Image-to-image translation has been revolutionized with gan-based methods.
However, existing methods lack the ability to preserve the identity of the
source domain. As a result, synthesized images can often over-adapt to the
reference domain, losing important structural characteristics