nakamura196/yolov11x-ukiyoe-face

YOLOv11x Ukiyo-e Face

Model Description

A YOLOv11x detector fine-tuned to localize faces in ukiyo-e prints (浮世絵). Photographic face detectors (Haar, RetinaFace, YuNet) struggle with the stylized faces of Japanese woodblock prints. This model is trained on the ARC Ukiyo-e Faces dataset (Ritsumeikan ARC × CODH), giving strong recall on ukiyo-e portraiture.

• Single class: face
• Trained at imgsz=1536 for robustness to small faces in dense group compositions.

This is the ukiyo-e counterpart to nakamura196/yolov11x-kaokore-face (trained on the KaoKore dataset).

Intended Uses

• Face detection / localization in ukiyo-e and related Japanese prints
• Building face-image collections (顔貌コレクション) from digitized artworks
• Pre-processing for downstream attribution / attribute classification

How to Use

from ultralytics import YOLO

model = YOLO("nakamura196/yolov11x-ukiyoe-face")

# imgsz=1536 matches the training resolution (better recall on small faces)
results = model.predict("your_image.jpg", imgsz=1536, conf=0.25, iou=0.45)

for result in results:
    print(result.boxes)

Download the weights directly:

from huggingface_hub import hf_hub_download

pt_path   = hf_hub_download(repo_id="nakamura196/yolov11x-ukiyoe-face", filename="best.pt")
onnx_path = hf_hub_download(repo_id="nakamura196/yolov11x-ukiyoe-face", filename="best.onnx")

The best.onnx export (input shape 1×3×1536×1536, opset 12) is suitable for ONNX Runtime (incl. onnxruntime-web in the browser).

Training Data

Fine-tuned on the ARC Ukiyo-e Faces dataset (rois-codh/arc-ukiyoe-faces). Face bounding boxes are derived from the dataset's normalized face annotations; images are sourced from the Ritsumeikan ARC ukiyo-e database. To avoid leakage, prints belonging to the same series/polyptych (揃物) are kept within a single split.

Split	Images
train	5,980
val	744
test	751

Evaluation

Validation metrics on the held-out split (best epoch):

Metric	Value
mAP@50	0.931
mAP@50-95	0.634
Precision	0.839
Recall	0.922

Model Architecture

• Base Model: YOLOv11x (Ultralytics, extra-large variant)
• Task: Object Detection (1 class: face)
• Training resolution: 1536
• Framework: Ultralytics
• Formats: PyTorch (best.pt), ONNX (best.onnx)

Limitations

• Optimized for ukiyo-e / Japanese prints; not intended for photographic face detection.
• For very high-resolution scans with many small faces, consider tiled inference (SAHI).
• Performance varies with style, period, and scan quality.

Acknowledgments & Dataset Attribution

This model is a derivative work of the ARC Ukiyo-e Faces Dataset, which is licensed under CC BY 4.0. This model uses both the dataset's face annotations and images sourced from the Ritsumeikan ARC database, so the dataset's required credits are reproduced below:

"ARC Ukiyo-e Faces Dataset" (Created by Yingtao Tian, ROIS-DS CODH; Collected from ARC), https://doi.org/10.20676/00000394

Art Research Center, Ritsumeikan University (2020): ARC Ukiyo-e database. Informatics Research Data Repository, National Institute of Informatics. (dataset). https://doi.org/10.32130/rdata.2.1

Metadata from the ARC Ukiyo-e database is distributed within the ARC Ukiyo-e Faces Dataset with permission from the Art Research Center, Ritsumeikan University.

Citation

If you use this model, please cite the ARC Ukiyo-e Faces dataset paper:

@inproceedings{tian2021ukiyoe,
  author    = {Yingtao Tian and Tarin Clanuwat and Chikahiko Suzuki and Asanobu Kitamoto},
  title     = {Ukiyo-e Analysis and Creativity with Attribute and Geometry Annotation},
  booktitle = {Proceedings of the International Conference on Computational Creativity (ICCC)},
  year      = {2021}
}

License

This model is released under AGPL-3.0, inherited from its base model Ultralytics YOLOv11. The training data (ARC Ukiyo-e Faces Dataset) is licensed under CC BY 4.0 (Created by Yingtao Tian and ROIS-DS CODH; collected from ARC); the attribution above is required under that license. Please also observe the terms of the Ritsumeikan ARC Ukiyo-e database.