Here's the thing about augmentation—it's still debated throughout the community on whether it's an absolute requirement for every machine learning project. I've supplied some follow up papers for your reading leisure to understand more modern discussions on data augmentation. In data poor environments, it's generally accepted that carefully selected augmentation can improve accuracy but cannot be used in place of actual data. In data-rich environments, augmentation can be applied more judiciously and will generally improve performance. Data augmentation, however- that is randomly chosen without benchmarks for your learners- can lead to decreased performance for the learners in the long run.

Data augmentation is a big topic in deep learning but is largely not discussed in scholarly journals for GANs. There are quite a few articles we can recommend to learn about vanilla data augmentation (especially for images) seen in the following examples:

• Return of the Devil in the Details: Delving Deep into Convolutional Nets: Example of a paper where they do a rigorous exploration of augmentation around their problems space: https://arxiv.org/pdf/1405.3531.pdf.
• Lecture: Data Augmentation case studies for deep convolutional models https://www.coursera.org/learn/convolutional-neural-networks/lecture/AYzbX/data-augmentation.

• imgaug library: Details on the imgaug library: https://media.readthedocs.org/pdf/imgaug/latest/imgaug.pdf.