IMPORTANT: NO TRIGGER WORD. For best results, do NOT mention "biting lip" in prompt (see below)
STRENGTH: 1.0 usually good, up or down as needed
Version 2 is much improved!
Reason for no trigger: klein interprets "biting" more like "feeding", which almost always involves a hand. Instead of forcing the lora to remove the hand like I did in V1, it made more sense to train without mention of biting or lips.
I tried doing side/middle in the same lora but they often got mixed or combined, separate loras is much cleaner and almost always gives expected results.
POTENTIAL ISSUES:
affects all mouths in image
sometimes teeth are janky, more of a klein problem
no control over which side of lip to bite, klein isn't great at left/right body parts anyway so not worth trying
Description
FAQ
Comments (3)
good work keep up!
May I ask how many steps of training can achieve such high-quality results? Thank you very much.
number of steps is tricky to estimate what will work best, it's probably not the best metric considering how much it depends on size/quality of dataset and what you're training, not to mention batch size and other training settings. this one was around 1000 steps though. best thing to do is just test checkpoints along the way and see for yourself what looks best. version 1 of this lora was only 240 steps and the checkpoints after that were terrible in comparison