Lucas Freire Sangoi

I know it's been months since this issue was opened, but yesterday I had the same problem and ended up finding the solution. It turns out there is no global...

@DarkAlchy, sorry, the solution wasn't entirely developed by me, I used an idea that was sent to bmaltais's repository, but with some adjustments because of the many updates that Kohya's...

> Thank you for opening this PR. I'm not sure `text_encoder.requires_grad_(False)` stops the training for Text Encoder. It may be needed to set `requires_grad_` for `network`. I will test it...

I've developed a much simpler modification to halt the text encoder. By leveraging the principle of underflow, I've implemented a mechanism where, upon reaching the designated stopping step for the...

As far as I know, which isn't much, I think that max grad and the scale weight normal (if that's what you're referring to) don't modify the weights, they modify...

Btw, I think I used an old version of train network in this PR 😂 too much lines changed but the workaround is like 10~20 lines

Hey! I'd like to really thank you for this feature, it's absolutely incredible. However, I have a question that came up regarding the Text Encoder. I understood the principle that,...

I got it, I even realized some things in the trainings that indicate that an eventual adjustment in the LoRA weight wouldn't help with my 'problem'. But now I'm going...

> > I got it, I even realized some things in the trainings that indicate that an eventual adjustment in the LoRA weight wouldn't help with my 'problem'. > >...