IncreFA: Breaking the Static Wall of Generative Model Attribution

Official implementation of the paper "IncreFA: Breaking the Static Wall of Generative Model Attribution" (CVPR 2026).

[Paper] [Dataset]

---

📢 News

• [2026-04] IABench dataset (512GB) is now available on ModelScope!
• [2026-03] IncreFA has been accepted by CVPR 2026.

---

🛠️ Code Maintenance

Important

We are currently cleaning up and refactoring the training and evaluation code to ensure reproducibility. The full source code, including pre-trained weights and configuration files, will be released sequentially. Please Stay Tuned! 🌟

---

📊 IABench Dataset

IABench (Incremental Attribution Benchmark) is a large-scale dataset specifically designed for generative model attribution in incremental learning scenarios.

• Scale: ~570,000 images, 512GB.
• Coverage: 28 generative models (2022–2025).
• Format: Optimized Arrow binary format for high-performance I/O and seamless streaming.

📥 Data Access

The dataset is hosted on ModelScope. You can load it directly using the modelscope library without manual decompression.

1. Installation

pip install modelscope datasets

✒️ Citation

If you find our work or the IABench dataset useful for your research, please cite:

@inproceedings{qin2026increfa,
  title={IncreFA: Breaking the Static Wall of Generative Model Attribution},
  author={Haotian Qin, Dongliang Chang, Yueying Gao, Lei Chen, and Zhanyu Ma},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}

---

📩 Contact

For any questions regarding the code or dataset, please open an issue or contact Haotian Qin at qinhaotian@bupt.edu.cn.