To utilize the model, place the file in models>lora and then call it in your negative prompt but weight it positive.
Note: This doesn't work if you put it in your positive prompt with a negative weight.
This is an experimental model made to correct bad hands, mutated hands, fused fingers, etc.
Training Data: I used renders of bad/mangled/unrealistic hands in various poses from men, women, and groups of people in different aesthetic styles.
Hypothesis: Once I've trained a model to create 'unrealistic' or 'bad' hands specifically, I could use it in negative weights to filter out the bad and unnatural hands
To utilize it, download the file, place the file in models>lora and then call it in your negative prompt.
I recommend adjusting the weight a bit to see how it impacts your composition. Disclaimer, it's not perfect but it does seem to work well in my tests. Please let me know your feedback
I recommend adjusting the weight a bit to see how it impacts your composition. Disclaimer, it's not perfect but it does seem to work well in my tests.
Non cherry picked renders of consecutive seed images using the lora at various weights as requested by anonymous internet experts:
Description
FAQ
Comments (10)
Thanks for this! :) Will come in....handy
Why put it in the Lora folder if it is a lyco?
A1111 supports lyco natively since 1.5.0, you should put it with other loras and it will sort it out
@rkfg So is the Lycoris folder redundant then?
@Honest_Abe Yes, I moved everything there to Lora and removed it together with the extension. Try it!
@Honest_Abe yeah when you run your startup in a1111... it now gives you a warning saying lyco is no longer supported via the folder
how to use it? i've put the file in the models/lora but i can't find it inside the a1111 lora
💖 👍
Hate to break it to you, but I'm pretty sure this doesn't work. From the official documentation of Stable Diffusion in the features section under Lora: "... LoRA cannot be added to the negative prompt".
I tried setting the weight to 100000000000000000000000000000000000000000000000000000000000000000000000000000000000000 (which would usually just break any image) to see if it had any effect at all, but it didn't. I just got a normal image of a lady holding a cup of coffee.
In this guide it says the lora name is still processed as a keyword so that's probably why you seem to see an effect.
just prompt "lively-and-over-excited-having-fun" and this thing will never again work on any image....