YOLOv8 Anime Face Detector

Yolov8_animeface is a detection model that identifies anime faces in images, based on the YOLOv8x architecture and fine-tuned on a custom dataset of 10,000 manually annotated images from Safebooru.

Architecture and Training

The model uses the YOLOv8x6 variant (a larger, non-standard input size variant of YOLOv8x) and was trained for 300 epochs at 1280x1280 pixel resolution. Training took approximately 110 hours on an RTX A4000 GPU. The dataset was split 70% training, 20% validation, and 10% testing.

Performance

On the test set (1,002 images with 1,562 instances), the model achieves:

Metric	Value
Box Precision (P)	0.957
Recall (R)	0.924
mAP50	0.955
mAP50-95	0.534

Inference speed is 81.9ms per image at 1280x1280 resolution, with 1.3ms preprocessing and 0.8ms postprocessing.

Comparison with YOLOv5-based Model

Compared to the earlier YOLOv5-based anime face detector (zymk9/yolov5_anime, now unavailable), this model shows significant improvements on the same test set:

Metric	YOLOv8-animeface	YOLOv5-anime
Precision	0.956	0.778
Recall	0.919	0.685
mAP50	0.953	0.633
mAP50-95	0.532	0.232

The YOLOv8-based model demonstrates higher precision and recall with fewer false positives, though the YOLOv5 model showed higher confidence in its predictions.

Visual Results

Comparison of manual annotations (left) and model predictions (right):

The base YOLOv8x model (from which this is fine-tuned) has 68.2M parameters and 257.8B FLOPs, achieving 53.9 mAP50-95 on COCO at 640px input. This fine-tuned model trades some general object detection capability for specialized anime face detection performance at higher resolution.