DiFF is a high-quality, large-scale facial forgery image dataset jointly developed by Shandong University, National University of Singapore and other institutions. It is generated based on the diffusion model and contains more than 500,000 images. These images are generated in 4 different environments using 13 cutting-edge technologies.

The dataset uses 30,000 carefully collected text and visual cues to ensure that the generated images have high authenticity and semantic consistency. The original images come from public datasets such as VoxCeleb2 and CelebA, containing 23,661 original images of 1,070 different identities. The DiFF dataset is suitable for facial forgery detection, adversarial attacks and defenses against deep forgeries, and other related computer vision task training.