Adversarial networks are super cool. The only theoretical issue I have with them, is how do you measure their quality or how much they are overfitting? The gene

Adversarial networks are super cool. The only theoretical issue I have with them, is how do you measure their quality or how much they are overfitting? The generative net could learn to produce perfect copies of images from the training set, and that would optimally satisfy the loss function. Of course the advantage of adversarial nets is they can use the unlimited amounts of training data you can find on the internet, so overfitting isn't a huge concern. But still, unlike normal machine learning, you can't make a validation set and test how well they are performing on it. And that limits the ability to set their hyperparameters optimally.