Multimodal Conditional Information Bottleneck for Generalizable AI-Generated Image Detection

Official implement of Multimodal Conditional Information Bottleneck for Generalizable AI-Generated Image Detection.

Haotian Qin¹, Dongliang Chang^1*, Yueying Gao¹, Bingyao Yu², Lei Chen², Zhanyu Ma¹

¹Beijing University of Posts and Telecommunications, Beijing, China
²Tsinghua University, Beijing, China

*Corresponding author.

News

• June 2025: Released training code and GenImage caption annotations. The initial codebase is now available for public use, including scripts for training the model and annotations for the GenImage dataset.

TODO

• Validation Code: Implement and release validation scripts to evaluate the model's performance on various datasets.
• Inference Code: Develop and share inference scripts for applying the trained model to new data for AI-generated image detection.
• Plug-and-Play: Add a modular, plug-and-play implementation of the MMCIB framework to facilitate easy integration into other networks.

Installation

To get started with the project, follow these steps:

1. Clone the repository:

   git clone https://github.com/Ant0ny44/InfoFD.git

2. Install dependencies:

   conda env create -f environment.yml -n infoFD

3. Modify the env.ini.

   [WANDB]
   TOKEN=YOUR_WANDB_TOKEN_HERE

4. Change the data path in configs/EP1.yml:

    data:
   
   
    # This is used to specify the cache path for training/validation/test data. 
    # You can also directly provide the path to preprocessed training/validation/test data here.
    # When using cached training, if the specified file path does not exist, 
    # preprocessing will be performed first, and the resulting data will be stored in the corresponding cache path.
    # Note that for training data, captions corresponding to the images are required.
    
    train_root_cache:  TRAIN_ROOT_CACHE_PATH 
    val_root_cache: VAL_ROOT_CACHE_PATH
    test_root_cache: TEST_ROOT_CACHE_PATH
   
    train_root: GENIMAGE_TRAIN_PATH
    train_captions_path: ./data/genImage_train_captions.json
    val_root: GENIMAGE_VAL_PATH
    test_root: GENIMAGE_TEST_PATH
   
   
    prompts: True
    shuffle: True
    num_workers: 14
    batch_size:  512
    ...

Usage

Training

The training code is available in the train.py directory. To train the model, run:

bash scripts/train_EP1.sh

To run the statistic results, run:

bash scripts/train_EP1_stas.sh

See the configs/ directory for model details.

GenImage Training Caption Annotations

The GenImage training caption annotations are available in the data/genImage_train_captions.json directory. These annotations provide textual descriptions for the GenImage dataset, generating by InternVL.

Contact

Thank you for your interest. We're currently finalizing the code organization. If you have any questions, please don't hesitate to reach out at ant0ny@163.com.

Citation

If you use this code or the GenImage annotations in your research, please cite our work using the following BibTeX entry:

@article{qin2025multimodal,
  title={Multimodal Conditional Information Bottleneck for Generalizable AI-Generated Image Detection},
  author={Qin, Haotian and Chang, Dongliang and Gao, Yueying and Yu, Bingyao and Chen, Lei and Ma, Zhanyu},
  journal={arXiv preprint arXiv:2505.15217},
  year={2025}
}

License

This project is licensed under the MIT License. See the LICENSE file for details.