The recent surge in machine learning (ML) methods has opened new possibilities for improving the resolution of coarse global climate model (GCM) simulation outputs through downscaling. These advancements aim to capture critical aspects of climate variability and change at regional scales, offering computationally efficient alternatives to traditional regional climate models (RCMs). This study presents a comparative analysis of precipitation outputs from state-of-the-art machine learning models including a Vision Transformer, Diffusion Model, Spherical Fourier Neural Operator and a Long Short-Term Memory model, and RCM simulations from the CORDEX-Australasia ensemble. Using a standardised benchmark framework, we evaluate model skills across diverse precipitation characteristics, leveraging Australia’s highly variable and geographically diverse climate as a test bed. Our findings indicate that ML-based climate downscaling performs within the range of dynamical models across all evaluated metrics and climate regimes, with the added advantage of significantly reduced computational costs. Notably, the optimal ML model varies depending on the metric and climate regime, underscoring the value of an ensemble approach. State-of-the-art ML models have the potential to augment regional precipitation ensembles, complementing dynamical downscaling to better quantify and characterise precipitation changes under climate variability and change.