The claim also applies to GANs as you can simply use a masking function to indicate which inputs were used for each optimization timestep, like the author suggests for stochastic gradient descent in remark 5.