Probabilistic Precipitation Nowcasting with Rectified Flow Transformers

Researchers introduced FREUD, a frame-wise encoder and unified decoder for probabilistic precipitation nowcasting.

The model uses rectified flow transformers to preserve uncertainty, support continuous updates, and generate short-term weather forecasts.

FREUD reported state-of-the-art results on the SEVIR benchmark, with stronger performance as the model scales.

The approach could improve severe-weather decision-making by better capturing uncertainty in extreme precipitation events.