Tokenized image synthesis is a generative modeling paradigm in which images are created by representing visual data as a collection of discrete tokens drawn from a learned codebook and using generative models to predict these tokens before reconstructing them into pixels. In this framework, an encoder and a vector quantization mechanism compress continuous visual features into a grid or sequence of discrete categorical indices. A downstream generative architecture, such as an autoregressive transformer, a masked generative model, or a discrete diffusion model, is then trained to learn the probability distribution of these tokens and generate new valid sequences. Finally, a paired decoder maps the predicted token representations back into high-fidelity continuous images, adapting discrete sequence modeling techniques from natural language processing to visual generation tasks.
