Summary

In this chapter, we entered the realm of high-definition image generation with ProGAN. ProGAN first trains on low-resolution images before moving on to higher-resolution images. The network training becomes more stable by growing the network progressively. This lays the foundation for high-fidelity image generation, as this coarse-to-fine training method is adopted by other GANs. For example, pix2pixHD has two generators at two different scales, where the coarse generator is pre-trained before both are trained together. We have also learned about equalized learning rates, minibatch statistics, and pixel normalization, which are also used in StyleGAN.

With the use of the AdaIN layer from style transfer in the generator, not only does StyleGAN produce better-quality images, but this also allows control of features when mixing styles. By injecting different style code and noise at different scales, we can control both the global and fine details of an image. StyleGAN achieved...