Adjust the contrast of the output images by normalizing the pixel values or applying a transfer learning-based post-processing step. This changes the focus from solely optimizing for similarity in Euclidean space to also considering visual aesthetics, which may lead to slightly reduced diversity but improved image quality.