(for bounty)
V0.7
This is a flexibility sidegrade focusing on making the lora play nice with other loras and style heavy base models. The data set was further improved a bit, but the model was made smaller and weaker for compatibility so do not expect much improvement in consistency
Main trigger is still “selfnosepinch,” but...
Since now it doesn’t fight base model as much, might want to try things like “nose pinch, hand on nose,” etc, sometimes it helps, sometimes it might lead to more deformities though, and "nose pinch" will often add another person or a floating hand doing the pinching if you do not also use the "solo," tag.
Also things like “finger in mouth” might help from my limited testing.
v0.5:
Cleaned up and augmented dataset with improved tagging
Half the size and about twice as effective as v0.1
I will still continue to experiment and possibly post updates if results look promising, though I might have to break it up into several loras for different hand positions for better results in the future...
v0.1:
WARNING: This is my first attempt at a concept lora! It does work, but is quite unstable so may take quite a few generations before you get something useable.
Also I accidentally forgot to decrease network size, so this one is probably larger than it needs to be. I will make future versions smaller if it does not make it significantly less effective in testing.
I will be returning to this later to improve, but want to get a few other updates and character/style loras out first, so decided to post it as-is if anyone wants to play around with it for fun, memes, and science.
Trigger word is:
selfnosepinch,
can also add "nose pinch, hand on own nose, " to help it, but sometimes adding that also increases the chance of getting deformed hands/nose/mouth etc (whether it helps or hurts seems to vary from merge to merge as well), so try without them too.