Home/ Tool index/ Latent Space Compressor
Launch ready Session 2 · Images Interactive

Latent Space Compressor

Image models don't paint pixel by pixel — a VAE crushes the picture into a tiny block of numbers, then decodes it back. Shrink the latent grid and watch detail dissolve; sample from noise and watch it hallucinate.

tools/latent-space-compressor/

Live preview · launch for the interactive version

§ A · What it makes visible

three hidden mechanisms
Fig. 01

Compression

Drag the latent size down and watch a clean image collapse into a blocky code.

Fig. 02

Decoding

The decoder rebuilds a full image from the shorthand — fidelity bounded by the code's size.

Fig. 03

Generative sampling

Fill the latent grid with random noise and the decoder hallucinates something new from nothing.

§ B · How to investigate it

run it like an experiment, not a toy
01 · Predict

Before you compress

Predict which preset survives compression best: gradients, hard edges, or fine detail.

guess: gradients survive
02 · Change one thing

Move the latent slider

Hold the preset; change only the compression ratio.

0% → 72%
03 · Compare evidence

Read what dissolves

Which features blur first? Which hold longest?

edges and detail go first
04 · Name it

Name the trade

Say what compression bought and what it cost.

compression = capacity vs. detail

§ C · Debrief questions

after the investigation
Why don't generative models store pixels?
Which kinds of images compress cleanly, and why?
What is the decoder doing when it samples from noise?
Where does 'latent space' stop being a metaphor?

§ D · Related

pairs well with · use in context