Celeb-DF (v2): A New Dataset for DeepFake Forensics

Yuezun Li1, Xin Yang1, Pu Sun2, Honggang Qi2 and Siwei Lyu1
1 University at Buffalo, State University of New York, USA
2 University of Chinese Academy of Sciences, China

Github Paper Celeb-DF (v1)

Green box: Real images, Red box: Corresponding DeepFake images.


author = {Yuezun Li and Pu Sun and Honggang Qi and Siwei Lyu},
booktitle = {IEEE Conference on Computer Vision and Patten Recognition (CVPR)},
title = {{Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics}},
address = {Seattle, WA, United States},
year = {2020}, }

What is New?

Celeb-DF (v2) dataset contains real and DeepFake synthesized videos having similar visual quality on par with those circulated online. The Celeb-DF (v2) dataset is greatly extended from our previous Celeb-DF (v1), which only contains 795 DeepFake videos. To date, Celeb-DF includes 590 original videos collected from YouTube with subjects of different ages, ethic groups and genders, and 5639 corresponding DeepFake videos.

Comparison with Other Datasets

See paper for details.

Average AUC of Detection Performance on Each Dataset

See paper for details.

Video Demo


For more information about dowloading Celeb-DF (v2) dataset, see our Github


- Check here for instructions of previous Celeb-DF (v1) dataset.