！日本語の説明は下部にあります！
This model detects NSFW areas (segments) in anime images. One model supports multiple types of detection.
Detectable Regions: nipples, pussy, anus, penis, testicles, x-ray, cross-section
Recommended Threshold: 0.5 ~ 0.8
Base Model: yolov11s-seg.pt
🗂️Install Location
Stable Diffusion WebUI/reForge -> models/adetailer
ComfyUI -> models/ultralytics/segm
❓FAQ
How to Deal with the Issue of Detected Areas Being Replaced by Another Object
If the detailer prompt is empty, enter the target prompt.
For example, if a nipples is being replaced with pussy, enter "nipples" in the prompt.
Lower the denoise value.
Expand the inpainting region.
WebUI (ADetailer): Increase "Mask erosion (-) / dilation (+)"
If "Inpaint only masked" is enabled, increase "Inpaint only masked padding, pixels"
ComfyUI (SEGS): Increase "crop_factor"
Troubleshooting “UltralyticsDetectorProvider Weights only load failed…” in ComfyUI
⚠️This workaround is required for v4 and earlier
Install the latest ComfyUI-Impact-Subpack and restart ComfyUI.
Open the file
<COMFYUI_INSTALL_PATH>\user\default\ComfyUI-Impact-Subpack\model-whitelist.txtin a text editor (e.g. Notepad).Add the following line to the end of the file, then save:
ntd11_anime_nsfw_segm_v4_all.pt
ntd11_anime_nsfw_segm_v4_anus.pt
ntd11_anime_nsfw_segm_v4_cross-section.pt
ntd11_anime_nsfw_segm_v4_nipples.pt
ntd11_anime_nsfw_segm_v4_penis.pt
ntd11_anime_nsfw_segm_v4_pussy.pt
ntd11_anime_nsfw_segm_v4_testicles.pt
ntd11_anime_nsfw_segm_v4_x-ray.ptRestart ComfyUI.
References
ComfyUI-Impact-Subpack README, Pull request
📄Note
ComfyUI allows you to filter detection class while using the "All" model.
Trained on images generated by Illustrious/noob derived models such as WAI-NSFW-illustrious-SDXL, Hassaku XL (Illustrious), Raehoshi illust XL, paruparu
Cannot be detected in monochrome images or comics
Cannot be detected in realistic images
⚙Training Settings (v5)
task="segment"
epochs=200
lr0=0.02
lrf=0.05
seed=0
imgsz=1024
batch=-1
close_mosaic=0
mosaic=0.0
erasing=0.0
scale=0.0For description of each setting, see: https://docs.ultralytics.com/modes/train/#train-settings
📢Past Announcements
Announcement for Stable Diffusion WebUI / reForge Users (2025/06/27)
I have an announcement for those using Stable Diffusion WebUI or reForge.
If you’re interested, please check out the details below—and I’d really appreciate any feedback or bug reports you may have!
When using my detection model with Stable Diffusion WebUI or reForge, the typical approach is to use it with the ADetailer extension.
While ADetailer offers many features, a few key functions were missing for making full use of my model.
To address that, I forked the original ADetailer and added the following features:
Class name filter: You can now filter detection results by specific class names
Class-specific prompts support: You can assign different prompts for each detected class using the
[CLASS=name]syntaxBounding box mask option: You can choose to use the bounding box as the mask for segmentation models, instead of the segmentation mask itself
You can find the repository here: https://github.com/newtextdoc1111/adetailer
Please see the README for installation instructions and more details on the new features.
My hope is that with class filtering, there will be no need to use anything other than the “ALL” model, and that the [CLASS=name] syntax will help reduce cases where unintended objects are drawn during inpainting.
Honestly, not having to train individual models for each release in the future would be a huge relief—this is one of the main reasons I implemented these changes 😂
If you’re someone who would still prefer to have individual models in future releases, feel free to share your specific use case. I may be able to offer an alternative solution (ComfyUI users are welcome too!).
