Tiny Autoencoder

Demo of a Tiny Autoencoder that compresses and reconstructs images

01 32 × 32 image 3,072 RGB values
02 512-value latent code — 6× smaller
03 32 × 32 reconstruction model reconstruction

Upload an image.

02Encoderfinds patterns

waiting

03Latent space

512 values6× smaller

04Decoderrebuilds detail

waiting

05Reconstruction
The autoencoder's 32 by 32 reconstruction
✦output appears here

the model's best guess

The model is trained on a limited dataset (CIFAR-10), so it may generate imperfect reconstructions for out-of-distribution images.