Kingma & Welling (2014) introduced the variational autoencoder, which trains an encoder to map data into a continuous probabilistic latent space and a decoder to reconstruct it — enabling both compression and the generation of new samples. The key contribution was a practical way to train such models with gradient descent (the reparameterisation trick).
VAEs influenced later generative methods and underpin parts of modern systems such as latent diffusion image generators.