ShyFoo/CountHallu-counting_model-ToyShape

CountHallu — ToyShape Counting Model

Counting model from Counting Hallucinations in Diffusion Models (arXiv:2510.13080). It scores images generated by a diffusion model trained on ToyShape to decide whether each sample is counting-correct or a counting hallucination.

Architecture & checkpoint

• ResNet-50 (torchvision, ImageNet-1k V2 init) with the final layer replaced by a 3-output regression head — one predicted instance count per shape class {triangle, square, pentagon}.
• Ships a single model.pth (a plain state_dict).
• Decision rule: round the 3 predictions; a sample is a hallucination if any class ≥ 2 or all classes are 0 (valid ToyShape images have at most one instance per class and at least one shape).

Usage

Inputs are RGB images normalised to [-1, 1] (ToTensor + Normalize([0.5]*3, [0.5]*3)).

import torch
from huggingface_hub import hf_hub_download
from counthallu.models.counting import CountingRegressor

ckpt = hf_hub_download("ShyFoo/CountHallu-counting_model-ToyShape", "model.pth")
model = CountingRegressor(num_classes=3)
model.load_state_dict(torch.load(ckpt, map_location="cpu"))
model.eval()

Or let the evaluation protocol fetch it for you:

from counthallu.utils import load_counting_model
model, model_type, _, _ = load_counting_model(
    "toyshape", use_hub_model=True,
    repo_id="ShyFoo/CountHallu-counting_model-ToyShape"
)

See the CountHallu repository for the full evaluation protocol.

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}
}