adetailer is a collection of YOLOv8 and YOLOv9 object detection models optimized for detecting detailed features such as faces, hands, clothing, and full-body silhouettes. Developed by Bingsu using the Ultralytics YOLO framework, the models are trained on a variety of datasets including WIDER Face, DeepFashion2, and anime segmentation sets. It supports a wide range of detection targets from realistic human faces and hands to anime characters and fashion garments. The models come in various sizes (nano, small, medium) offering different trade-offs between speed and accuracy, with the best face model (face_yolov9c.pt) achieving mAP50 of 0.748. Pretrained weights are available and can be loaded directly using the Hugging Face Hub and Ultralytics’ YOLO() interface. Despite security warnings related to pickle safety (due to getattr usage), the models are safe to use if sourced from trusted repositories.
Features
- Multiple YOLOv8 and YOLOv9 model variants for different precision/speed needs
- Specialized detectors for face, hand, person segmentation, and clothing
- Trained on WIDER Face, DeepFashion2, and anime-specific datasets
- Compatible with Ultralytics YOLO interface and Hugging Face Hub
- Provides both bounding box and segmentation mask outputs
- Face model (YOLOv9c) reaches 0.748 mAP50 for realistic face detection
- Open-source under Apache-2.0 license
- Clear documentation on usage and integration with cv2 and PIL