The appropriate use of objective functions is a fundamental research topic in the field of AI. Utilizing an objective function that aligns with the intended goal is key to efficient AI. No matter how advanced the neural architecture, a poorly chosen objective function leads to inefficient learning in neural networks. For visual information processing tasks where both input and output are image data, distortion measures (e.g., mean squared error) are widely used for both training and evaluating neural networks. However, there is a known trade-off between distortion and perceptual quality. We have studied objective functions for neural networks to balance perceptual quality and distortion (Cheon et al., 2018; Choi et al., 2020). Additionally, we have explored how to use language information for training neural networks to enhance perceptual quality of generated images (Lee et al., 2024).