Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion

2026-10-08

AI Summary

This study addresses the limitations of existing diffusion models for precipitation forecasting, which rely on multiple independent components and incur high computational costs due to iterative denoising. To this end, we propose JWS, a single-stage, end-to-end diffusion model that operates directly in radar space. By eliminating lossy compression to mitigate uncertainty, JWS integrates masked asynchronous diffusion, timestep sampling, and a scoring rule objective function to enable efficient generation in fewer steps while streamlining both training and inference pipelines. Experimental results demonstrate that JWS achieves state-of-the-art performance on benchmarks such as SEVIR. Compared to baseline models, it attains over 17× inference acceleration with fewer parameters, substantially reducing computational overhead. This work establishes an efficient new paradigm for probabilistic precipitation nowcasting.

Abstract

Generative diffusion models are well-suited for probabilistic precipitation nowcasting, but existing approaches often rely on separately trained compression or deterministic forecasting components and remain costly at inference due to iterative denoising. We introduce Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model which addresses both issues by forecasting directly in radar space and enabling few-step generation. Radar-space modeling greatly simplifies training and inference and eliminates uncertainty arising from lossy compression. JWS combines Masked Asynchronous Diffusion, a timestep-sampling scheme that preserves clean context while adapting diffusion training to high-dimensional spatio-temporal data, with a simple scoring-rule objective that aligns training with probabilistic forecasting and unlocks few-step generation. On the SEVIR and MeteoNet benchmarks, JWS achieves state-of-the-art probabilistic forecasting performance at reduced training and inference cost. Even our smallest model remains competitive using substantially fewer parameters and more than 17x faster inference.

Problem

• precipitation nowcasting
• diffusion models
• probabilistic forecasting
• inference cost
• end-to-end

Innovation

• End-to-end diffusion model
• Radar-space modeling
• Masked Asynchronous Diffusion
• Scoring-rule objective
• Few-step generation