Extreme Event Likelihoods with Guided Generative Models | NVIDIA Technical Blog

This post details the practical application of guided diffusion models for estimating the likelihood of rare extreme events, specifically using NVIDIA cBottle and Earth2Studio for tropical cyclone risk estimation. It illustrates how to steer a diffusion model towards desired outcomes (e.g., a tropical cyclone near Florida) and then use the odds-ratio diagnostic to reweight guided samples. This method enables importance sampling, significantly reducing the standard error of probability estimates compared to traditional Monte Carlo methods. The post provides a code example for running guided sampling and calculating the log-odds ratio, highlighting the potential for broader applications beyond climate science.