Tool 21 · Session 2 · Images · Latent compression

Latent Space Compressor

Image models never paint pixel-by-pixel — that would melt the processor. A Variational Autoencoder first crushes the picture into a tiny block of numbers (the latent space), then a decoder reconstitutes a full image from that shorthand. Shrink the latent grid and watch the picture dissolve.

Big idea: generative models store a compressed code, not pixels — then rebuild the rest.

64×64

Latent grid

12,288

Stored values

Encode

Decoder source

fig. 1 — latent code & decoded reconstruction

the bouillon cube

decoder output · faithful

fig. 2 — original frame, pre-encode