Generative Adversarial Network Simulator
Watch a generator and a discriminator play a real adversarial minimax game: the generator's Gaussian distribution is trained live to fool a small neural discriminator, with mode collapse, learning rates and the discriminator's decision field all visible in 3D.
A Generative Adversarial Network trains two opposing networks at once: a generator that tries to manufacture convincing fake data, and a discriminator that tries to catch it in the act. This simulator runs a genuine miniature version of that game in your browser — a Gaussian generator and a small trainable neural discriminator, both updated with real gradient steps every frame — so you can watch the adversarial feedback loop converge, oscillate, or collapse onto a single mode of the target distribution.
A live adversarial minimax game between a Gaussian generator and a small trainable neural discriminator: watch the two-point-cloud distributions converge, oscillate or fall into mode collapse as you tune both learning rates.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install