Click in latent space to decode shapes — interpolate between learned representations
Latent Space (z₁, z₂) — click to decode
Decoded Output
Interpolate A → B
VAE Architecture
Encoder: x → μ, σ
Reparameterize: z = μ + εσ
Decoder: z → x̂
Reconstruction—
KL Divergence—
Total ELBO—
z₁ = size/scale | z₂ = rotation
z = (0.00, 0.00)
The VAE learns a smooth latent manifold. The KL term forces z~N(0,I), enabling interpolation. Points near each other in latent space decode to similar shapes. The reparameterization trick allows gradients to flow through the sampling step.