license: ecl-2.0
task_categories:
- image-classification
- image-segmentation
- image-feature-extraction
language:
- en
pretty_name: UniFaceForge Real Images
size_categories:
- 100K<n<1M
To comprehensively evaluate our UniFaceForge model across all types of face forgery attacks, we created a diverse dataset combining real face images from the CelebA, FFHQ datasets and manipulated images including traditional Face morphing, Face swapping with varying blending degrees, regional retouching and face attributes shifting, GAN-based alterations such as StyleGAN2-based attribute modifications, StarGAN v2 style transfer, STGAN-generated identity transformations, and advanced synthetic faces produced by Stable Diffusion v2.1, Midjourney v5, and DALL-E 2. Overall, our dataset comprises 100,000 authentic images and 1,87,185 manipulated images distributed as 79,862 traditional, 92,247 GAN-based, 15,076 advanced synthetic forgeries, organized with a 70% training, 15% validation, and 15% testing split.