アニメ画像のNSFW領域(セグメント)を検出するモデルです。一つのモデルで複数種類の検出に対応しています。
検出可能な領域： nipples, pussy, anus, penis, testicles, x-ray, cross-section
推奨閾値: 0.5 ~ 0.8
ベースモデル: yolov11s-seg.pt
🗂️インストール場所
Stable Diffusion WebUI/reForge -> models/adetailer
ComfyUI -> models/ultralytics/segm
❓FAQ
検出された領域が別のオブジェクトに描き換えられてしまう問題の対処法について
Detailer用プロンプトが空の場合、対象のプロンプトを入力する
nipplesがpussyに描き変わる場合、プロンプトに"nipples"と入力する
Denoseを下げる
インペイント領域を拡張する
WebUI (ADetailer): "Mask erosion (-) / dilation (+)" を増やす
ComfyUI (SEGS): "crop_factor" を増やす
ComfyUIで生成実行時に「UltralyticsDetectorProvider Weights only load failed. ...」と表示される問題の対処法について
⚠️v4以前のモデルで必要な対処法です
最新の ComfyUI-Impact-Subpack をインストールし、ComfyUIを再起動
<COMFYUIインストールパス>\user\default\ComfyUI-Impact-Subpack\model-whitelist.txtが作成されているので、メモ帳などで開く以下のテキストを追記して保存する
ntd11_anime_nsfw_segm_v4_all.pt
ntd11_anime_nsfw_segm_v4_anus.pt
ntd11_anime_nsfw_segm_v4_cross-section.pt
ntd11_anime_nsfw_segm_v4_nipples.pt
ntd11_anime_nsfw_segm_v4_penis.pt
ntd11_anime_nsfw_segm_v4_pussy.pt
ntd11_anime_nsfw_segm_v4_testicles.pt
ntd11_anime_nsfw_segm_v4_x-ray.ptComfyUIを再起動する
参考資料
📄メモ
ComfyUIの場合、"All" モデルを使いつつ検出したい種類をフィルタリング可能です
WAI-NSFW-illustrious-SDXL, Hassaku XL (Illustrious), Raehoshi illust XL, paruparu等のIllustrious/noob派生モデルで生成した画像で学習しています
モノクロ画像やコミックでは検出できません
リアル調の画像では検出できません
⚙学習設定 (v5)
task="segment"
epochs=200
lr0=0.02
lrf=0.05
seed=0
imgsz=1024
batch=-1
close_mosaic=0
mosaic=0.0
erasing=0.0
scale=0.0各設定の説明についてはこちら： https://docs.ultralytics.com/modes/train/#train-settings
📢過去のお知らせ
Stable Diffusion WebUI / reForge をお使いの方へのお知らせ（2025/06/27）
Stable Diffusion WebUI / reForge をお使いの方にお知らせがあります。
気になった方は以下の内容を確認し、感想や不具合の報告を頂けると嬉しいです！
Stable Diffusion WebUI / reForge （以下WebUIに統一）で私の検出モデルを利用する場合、ADetailerの拡張機能を使う方法が一般的です。
ADetailerは多くの機能がありますが、私の検出モデルを便利に活用するための機能がいくつか不足していました。
そこで、元のADetailerをフォークし以下の機能を追加しました。
クラスフィルター：検出したい対象をクラス名でフィルタリングする機能
クラス毎のプロンプト指定：
[CLASS=name]の構文で検出したクラス毎に個別のプロンプトを指定する機能バウンディングマスクオプション：セグメント検出モデルでもバウディングボックス形状でマスクを生成するためのオプション
リポジトリは https://github.com/newtextdoc1111/adetailer にあります。追加機能の詳細とインストール方法はREADMEを確認してください。
私の想定では、クラスフィルターがあれば「ALL」モデル以外不要になり、 [CLASS=name] 構文で適切なプロンプトを与えればインペイント後に別のオブジェクトが描画される事故も減るはずです。
特に、今後のリリースで複数のモデルファイルを学習しなくて済むと本当に助かります。このために実装したと言っても過言ではありません😂
今後のリリースでも個別モデルを残してもらいたい、と考えている方は具体的なユースケースを教えてもらえれば別の解決策を提案できるかもしれません。（ComfyUIユーザーの方からでも大丈夫です）。
Description
Increased the dataset for
x-rayandcross-section, resulting in improved detection accuracy. Individual models are also available for these cases.Expanded the variety of compositions and poses, leading to an overall improvement in detection accuracy.
Notes:
If you're using ComfyUI, make sure to add the model file name to
model-whitelist.txt. For more details, please see the FAQ section in the model description.
