Yep, this is a current area of research for content generation.

I think most current approaches build some transform to a latent space and then compare generated images with their nearest neighbors in the training set. If they're identical then your network just learned to reproduce the dataset.