omriav/blended-diffusion

weights optimization

fido20160817 opened this issue · 1 comments

Hi, when optimizing the UNet, why "qkv", "norm" and "proj" are chosen to be optimized? Any reason behind this? why not the whole model or other elements of the model? Thanks.

omriav commented

Hi,

I did not optimize the UNet weights in this project, they are kept constant.
There was a need to enable the gradient calculation for them in order to perform the classidier-guidance.