The really cool thing about LoRA is that because its an additive on top of base #stablediffusion, you can use it to interpolate in weightspace!
Here I'm interpolating between a base SD 2.1 output and LoRA adjusted SD (trained on a set of imgs of myself) in checkpoint weightspace
Conversation
You can now essentially take any SD output image and make it look like you!
I also see that LoRA finetuning of the text-encoder (which is what I added myself this weekend, is now also in the main repo):
github.com/cloneofsimo/lo
Replying to
wow! so cool,
can you share the # of pics, steps and LR you used? I tried a few times yesterday and got 'meh' results.. haven't tried it with the text-encoder yet
1
2
Replying to
20 pictures, 8000 steps, LR 1e-4, and I used Clip-interrogator to get a decent prompt for every source img, this greatly helps against overfitting
2
2
13
Show replies



