Watch a generator evolve from pure noise to convincing images while the discriminator fights back. Understand adversarial training, Nash equilibrium, and why GANs are both powerful and notoriously tricky to train.
Sign in to save progress
Adversarial Training Architecture
Epoch 0 / 30
Pure noise: G hasn't learned anything yet
Watch losses converge toward log(2) ≈ 0.693, the Nash equilibrium
GANs are notoriously difficult to train. Expand each problem to understand why GANs fail and how researchers work around them.
Mode Collapse
G generates only one type of output
Training Instability
D crushes G before G can learn
Memorization
G copies training data instead of generalizing
GANs were invented by Ian Goodfellow in 2014 at a bar after a debate with colleagues about generative models. He coded up the first working version that same night. Successors like StyleGAN pushed the adversarial idea to photorealistic faces. Today's image generators such as DALL-E and Stable Diffusion are a different lineage: DALL-E 1 was autoregressive, while DALL-E 2/3 and Stable Diffusion are diffusion models, which largely replaced GANs for image generation, though GAN-style adversarial losses still appear inside some of their components.