An intercomparison of generative machine learning methods for downscaling precipitation at fine spatial scales

Machine learning (ML) offers a computationally efficient approach for generating large ensembles of high-resolution climate projections, but deterministic ML methods often smooth fine-scale structures and underestimate extremes. While stochastic generative models show promise, few studies have compared their skill under both present-day and future climates. This study compares Generative Adversarial Networks (GANs), flow matching and diffusion models across multiple configurations for downscaling daily precipitation from a regional climate model (RCM) over New Zealand. Model skill is assessed across spatial structure, distributional metrics, climatological means, extremes, ensemble calibration, and climate change signals. Unlike GANs, diffusion and flow matching models generate predictions through many sequential steps. Here we show that using higher-order differential equation solvers, the number of steps required can be reduced with only a minor reduction in skill, heavily reducing the computational burden for downscaling large ensembles, which may have otherwise prevented their use in operational settings. Overall, GANs, flow matching and diffusion perform competitively across most metrics, except that diffusion and flow matching produce higher-fidelity predictions and better-calibrated ensembles compared to GANs -which are under-dispersive. Most approaches capture mean precipitation signals reasonably well, but underestimate end-of-century climate change signals of extreme precipitation, despite being trained on RCM simulations spanning the future period. Only one GAN and one flow matching configuration can reproduce this change signal reliably. These results highlight the importance of evaluating model performance across a comprehensive set of metrics, and that neither visual realism nor good skill on standard metrics guarantee skill in predicting climate change signals.