search

LeapLabTHU / Agent-Attention

Official repository of Agent Attention (ECCV2024)
473 stars 35 forks source link

meaning and method #25

Closed john09282922 closed 6 months ago

john09282922 commented 7 months ago

Thanks for the information. I have additional question. What is that meaning of remove and apply patch? also sx, sy, ratio? agentsd is your model?
agentsd.remove_patch(self.model) agentsd.apply_patch(model, sx=4, sy=4, ratio=0.4, agent_ratio=0.95)

actually, I would like to apply your agent attention module to ddim from guided diffusion model.

thanks, jungmin

john09282922 commented 7 months ago

Also, I have a question for sentence in your paper "we slightly increase the scale used for the second Softmax attention, i.e., agent broadcast." what is that scale value? can you make the link to your code? And Is it not possible to use DWC in diffusion model without training?

tian-qing001 commented 6 months ago

Hi @john09282922, our code for agentsd is developed on top of ToMeSD, where you can refer to the paper for explanations of various variables. The scale value is referenced as self.scale in this line and this line. It is not possible to use randomly initialized DWC without training.