What is mode collapse in GANs?

Mode collapse is when the generator produces only a narrow variety of outputs instead of diverse results, a common GAN training problem.

How do GANs compare to diffusion models?

Both generate images, but diffusion models have largely overtaken GANs for high-quality image generation due to more stable training and better results.

Advanced🖼️ 20 slides⏱ 4 minDeep Learning

🎭 GANs — Generative Adversarial Networks

Generative Adversarial Networks (GANs) generate realistic images and data by pitting two networks against each other. This visual guide explains the generator and discriminator, adversarial training, common challenges, and GAN applications.

Slide 1 / 20

What Is a GAN?

Two networks compete to generate realistic data.

Slide 2 / 20

The Big Idea

A forger vs a detective, improving together.

Slide 3 / 20

The Generator

Creates fake data trying to look real.

Slide 4 / 20

The Discriminator

Judges whether data is real or fake.

Slide 5 / 20

Adversarial Training

They improve by competing each round.

Slide 6 / 20

The Min-Max Game

Generator fools; discriminator catches.

Slide 7 / 20

Starting from Noise

Generator turns random noise into images.

Slide 8 / 20

Reaching Balance

Training ends when fakes look real.

Slide 9 / 20

Mode Collapse

Generator produces too little variety.

Slide 10 / 20

Training Instability

GANs are notoriously tricky to train.

Slide 11 / 20

Image Generation

GANs create realistic faces and art.

Slide 12 / 20

Super-Resolution

Upscale low-res images sharply.

Slide 13 / 20

Image-to-Image

Turn sketches into photos.

Slide 14 / 20

Style Transfer

Apply one image’s style to another.

Slide 15 / 20

Data Augmentation

Generate extra training data.

Slide 16 / 20

Deepfakes

GANs can create realistic fake media.

Slide 17 / 20

Ethical Concerns

Misuse for misinformation is a risk.

Slide 18 / 20

GANs vs Diffusion

Diffusion models now lead image generation.

Slide 19 / 20

Famous GANs

StyleGAN and CycleGAN.

Slide 20 / 20

Getting Started

Train a simple GAN on a small image set.

Frequently Asked Questions

A Generative Adversarial Network (GAN) is a deep learning model where a generator creates fake data and a discriminator judges it, improving together to produce realistic output.