Announcement_5

New preprint out: Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Training! We propose a simple modification to the forward masking process of Masked Diffusion Models that speeds up training by up to 2.5x.