Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning

Yiheng Li, Zicahng Tan, Guoqing Xu, Zhen Lei, Xu Zhou and Yang Yang

MAIS&CASIA, UCAS, Sangfor

Introduction

This repository is an official implementation of IAPL, codes and weight will be released after paper accepted.

News

• [2026/3/4] Codes and pre-trained weights are released.
• [2026/2/21] Our paper is accepted by CVPR 2026.

Methods

Visualization

Environment Setting

pip install -r requirements.txt

Data Preparation

Download UniversalFakeDetect and GenImage Datasets.

Organize the directory structure as follows:

Datasets
└── UniversalFakeDetect
    └── train
          ├── car
          ├── horse
          │      .
          │      .
    └── test					
          ├── progan	
          │── cyclegan   	
          │── biggan
          │      .
          │      .

└── GenImage
    └── train
          ├── SDv14
              ├── 0_real
              ├── 1_fake

    └── test					
          ├── ADM
              ├── 0_real
              ├── 1_fake
          │── BigGAN   	
          │── glide
          │      .
          │      .

Experiments on 4-Class ProGAN

Training:

sh run_universalfake.sh

Testing on universalfakedetect:

sh tta_universalfake.sh

Testing on Chameleon:

sh tta_chameleon.sh

Results:

Benchmark	mACC(%)	mAP(%)
UniversalFakeDetect	95.61	99.32
Chameleon	60.70	50.43

Experiments on SD v1.4

Training:

sh run_genimage.sh

Testing on GenImage:

sh tta_genimage.sh

Testing on Chameleon:

sh tta_chameleon_sdv1.4.sh

Results:

Benchmark	mACC(%)	mAP(%)
GenImage	96.7	99.5
Chameleon	75.09	64.69

Pre-trained Models

We release the pre-trained models on ModelScope

Acknowledgement

We sincerely thank the following repos: UniversalFakeDetect, FatFormer, AIDE and TPT.

Citation

@article{li2025towards,
  title={Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning},
  author={Li, Yiheng and Tan, Zichang and Xu, Guoqing and Lei, Zhen and Zhou, Xu and Yang, Yang},
  journal={arXiv preprint arXiv:2508.01603},
  year={2025}}