---
x-ray,cross-sectionのデータセットを増やし、検出率を向上させました。また、個別モデルも用意しました構図やポーズのバリエーションを増やし、全体的な検出率が向上しました
メモ：
ComfyUIをお使いの方は
model-whitelist.txtにモデルファイル名を追記する必要があります。詳しくはモデル説明欄のFAQを確認してください
FAQ
Comments (50)
when close-up, lower body only, it cannot be detected
Thank you for your feedback!
I'll check this case on my end. If the detection accuracy is low, I’ll work on improving it in a future update.
If you notice anything else, please feel free to leave a comment anytime.
Would you be interested in creating a model specifically for the chastity cage
Sorry, but I can't take on requests at the moment as I'm currently prioritizing improving the detection accuracy of this model.
However, if you're interested in creating a detection model yourself, I'd be happy to offer technical advice.
Sometimes it makes these weird puffyahh nipples that look ugly af, needs fixing.
If the appearance becomes strange after running the Detailer, it's often because another body part is being drawn instead (e.g., nipples being replaced with something like genitals).
Please check the FAQ section under “How to Deal with the Issue of Detected Areas Being Replaced by Another Object” and try adjusting the suggested parameters—this might help resolve the issue.
If it still doesn’t improve after that, feel free to share an image with metadata or a screenshot of the Detailer settings. I might be able to give more accurate advice based on that.
@newtextdoc11996 Easy fix for that would be to actually have an order of detection, and then user can use [sep] tags in its prompt field so it knows that its inpainting. (If detection order is even possible?)
@KoujiAI I actually looked into the approach you suggested when I was preparing to release the very first version of this model. If it were possible to control ADetailer’s behavior in that way, I could’ve just released the “All” model alone, which would have made the release process much simpler.
However, since ADetailer doesn’t provide any way to filter or sort by class (i.e., the detected body part), I ultimately concluded that it wasn’t feasible. That’s why I decided to provide separate models that detect only one class each.
As an alternative, ADetailer allows you to add multiple tabs. By creating a tab for each detection target and assigning a different detection model and prompt to each one, you might be able to reduce the issue to some extent.
@newtextdoc11996 I understand what you mean, but the problem is, 4 of my adetailer slots are occupied, so i only got 1 to play with (which is your all in one). But generally im very happy with it, and its very accurate, if i prompt for 'sex' or 'after sex' it will give better results than just leaving it blank. (thats just a tip for others)
@KoujiAI I'm glad to hear you're satisfied with the detection accuracy!
Regarding the number of ADetailer tabs, you can actually increase it up to 15. You can find the setting under Settings → [Uncategorized] ADetailer → Max tabs.
If you increase the number of tabs based on how many individual detection models you want to use, and assign appropriate settings to each one, it might help you get the results you're aiming for.
You probably already know this, but sharing just in case it helps someone else too — when you increase the number of tabs, it can be a bit tedious to reconfigure ADetailer every time you launch the WebUI.
After setting everything up, you can go to Settings → [Others] Defaults and click "Apply" to save your setup as the default. That way, the same configuration will be loaded automatically next time.
@newtextdoc11996 Thats actually genious, thank you :)
I think you should also make this model for more realistic i mean 2.5D or CG style images. This is a great project btw.
Thank you for your kind words!
Regarding the "more realistic" style — do you mean something like the following models?
- https://civitai.com/models/1110783/ilustmix
- https://civitai.com/models/1025051/illustrij
I tested a few showcase images from these models, and the detection accuracy wasn’t bad.
If you had a different style in mind, please feel free to share the model or some example images.
@newtextdoc11996 The Style i am talking about is anime but 2.5D not fully CG like you mentioned in your comment .
for examples i couldn't find the exact model for these types of image but for i found great SFW examples here : Please Check out this post https://civitai.com/posts/17699440
Also for NSFW examples you can check Nyako on Pixiv.
I am still searching for better models that can make these type of generation if you find a good model , you can ,mention it here. I checked your model on these types of images its not detecting a thing but its working great for 2D anime images.
Also if you need prompts for what style i am talking about i guess it will be 2.5D or 3D , shiny_skin, smooth shading.
