🕳️ GAN Mode Collapse Visualizer

Cat & mouse: a weak Discriminator lets the Generator collapse into a single black hole.

📋 What Am I Looking At? (Read This First)
  1. 🟢 Green dots = REAL data — actual training examples (e.g. real photos). Fixed, never move.
  2. 🔵 Blue dots = the Generator's FAKE data — a neural net trying to invent examples good enough to pass as real.
  3. 🔴 Red sweeping radar = the Discriminator — a second neural net whose only job is to catch which dots are fake.
  4. Press ▶ Watch It Evolve — Generator and Discriminator train against each other, live. Drag either LR slider WHILE it's running — the simulation reacts on the very next step, no need to stop it.
  5. Mode collapse = the Generator gives up trying different fakes and just spams ONE fake it knows fools the Discriminator — that's the black hole you'll see form.
Step 0
Diversity (spread) = 2.20
Press ▶ Watch It Evolve to begin.
Cat & Mouse Topography
Minibatch Discrimination Off
On
🎭 FORGERPress Watch It Evolve — I'll keep dropping fake dots wherever the Discriminator seems weakest.
🧑‍🏫 PROF. TORCHMode collapse happens when the Generator finds ONE spot that reliably fools the Discriminator and stops exploring anywhere else.