Accurate representation of oceanic conditions is fundamental for reliable climate modeling, weather forecasting, and environmental monitoring. However, ocean models and observational datasets often exhibit systematic biases due to limitations in model physics, parameterizations, resolution, or observational coverage. In this work, we propose a diffusion model for bias correction. We systematically evaluated its performance for Sea Surface Temperature on the oceanic sea surface temperature generation by varying different hyperparameters in the U-Net architecture. The model is trained to denoise simulated data and reconstruct the SST field guided by reanalysis data. Our results demonstrate that increasing the base channel’s depth significantly improves the model’s performance, with improvements in convergence speed, reconstruction accuracy, and spatial detail retention. Quantitative metrics such as root mean squared error (RMSE), Pearson’s correlation coefficient (PCC), and coefficient of determination (R 2 ) show notable gains up to a base channel depth of 64, beyond which performance gains plateau. A detailed temporal generalization analysis using seasonal batches every two months confirms the robustness of the model in varying SST regimes. At the same time, qualitative visualizations show sharp and coherent reconstructions with minimal error. The study highlights the trade-off between model complexity and performance and identifies 64 base channels as a computationally efficient and accurate configuration for SST modeling using diffusion-based generative methods.