This uses a particular type of GAN of which it is possible to calculate the likelihood. So it can be trained by either adversarial minmax or by max likelihood.

This uses a particular type of GAN of which it is possible to calculate the likelihood. So it can be trained by either adversarial minmax or by max likelihood. The restriction on the GAN is that it is volume preserving.