ProGAN

ProGAN stabilizes high-resolution image synthesis by incrementally growing generator and discriminator layers during training.

Developed by NVIDIA researchers in 2017, Progressive Growing of GANs (ProGAN) solves the instability of high-dimensional generative modeling. The architecture starts by training on 4x4 pixel images, then doubles resolution in stages until reaching 1024x1024. This phased approach (using pixel-wise normalization and a minimax loss function) prevents the generator from collapsing during early training. By focusing on coarse features before fine textures, ProGAN produces photorealistic faces with unprecedented clarity and reduced training time on hardware like the Tesla V100.

https://arxiv.org/abs/1710.10196