How do I finetune the attention layers of the unet?
#3
by
Meeeeq
- opened
There are optimized dreambooth repositories, but I don't know if I have access to this feature. Do you have any published code that demontrates this finetuning the attention layers? Thx
Also curious...
@Meeeeq check my implementation
https://github.com/kopyl/diffusers/blob/main/examples/text_to_image/train_text_to_image.py
Just add --train_attention_only argument when run train_text_to_image.py
Now i'm curious why did the author trained attention-only first then full unet instrad of training full unet from the beginning...