LTH14 / rcg

PyTorch implementation of RCG https://arxiv.org/abs/2312.03701

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About crossattn in RDM

umnooob opened this issue · comments

Thanks for your impressive work!

I notice that in the config and comments, crossattn is used for context to condition input. However, simply addition op is used in ResBlock. Any performance differences between them?

And it would be nice if you could provide the original crossattn implementation.

Thanks for your interest! I directly modify LDM's code for RDM, so there might be some legacy issues (crossattn is one of them). To clarify, cross-attention is not used in RDM. We only use simple addition when incorporating conditioning.

Thanks for your reply!