diffusers
diffusers copied to clipboard
examples\controlnet add only_mid_control=true and sd_locked=false
Is your feature request related to a problem? Please describe. The original controlnet training tutorial mentions the options
-
only_mid_control=True
: This can be helpful when your computation power is limited and want to speed up the training, or when you want to facilitate the "global" context learning. -
sd_locked=False
: This will unlock some layers in SD and you will train them as a whole.
Describe the solution you'd like Please make the option available as well.
Describe alternatives you've considered Going back to the original training script without any optimizations implemented.
cc @williamberman @patil-suraj
Yeah these are neat ideas! I think these would be good options for community PRs, happy to open up to the community :)
Have any updates about this project ?
How to Implement SD_ Locked=False in Diffusers?
I would like to try this issue. Seemingly, It can be done by set different gradient recording switch and keep it compatible with accelerate. @williamberman @patrickvonplaten @patil-suraj
Seems pretty straightforward, just add an option to freeze/unfreeze certain weights.
Hi @ValMystletainn, are you still on it? :) Otherwise I'd tackle it.
Okay, I'd tackle this :) Eta 1 week.
This is not an option still in current controlnet train script