license: agpl-3.0
pipeline_tag: object-detection
tags:
- yolo
- finger
- detection
CountHallu — RealHand Counting Model
Finger detector from Counting Hallucinations in Diffusion Models (arXiv:2510.13080). A YOLO-based detector is used to count fingers in generated/real hand images, enabling the evaluation to flag hands with an incorrect number of fingers as counting hallucinations.
Architecture & checkpoint
- An Ultralytics YOLO-v13 object detector, ships a single
model.pt. - One class,
0 = finger; a correct hand has exactly 5 detected fingers. - Inference settings used in the paper:
imgsz=640, conf=0.25, iou=0.1.
Two-stage pipeline. RealHand scoring first passes each image through the quality classifier
CountHallu-quality_cls_model-RealHand, which filters out non-countable (visually failed) images; only clean images reach this finger detector. You need both models to calculate the counting hallucination rates in the RealHand dataset.
Usage
See the CountHallu repository for the full evaluation protocol.
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("ShyFoo/CountHallu-counting_model-RealHand", "model.pt")
model = YOLO(ckpt, task="detect")
results = model.predict(source="hand.png", imgsz=640, conf=0.25, iou=0.1)
num_fingers = len(results[0].boxes)
Or let the evaluation protocol fetch it for you:
from counthallu.utils import load_counting_model
model, model_type, ref_counts, target_classes = load_counting_model(
"realhand", use_hub_model=True,
repo_id="ShyFoo/CountHallu-counting_model-RealHand"
)
License
This detector is trained with Ultralytics YOLO-v13 on the CountHallu-RealHand dataset, which is AGPL-3.0.
Citation
@article{fu2025counting,
title={Counting Hallucinations in Diffusion Models},
author={Fu, Shuai and Zhou, Jian and Chen, Qi and Jing, Huang and Nguyen, Huy Anh and Liu, Xiaohan and Zeng, Zhixiong and Ma, Lin and Zhang, Quanshi and Wu, Qi},
journal={arXiv preprint arXiv:2510.13080},
year={2025}
}