Autoencoders and Variational Autoencoders (VAEs)

An autoencoder squeezes data through a narrow bottleneck and tries to rebuild it. To succeed, it must learn the most important features. A Variational Autoencoder (VAE) makes that bottleneck a smooth, organised space, so you can pick a new point in it and generate something new. VAEs are also a key part of modern image generators: Stable Diffusion runs its diffusion inside a VAE’s compressed space.

🎬 Packing a suitcase