Generative adversarial networks are a type of generative model. It is close in spirit to Variational autoencoders, but has key differences. The main one is the way the model is trained, which uses an adversarial equilibrium between training a generator and training a discriminator.
(Richardson and Weiss 2020) This paper seems to show that image-to-image translation models are ill-posed and imply the image transformation should always be very local.
The authors propose a simple linear model for solving complex image-to-image translation tasks which seems to yield competitive results.
- Richardson, Eitan, and Yair Weiss. July 24, 2020. "The Surprising Effectiveness of Linear Unsupervised Image-to-Image Translation". arXiv:2007.12568 [Cs]. http://arxiv.org/abs/2007.12568.