A LoRA I trained a few months ago on a large collection of high-resolution images of her face.

I found that it works far better to generate an image, and then inpaint the face with the LoRA, rather than using it from the beginning. When inpainting use a weight like 0.7 or weaker.

I got the best results, when inpainting in combination with ControlNet, and then repeating that a few times.

At the end send it to img2img with a very low denoising strength, to smooth out edges.