*READ CAREFULLY!
So, folks, since I wasn't doing well with other LoRAs for “perfect feet”, I created my own LoRA – with moderate success. It remains difficult to use. Although I have put countless hours of work into it ...
There are several problems:
True, you can get pretty good close-ups with this LoRA, for example if you want to show off a fancy nail polish. But normally you don't just want to see the feet or legs, but the whole model! And that means you're dealing with full body shots, which quickly lead to errors with Stable Diffusion SDXL, even in the standard resolution of 1024 x 1024.
If you break out of the standard resolution, e.g. if you take portrait or landscape shots in 1024 x YYYY or ZZZZ x 1024, Stable Diffusion is forced to create new areas from the image noise. This can go well – or you get three legs or even something more monstrous. It seems to be incredibly difficult for the AI to understand where both hands are, where both legs are, where the feet are.
So what to do?
A) Don't use the LoRA at all. Yes, you read that right. Sometimes the results just don't get any better.
B) Do not start with a LoRA weight of 1, but start lower and then increase the strength.
C) Test another training EPOCH than 6 (patreon only)
D) Raise or lower the CFG scale, use different samplers, different seeds, different sampling steps.
E) Only use the LoRA in a second or third HighRES pass.
F) Test the ADetailer.
G) Try ControlNet.
H) Use the "PAG (Perturbed-Attention Guidance)" add-on.
I) Use KOHYA HIRES.FIX, especially for images outside the standard resolution.
J) Experiment. Experiment. Experiment!
However, I release it into the wild. Looking forward to your results.
Luv <3 Frank
This is an add-on LoRA working best with my other SDXL LoRAs for Skin, Hands & Eyes etc. etc. – trained on female feet as close-ups or in full body shots.
Here I share "Epoch 6" for free. Have fun. <3 If you need more flexibility, please join my PATREON page, where you get all training steps and the final LoRA and how to use it: