README.md

---
license: apache-2.0
pipeline_tag: image-text-to-text
---

# FakeReasoning

FakeReasoning is a forgery detection and reasoning framework designed to accurately detect AI-generated images and provide reliable reasoning over forgery attributes. It formulates detection and explanation as a unified Forgery Detection and Reasoning task (FDR-Task).

- **Project Page:** [https://pris-cv.github.io/FakeReasoning/](https://pris-cv.github.io/FakeReasoning/)
- **Paper:** [Toward Generalizable Forgery Detection and Reasoning](https://huggingface.co/papers/2503.21210)
- **Code:** [https://github.com/PRIS-CV/FakeReasoning](https://github.com/PRIS-CV/FakeReasoning)

## Model Details

FakeReasoning consists of three key components:
1.  **Dual-branch visual encoder:** Integrates CLIP and DINO to capture both high-level semantics and low-level artifacts.
2.  **Forgery-Aware Feature Fusion Module:** Leverages DINO's attention maps and cross-attention mechanisms to guide the model toward forgery-related clues.
3.  **Classification Probability Mapper:** Couples language modeling and forgery detection, enhancing overall performance.

The model was trained on the **MMFR-Dataset**, a large-scale dataset containing 120K images across 10 generative models with 378K reasoning annotations.

## Sample Usage

To use the model, please follow the installation instructions in the [official repository](https://github.com/PRIS-CV/FakeReasoning). You can then run inference using the following commands:

```bash
cd LLaVA/forgery_eval
export DINO_PATH='path_to_dinov2-main'
export DINO_WEIGHT='path_to_dinov2_vitl14_pretrain.pth'

python inference.py \
--model-path AnnaGao/FakeReasoning \
--img_path path_to_your_image.png
```

Note: Inference and evaluation require at least 30 GB of GPU memory on a single GPU.

## Citation

```bibtex
@article{gao2025fakereasoning,
  title={FakeReasoning: Towards Generalizable Forgery Detection and Reasoning},
  author={Gao, Yueying and Chang, Dongliang and Yu, Bingyao and Qin, Haotian and Chen, Lei and Liang, Kongming and Ma, Zhanyu},
  journal={arXiv preprint arXiv:2503.21210},
  year={2025},
  url={https://arxiv.org/abs/2503.21210}
}
```