GANs — Generative Adversarial Networks

GANs generate realistic images by pitting two neural networks against each other.

How GANs Work

Generator (G): Creates fake images from random noise
Discriminator (D): Distinguishes real from fake images

Training Loop:
1. G generates fake images
2. D classifies real vs fake
3. D improves at detection
4. G improves at fooling D
5. Repeat until G generates realistic images

Analogy: Counterfeiter (G) vs Detective (D)
Counterfeiter gets better → Detective gets better → Counterfeiter gets better...

Loss Functions

Discriminator loss:
L_D = -[log(D(real)) + log(1 - D(G(z)))]

Generator loss:
L_G = -log(D(G(z)))

Minimax game:
min_G max_D V(D,G) = E[log D(x)] + E[log(1 - D(G(z)))]

DCGAN (Deep Convolutional GAN)

Architecture guidelines:
├─ Replace pooling with strided convolutions (D)
├─ Use batch normalization
├─ Remove fully connected layers
├─ Use ReLU in G (except output: Tanh)
└─ Use LeakyReLU in D

class Generator(nn.Module):
    def __init__(self, latent_dim=100):
        super().__init__()
        self.gen = nn.Sequential(
            nn.ConvTranspose2d(latent_dim, 512, 4, 1, 0),
            nn.BatchNorm2d(512), nn.ReLU(),
            nn.ConvTranspose2d(512, 256, 4, 2, 1),
            nn.BatchNorm2d(256), nn.ReLU(),
            nn.ConvTranspose2d(256, 128, 4, 2, 1),
            nn.BatchNorm2d(128), nn.ReLU(),
            nn.ConvTranspose2d(128, 3, 4, 2, 1),
            nn.Tanh()
        )
    def forward(self, z):
        return self.gen(z.view(-1, 100, 1, 1))

GAN Variants

StyleGAN: Style transfer, high-res faces
Pix2Pix: Paired image translation (sketch → photo)
CycleGAN: Unpaired image translation (horse → zebra)
ProGAN: Progressive growing for high resolution

Key Takeaways

1. GANs consist of Generator vs Discriminator in adversarial training
2. Training is unstable — requires careful balancing
3. Mode collapse — G produces limited variety
4. DCGAN established stable convolutional architecture
5. StyleGAN produces photorealistic faces
6. Wasserstein loss (WGAN) improves training stability
7. GANs are being replaced by diffusion models for many tasks
8. GANs still useful for style transfer and image editing