This dataset is derived from FaceForensics++ and is distributed under the FaceForensics Terms of Use. By requesting access you agree to be bound by those terms, including using the data only for non-commercial research and educational purposes.

FaceForensicsC23

FaceForensics++ (paper) videos split into train/test for DeepGuard, together with the face crops, 5-point facial landmarks and face bounding boxes that DeepGuard extracted from the train videos (preprocess/ff++).

Source: The videos and the Kaggle CSVs (csv/) come from the Kaggle dataset FaceForensics++ (c23) (xdxd003/ff-c23). Everything else was produced by DeepGuard from those videos.

Path	Split	Content
`frames/` (Dataset Viewer)	`train`	Preprocessed face crops of the train videos, with landmarks, boxes and labels
`videos/{split}/{source}.zip`	`train`, `test`	Original `.mp4` videos, one zip per split and manipulation method
`raw/`	`train`	Preprocessing outputs and metadata CSVs to restore the DeepGuard `DATA_ROOT`

Overview

FaceForensics++ is a forensics dataset of 1,000 original video sequences sourced from 977 YouTube videos, manipulated with automated face manipulation methods. All videos contain a trackable, mostly frontal face without occlusion. The videos are the c23 (light compression) release taken from Kaggle xdxd003/ff-c23; this dataset uses the five manipulation methods below and excludes DeepFakeDetection.

Source	Real/Fake	Description
`Deepfakes`	Fake	Deep learning-based face replacement using autoencoders
`Face2Face`	Fake	Facial reenactment that transfers expressions from source to target
`FaceSwap`	Fake	Graphics-based face replacement using traditional algorithms
`FaceShifter`	Fake	High-fidelity face swapping with robust occlusion handling
`NeuralTextures`	Fake	Facial reenactment using learned neural textures to modify expressions
`original`	Real	Unaltered, authentic videos collected from YouTube

Terms of Use

This is FaceForensics++ data and data derived from it, and it remains subject to the FaceForensics Terms of Use. Access is granted only after you agree to those terms. Do not share the data with anyone who has not agreed to them.

Videos (`videos/`)

FaceForensics++ has no fixed test set, so DeepGuard splits the videos 8:2 into train and test with train.json / test.json. videos/{split}/{source}.zip holds {source}/{d_vid}.mp4; the video-level metadata is raw/train_metadata.csv and raw/test_metadata.csv.

train

Label	Source	Videos
REAL	original	796
FAKE	Deepfakes	796
FAKE	Face2Face	798
FAKE	FaceShifter	810
FAKE	FaceSwap	793
FAKE	NeuralTextures	807
Total		4,800

test

Label	Source	Videos
REAL	original	204
FAKE	Deepfakes	204
FAKE	Face2Face	202
FAKE	FaceShifter	190
FAKE	FaceSwap	207
FAKE	NeuralTextures	193
Total		1,200

Column	Description
`label`	`REAL` or `FAKE`
`frame_cnt`	Total number of frames in the video
`d_vid`	Video file name within its method folder
`source`	Manipulation method, or `original`
`ori_vid`	Source real video ID used for the manipulation
`ori_frame_cnt`	Total number of frames in the source real video
`vid`	Unique video ID with method prefix (`0_` Deepfakes, `1_` Face2Face, `2_` FaceShifter, `3_` FaceSwap, `4_` NeuralTextures, none for original)

Preprocessed Frames (`frames/`)

Only the train videos were preprocessed. DeepGuard runs face detection on the fly for test videos at inference time, so the test split has no frames.

Label	Source	Videos	Frames
REAL	original	794	15,847
FAKE	Deepfakes	622	12,422
FAKE	Face2Face	643	12,685
FAKE	FaceShifter	644	12,847
FAKE	FaceSwap	629	12,416
FAKE	NeuralTextures	649	12,808
Total		3,981	79,025

Preprocessing

Detect original face — faces are detected only on real videos with yolov8n-face; fake videos reuse the boxes of their source real video, which cuts detection work by about 80%. Frames are rescaled by their longest side before detection (< 300px: x2.0, 300–700px: x1.0, 700–1500px: x0.5, > 1500px: x0.33). 20 frames per video, random frame offset 5, confidence threshold 0.5, minimum face area ratio 0.01.
Crop face & extract landmarks — faces are cropped with a 0.2 margin (no jitter) and 5 facial landmarks (eyes, nose, mouth corners) are detected on each crop. Only frames with detected landmarks are included.

Fields

Field	Description
`image`	Face crop (PNG)
`landmarks`	5 facial landmarks `[[x, y], ...]` in crop coordinates
`box`	Face box `[xmin, ymin, xmax, ymax]` in the original frame
`label`	`REAL` (0) or `FAKE` (1)
`source`	Manipulation method, or `original`
`vid`	Video ID with method prefix
`ori_vid`	Source real video ID
`frame_idx`	Frame index in the video

Usage

from datasets import load_dataset

frames = load_dataset("KoreaPeter/faceforensics-c23", split="train")

Restore the DeepGuard `DATA_ROOT` layout

raw/ holds the original preprocessing outputs (crops.tar, boxes.tar, landmarks.tar), the metadata CSVs (train_metadata.csv, test_metadata.csv, train_frame_metadata.csv) and the Kaggle CSVs (csv/).

hf download KoreaPeter/faceforensics-c23 --repo-type dataset --include "raw/*" "videos/*" --local-dir DATA_ROOT
cd DATA_ROOT
for f in raw/*.tar; do tar -xf "$f"; done
mv raw/*.csv . && mv raw/csv .
for f in videos/*/*.zip; do unzip -q "$f"; done

DATA_ROOT/
├── Deepfakes/ … original/{d_vid}.mp4
├── boxes/{ori_vid}.json
├── crops/{vid}/{frame_idx}.png
├── landmarks/{vid}/{frame_idx}.npy
├── train_metadata.csv
├── test_metadata.csv
└── train_frame_metadata.csv

DeepGuard Models & Demos

Models (MS-EffGCViT) trained with DeepGuard, also available as transformers repos:

Model	FaceForensics++	Celeb-DF-v2	KoDF
ms_eff_gcvit_b0	link	link	link
ms_eff_gcvit_b5	link	link	link

Spaces:

Citation

@inproceedings{roessler2019faceforensicspp,
  author    = {Andreas R\"ossler and Davide Cozzolino and Luisa Verdoliva and Christian Riess and Justus Thies and Matthias Nie{\ss}ner},
  title     = {Face{F}orensics++: Learning to Detect Manipulated Facial Images},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year      = {2019}
}

Downloads last month: 29

Total file size:

30.5 GB

Paper for KoreaPeter/faceforensics-c23

FaceForensics++: Learning to Detect Manipulated Facial Images

Paper • 1901.08971 • Published Jan 25, 2019 • 1

Access to FaceForensicsC23