Process - Use cut&paste training images to convince PonyXL to grab arm/shoulder instead of breasts. Then generate outputs and thicken the dataset for v2. First generation was trained using Blair Dame PonyXL character.
v2 - tends to add armbands to characters who wouldn't normally have them. So negative prompt "armbands" avoids adding armbands or noise around the point of contact.
v2 - tends to "steal" gloves from character LORA. You could avoid this with negative prompt. Or you can even add a LORA and try to get the gloves that you want on the POV hands. (v1 was super basic and tended to use hands that looked like the training data hands, and v2 seems to more fully recognize that these things really are hands)
Description
Tried v1 with other LORAs and generated some ok results. Re-using and seeing how that goes.
"wristgrab" and "resting on hands" are very rare. If they come up more, I might try to make them more robust.
FAQ
Comments (1)
Not sure if this is a conclusive finding, but I feel like v2 has a stronger sense of what a "POV hand" is, and sometimes connects to targets easier, but it also is prone to steal the armwear of the subject (and theor LORA) and put it on the viewer's arms. I feel like v1 was "dumb" and just used the same few hands without the dressing up step.
So yeah, v2 may be a weird doppleganger POV thing (always has been???). "Naked arms for everyone" is the simplest hack, but it would also be nice to have a way of varying the POV arm size.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.