We propose OmniCast, a scalable and skillful probabilistic model that unifies weather forecasting across timescales.

A sequence-level EDM-derived diffusion model using progressively increasing noise levels is designed for probabilistic forecasting of complex dynamics like weather.