The authors build on JiT-style transformers that generate directly in pixel space using large image patches.
Pixel diffusion benefits from adapting its prediction target by data direction
MFlow learns when a pixel-space diffusion transformer should predict clean data or velocity, while a small convolutional refiner corrects missing high-frequency detail.
Big Tech
Soumik Mukhopadhyay · Long Mai · Abhinav Shrivastava · Aniruddha Mahapatra
University of Maryland, College Park · Adobe Research
Research Digest··3 min read
Mukhopadhyay et al.
§