OmniDFA

Few-Shot Synthetic Image Attribution: Identifying Unseen Generators with Limited Samples

<p align="center"> <a href="https://arxiv.org/abs/2509.25682"><img src="https://img.shields.io/badge/arXiv-paper-red?style=flat&logo=arXiv" alt="Paper" height="25"></a> <a href="https://github.com/teheperinko541/OmniDFA"><img src="https://img.shields.io/badge/GitHub-code-black?style=flat&logo=github" alt="Code" height="25"></a> <a href="https://huggingface.co/datasets/MoeNew/OmniFake"><img src="https://img.shields.io/badge/Dataset-OmniFake-blue?style=flat&logo=huggingface&logoColor=white" alt="Dataset" height="25"></a> </p>

OmniDFA (Omni Detector and Few-shot Attributor) is a unified framework for AI-generated image (AIGI) detection and few-shot source attribution. It simultaneously handles two tasks:

• Authenticity Detection: distinguishes real images from AI-generated ones via a one-class hypersphere formulation, achieving state-of-the-art generalization on unseen generators.
• Few-Shot Source Attribution: given K examples from N generator classes, identifies the source of a query image via prototype-based classification — with no retraining required.

Trained and evaluated on OmniFake, a large-scale dataset of 1.17 million images from 45 distinct generators.

Model Weights

This repository contains four checkpoints corresponding to the three cross-validation folds and the zero-shot evaluation setting.

For full evaluation scripts, see the GitHub repository.

Citation

@article{omnidfa2026,
  title={Few-Shot Synthetic Image Attribution: Identifying Unseen Generators with Limited Samples},
  author={Shiyu Wu and Shuyan Li and Jing Li and Jing Liu and Yequan Wang},
  journal={arXiv preprint arXiv:2509.25682},
  year={2026}
}