GitHub - AniSundar18/AlignedForensics (original) (raw)

Improving Detectors by Mitigating Spurious Correlations

Official repository of Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025) and Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025). The specific instructions for training the models and reproducing the plots can be found in the training_code and testing_code directories.

Pretrained Models

Our pre-trained fake image detectors can be found below,

• Ours (ResNet-50 trained without sync'd batches)
• Ours-Sync (ResNet-50 trained with sync'd batches)
• Ours (shaders) (ResNet-50 trained on shaders)
• Ours-Sync (shaders) (ResNet-50 trained on shaders, with sync'd batches)

Our Detectors from the ICML paper (Stay-Positive)

• Corvi + (ResNet-50)
• Rajan/Ours + (ResNet-50)

Evaluation Dataset

In order to download the dataset that we use for our evaluation, please use the following link. For the dataset used in the ICML paper, which includes additional images from FLUX, Wuerstchen and aMUSEd, please use the following link.

Citation

If you find this code useful in your research, consider citing our work:

@inproceedings{
rajan2025aligned,
title={Aligned Datasets Improve Detection of Latent Diffusion-Generated Images},
author={Anirudh Sundara Rajan and Utkarsh Ojha and Jedidiah Schloesser and Yong Jae Lee},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=doBkiqESYq}
}

@inproceedings{
rajan2025staypositive,
title={Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection},
author={{Anirudh Sundara Rajan and Yong Jae Lee},
booktitle={Forty-second International Conference on Machine Learning},
year={2025},
url={https://openreview.net/forum?id=VNLmfMJi3w}
}

Acknowledgements

We thank the authors of On the detection of synthetic images generated by diffusion models and Raising the Bar of AI-generated Image Detection with CLIP for releasing their source code.