This model is nipples detection model for ADetailer / ComfyUI-Impact-Pack extension.
Segmentation: Detects the edges of nipples and areolae
nipples_v2_yolov11s-seg.pt
nipples_yolov8s-seg.pt
BBox: Detects nipples with bounding box.
nipples_yolov8s.pt
Notice:
If nipples_v2_yolov11s-seg.pt (Segmentation model) doesn't work, please re-install ultralytics 8.3.216
https://github.com/Haoming02/sd-webui-forge-classic/issues/375#issuecomment-3563484160
https://github.com/Bing-su/adetailer/issues/817#issuecomment-3462595521
nipples_yolov8s-seg.pt and nipples_yolov8s.pt are old YOLO model, so ComfyUI with PyTorch 2.6 and later will prevent these models from loading for security reasons.
(In the case of ADetailer, the problem does not occur because ADetailer loads the model with weights_only=False forcibly)
The workaround is as follows:
Use nipples_v2_yolov11s-seg.pt (recommended)
Add model name to <user_directory>/default/ComfyUI-Impact-Subpack/model-whitelist.txt (refer https://github.com/ltdrdata/ComfyUI-Impact-Subpack#model-loading-configuration-related-to-weights_only)
Usage (ADetailer):
Put this model into ADetailer model folder. (in the case of AUTOMATIC1111 and Forge, "models/adetailer")
Relaunch webui.
Enable Adetailer and select this model.
Adjust threshold, denoising and padding as needed.
Usage (ComfyUI):
Install ComfyUI-Impact-Pack and ComfyUI-Impact-Subpack
Put this model into "/models/ultralytics/segm" (Segmentation model) or "/models/ultralytics/bbox" (BBox model)
Relaunch ComfyUI.
Create workflow using UltralyticsDetectorProvider node and Detailer node. (or drop ComfyUI_sample_workflow.png into ComfyUI and modify embedded workflow)
Adjust threshold, denoise, crop_factor, feather and so on as needed.
ADetailer configuration sample:
The recommended values vary depending on the model, character orientation, and breast size.
Detection model confidence threshold: 0.5
Inpaint mask blur: 8-16
Inpaint denoising strength: 0.3-0.4
Inpaint only masked padding, pixels: 48-96
Notes:
BBox model was trained with anime images only, so use for realistic images is not supported.
Segmentation model was trained with anime and realistic images.
If you get "RuntimeWarning: invalid value encountered in cast x_sample = x_sample.astype(np.uint8)" error message and get black masked image (after that, you only get whole black images unless restart webui) using ADetailer, PyTorch 2.7 and cu128, try changing Ultralytics device to CPU. Reference.
Description
Segmentation model added.
FAQ
Comments (16)
Good job, in my experience it detects half of the images all nipples. So in my case it works 50%. I hope for a new update, it is very good.
@JuanpiAMVs Can you give me the checkpoint name and the prompts regarding the breasts you are using?
For whatever reason I cant get this to work on ComfyUI. The detailer runs but nothing is detected. I'm trying with the segs model with realistic pictures, but no luck. Any recommended workflow? Thanks!
@tiosolid I added a ComfyUI sample workflow using nipples_yolov8s-seg.pt to the model's zip archive. Please re-download it. I don't use ComfyUI much, so I don't know the optimal settings. However, I've confirmed that it can be detected.
@nuko_masshigura wow, thank you very much for your help. I will check it out!
意見ですが、
１、乳首が真下か真上に向いている場合、人の目でははっきり判断できる乳首は、スコアが極めて低い（0.3以下）です。特に(leaning forward, pov)などが含んでいる場合、この問題が頻発です。
２、ほかの乳首系のLoraを使用していなくても、(huge nipples)など、乳首を大きくするプロンプトを使用する場合、スコアは低くなります。
３、(white gloves)、(white pantyhose)または(white thighhighs)を使用する場合、高いスコア（0.5以上）で間違って肘、膝を乳首だと認識する不具合があります。場合により、へそ、口も間違って高いスコアで乳首に認識されます。問題３は問題１＆２と一緒に存在する場合、本当に厄介です。
ご検討のほどよろしくお願いいたします。
@hfsztt 貴重なご意見ありがとうございます。検出できないというコメントはよくいただくのですが、どのような条件のものが検出できないのかが書かれてないため、どう対応したらよいか困っていました。
上記の問題を改善できないか検討してみます。
Segmentation model only works on front-on or close to front-on nipple, failed to detect sideboob nipples and perky/pointy nipples view from side. Hope you add other views and perky/pointy nipple to the training data. Thanks.
The statement is incorrect, and you have not specified your threshold parameter value. By default, the detectors have a value of 0.5. Reduce it to 0.3/0.2. If the number of recognized zones is more than necessary, use the SEGS filter (ordered), as Dr. Lt. Data showed in his video.
@windlike_gust problem is it detects many many things as nipples if you reduce it to 0.2. at that range it will sometimes detect 10 things as nipples
@Wizzi In order to separate out unnecessary ones, there are three sets of filters (if we are talking about ComfyUI).
@windlike_gust i am aware, i am not using comfy though.
i am using lustify end game checkpoint
and comfyui somehow i dont know what the best settings are it detects the nipples but somehow the results are slightly improved
Ive been using this for weeks, i love it but my only problem is detection range for side angles is pretty low, around 0.5. And when you lover detection to that low, it starts detecting point noses and anus as nipples, and if the model has shiny skin it also can detect sunlight and sparkles on breasts as nipples as well.
So if there is ever gonna be a v2 i would wish to see either better side angle detection so can keep detection range higher or better prevention for anus/nose detection. And same with see-through.
SEGM wont run in ComfyUI because of dangerous code detection.
@whateverr Please use a new model or add the model name to model-whitelist.txt. (please refer Notice).