Contributions
Primary contribution:
Propose a training methodology for GANs where start with low-resolution images, and then progressively increase the resolution by adding layers to the networks.
Others:
Increasing variation using minibatch standard deviation
A new normalization in G and D
Propose sliced Wasserstein distance (SWD) to estimate the statistical similarity.
progressive training
Training frame
Transition from 16x16 to 32x32
Benefits of progressive Training
More Stable:
Early on, the generation of smaller images is substantially more stable because there is less class information and fewer modes. By increasing the resolution little by little we are continuously asking a much simpler question compared to the end goal of discovering a mapping from latent vectors.
The reduced training time:
With progressively growing GANs most of the iterations are done at lower resolutions, and comparable result quality is often obtained up to 2–6 times faster, depending on the final output resolution.
sliced Wasserstein distance
Motivation
The previous methods like MS-SSIM do not directly assess image quality in terms of similarity to the training set.
Therefore, they propose a new metric sliced Wasserstein distance which will measure the distance between training set and generated samples.
Experiments
Convergence and Training speed
Experiments
Celeb High quality
lsun
CIFAR10
