SinGAN

SinGAN is a type of GAN aimed to be trained on a single image. It makes use of generators and discriminators at different scales and builds output images with a sequence of generators, discriminators, and upsampling. The images that discriminators see are sliced randomly, which prevents the discriminator from remembering the big picture but focus on details at each scale ^[@shahamSinGANLearning2019].

The SinGAN architecture ^[@shahamSinGANLearning2019]:

^[@awiszusWorldGANGenerativeModel2021] and ^[@awiszusTOADGANCoherentStyle2020] have adapted the SinGAN approach to generate discrete data, which is a challenge for GANs that favour continuity.

Architecture for the TOAD-GAN model used to generate level for Super Mario Bros @awiszusTOADGANCoherentStyle2020:

and World-GAN for Minecraft worlds@awiszusWorldGANGenerativeModel2021: