In this video, I explain how the MAR approach and how we can generate new images by using autoregressive models but modeling it with diffusion loss in continuous space instead of commonly used cross entropy.
Paper link: https://arxiv.org/abs/2406.11838
Dr. Song's explanation of Diffusion and Score-based Generative Models: • Diffusion and Score-Based Generative Models
Table of Content:
00:00 Intro
00:23 Autoregressive models in NLP
04:19 Switching into images
07:39 Autoregression in continuous space
15:00 Unifying Autoregressive and Masked Generative Models
20:49 Final figure
21:35 Outro