a Variational Autoencoders is a generative model. a VAE is a particular autoencoder :
- which aims to capture the latent space of our input (by optimizing the reconstruction from the latent space to the input).
- whose encodings distribution is regularized during the training (in order to ensure that its latent space follows a given distribution.)
- uses the variational inference method (approximation of the theoretical loss by the NELBO/ELBO)
As such - when trained - this process to randomly sample latent representation to generate some new data.