This repository provides the official PyTorch implementation of the MAR model with DiffLoss for autoregressive image generation without vector quantization, as presented at NeurIPS 2024. The approach generates images token-by-token using a continuous diffusion loss instead of discrete quantization, combining benefits of both autoregressive and diffusion approaches. It includes pre-trained class-conditional models for ImageNet 256x256 and supports distributed training via PyTorch DDP.