☁️ Variational Autoencoder Visualizer

Watch a rigid dot become a probability cloud — then drag through it to generate new data.

📋 What Am I Looking At? (Read This First)
  1. The Latent Space Map is a 2D summary of everything the model has learned to represent. Standard AE = one fixed point per class. VAE = a fuzzy cloud (Gaussian probability distribution) per class.
  2. The Reparameterization Trick diagram shows why VAEs need a special trick: randomly Sampling a point is NOT differentiable — gradients can't flow backward through pure randomness.
  3. The KL Divergence (β) slider is a literal "rubber band" pulling every cloud's center toward the origin (0,0) and normalizing its size — that's what keeps the latent space smooth instead of scattered chaos.
  4. Drag the pink crosshair anywhere in the map — the Output Window morphs live between Cat/Dog/Car based on exactly where you are.
Mode: Standard AE
KL Weight (β) = 0.00
Latent pos = (−2.0, 2.0)
Pick a mode below, then explore the latent map.
Latent Space Map — Standard AE (dots) vs VAE (clouds) Output Window Reparameterization Trick
Gradient Flow Off
Show Gradient Flow (Reparam Trick)
☁️ CLOUDYSwitch to VAE mode — watch each rigid dot puff up into a fuzzy probability cloud.
🧑‍🏫 PROF. TORCHA Standard AE only ever sees the exact points it memorized — a VAE learns the whole neighborhood around them.