Cloud imagery generation by physics-inspired motion-constrained diffusion networks
This work provides a robust and efficient novel approach for integrating generative models with physical constraints to achieve high-quality spatiotemporal forecasting.