TRANSFORMER BASED DIFFUSION MODEL IS 90% CHEAPER TO TRAIN THAN UNET BASED MODELS

Summary from the TLDR AI newsletter

PixArt is a new text to image model that uses T5 text encodings, cross attention, and a diffusion transformer to great results at a fraction of the compute cost of comparable models.