Anzhc's YOLOs
Anzhc/Anzhcs_YOLOs
published Feb 2024 · updated Feb 2026
Anzhc's YOLOs is a collection of object detection and instance segmentation models for face, eyes, head/hair, and breasts, trained on custom datasets for anime and real images.
specs
| Task | Object Detection & Instance Segmentation |
| Architecture | Ultralytics YOLOv8 / YOLO11 (nano, small, medium variants) |
| License | AGPL-3.0 |
| Training Data | Custom annotated datasets (face ~500–1030 images, eyes ~500, head/hair ~3180, breasts ~2000, breast size ~16100) |
about this model
Anzhc/Anzhcs_YOLOs is a collection of YOLO-based object detection and instance segmentation models specialized in human-centric tasks such as face, eyes, head-hair, and breast segmentation and classification. All models are trained on custom-annotated datasets and are optimized for both illustration and real imagery.
Face Segmentation
The universal face segmentation models (v2, v3, v4) detect and segment faces with increasing accuracy. Version 4, trained at 640px on ~1,030 images, achieves mAP 50 of 0.835 (box) and 0.800 (mask). Gender-specific real-photo models are also available, with the ManFace model attaining mAP 50 of 0.883 (box, mask).
| Model | mAP 50 (box) | mAP 50 (mask) | Training Res. |
|---|---|---|---|
| Anzhc Face seg 640 v4 y11n | 0.835 | 0.800 | 640 |
| Anzhcs ManFace v02 1024 y8n | 0.883 | 0.883 | 1024 |
Eyes Segmentation
This model specializes in detecting anime eyes (sclera area) for inpainting workflows. At 1024px training resolution it achieves mAP 50 of 0.925 (box) and 0.868 (mask).
Head+Hair Segmentation
One model each at y8n and y8m sizes. The y8m variant scores mAP 50 of 0.867 (box) and 0.862 (mask) at 640px, trained on ~3,180 images covering both illustration and real content.
Breast Segmentation and Classification
The breast segmentation model (v1, 1024m) achieves mAP 50 of 0.782 (box) and 0.775 (mask) on illustration data. A separate breast size detection/classification model covers 15 size classes using a custom body-proportion scale; +-1 class accuracy is the recommended metric for evaluation.
best for
- ·Automated face masking and inpainting in Stable Diffusion workflows (ADetailer)
- ·Anime character head/hair segmentation for likeness inpainting
- ·Breast size classification for content tagging or moderation
FAQ
It includes YOLO-based detectors/segmenters for faces, eyes, head/hair, and breasts, trained on custom datasets for both anime/illustration and real photos.
The v4 y11n model achieves mAP50 0.835 (box) and 0.800 (mask) on face detection. Gendered real-face models (ManFace, WomanFace) achieve mAP50 0.82–0.88.
The models are released under the AGPL-3.0 license.
Use the gigarouter OpenAI-compatible endpoint with your API key, specifying the model ID as Anzhc/Anzhcs_YOLOs.
Some models target only real photos (e.g., ManFace, WomanFace), while others (face v2–v4) are trained on both illustration and real faces. The breast size model also works on real images.
We're benchmarking and onboarding Anzhc's YOLOs as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.