The problem of rendering bad hands has always been a challenge for AI drawing.
This LORA provides a solution approach, primarily adjusting the hands using different weights while maximizing the preservation of the overall picture structure, thereby increasing the likelihood of generating good hands.
LORA demonstrates good generalization, being compatible with realistic, 2.5D, and anime base models.
While LORA optimizes hands to a certain extent, it cannot guarantee generating good hands every time.
If you can use graphic design software such as Photoshop (PS), these approximate images can provide more reference materials for hands, enhancing the overall image.
Although it can also be used with the XL model, the performance is not satisfactory, and it can only generate minor changes.
I hope you like to use it.
The biggest support for me is posting more pictures, :)
Description
FAQ
Comments (5)
Based on the preview images alone, the lora seems to work best (i.e. most consistently) with 0.3; it doesn't perfectly fix hands, but does a decent job of it across all prompts.
But that's good enough; it makes hands recognizable for the new depth hand refiner preprocessor for Controlnet Depth to actually work its magic. See, a source of frustration in using this feature is that sometimes, hands are just too mangled; this lora might just be able to solve that problem.
I'll do tests when I can and post results here.
OK
Any progress on those tests?
Edit: Thank you for fixing this. All uploaders should follow your example and filter unrelated garbage.
The images are not showing up in my browser.
Thanks for the tip! The filter has been set.