Generating bums people with a deep convolutional GAN

Dataset: https://www.kaggle.com/datasets/greatgamedota/ffhq-face-data-set

Trying to squeeze every last drop of quality from DCGAN architecture with mininmal effort.

Contains:

• standard tricks (reflect padding, betas);
• normalization and Tanh() output;
• some augmentations;
• weight initialization according to DCGAN paper;
• SiLU;
• PixelNorm (akin to ProGAN/StyleGAN)