This version, v3, was retrained from scratch and the dataset altered, keeping the masking technique. The hope was that changing a few training parameters would avoid the masking bleed through, and it seemed to have worked.
Weight of 0.6 can work if you've not added any character LoRA, 0.8 can hit the mark if you have, I guess it would also depend on that particular LoRA and certainly its weight. Though giving more details to the object's effect on the lips and mouth can obviate the need for higher weights. This is just opinion from my preliminary testing.
Triggers that usually work are lip retractor, dental retractor, cheek retractor and just retractor, the latter sometimes giving a very different appearance to the object.
The object can be separated at its connected area, creating two mirrored items if you like, and can also be colored separately.
