It's true that there aren't too many practical applications of GAN's yet, but I'd argue that transfer learning is already pretty powerful. It's fairly commonplace to approach a compute vision task by starting with VGG/AlexNet/etc and fine-tuning it on a relatively small dataset.
