Use the official FaceForensics++ distribution. The default protocol expects the original, Deepfakes, Face2Face, FaceSwap, and NeuralTextures C23 video folders. The code fetches or reads the official train/validation/test split JSON files under protocols/faceforensics_splits/.
Expected local structure:
data/FaceForensics++/
├── original/
├── Deepfakes/
├── Face2Face/
├── FaceSwap/
└── NeuralTextures/
If additional folders such as DeepFakeDetection or FaceShifter are present locally, the standard protocol ignores them unless the configuration is intentionally changed.
Use the official Celeb-DF v2 release. The external evaluation uses the official testing list when available.
Expected local structure:
data/Celeb_DF_v2/
├── Celeb-real/
├── Celeb-synthesis/
├── YouTube-real/
└── List_of_testing_videos.txt
Use the official WildDeepfake split folders. In this implementation, WildDeepfake is evaluated as cached sample-level face images.
Expected local structure:
data/WildDeepfake/
├── train/real/
├── train/fake/
├── valid/real/
├── valid/fake/
├── test/real/
└── test/fake/