What is the idea behind Cycle-GAN? What losses are used in it?

Cycle-GAN is an architecture of generative adversarial networks designed for learning to translate images from one domain to another without paired examples (unsupervised image-to-image translation).

The main idea is to ensure cyclic consistency: if an image is transformed from domain A to domain B, and then back to A, the result should be close to the original image.

The following types of losses are used:

• Adversarial Loss: encourages generated images to resemble real images from the target domain.

• Cycle Consistency Loss: penalizes discrepancies between the original image and the image reconstructed after double transformation (A→B→A and B→A→B).

• Identity Loss: optional, helps to preserve color palette and structure when the image already belongs to the target domain.

Thus, Cycle-GAN allows training transformation models between domains without the need for paired images.