Better is to use optimised multipack sampling and 3d attention.

→ View original post on X — @jeremyphoward