A LoRA specifically trained to counteract the annoying "sameface" phenomena exhibited by SDXL and Pony models, where every generation produces essentially the same girl cosplaying as someone else.
You know the type: Caucasian, freckles, long brunette hair, green-brown eyes, full lips... just 1girl.
This LoRA greatly reduces the likelihood of getting the generic 1girl sameface. Instead you'll get a different random girl, different eye and hair color, perhaps even a different ethnicity.
While this model certainly does not completely eliminate the problem, I do find that it introduces a bit of extra variation in the faces by pushing the model away from the baked in sameface.
Note: This LoRA is intended to be used at negative weights.
Description
FAQ
Comments (2)
Still experimenting. v3 has a tendency to reduce detail and bias towards a 2.5D style, whereas v1 was adding contrast and artifacts. Really hard to get this right without any side-effects.
I never thought randomizing faces was possible and here you are proving otherwise.