Perceptual Loss

Perceptual loss methods compare outputs using features or quality models that capture structure, texture, sharpness, or sensory sensitivity, rather than relying only on pixel-wise errors such as mean squared error. They are used to train image restoration and super-resolution, generative models, speech enhancement, medical-image synthesis, and 3D reconstruction, with approaches ranging from pretrained feature losses such as VGG or domain-specific encoders to learned quality predictors and perceptually weighted signal errors. Current work adapts these representations to particular domains and combines perceptual objectives with distortion, structural, or rate constraints. A central tension is that better perceptual scores do not always mean outputs are more useful or preferred: metric optimization can introduce artifacts, hurt scientific fidelity, or disagree with listening tests and human ratings.