In the original GAN paper (table 2), why is it mentioned that you can sample deep directed graphical models without a Markov chain (well, they say without difficulties, but others list MCMC as a difficulty).
I was wondering how this is done because I have only seen MCMC based approaches.
