Accurate forecasting of Convective Available Potential Energy (CAPE) is critical for severe weather prediction. However, the operational GFS model exhibits a persistent low-CAPE bias. In this work, we apply a two-step regression–diffusion model (NVIDIA CorrDiff) to address this issue. Our results indicate that while a standard U-Net can successfully reduce the bulk systematic bias, the generated output remains overly smoothed. This occurs because, for data with long-tailed statistical distributions such as CAPE, standard models trained on mean squared error fail to capture rare, high-magnitude events. In contrast, generative diffusion models can reproduce realistic, small-scale features similar to the ground truth by learning to reverse a noise-corruption process through a series of iterative denoising steps. Our study begins with bias correction for the 24 h forecast. We then extend this by applying the model—trained solely on 24 h data—to correct forecasts of up to 120 h. This strategy leverages our finding that the GFS forecast bias is highly persistent over time. Furthermore, our examination of CAPE’s joint Probability Density Functions emphasizes the necessity of matching machine learning models to the target variable’s statistical properties. Ultimately, the effectiveness of CorrDiff highlights its potential for other challenging applications involving small-scale phenomena with long-tailed distributions.