image image | mask image | label int64 | label_name string | domain_role string | source_dataset string | manipulation_type string | image_id string |
|---|---|---|---|---|---|---|---|
0 | authentic | model1_in_domain_msccnet | FaceForensics++_c23_real | real_youtube | model1_in_domain_msccnet_ff_real_953_112 | ||
0 | authentic | model1_in_domain_msccnet | FaceForensics++_c23_real | real_youtube | model1_in_domain_msccnet_ff_real_012_116 | ||
0 | authentic | model1_in_domain_msccnet | FaceForensics++_c23_real | real_youtube | model1_in_domain_msccnet_ff_real_078_168 | ||
0 | authentic | model1_in_domain_msccnet | FaceForensics++_c23_real | real_youtube | model1_in_domain_msccnet_ff_real_623_113 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_Deepfakes | Deepfakes | model1_in_domain_msccnet_ff_deepfakes_919_015_275 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_Deepfakes | Deepfakes | model1_in_domain_msccnet_ff_deepfakes_367_371_334 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_Face2Face | Face2Face | model1_in_domain_msccnet_ff_face2face_847_906_200 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_Face2Face | Face2Face | model1_in_domain_msccnet_ff_face2face_529_633_263 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_FaceSwap | FaceSwap | model1_in_domain_msccnet_ff_faceswap_418_507_164 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_FaceSwap | FaceSwap | model1_in_domain_msccnet_ff_faceswap_227_169_227 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_NeuralTextures | NeuralTextures | model1_in_domain_msccnet_ff_neuraltextures_389_480_161 | ||
1 | tampered | model1_in_domain_msccnet | FaceForensics++_c23_NeuralTextures | NeuralTextures | model1_in_domain_msccnet_ff_neuraltextures_821_812_217 |
Face Forgery Domain Evaluation (face-forgery-domain-eval)
Curated human-face forgery detection and pixel-level localization evaluation dataset organized by domain alignment for MSCCNet (Model 1) and OmniFD (Model 2).
Splits
| Split | Samples | Authentic / Tampered | Origin | Domain Role |
|---|---|---|---|---|
model1_in_domain |
12 | 4 real / 8 fake | FaceForensics++ C23 (test.json pairs 0..34) |
Model 1 (MSCCNet) In-Domain + Model 2 (OmniFD FF++) aligned |
model2_in_domain |
12 | 4 real / 8 fake | FaceForensics++ C23 (test.json pairs 35..69, disjoint) |
Model 2 (OmniFD ff_image / ff_spatial) In-Domain |
out_of_domain |
12 | 8 real / 4 fake | nielsr/CelebA-faces (6) + Celeb-DF-v2 (6) |
Out-of-Domain human faces for both models |
Source Provenance & Unavailable Datasets Note
- FaceForensics++ C23 & Celeb-DF-v2 frames/masks were extracted from
Ella4839/DeepfakeBench. Official FF++ binary masks are used directly for manipulated samples inmodel1_in_domainandmodel2_in_domain;Celeb-DF-v2masks are computed via SSIM structural diff + morphological hull between target real and synthesized frames. - Random In-the-Wild Faces were sampled from
nielsr/CelebA-faces(trainsplit) with all-zero ground-truth masks. - Upstream unavailability note: ForgeryNet
public_test_images.taron Google Drive currently hits download quota limits, and the upstream FMLD SharePoint link returns HTTP 404. Because OmniFD officially releases pretrained checkpoints for FaceForensics++ (omnifd_swin-s_20k_ff_image.pth,omnifd_swin-s_20k_ff_spatial.pth) alongside ForgeryNet and FMLD,model2_in_domainuses disjoint FaceForensics++ C23 test pairs directly aligned with OmniFD's FF++ test list.
Schema
image: RGB human face image (PIL.Image)mask: Grayscale binary manipulation mask (0= authentic pixel,255= manipulated pixel)label:0(authentic) or1(tampered)label_name:"authentic"or"tampered"domain_role: split domain designationsource_dataset: upstream source sub-datasetmanipulation_type: manipulation family (real_youtube,Deepfakes,Face2Face,FaceSwap,NeuralTextures,real_portrait,real_celebrity_frame,CelebDF_faceswap)image_id: unique sample identifier
- Downloads last month
- 41