Towards Generalizable Forgery Detection and Reasoning

Achieving accurate and interpretable AI-generated images detection via MLLMs' visual reasoning capability.

ArXiv Paper GitHub Code Model Model Dataset Dataset

Authors

Yueying Gao1, Dongliang Chang1, Bingyao Yu2, Haotian Qin1, Muxi Diao1, Lei Chen2, Kongming Liang1, Zhanyu Ma1
1 PRIS, Beijing University of Posts and Telecommunications
2 Tsinghua University

Contributions

FakeReasoning: A forgery detection and reasoning framework, providing accurate detection with structured and reliable reasoning over forgery attributes.
MMFR-Dataset: A multi-modal forgery reasoning dataset, containing 120K training images and 20K evaluation images annotated with detailed reasoning over forgery attributes.

Detection and Reasoning Performance of FakeReasoning

FakeReasoning conducts forgery reasoning task in structured stages (summary, caption, reasoning and conclusion) and hierarchical steps (low-level and high-level), leading to accurate and interpretable detection.

GAN-generated image
DM-generated image
Real image I
Real image II

MMFR-Dataset

Constructed with advanced GPT-4o, MMFR-Dataset contains over 100,000 images with over 310,000 reasoning annotations as its training set.
Here we illustrate several images and according annotations from reasoning stage.

Evaluation sets include 20,000 images with over 66,000 reasoning annotations across 10 up-to-data generative models and are balanced in terms of authenticity.
Here we illustrate several real and fake images from each evaluation set.

Cite Our Work

@article{gao2025fakereasoning,
  title={FakeReasoning: Towards Generalizable Forgery Detection and Reasoning},
  author={Gao, Yueying and Chang, Dongliang and Yu, Bingyao and Qin, Haotian and Chen, Lei and Liang, Kongming and Ma, Zhanyu},
  journal={arXiv preprint arXiv:2503.21210},
  year={2025},
  url={https://arxiv.org/abs/2503.21210}
}