Comment by rafaelero

They are very likely using VQVAE to create a dictionary of tokens and then just converting images into them with an encoder.