New models, woman and man face detection: Basically sponsored by @girlsthatdontexist, I actually don't know if they want to be mentioned, they can comment to claim that they did if so. So far anonymous and very charitable member of community. I was really lazy about posting those, but they persisted that i should, so thank them for those 2 new models:
Man face detection - tries to detect only real male faces in image.
Woman face detection - tries to detect only real female faces in image.
It was trained only on realistic images, so is unusable for anime. Can't promise i'll expand it in future, but possibility is there.

They, and all models going forward will be trained on Ultralytics >= 8.3, please update package in your webui of choice to use it, since most likely you are using 8.1 or 8.2, which will not work.

If you are having issues:

1. Activate your venv

2. pip install dill

3. pip install --upgrade ultralytics

This should fix generic problems.

Small segmentation model aiming to create accurate masks of face for improved inpainting quality with adetailer extension.

V2: Tried to improve colored skin detection. Possibly improved detection from side.
Removed female/male separation, since i forgor which class was what, it's simpler that way.

This time around im releasing a small lineup of models, in order of quality you'll get in a1111: v2 640, v2 768 MS, v2 768, v2 1024

Use v2 640, if you need it for adetailer. Rest are situational and for out of automatic purposes, since adetailer extension doesn't support other than 640 resolution right now, it will use it with any of model listed above.

Can also be downloaded at HF: https://huggingface.co/Anzhc/Anzhcs_YOLOs

Script im using: https://github.com/Anzhc/Training-script-for-Ultralytics-YOLO

Usage

Put into `automatic1111/models/adetailer`, or any other place that you know will ork with YOLOv8 models.

Use.

What it does?

Instead of just finding face, it tries to create a mask covering face, which narrows down area of inpainting, which leads to better results, due to attention of model not spreading to anything else.

Yolo face mask:

My face mask:

Due to it being very small(in dataset), and experimental in nature, it's not generalized enough, and will have issues, so be advised.

P.S. In case of tolo face, WHOLE box is inpainted, while my model creates mask to inpaint.

In some cases, particularly close-ups and face shots, standard yolo face model will create hard burn

Because area of inpaint is almost whole image. In case of my model it tries to localize area, which doesn't create such burn.

But yes, here you can see that it's not quite adept with close-ups, and chops it quite a bit.

This is really easy to improve, but im busy with work and other stuff to really do that.