ATTENTION: IF YOU HAVE DOWNLOADED MODELS BEFORE ALREADY - RE-DOWNLOAD THEM PLEASE. I SHOULD HAVE UPDATED THEM WITH NO DILL VERSION(EXCEPT MODEL THAT WAS CHOSEN AS NEW VERSION UPDATE, SO YOU CAN RECEIVE NOTIFICATIONS.

VERY EARLY PROTOTYPE, BARELY WORKS AS INTENDED.

Small segmentation model aiming(and pretty much failing, kek) to create accurate masks of manga panels for improved inpainting quality with adetailer extension.

Particular behaviour to expect: Covrage of whole panel + extruding parts, like cross-panel objects and characters(which it will miserably fail to do btw). Extruding speech bubbles will be excluded from selection. Speech bubbles that are fully in panel will not be excluded.

Likely to perform quite a bit better at higher resolution, but examples are made at 640.

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 work with YOLOv8 models.

Use.

What is it for?

I use it as utility model, not necesserily for A1111. It can help you improve quality of comic pages, sequences, etc. if you like to do that sort of thing. It will try to find all manga panels with their specific borders.

Here is example from Jobless Reincarnation manga:

(Yes, it fails even fairly simple rectangles due to the way i decided to work on it, but it will get better with time(if i ever would update it, lmao(making those datasets are boring)))

Core idea is to take into consideration not only panels, but also objects related to them, if they go off-panel, while excluding objects that go off-panel, but related to other one.