This repository provides a PyTorch implementation of the Autoregressive Image Generation architecture from the paper ‘without Vector Quantization’. It includes models for both sequential data and image generation treated as sequences of patches, along with training utilities. The implementation supports flow matching as an alternative approach.