Predicting visual tokens from a learned vocabulary to reconstruct visual content as part of the same framework as text generation.