WenhaoWang/ScaleDF

---
license: cc-by-nc-sa-4.0
task_categories:
- feature-extraction
language:
- en
tags:
- deepfake-detection
- face-generation
- ai-safety
- trust-and-safety
pretty_name: ScaleDF
size_categories:
- 10M<n<100M
---

# Summary
This is the dataset proposed in our paper [**Scaling Laws for Deepfake Detection**](https://arxiv.org/abs/2510.16320).
<p align="center">
<img src="https://huggingface.co/datasets/WenhaoWang/ScaleDF/resolve/main/assets/teasor.png" width="800">
</p>

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
...
*unprocessed_fake
3dswap_output_v2.tar # The tar files contain unprocessed fake images or videos. We release unprocessed ones for FS, FR, FE, and TF (except FF, which is too large).
AMatrix_output_v2.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

```python
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
<p align="center">
<img src="https://huggingface.co/datasets/WenhaoWang/ScaleDF/resolve/main/assets/compare.png" width="800">
</p>

# Observed scaling laws

<p align="center">
<img src="https://huggingface.co/datasets/WenhaoWang/ScaleDF/resolve/main/assets/scaling_laws.png" width="800">
</p>

# Included real datasets

<p align="center">
<img src="https://huggingface.co/datasets/WenhaoWang/ScaleDF/resolve/main/assets/real_dataset.png" width="800">
</p>

# Included deepfake methods

<p align="center">
<img src="https://huggingface.co/datasets/WenhaoWang/ScaleDF/resolve/main/assets/fake_method.png" width="800">
</p>

# License

Our ScaleDF is released under the [CC BY-NC-SA 4.0 license](https://creativecommons.org/licenses/by-nc-sa/4.0/deed.en).

# 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 ([e-mail]).