What is the main architectural difference between DCGAN and WGAN? For which problems each models can be more useful than the other one?
DCGAN is more about network architecture alterations, while WGAN is an change to the loss function. There's nothing stopping you from using the DCGAN architecture with the WGAN objective function: all this means is minimizing an approximate Wasserstein loss, rather than a Jensen-Shannon divergence, using a particular network architecture. The WGAN (or its followups, e.g. WGAN-GP) objective is agnostic to the architecture. The only thing (that I can think of) that you need to watch for is the use of batch norm: DCGAN recommends putting it everywhere, but (at least for WGAN-GP) it messes with statistics of the critic regularization.