🎭 GAN Generator Playground

Understanding GANs

Generative Adversarial Networks (GANs) are a revolutionary approach to generative modeling. Two neural networks compete: a generator creates fake data, while a discriminator tries to distinguish fake from real. This adversarial training produces remarkably realistic generated samples.

The Two-Player Game

• Generator (G):
  • Input: Random noise vector z (latent space)
  • Output: Fake image G(z)
  • Goal: Fool the discriminator
  • Learns mapping from noise to realistic images
• Discriminator (D):
  • Input: Image (real or fake)
  • Output: Probability that image is real
  • Goal: Correctly classify real vs fake
  • Acts as learned loss function for generator

Training Dynamics

GANs train through minimax game:

• Discriminator Training:
  • Maximize D(real) - encourage high probability for real images
  • Minimize D(fake) - encourage low probability for fake images
  • Standard binary classification
• Generator Training:
  • Maximize D(G(z)) - fool discriminator
  • No access to real data
  • Backprop through discriminator (frozen weights)
• Equilibrium:
  • Ideally: D(x) = 0.5 for all x
  • Generator produces perfect fakes
  • Discriminator can't tell difference

Training Challenges

• Mode Collapse:
  • Generator produces limited variety
  • Ignores parts of data distribution
  • Solutions: Minibatch discrimination, unrolled GANs
• Training Instability:
  • Generator and discriminator must stay balanced
  • If D too strong, G gets no useful gradient
  • If G too strong, D fails to provide signal
  • Solutions: Careful learning rates, architectural tricks
• Vanishing Gradients:
  • When D is perfect, G gradients vanish
  • Solution: Non-saturating loss, Wasserstein loss
• Evaluation:
  • No single metric for quality
  • Inception Score, FID popular
  • Often requires human evaluation

GAN Architectures

• DCGAN (Deep Convolutional GAN):
  • First stable architecture (2015)
  • All-convolutional (no fully connected)
  • Batch normalization
  • ReLU (G), LeakyReLU (D)
  • Foundation for many variants
• WGAN (Wasserstein GAN):
  • Uses Wasserstein distance
  • More stable training
  • Meaningful loss metric
  • Weight clipping or gradient penalty
• StyleGAN:
  • NVIDIA's state-of-the-art
  • Style-based generator
  • Incredibly realistic faces
  • Controllable generation
• Conditional GAN (cGAN):
  • Condition on class labels
  • Controlled generation
  • Both G and D see condition
• CycleGAN:
  • Unpaired image-to-image translation
  • Horses ↔ Zebras, Photos ↔ Paintings
  • Cycle consistency loss
• Pix2Pix:
  • Paired image-to-image translation
  • Sketches → Photos
  • U-Net generator

Applications

• Image Generation: Create realistic faces, art, designs
• Data Augmentation: Generate training examples
• Image-to-Image Translation: Day→Night, Summer→Winter
• Super-Resolution: Enhance low-res images
• Text-to-Image: DALL-E, Stable Diffusion
• Video Generation: Create video frames
• 3D Object Generation: Create 3D models
• Drug Discovery: Generate molecular structures
• Music Generation: Compose audio

Training Tips

• Use DCGAN architecture as starting point
• Batch normalization except in D's output and G's input
• LeakyReLU (α=0.2) in discriminator
• ReLU in generator (except output: tanh)
• Adam optimizer (β1=0.5 for stability)
• Different learning rates for G and D (often D lower)
• Train D more steps per G step initially
• Monitor generated samples visually
• Use label smoothing (0.9 instead of 1.0)
• Add noise to discriminator inputs

Evaluation Metrics

• Inception Score (IS):
  • Uses pre-trained Inception network
  • Measures quality and diversity
  • Higher is better
  • Can be gamed
• Fréchet Inception Distance (FID):
  • Compare distributions of real and generated
  • Lower is better
  • More reliable than IS
  • Current standard
• Precision & Recall:
  • Precision: Generated samples look real
  • Recall: Covers all modes of real data

GAN Variants

• Progressive GAN: Grow resolution during training
• BigGAN: Large-scale, high-fidelity generation
• Self-Attention GAN (SAGAN): Attention for long-range dependencies
• Spectral Normalization GAN: Stabilize discriminator

Latent Space Exploration

• Interpolation: Smooth transitions between samples
• Arithmetic: Smiling woman - neutral woman + neutral man = smiling man
• Disentanglement: Control specific attributes independently

Beyond Images

• MusicVAE/GANSynth: Music generation
• MolGAN: Molecular generation
• TimeGAN: Time series generation
• StackGAN: Text-to-image

Experiment with the Playground

Use the interactive tool above to:

• Generate samples from random noise
• Watch generator and discriminator compete
• Explore latent space with sliders
• See training dynamics over time
• Understand adversarial training

GANs opened entirely new possibilities in AI - creating rather than just recognizing. They're behind many of the most impressive AI demos you've seen!