High-resolution precipitation is critical for hydrologic and climate-risk applications, but climate models are too coarse to resolve storm-scale extremes. We compare a deterministic U-NET with two generative models, a Wasserstein Generative Adversarial Network (WGAN) and a diffusion model, for 8× and 16× precipitation downscaling using ERA5-Land. U-NET is stable but smooths extremes, whereas generative models better capture variability and heavy tails while introducing greater uncertainty.