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

Surprisingly or not surprisingly, this model runs only 1.5x as slow on my machine despite having 5x number of parameters. Visually, it produces fewer "duplicates".

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