How does the pix2pixHD model improve upon the traditional conditional GAN loss, and what is the total loss function used in this study?

The pix2pixHD model improves the traditional conditional GAN loss by adding feature matching loss and perceptual (VGG) loss to the standard GAN loss. The total loss used in the study is the sum over multi-scale discriminators of the GAN loss plus weighted sums of feature matching and VGG losses, with both weights set to 10.

In pix2pixHD, the basic conditional GAN loss alone is not enough to produce realistic high-resolution images, so the model augments it with two additional terms. Feature matching loss compares intermediate discriminator features between real and generated images, encouraging the generator to match the learned feature representations. Perceptual loss (VGG loss) compares high-level features extracted by a pretrained VGG network, preserving perceptual quality rather than just pixel accuracy. The combined objective is L_total(G, D_k) = sum_k L_GAN(G, D_k) + lambda_FM sum_k L_FM(G, D_k) + lambda_VGG L_VGG(y, G), where k indexes the multiple discriminators operating at different scales. In this study, the loss weights were set to lambda_FM = 10 and lambda_VGG = 10.