DeeCLIP: A Robust and Generalizable Transformer-Based Framework for Detecting AI-Generated Images

☀️ If you find this work useful for your research, please kindly star our repo and cite our paper! ☀️

TODO

We are working hard on the following items:

• Release ArXiv paper
• Release training scripts
• Release inference scripts
• Release checkpoints

DeeCLIP

DeeCLIP is a novel framework for detecting AI-generated images using a combination of CLIP-ViT and fusion learning. As generative models continue to improve, detecting synthetic images becomes more challenging due to poor generalization and vulnerability to common image manipulations. DeeCLIP tackles these issues with a robust and adaptable architecture.

🔍 Overview

DeeCLIP enhances the detection of AI-generated images by integrating:

• ✅ Zero-shot learning via CLIP-ViT for model-agnostic generalization.
• 🧠 LoRA-based fine-tuning to efficiently adapt without forgetting.
• 🔀 DeeFuser: a custom fusion module combining deep and shallow features for post-processing robustness.
• 🎯 Triplet loss to structure the embedding space and improve the real vs. fake discrimination.

Requirements

pip install -r requirements.txt

Download Model Weights

To use the model, download the weight and save it in the weights folder.

Automatic Download (Command Line)

Run the following command in your terminal:

mkdir -p weights && wget -O weights/deeclip_weight_complete_with_lora_5.pth "https://www.dropbox.com/scl/fi/ttiqnbxu8atz4on5gqvgd/deeclip_weight_complete_with_lora_5.pth?rlkey=6xznuvriabkqfdcofhi1pbihu&st=fk02k7hf&dl=1"

SOTA Detection Methods

• C2P-CLIP
• RINE
• FatFormer
• AntifakePrompt
• Bi-LORA

Evaluation

To run the test on specific dataset, use the following command:

python testing.py

T-SNE Plot of feature distribution

Generalization Performance of Methods Trained on ProGAN

Results show accuracy (%) on real and synthetic data subsets.

📖 Citation

if you make use of our work, please cite our paper

@misc{keita2025deecliprobustgeneralizabletransformerbased,
      title={DeeCLIP: A Robust and Generalizable Transformer-Based Framework for Detecting AI-Generated Images}, 
      author={Mamadou Keita and Wassim Hamidouche and Hessen Bougueffa Eutamene and Abdelmalik Taleb-Ahmed and Abdenour Hadid},
      year={2025},
      eprint={2504.19876},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.19876}, 
}