anime-face-detector-hrnetv2

An anime facial landmark model (HRNetV2-W18 with a heatmap head) that predicts 28 landmark points per face, from hysts/anime-face-detector. It runs on face regions found by one of the companion detectors (see below), with flip test and unbiased (DARK) heatmap decoding.

The model was trained by hysts in 2021 using mmpose. model.safetensors contains the exact weights of the original release (mmpose_anime-face_hrnetv2.pth, v0.0.1); the source file name and its sha256 are recorded in the safetensors metadata. Since v0.1.0 the anime-face-detector package runs this model in plain PyTorch, without the mmpose runtime.

Usage

pip install anime-face-detector

import cv2
from anime_face_detector import create_detector

detector = create_detector("yolov3")  # face detector + this landmark model
image = cv2.imread("input.jpg")
preds = detector(image)  # bounding boxes and 28 landmark points per face

The landmark definition is illustrated in the GitHub repository.

License

The weights are released under the MIT License, like the code in the GitHub repository. They were trained by hysts; as stated in the license, they are provided as is, without warranty of any kind, including with respect to the provenance of the training data.

Related models

• hysts/anime-face-detector-yolov3: the default face detector
• hysts/anime-face-detector-faster-rcnn: a more accurate, heavier face detector