A variation of my fourth GAN trained on the same day.
Intended to test the possibility of running large CNN models (especially with 3x3 convolutional layers) in web browsers in real time.

This is slower and visually worse than the previous one due to convolutional layers at the end smoothing out high-frequency details.

==== Generator (314k params) ====
Latent - 32
Dense layer - 768x32+768
 => 768, LeakyReLU(0.1)
 => 48x4x4
Conv2d - 48x64x3x3
 => 64x4x4, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 64x64x4x4
 => 64x8x8, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 64x64x4x4
 => 64x16x16, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 64x64x4x4
 => 64x32x32, BatchNorm2d, LeakyReLU(0.1)
ConvTranspose2d - 64x48x4x4
 => 48x64x64, BatchNorm2d, LeakyReLU(0.1)
Conv2d - 32x48x3x3
 => 32x64x64, BatchNorm2d, LeakyReLU(0.1)
Conv2d - 3x48x3x3
 => 3x64x64, Tanh*0.5+0.5