Also i noticed these types of style uses a mixture of :
https://civitai.com/models/943607/748cm-or-style-for-illustriousnoobai-075?modelVersionId=1056404
https://civitai.com/models/176554/usnr-style?modelVersionId=1552087
https://civitai.com/models/856285/pony-peoples-works-v1-v6?modelVersionId=1036362
https://civitai.com/models/915918/moriimee-gothic-niji-or-lora-style?modelVersionId=1244133
these were the infos i could provide you at this moment . I will let you know if i find anything else in the future
@TsukikageStudio Thank you for the detailed information. I believe I now have a good general understanding of the style you're referring to.
As you mentioned, I haven’t seen any checkpoint models that can fully reproduce a style like "Nyako." The "coco-style-Illu-XL" model does indeed seem quite similar.
As for the next update of my model, I’m focusing on improving detection accuracy for various actions using images generated with models like WAI and Hassaku. The priority is more on action diversity rather than style diversity.
So, I’m sorry, but I can’t promise improved detection accuracy for the styles you shared at this time.
That said, if models in this style become more popular in the future, I might consider adding support for them.
@newtextdoc11996 No problem and thanks, from a comment where you you mentioned how can someone make their own model like this, and a little bit help from internet i managed to train my own model
I was wondering if you could release bbox versions as well, thanks
I'm sorry, but I don't currently plan to release bbox versions of the model, for the following two reasons:
1. Training time
Each model is currently trained from scratch individually. Releasing both segment and bbox versions would effectively double the training time, bringing the total to about a full day. So rather than spending extra time on training, I'd prefer to explore alternative solutions.
2. BBox can be obtained from segment models
In principle, you can still extract bounding boxes from segment models. This works in ComfyUI (Impact-Pack), although unfortunately ADetailer doesn't offer that kind of option.
Lastly, if you don't mind, I'd like to ask you a quick question to help guide future releases. Feel free to ignore this if it's not something you're interested in.
I checked ADetailer’s source code, and it looks like it wouldn’t be too difficult to add both class-based filtering and bbox detection from segment models. Would you be interested in a version of ADetailer that includes those features?
@newtextdoc11996 I cant tell you wether or not im interested in it, since i dont know what those are, or how do they compare to bboxes. Are they the same thing in the end?
@KoujiAI Thanks for your reply!
Regarding your question — the results should be almost the same as when using a bbox detection model.
To give an example of how it would work if this feature were implemented: imagine there’s a checkbox in the ADetailer UI labeled “Detect as bbox.” If you check that box, the detected area would be treated as a rectangular region instead of a pixel-level mask, and the entire region would be inpainted accordingly.
@newtextdoc11996 Yes illd love that in that case :)
@KoujiAI Thank you!
Once I’m able to implement it, I’ll make an announcement for everyone. If you happen to have some time then, I’d be happy if you could take a look at how it works.
@newtextdoc11996 Sure, im always generating big batches of images, i can let you know.
@newtextdoc11996 : Thanks for providing these detectors - I am sure there was a lot of hard work involved in creating them! Do you see a possibility of updating them to a newer version / internal format that does not reqiure the workaround you mentioned for ComfyUI? Any reply would be deeply appreciated! :-D
Thank you for your kind words!
I previously looked into whether it would be possible to run this without the workaround in ComfyUI. As a result, I concluded that it would be difficult to address this on the model side for the following two reasons:
1. There is no option during training to prevent Python code from being embedded into the model.
2. According to an issue on the training tool (Ultralytics), the problem still hasn't been resolved on their side.
Issue URL: https://github.com/ultralytics/ultralytics/issues/19824
I also tried converting the trained model (.pt) into the more secure ONNX format to see if it would work, but it failed during detection due to runtime errors.
I’ve also considered updating the base model from YOLOv8 to YOLOv11 or YOLOv12, but this approach seems unlikely to resolve the issue, so I haven’t tried it yet.
In conclusion, at this point, I haven’t found a way to make it work without a workaround.
If you have any questions or suggestions regarding the above, please feel free to reach out.
Hello everyone!
I have an announcement for those using Stable Diffusion WebUI or reForge.
Details are now available in the Model Detail section, under the heading: "Announcement for Stable Diffusion WebUI / reForge Users (2025/06/27)"
Please take a look—and I’d be happy to hear your feedback or bug reports in the comments or via chat!
Lastly, thank you all for your kind comments and reviews so far. Updates may not be frequent, but I’ll continue improving the detection accuracy over time.
