Closed v0xie closed 1 month ago
Adds a new method of combining the guidance from PAG and CFG.
Derives from "High-fidelity Person-centric Subject-to-Image Synthesis": https://arxiv.org/abs/2311.10329
In the paper they are combining the guidance from two different models, so I thought we could apply that to PAG since it's doing pretty much the same thing.
A couple of examples:
High CFG scales:
Greater than 512px for SD1.5:
Normal CFG scale, high PAG scale:
Adds a new method of combining the guidance from PAG and CFG.
Derives from "High-fidelity Person-centric Subject-to-Image Synthesis": https://arxiv.org/abs/2311.10329
In the paper they are combining the guidance from two different models, so I thought we could apply that to PAG since it's doing pretty much the same thing.
A couple of examples:
High CFG scales:
Greater than 512px for SD1.5:
Normal CFG scale, high PAG scale: