Seems like you could validate using random points in the latent space, and crudely verifying that the nearest neighbors in image space aren't similar.