---
皆さんこんにちは！
Stable Diffusion WebUI / reForge をお使いの方にお知らせがあります。
詳細はモデルの紹介ページに「Stable Diffusion WebUI / reForge をお使いの方へのお知らせ（2025/06/27）」を追加したのでそちらをご確認ください。
内容を確認し、コメント欄やチャットで感想・不具合報告を頂けると嬉しいです。
最後に、これまでの皆さんからの温かいコメント、レビューありがとうございます。不定期になりますが、これからも検出精度の向上を目指して更新を継続していきます。
github link does not work. But my first question before i try is, how does it handle LORAs? can you only apply LORAs to each class or does it only work with prompts.
@Wizzi Oops, sorry about that—the GitHub link got messed up when I copied and pasted it. I've fixed it now.
> how does it handle LORAs?
You can apply a different LoRA to each class individually.
@newtextdoc11996 sry couldn't make it work. Had to remove it to be able to gen again. I will probably try it on my reforge instead of a111 when i find time for it.
@Wizzi Thanks for the feedback!
If there were any error messages in the console, it would be very helpful if you could share them.
I hadn't tested it on a1111 on my end, so I'll try it out later.
was testing this out compared to the original adetailer extentsion and it seems this one makes the deatiled face a noticably darker than the original , tested with same parameters, settings, etc
@Ru5144 this is not a face detailer tho did you type it in a wrong thread?
@Wizzi i meant the new github repository this person did for adetailer that adds the class name filter, it essentially works like the original adetailer extension but when used in comparasion, faces for example come out darker than it would when compared with original adetailer fork
@Ru5144 ah i see. that usually happens when you have low mask blue. maybe default setting is low or something.
Really great detection
Thank you for the kind words!
I've been able to keep updating the model thanks to all the feedback from everyone.
Works nice depending on the pose/angle, but it has a problem with downward shots like this https://civitai.com/models/1487610/penis-pointing-toward-viewer-69-pov?modelVersionId=1682721
It also cannot detect nipples on flat chests/men though im not sure if thats because im using it for furry only.
Thank you for the feedback!
Regarding the composition shown in the LoRA image you shared, as well as flat chests females, those types of angles weren't included in the dataset—I'll plan to add them in a future update.
As for detecting male nipples, I'm currently not planning to support that for two main reasons:
- It would require checking through the existing dataset to identify and annotate images that include male nipples.
- I'm not sure how much demand there is for this use case.
That said, I'd love to hear more about how you're using the model in relation to this—if you're open to sharing, it would be helpful for future updates.
@newtextdoc11996
You're welcome. For the male nipples, I was including them in with flat chests since they're both rather similar looking as long as the male isn't muscular or has pecs. Skinny and effeminate males were what I was planning on using it for.
@DestructiveForce
Thank you for your response.
Given the use case you described, it's possible that detection will work in a future version, as I plan to include images with flat chests in the dataset.
I’ll mention it in the version information when the update is released.
For a downward shooting angle like this, is it possible to rotate the image first, adjust it to an upward direction, and then modify it with AD？
Please add pubic hair for both male/female too
Thank you for your feedback.
Sorry, but I can't take on your request at the moment as I'm prioritizing improving the detection accuracy of this model.
@newtextdoc11996 I'll patiently wait until you consider release them ! please just keep them in mind later then. I wanna make the pubic hair not being real-style effected
This works fine for most cases, but it seems to cannot detect anal object insertion, is there a plan to support this?
Thank you for the feedback.
That case was included in the dataset, but since there were only a few images, the detection rate was relatively low. I plan to improve this in a future version.
@newtextdoc11996 Awesome! Excited for what's next.
newtextdoc11996 you could add it to the website https://www.seaart.ai/ pls
Sorry for bothering you, but I have a problem: I used your anus detector for long and it worked perfectly, but lately it almost always inpaints a pussy over an anus (even if the box for the inpaint says 'anus'). At the beginning, I thought that the problem was in the specific prompt I used for the anus tag in Adetailer, but I checked it and there is no other keywords other than 'anus' and the usual quality tags (like anime screencap, masterpiece, etc.). Do you have any idea why this is happening?
Is it possible that the ADetailer parameters or the version of the detection model are different now compared to when it was working well?
Also, have you tried the method described in the FAQ under “How to Deal with the Issue of Detected Areas Being Replaced by Another Object,” and did it fail to improve the issue?
If the cause cannot be identified, sharing the specific steps you have already tried might help me provide more accurate advice.