We consider the following generator architecture.
Resnet architecture introduced by Kaiming et al. [11] allows for deeper networks using skip connections.
Unet architecture introduced by Olaf et al. [9] has longer skip connections from encoder to decoder stage.
Resnet Architecture has skip connections which allows for deeper networks by negating vanishing gradient problem.
Image : Wikipedia
Unet Architecture has longer skip connections from encoder layers to decoder layers. This allows the network to understand the features extracted by encoders during the decoding stages. Therefore, Unet proves to be beneficial in Image to Image translation problems.