“Progressive Growing of GANs for Improved Quality, Stability, and Variation,” posted to arXiv on October 27, 2017 by Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen of NVIDIA, introduced a training recipe that let generative adversarial networks reach much higher resolutions than before. Training a GAN to produce a large, detailed image all at once was unstable; ProGAN avoided this by building up to it gradually.
The method starts with both the generator and discriminator operating at a tiny resolution, learning the coarse structure of images. As training proceeds, new layers are smoothly faded in to handle progressively finer detail, doubling the resolution step by step until the networks reach megapixel scale. Because each new level of detail builds on an already-stable foundation, training stays well-behaved, runs faster, and produces more varied outputs. The authors demonstrated convincing 1024-by-1024 face images trained on the CelebA dataset.
ProGAN was the direct predecessor to StyleGAN, and its progressive idea shaped the way researchers thought about scaling generative models. For a general reader, it illustrates a recurring engineering pattern in AI: rather than attacking the hardest version of a problem head-on, you solve an easy version first and grow the difficulty in controlled steps, which often makes an otherwise intractable training process succeed.