Announcement
HiT learns its own image hierarchy from pixels, its creators say
The team says its image generator groups patches dynamically, without a separately trained encoder or region labels.
TLDR
HiT’s creators say the background can take a back seat: it collapses into a few groups while the main subject gets many. The team says the model learns this hierarchy directly from pixels, without masks or region labels. It also reports 2.4× less training compute than JiT-H/16 at matched quality.
Combined views
24.2K
9 Sources, first seen ago

Comments