kd13/deepfakebench-unfiltered

DeepfakeBench Unfiltered

Face images from the datasets of DeepfakeBench (NeurIPS 2023 D&B), one subset per source dataset, each split 70/15/15 train/validation/test.

• Export ID 58bbde81a437004c · scope auto · 818,847 unique images (850,796 rows; FF++ real frames are also included in FaceShifter).
• Original image bytes, no resizing, re-encoding, sampling or quality filtering. Masks, landmarks and videos are not included.

Subsets

subset	train	validation	test	total	real	FS	FR	EFS	FE	PA
FF++	111,781	23,929	23,916	159,626	31,949	63,797	63,880	0	0	0
FaceShifter	44,699	9,589	9,585	63,873	31,949	31,924	0	0	0	0
DFD	74,523	14,026	13,992	102,541	10,741	91,800	0	0	0	0
DFDCP	100,294	21,492	21,492	143,278	28,326	114,952	0	0	0	0
DFDC	92,481	19,818	19,817	132,116	63,265	68,851	0	0	0	0
Celeb-DF-v1	29,612	4,341	4,359	38,312	12,950	25,362	0	0	0	0
Celeb-DF-v2	150,248	28,847	28,856	207,951	28,174	179,777	0	0	0	0
UADFV	2,167	442	490	3,099	1,548	1,551	0	0	0	0

Labels

class_id	class_name	binary_label	binary_name	unique images
0	real	0	real	176,953
1	FS	1	fake	578,014
2	FR	1	fake	63,880
3	EFS	1	fake	0
4	FE	1	fake	0
5	PA	1	fake	0

• FS (face swap): FF++ Deepfakes & FaceSwap, FaceShifter, DFD, Celeb-DF, UADFV, DFDC, DFDCP, DeeperForensics (DF-VAE).
• FR (face reenactment): FF++ Face2Face & NeuralTextures.
• EFS / FE / PA are part of the schema but none of these nine sources contain them (0 rows; nothing is relabelled to fill them).
• DFDC/DFDCP fakes are mapped to FS at dataset level; the benchmark has no per-frame generator identity.

Splits

Splits are created by this collector, not taken from the original papers. Inside every subset the ratio is applied separately for each category (real, FS, FR, …), and whole groups move together:

• all frames of one video;
• FF++/FaceShifter: the two youtube ids of a pair (real 000, 003 and every fake 000_003, 003_000);
• DFD: a fake and its target real actor video; Celeb-DF: a fake and its target real video (shared by v1 and v2); UADFV: 0000 and 0000_fake.

A group keeps the same split in every subset it appears in. group_id is a column, and the original DeepfakeBench split memberships are kept in original_splits (JSON list). Exact per-class counts: split_counts.tsv, split_report.json.

Usage

from datasets import load_dataset

ds = load_dataset("kd13/deepfakebench-unfiltered", "FF++")            # DatasetDict: train / validation / test
celeb = load_dataset("kd13/deepfakebench-unfiltered", "Celeb-DF-v2", split="test")
stream = load_dataset("kd13/deepfakebench-unfiltered", "all", split="train", streaming=True)

row = ds["train"][0]
row["image"], row["class_id"], row["class_name"], row["binary_label"], row["binary_name"]

The all config unions every subset (column subset). FF++ real frames appear there twice (as FF++ and FaceShifter), always in the same split; deduplicate on sha256/source_path if needed.

Coverage and caveats

• Official DeepfakeBench RGB release: FF++/FaceShifter/DFD are c23 only; DFDC is the test set only.
• DeeperForensics-1.0 is not in the official release; it is present only if the uploader supplied it.
• Audit: every supplied image was decoded and hashed, and every in-scope reference frame was matched. This does not certify every frame of every original upstream dataset. See audit_summary.json.

License and terms

Every source dataset keeps its own license and terms of use (FaceForensics++, DFD, DFDC/DFDCP, Celeb-DF, DeeperForensics, UADFV). Most allow research use only, and several forbid redistribution without agreement. Check them before you use or share this data. The DeepfakeBench code is CC BY-NC 4.0, which does not grant rights to the images.

Citation

@inproceedings{DeepfakeBench_YAN_NEURIPS2023,
  author = {Yan, Zhiyuan and Zhang, Yong and Yuan, Xinhang and Lyu, Siwei and Wu, Baoyuan},
  booktitle = {Advances in Neural Information Processing Systems},
  title = {DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection},
  volume = {36}, pages = {4534--4565}, year = {2023}
}