This is a face detection model for girls created specifically for ADetailer.
It was trained using around 500 anime-style images.
If you're looking to detect realistic female faces, please use Anzhc's model instead:
https://civarchive.com/models/293448/anzhcs-face-segmentation-prototype-or-yolov8-or-adetailer-model
I developed this model to fill a niche, focusing solely on detecting anime girls' faces.
Since it uses a segmentation-based approach, it has the advantage of avoiding nearby people or hats being mistakenly included, which is often an issue with bounding box methods.
It also excludes male faces, ensuring you can avoid the all-too-common tragedy of accidentally "repainting" a male character’s head with a female face.
Tagging the data was like the most indulgent kind of Zen—simultaneously painstaking and oddly meditative.
I'm not too keen on the idea, but if there's enough demand, I might consider making a face detection model for anime-style male characters as well.
Q&A
Q. A cute boy is mistakenly detected as a girl.
A. Increase the first option under ADetailer’s [Detection] section, [Detection model confidence threshold], to around 0.7.
This will make the model more selective when detecting girls.
Be cautious, as setting it too high (e.g., above 0.85) might result in actual girls not being detected.
Q. Eyebrows are not included in the correction area.
A. Due to the segmentation method, eyebrows can occasionally be treated as part of the face boundary.
Try increasing the [Mask erosion (-) / dilation (+)] setting under [Mask Preprocessing] to around 24.
This should help include the eyebrows in the correction area.
Happy New Year in advance!
Any help to purchase a PC for machine learning would be greatly appreciated.
https://www.patreon.com/c/user?u=44405057
アニメガール用の顔検知モデル
これは ADetailer で使用できるモデルです。
アニメの画像約500枚でトレーニングされました。
リアルな女性の顔検知を行う場合は、Anzhc氏のモデルを使用してください。
https://civarchive.com/models/293448/anzhcs-face-segmentation-prototype-or-yolov8-or-adetailer-model
私はこのニッチを埋めるために、アニメガールに特化した顔検知を制作しました。
セグメント方式で分割するため、隣にいる人や帽子を巻き込まないことが、bbox方式にはない利点です。
また男性を検知から除外するため、男性の頭を女性用のADetailerで上書きしてしまうといったよくある悲劇を防ぐことができます。
タグ付けの作業は地上で最も煩悩にまみれた禅のようでした。
あまり食指が動きませんが、もし需要があればアニメの男性の顔検知モデルも制作するかもしれません。
Q&A
Q.可愛らしい男の子が誤って女の子として検出されてしまう
A.ADetailerの[Detection]の最初の項目[Detection model confidence threshold]を0.7程度に上げてください。
より厳密に女の子を探します。
上げすぎると本当の女の子も見つからなくなるので0.85くらいが上限です。
Q.眉毛が修正範囲に含まれない
A.セグメントの都合上、まれに眉毛が顔の境界線になってしまうことがあります。
[Mask Preprocessing]の項目にある[Mask erosion (-) / dilation (+)]を24程度に上げると眉毛も含めることができます。
よいお年を！
Description
bbox
FAQ
Comments (27)
it's so good!
well done
Thanks! Works great. Do you mind also trying to make an anime hand adetailer?
Could you make feet segmentation for aDetailer please?
Anzhc's model seems to be gone, any chance you could make this for realistic as well?
is this can use in forge?
It is quite a good detector, would not bother to detect things like faceless male head anymore. However, sometimes it does not detect face even the threshold is as low as 0.3. Hope there can be an updated version!
This Changes the Artstyle too much, to a certain style. Was this trained on only 1 artist? Can you please do a new version.
this is only trained to detect not denoise/generate like lora and checkpoints.
The way this was trained was they gave it a bunch of images on anime girls, then they mark the faces for the yolo model to train off of, so when you use it all it does is detect the face, and then uses whatever checkpoint and loras(if you are using any) to inpaint/generate over the face (for this detection model.
This should not affect the style.
I would say that it has not been trained with all the anime girls that exist, because I tried it on Asuna Yuuki and it changed my style completely, if it were me I would use this tool only for anime girls in general
@Aniky This can't change the artstyle in any way. Its only function is to detect and mask anime girl's faces. Your ckpts, loras, controlnets, etc are the ones changing the style.
@HoloLeon actually you can do it, just write in the positive prompt of ADetailer the name of the girl, and the facial expression, if you leave the prompt empty, it will only change the natural style and expression
It's a detection model. It literally does not "change" anything. It doesn't generate anything.
@QuattroBajina2 As other people already confirmed, It changes the Artstyle too much.
@iLikeCookies Let's review, shall we?: It's a detection model. It literally does not "change" anything. It doesn't generate anything.
People know how to generate images, etc. but don't actually understand how it works.
these models are object‐detection model that tells you where objects are by drawing boxes or masks. It does not change any pixels. Whatever you use to generate is just inpainting the location the detection model selects. That's it.
If you are getting messed up images etc. your settings are bad. lower the denoise etc. again, this is just masking the location, whatever you use to generate is inpainting the location.
こんな神機能を見過ごしていたなんて…
The best face detector so far, but sadly it can't detect upside-down faces
How do I use it?I have added it to the AD folder.
ComfyUI\models\ultralytics\segm
Not work with furry girls
the bbox version is better for me because the face mask doesn't always cover the whole face
It struggles to detect certain faces in profile. Specifically ones with narrow eyes and mature face shape.
When using segment detect, it does detected the face, but can't re-draw, why is that?
検出されるが再描画されないゴミ