Variational Autoencoder

Variational Autoencoders, introduced by Kingma and Welling in 2014, represent a crucial bridge between autoencoders and modern generative AI. While standard autoencoders simply learn to compress and reconstruct data, VAEs add a probabilistic framework that makes the latent space smooth, continuous, and suitable for generation. The key innovation is that the encoder does not map inputs to single points in latent space but to probability distributions (specifically, Gaussian distributions parameterized by a mean and variance). During training, the model is encouraged to keep these distributions close to a standard normal distribution through a KL divergence penalty, ensuring the latent space is well-structured with no holes or gaps. New data is generated by sampling a point from the latent space and passing it through the decoder. The smoothness of the latent space means that nearby points decode to similar outputs, enabling meaningful interpolation and manipulation. VAEs were among the first deep generative models capable of producing realistic images and have influenced the development of subsequent architectures. While VAEs tend to produce slightly blurry outputs compared to GANs and diffusion models, they offer advantages in training stability, latent space structure, and principled probabilistic framework. VAEs are important components in latent diffusion models like Stable Diffusion, where they provide the encoder-decoder that maps between pixel space and the latent space where diffusion occurs.