The following diagram shows shows our results in one spot:

As you may have noticed, these results show us that there's some work left to do. You can see that we stopped our process intentionally at epoch 50, rather than 100. Why? Well, there are many knobs and parameters that can be tuned to get different results, and it appears this generator has started to overfit to certain types of structures. We may be able to enforce additional generalization, but there are some incredibly hard-to-see examples in this dataset.

The generator will only learn what it trains from. In this case, then, it may be of use to train a smaller subset of data and learn a certain type of structure.