Rethinking Training Dynamics in Scale-wise Autoregressive Generation
IntermediateGengze Zhou, Chongjian Ge et al.Dec 6arXiv
This paper fixes two big problems in image-making AI that builds pictures step by step: it often practices with perfect answers (teacher forcing) but must perform using its own imperfect guesses later, and the earliest coarse steps are much harder than the later fine steps.
#visual autoregressive modeling#next-scale prediction#exposure bias