GAN Playground

Train a generative adversarial network on 2D distributions. Watch generator and discriminator compete in real time with decision boundary visualization.

Target Distribution

Presets

Batch Size: 64

Noise Dim: 2

Generator LR: 0.010

Discriminator LR: 0.010

Speed: 5 steps/frame

Step: 0G Loss: 0.000D Loss: 0.000Coverage: 0%

Decision Boundary & Points

Loss Curves

Target Coverage0%

Blue in heatmap = discriminator thinks "real"

Red in heatmap = discriminator thinks "fake"

Try the "Mode Collapse" preset to see the generator fixate on a single mode. "D Dominance" shows what happens when the discriminator learns too fast.

Generator (2→16→16→2) and discriminator (2→16→16→1) implemented from scratch with backpropagation. No ML libraries — all training runs entirely in your browser.