868_1_rp.rar

: Standard AR models generate images in a fixed "raster" order (like reading a book), which limits their ability to understand the whole image at once. RAR introduces Randomized Autoregressive modeling , which randomly permutes the order of image tokens during training.

: The model starts with high randomness (permuted order) and gradually returns to the standard raster order as training progresses. 868_1_RP.rar

: It achieved a Frechet Inception Distance (FID) score of 1.48 on the ImageNet-256 benchmark, outperforming many leading diffusion-based and masked transformer models. : Standard AR models generate images in a