This is the SD1.5 version. Find the newer XL version here: https://civitai.com/models/148786
LoRA extraction of a dreambooth model I trained on 127 meticulously created images of this completely fictional woman. I don't know why this method worked the best, but it did. I created many facial expressions for the dataset, and the training turned out very well as a result. It's able to generate a wide range of emotion because of this.
IMPORTANT NOTE:
This model should be used with a denoising after process extension like After Detailer ( https://github.com/Bing-su/adetailer ) or similar for best quality. I believe this is because dataset was mostly trained on portraits. Full body images will have facial features be distorted. The first 2 images in the preview (you know the ones I mean) required it.
Background:
Sanaa is a character I made for a TTRPG a year and a half ago. It started with a single Artbreeder portrait and sorta snowballed from there. 3d model led to training which led to more training with the best images of successive generations. The dataset for this generation was 127 images. And in the interest of diversifying the current training available I'm uploading it.