Generative Adversarial Network (GAN)

Deep learning architecture where two neural networks compete against each other to generate realistic synthetic data.

-Introduced by Ian Goodfellow in 2014.

A GAN is a neural network system where:

• one model generates fake data,
• another model detects fake data,
• and both improve through adversarial competition.

Popular GAN Variants

Variant	Purpose
DCGAN	CNN-based GAN
CycleGAN	Image translation
StyleGAN	High-quality face generation
Conditional GAN	Controlled generation
Wasserstein GAN (WGAN)	Stable training

GANs in the Generative AI Landscape

• GAN models are known for potentially unstable training and less diversity in generation due to their adversarial training nature.
• VAE relies on a surrogate loss.
• Flow models have to use specialized architectures to construct reversible transform.
• Diffusion models are more stable and produce higher quality outputs, but are slower to generate content.

GAN vs Diffusion Models

GAN	Diffusion Models
Faster generation	Slower generation
Harder training	More stable
Sharp outputs	Better diversity
More instability	Higher realism

Applications of GANs

1. Image Generation
   • Human faces
   • Art
   • Landscapes
   • Avatars
2. Deepfakes
   • AI Art
3. Super Resolution: Improve image resolution and quality.
4. Image-to-Image Translation
   • Sketch → Photo
   • Day → Night
   • Black & White → Color
5. Data Augmentation: Generate synthetic training datasets.

GAN Architecture

flowchart TD

    A[Random Noise z] --> B[Generator]

    B --> C[Generated Fake Image]

    C --> D[Discriminator]

    E[Real Image Dataset] --> D

    D --> F{Real or Fake?}

    F -->|Real| G[Correct Classification]

    F -->|Fake| H[Generator Improves]

Core Components

Component	Role
Generator	Creates fake/generated data
Discriminator	Detects whether data is real or fake

Generator Objective

Generator = Counterfeit Artist

Example: The artist improves fake currency generation.

The Generator tries to create realistic synthetic data.

$$G(z) \rightarrow x_{fake}$$

Where:

• $z$ = Random latent vector
• $x_{fake}$ = Generated fake sample

Discriminator Objective

Discriminator = Police Detective

Example: The detective improves fake detection.

The Discriminator classifies whether data is real or fake.

$$D(x) \in [0,1]$$

Where:

• $0$ = Fake
• $1$ = Real

GAN Minimax Objective Function

GAN training is a minimax optimization problem.

$$\min_G \max_D V(D,G)= \mathbb{E}_{x \sim p_{data}(x)}[\log D(x)]+ \mathbb{E}_{z \sim p_z(z)}[\log(1 - D(G(z)))]$$

Where:

• $G$ = Generator
• $D$ = Discriminator
• $x$ = Real data sample
• $z$ = Random noise vector
• $p_{data}$ = Real data distribution
• $p_z$ = Noise distribution

---

GAN Training Flow

sequenceDiagram

    participant Z as Random Noise
    participant G as Generator
    participant D as Discriminator

    Z->>G: Generate Fake Sample

    G->>D: Fake Image

    D->>D: Predict Fake

    Note over D: Train Discriminator

    D-->>G: Feedback / Gradient

    Note over G: Improve Generator

Step 1: Train Discriminator

• Feed real images
• Feed fake/generated images
• Learn classification

Step 2: Train Generator

• Generate fake images
• Try to fool discriminator

Step 3: Repeat Iteratively

Both networks improve over time.

flowchart LR

    G[Generator] -->|Creates Fake Data| D[Discriminator]

    D -->|Detects Fake Data| G

    G -->|Improves Realism| D

    D -->|Improves Detection| G

---

Challenges in GANs

1. Mode Collapse : Generator produces limited variety.
2. Training Instability: GANs are difficult to balance and optimize.
3. Vanishing Gradients : Discriminator becomes too strong.
4. Evaluation Difficulty: Generated quality is hard to measure objectively.