Models & Architecture
Autoencoder
A neural network that learns to compress data into a low-dimensional representation and reconstruct it back to the original.
An autoencoder is a neural network with two parts: an encoder that compresses input into a compact latent representation, and a decoder that reconstructs the original from that latent. The bottleneck forces the model to learn meaningful features.
Autoencoders are used for dimensionality reduction, denoising, anomaly detection, and generating compact embeddings. Variants include Variational Autoencoders (VAEs) and Denoising Autoencoders.
Modern use: the VAE in Stable Diffusion compresses images into latent space, where the diffusion model actually operates — making generation tractable.
VAEs learn a probability distribution over latents, making them suitable for generation. They played a critical role in making modern diffusion models efficient enough to train and run.