Removed full face images from training data for LoRA, which were affecting other areas of the face (eyebrows in particular). This version is even more tricky to use, but when it works, adapts to a much wider range of faces/poses. As with v1.0, check out my examples.
You will still get the best results with face close-ups.
While Euler/Beta is a good start, try out different samplers and schedulers. See my example images. Also, with this version, strengths under 1.0 work better in some situations.
Removed full face images from training data for LoRA, which were affecting other areas of the face (eyebrows in particular). This version is even more tricky to use, but when it works, adapts to a much wider range of faces/poses. As with v1.0, check out my examples.
You will still get the best results with face close-ups.
While Euler/Beta is a good start, try out different samplers and schedulers. See my example images.
Also, with this version, strengths under 1.0 work better in some situations.