My third GAN after the previous one. Trained on FFHQ Anime with PyTorch and manually exported to WebGL 2.0.

This version uses a variational autoencoder, which produces results with higher variety and more accurate colors.

==== Generator ====
Latent - 32
Dense layer - 512x32
 => 512, LeakyReLU(0.1)
Reshape to 32x4x4
ConvTranspose2d - 32x64x4x4
 => 64x8x8, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 64x32x4x4
 => 32x16x16, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 32x16x4x4
 => 16x32x32, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 16x3x4x4
 => 3x64x64, Sigmoid