ProGAN overview

In a typical GAN setting, the generator output shape is fixed. In other words, the training image size does not change. If we want to try to double the image resolution, we add an additional upsampling layer to the generator architecture and start the training from scratch. People have tried and failed to increase image resolution by this brute-force method. The enlarged image resolution and network size increases the dimension space, making it more difficult to learn.

CNNs faced the same problem and solved it by using a batch normalization layer, but this doesn't work well with GANs. The idea of ProGAN is to not train all the layers simultaneously but start by training the lowest layer in both the generator and the discriminator, so that the layer's weights are stabilized before adding new layers. We can see it as pre-training the network with lower resolutions. This idea is the core innovation brought by ProGAN, as detailed in the academic paper Progressive...