Summary
Since the inception of GANs and VAEs in 2014, significant advancement has been made in 2D image generation. Generating high-fidelity images is still challenging in practice as it requires huge amounts of data, computing power, and hyperparameter tuning. However, as demonstrated by StyleGAN, it seems that we now have the technology to do this, especially in face generation.
In fact, at the time of writing this book, there haven't really been any major breakthroughs in this area since 2018. With this book, we have included all the important techniques leading to BigGAN. These techniques include the use of AdaIN and self-attention modules, which are now commonplace even in adjacent fields such as video synthesis. This gives us a solid foundation to explore other emerging generative technologies.
In this chapter, we looked back at the things we have learned and summarized them in different groups, such as losses and normalization techniques. We then looked at some...
