My first GAN. Trained on FFHQ Anime with PyTorch. Manually exported to WebGL 2.0.
I believe a GAN of this scale can do much better than this. The model may not be properly trained.

Latent - 32
Dense layer - 256x32, LeakyReLU(0.1)
Reshape to 16x4x4
Upsample - 16x4x4 => 16x8x8
Convolutional layer - 16x16x3x3, LeakyReLU(0.1)
Upsample - 16x8x8 => 16x16x16
Convolutional layer - 16x16x3x3, LeakyReLU(0.1)
Upsample - 16x16x16 => 16x32x32
Convolutional layer - 8x16x5x5, LeakyReLU(0.1)
Upsample - 8x32x32 => 8x64x64
Convolutional layer - 3x8x5x5, Sigmoid
Output - 3x64x64