Announcement
Level-of-Token Diffusion puts fine tokens where detail is needed
A post introducing the method says it puts coarse tokens elsewhere, rather than treating blank walls and faces alike.
TLDR
A post introducing Level-of-Token (LoT) Diffusion says diffusion models spend the same compute on a blank wall as on a face, even though it’s often possible to know where detail will be. The proposed method uses a multiresolution token layout, with fine tokens where detail is needed and coarse tokens elsewhere.
Combined views
14.5K
2 Sources, first seen ago
