PreprintCC BY 4.0arXiv
Diffusion Distillation for Efficient Weather Ensembles
Abstract
Diffusion models generate skillful weather ensembles but require costly iterative sampling. We introduce a supervised energy-distance distillation method that compresses a multi-step diffusion teacher into a single-step student by aligning student forecasts with teacher samples and ground-truth observations. Experiments on global forecasting and typhoon-track prediction show that our student outperforms existing distillation methods and preserves skill for extreme events. It matches or surpasses the teacher across key metrics using only one neural function evaluation per autoregressive step.
Related
- Value of Adding AI Weather Model Ensembles to a Multi-model Probabilistic Blend
- Distribution Matching Distillation for Continuous Diffusion Language Models
- DASH: Dual-Branch Score Distillation for Guidance-Calibrated Compact Diffusion Models
- POETS: Uncertainty-Aware LLM Optimization via Compute-Efficient Policy Ensembles
- Fast Probabilistic Wind Power Forecasting via One-Step Consistency Distillation and Memory-Efficient CRPS Optimization
BibTeX
@misc{yang2026diffusion,
title = {{Diffusion Distillation for Efficient Weather Ensembles}},
author = {Yiming Yang and Valentin Brekke and James Briant and Serge Guillas},
howpublished = {arXiv},
year = {2026},
month = aug,
doi = {10.48550/arxiv.2608.27728},
eprint = {2608.27728},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2608.27728},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Diffusion Distillation for Efficient Weather Ensembles
AU - Yiming Yang
AU - Valentin Brekke
AU - James Briant
AU - Serge Guillas
PY - 2026
DA - 2026/08/27
PB - arXiv
DO - 10.48550/arxiv.2608.27728
UR - https://arxiv.org/abs/2608.27728
AB - Diffusion models generate skillful weather ensembles but require costly iterative sampling. We introduce a supervised energy-distance distillation method that compresses a multi-step diffusion teacher into a single-step student by aligning student forecasts with teacher samples and ground-truth observations. Experiments on global forecasting and typhoon-track prediction show that our student outperforms existing distillation methods and preserves skill for extreme events. It matches or surpasses the teacher across key metrics using only one neural function evaluation per autoregressive step.
ER -