Awesome

DeCLIP

Official PyTorch Implementation of the Paper:

Ștefan Smeu, Elisabeta Oneață, Dan Oneață
DeCLIP: Decoding CLIP Representations for Deepfake Localization
WACV, 2025

Data

To set up your data, follow these steps:

1. Download the datasets:

   • Dolos Dataset: Follow instructions from Dolos GitHub repo
   • AutoSplice Dataset: Follow instructions from AutoSplice GitHub repo

2. Organize the data:

   After downloading, place the datasets in the datasets folder to match the following structure:

   ├── data/
   ├── datasets/
   │   ├── AutoSplice/
   │   ├── dolos_data/
   │   │   ├── celebahq/
   │   │   │   ├── fake/
   │   │   │   │   ├── lama/
   │   │   │   │   ├── ldm/
   │   │   │   │   ├── pluralistic/
   │   │   │   │   ├── repaint-p2-9k/
   │   │   │   ├── real/
   │   │   ├── ffhq/
   ├── models/
   ├── train.py
   ├── validate.py
   ├── ...
   

Installation

Main prerequisites:

• Python 3.10.14
• pytorch=2.2.2 (cuda 11.8)
• pytorch-cuda=11.8
• torchvision=0.17.2
• scikit-learn=1.3.2
• pandas==2.1.1
• numpy=1.26.4
• pillow=10.0.1
• seaborn=0.13.0
• matplotlib=3.7.1
• tensorboardX=2.6.2.2

Train

To train the models mentioned in the article, follow these steps:

1. Set up training and validation data paths in options/train_options.py or specify them as arguments when running the training routine.

2. Run the training command using the following template:

python train.py --name=<experiment_name> --train_dataset=<dataset> --arch=<architecture> --decoder_type=<decoder> --feature_layer=<layer> --fix_backbone --fully_supervised

Example commands:

Train on Repaint-P2:

python train.py --name=test_repaint --train_dataset=repaint-p2-9k --data_root_path=datasets/dolos_data/celebahq/ --arch=CLIP:ViT-L/14 --decoder_type=conv-20 --feature_layer=layer20 --fix_backbone --fully_supervised

Where:

• arch specifies the architecture, such as CLIP:RN50, CLIP:ViT-L/14, CLIP:xceptionnet, or CLIP:ViT-L/14,RN50.
• decoder_type can be linear, attention, conv-4, conv-12, or conv-20.
• feature_layer ranges from layer0 to layer23 for ViTs and from layer1 to layer4 for ResNets.

Exceptions:

• For CLIP:xceptionnet, features are always extracted from the 2nd block.
• For CLIP:ViT-L/14,RN50, the argument value specifies the layer from ViT; for RN50, features are always extracted from layer3.
• Use --fully_supervised for localization tasks. Omit it for image-level detection tasks.

Pretrained Models

We provide trained models for the networks which rely on ViT and ViT+RN50 backbones listed in the table below.

Additionally, one can download the checkpoints using gsutil from this GCS bucket. The weights are located in backbone_VIT and backbone_VIT+RN50 folders, where each checkpoints follows the naming convention: <backbone>_<feature_layer>_<decoder>_<training_dataset>, where training_dataset is lower-cased. For the case of features concatenated from ViT and RN50, a + charachter joins the 2 backbones and feature layers.

Evaluation

To evaluate a model, use the following template:

python validate.py --arch=CLIP:ViT-L/14 --ckpt=path/to/the/saved/mode/checkpoint/model_epoch_best.pth --result_folder=path/to/save/the/results --fully_supervised

License

This repository also integrates code from the following repositories:

@inproceedings{ojha2023fakedetect,
      title={Towards Universal Fake Image Detectors that Generalize Across Generative Models}, 
      author={Ojha, Utkarsh and Li, Yuheng and Lee, Yong Jae},
      booktitle={CVPR},
      year={2023},
}

@inproceedings{patchforensics,
  title={What makes fake images detectable? Understanding properties that generalize},
  author={Chai, Lucy and Bau, David and Lim, Ser-Nam and Isola, Phillip},
  booktitle={European Conference on Computer Vision},
  year={2020}
 }

Citation

If you find this work useful in your research, please cite it.

@InProceedings{DeCLIP,
    author    = {Smeu, Stefan and Oneata, Elisabeta and Oneata, Dan},
    title     = {DeCLIP: Decoding CLIP representations for deepfake localization},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    year      = {2025}
}