v1.3以上はyolo11n-segで学習したため、使用している仮想環境内のultralyticsのバージョンを8.3.0以上にアップデートしてください。
pip install -u ultralyticsイラストキャラの下腹部(belly)を検出するためのモデルでしたが、アノテーション中に胴体(torso)も含めた大雑把な範囲を指定したので、多分腹回りを検出するようになったモデルです。一応、腸骨から剣状突起を目安に範囲指定してあります。(タグで書くとhip bonesからunderboob辺りか?)
画像を100枚程度で作成したために、v1.0は誤検出やそもそも検出しないことが多々有ります。
実写は学習していないし試してもいないので、検出するか分かりません。
既知の検出が弱い構図など
上からの構図 (upside-down)
白黒やグレースケールの画像(v1.3で改善)腹の上に手が被さるポーズ
腹の上に服を着用している画像(透けていると多少検出します)
English:by Claude3.5 sonnet
This is a model designed to detect the lower abdomen (belly) of illustrated characters. However, during the annotation process, I specified rough areas that included the torso, so it probably became a model that detects the general midsection area. For reference, the specified range is roughly from the iliac bone to the xiphoid process. (hip bones-underboob)
Since it was created using only about 100 images, version 1.0 frequently has false detections or sometimes fails to detect at all.
The model hasn't been trained on or tested with real photographs, so I'm unsure if it would detect them.
Known weak points in detection include:
- Bird's eye view angles (upside-down)
- Black and white or grayscale images (Improved in v1.3)
- Pose with hands over belly
- Images where clothes are worn over the stomach (though it can detect somewhat through semi-transparent clothing)
Description
Excluded labia around (pussy/groin) from the area.
Remove duplicate images from the dataset
FAQ
Comments (3)
@aa4666lo : You did some very impressive detectors. Unfortunately, I am not able to get them to work in ComfyUI due to security restrictions. This is known issue, see:
https://github.com/ltdrdata/ComfyUI-Impact-Pack/issues/931
Do you see any chance you can redo the detection models in the new, safe format? Any help would be much appreciated! Thanks a lot! :-D
Thank you for the detailed feedback about the ComfyUI compatibility issue.
After extensive investigation, I must honestly report that creating a "new, safe format" for segmentation models that works seamlessly with ComfyUI Impact Pack is not currently feasible with existing technology.
⚠This information is only based on personal testing and is likely to contain errors.
Technical limitations identified:
1. ComfyUI Impact Pack's ONNX segmentation support is legacy/deprecated
2. Converting to BBOX detection would require completely rebuilding the dataset
3. No standardized "safe format" exists for segmentation models that satisfies both security requirements and ComfyUI compatibility
Current workaround:
The only practical solution is adding the model filename to ComfyUI's model-whitelist.txt, which bypasses the security restriction but doesn't address the underlying security concern you mentioned.
Conclusion:
While the model works perfectly with A1111/ADetailer, creating a truly secure ComfyUI-compatible segmentation format would require fundamental changes to either ComfyUI's architecture or PyTorch's security model - both beyond the scope of model distribution.
I apologize that I cannot provide the secure ComfyUI solution you requested.
@aa4666lo Understood. Thanks anyway for that very extensive, yet comprehensive summary. I appreciate the work you put into this! Best of luck! :-)












