model for detecting nsfw areas in images for detailing, censoring or de-censoring.

objects: nipple, vagina, penis, anus, pubic hair, female face & male face

on huggingface: https://huggingface.co/01miku/anime-nsfw-segm-yolo26

(if you don't want zip files ↑)

versions(yolo26):

xl(1280) - detailing images and other high accuracy, low speed jobs (low speed ie it takes 30 milisecond on a gpu, recommended default)

medium(1280) - a balance between quality and speed

nano(640) - speed for tasks like real time censoring

resolution is 1280 & 640 (the yolo library handles downscaling or upscaling images automatically)

recommendations:

confidence threshold - keep at 0.25 - 0.8

If you want to reduce the amount of detailing steps, you can combine masks (of the same class) with a node after filtering based on the object.

usage:

I will get into how the segmentation regions look for all the classes.

nipple: segmentation is on the (each individual) nipple and not the areola. If you intend to select the areola you should expand the mask outwards. the dataset does include some inverted nipples but the accuracy will not be as high as normal nipples.

female face & male face: selects visible parts of the face in 1 segmentation mask per face.

non-nsfw models often don't detect female or male when the face is partially obscured, which is why I included it in this model. use female face for girls & futanari, use male for men and traps.

in hindsight, I would have preferred to use feminine face and masculine face classes as this is an easier concept to teach but the ship has already sailed for that. I'm not reannotating 3000 images.

vagina: selects visible parts or edges of the vagina(s). vaginal prolapse is likely to be misidentified as nipple or penis.

penis: selects visible parts or edges of the penis(s) in 1 segmentation mask, in multiple segmentation masks (if:internal cross-section, middle of the penis is obscured but the base and head are not).

anus: selects visible parts or edges of the anus(s) in 1 segmentation mask.

pubic hair: selects all visible parts of pubic hair in multiple segmentation masks.

I spent a lot of time annotating images, so I would greatly appreciate your support! be it liking the model or leaving a review. Thanks!

this was trained using 3300 sfw and nsfw images with various styles including: anime, 3d, realistic, comic, manga, furry, uncensored images and censored images & gay hentai so it would have better real world performance. although it does have some limitations: performance is not consistent between styles, with manga (black&white) or 3d realistic images but it will work if you're in a pinch. It's mainly for ai generated images.