For the encoder just use whatever works for your normal vision models I think.

→ View original post on X — @jeremyphoward