WenhaoWang/ScaleDF

Summary

This is the dataset proposed in our paper Scaling Laws for Deepfake Detection.

ScaleDF is the largest dataset in the deepfake detection domain to date. It contains over 5.8 million real images from 51 different datasets (domains) and more than 8.8 million fake images generated by 102 deepfake methods.

Using ScaleDF, we observe power-law scaling similar to that shown in large language models (LLMs). Specifically, the average detection error follows a predictable power-law decay as either the number of real domains or the number of deepfake methods increases.

Directory

*DATA_PATH
    *ScaleDF
        *train
            000000AFAD.tar # The tar files starting with 000000 contain real faces.
            000000AVA.tar 
            ...
            AMatrix_faces.tar # The tar files starting without 000000 contain fake faces.
            AniPortrait_faces.tar
            ...
        *val
            000000300VW.tar # The tar files starting with 000000 contain real faces.
            000000GENKI-4K.tar
            ...
            3dSwap_faces.tar # The tar files starting without 000000 contain fake faces.
            DiffFace_faces.tar
            ...
    *Established_benchmarks
        *CDFv2.tar
        *DF40.tar
        *DeepFakeDetection.tar
        *DeepFakeFace.tar
        *ForgeryNet.tar
        *Wild_Deepfake.tar
        *ScaleDF.tar # We also adapt the ScaleDF validation set format to other established benchmarks and provide the adapted version here.

Download

Automatic

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="WenhaoWang/ScaleDF",
    repo_type="dataset"
)

Manually

wget https://huggingface.co/datasets/WenhaoWang/ScaleDF/resolve/main/ScaleDF/train/000000AFAD.tar # This is an example.

Compared to existing datasets

Observed scaling laws

Included real datasets

Included deepfake methods

License

Our ScaleDF is released under the CC BY-NC-SA 4.0 license.

Citation

@article{wang2025scaling,
  title={Scaling Laws for Deepfake Detection},
  author={Wang, Wenhao and Cai, Longqi and Xiao, Taihong and Wang, Yuxiao and Yang, Ming-Hsuan},
  journal={arXiv preprint arXiv:2510.16320},
  year={2025}
}

Contact

If you have any questions, feel free to contact Wenhao Wang ([email protected]).