Introduction
The first GANs used dense layers and produced small, blurry images. Three steps turned them into the photorealistic image generators that made headlines:
- DCGAN (2015) found a reliable recipe for building GANs from convolutional layers;
- progressive growing (ProGAN, 2017) trained GANs at low resolution first and gradually added detail, reaching 1024 × 1024 faces;
- StyleGAN (2018–2020) redesigned the generator so that "style" is injected at every scale, giving stunning quality and control over features like pose, hair and lighting.
Think of a painter's process: first rough shapes on a small canvas, then gradually larger brushes replaced by finer ones, adding details layer by layer. StyleGAN even lets you control each stage separately: the composition from one reference, the colours from another.