) in a single system that handles everything from simple before-after comparisons to understanding *what* changed semantically. The team also released RSITCD, a 300K+ multimodal dataset, and demonstrated stronger generalization and robustness than prior methods across multiple benchmarks.