How Do Generative Downscaling Models Use Noise? Insights from Conditional Precipitation Downscaling

Abstract

Generative deep learning models offer a promising framework for stochastic precipitation downscaling, but it remains unclear how they utilize large-scale conditioning inputs versus stochastic latent variables. This challenge limits physical interpretability and raises concerns about whether generated fine-scale variability reflects meaningful uncertainty or unconstrained randomness. Here, we investigate how different training objectives shape stochastic information usage in conditional precipitation downscaling, for a demanding task of 16× increase in spatial resolution. We analyze two models based on the same U-Net architecture: an MSE-trained U-Net and the same U-Net trained within a conditional Wasserstein generative adversarial network (WGAN). Using Integrated Gradients, noise ensemble sensitivity analysis, and layer-wise representation analysis, we quantify how each model uses low-resolution (LR) precipitation inputs and Gaussian noise to generate high-resolution precipitation fields. We find that the predicted mean precipitation is overwhelmingly attributed to the LR input in both models, whereas maximum intensity exhibits a relatively greater stochastic contribution. However, only the WGAN meaningfully exploits noise in practice to stochastically represent underdetermined aspects of the downscaling task. The MSE-trained U-Net exhibits comparatively limited and less organized utilization of stochastic input. Internal representation analysis further shows that these differences emerge primarily at the bottleneck and early decoder stages of the U-Net architecture. These results demonstrate that meaningful stochastic behavior in generative AI downscaling is not guaranteed by architectural design alone but is strongly shaped by the training objective. The study provides a framework for evaluating the physical credibility and trustworthiness of generative models in hydroclimatic applications.