NFA-ViT Pretrained Weights (Re-upload)

This repository contains the re-uploaded pretrained weights for NFA-ViT (Noise-guided Forgery Amplification Vision Transformer), introduced in the AAAI 2026 paper "Zooming In on Fakes: A Novel Dataset for Localized AI-Generated Image Detection with Forgery Amplification Approach".

This is NOT an officianl re-upload. The original weights were previously available only via Baidu Netdisk, that requires a Chinese phone number to create an account. Re-uppload distibuted under the original CC BY 4.0 license.

Model Weights

The initial weights of NFA-ViT consist of three parts:

File Name	Description
noiseprint.pth	DnCNN-based noise extractor weights
segformer_b2_backbone_weights.pth	Image backbone (SegFormer-B2)
segformer_b0_backbone_weights.pth	Noise backbone (SegFormer-B0)

Full NFA-ViT training weights fine-tuned on the BR-Gen dataset also provided.

Citation

If you use these weights, please cite the original paper using this BibTeX:

@article{cai2025zooming,
  title={Zooming In on Fakes: A Novel Dataset for Localized AI-Generated Image Detection with Forgery Amplification Approach},
  author={Cai, Lvpan and Wang, Haowei and Ji, Jiayi and ZhouMen, YanShu and Ma, Yiwei and Sun, Xiaoshuai and Cao, Liujuan and Ji, Rongrong},
  journal={Proceedings of the the AAAI Conference on Artificial Intelligence (AAAI)},
  year={2026}
}

Original Resources

• Paper: arXiv:2504.11922View
• Official Repository: clpbc/BR-Gen

Acknowledgments & Credits

Original Authors

The model and dataset were developed by:

Lvpan Cai, Haowei Wang, Jiayi Ji, Yanshu Zhoumen, Shen Chen, Taiping Yao, Xiaoshuai Sun (Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Xiamen University & Youtu Lab, Tencent)

Re-upload Credits

This re-upload was made possible thanks to Yuri The Romantic Dev (@the-romantic-dev, Linkedin), who discovered the paper and successfully retrieved the original weights from Baidu Netdisk, making them accessible to the international research community.