This repository provides a reproducible pipeline to compute the Video Fidelity Score (VFS) for a single real and a single fake video.
The method extracts facial crops, computes NR-IQA metrics (BRISQUE + deep IQA models), and summarizes the distributional difference (effect size) between real and fake to yield VFS.

Intuition: Lower VFS values indicate higher visual quality, with VFS ≈ 0 suggesting the fake video exhibits no visible difference from the pristine video.

Figure 1: Illustration of the proposed method. The VFS score is calculated by measuring the distributional difference in NR-IQA scores between fake and real videos.

The pipeline follows three major steps:

1. Pre-processing – Extract face crops from both videos using a dlib frontal face detector with padding and resize.
2. Quality Assessment – Compute NR-IQA metrics: BRISQUE and deep models (LIQE, MANIQA, MUSIQ, DBCNN, TReS).
3. VFS Computation – Compute standardized effect size (|Cohen’s d|) per metric to quantify perceptual similarity.

The example real and fake videos used in this repository are derived from the FaceForensics++ dataset.

Citation:
Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., & Nießner, M. (2019).
FaceForensics++: Learning to detect manipulated facial images.
Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 1-11.
https://github.com/ondyari/FaceForensics

All visual examples and cropped faces included here are used only for research and educational purposes in accordance with the dataset’s original license and terms of use.

3rd-party dataset DOI/URL: https://github.com/ondyari/FaceForensics

To install the Python dependencies:

pip install -r requirements.txt

🛠 Computing Infrastructure

Experiments were tested on:

• Windows 10 (x64) and Ubuntu 24.04 LTS
• MATLAB R2024a, Python 3.8 (Anaconda environment)
• Hardware: Intel i7 CPU or higher, 8 GB RAM minimum, optional NVIDIA GPU (e.g., RTX 1080 Ti)

Software dependencies

All codes in this repository are executed from a single MATLAB entry point (main.m). The MATLAB script automatically calls the Python scripts for deep NR-IQA metrics, so users do not need to run Python separately.

Once the paths to the Python environment (pythonPath) and metric folder (pyScriptsDir) are correctly set, running main.m will automatically execute both MATLAB- and Python-based components, generate the corresponding metric scores, and produce the final VFS output files.

1. Clone this repo and prepare your environment (MATLAB + Python).
2. Place one real and one fake video.
3. Run the main.m .

addpath('path/to/codes/folder');

realVideo = 'example_videos/real/video125.mp4';
fakeVideo = 'example_videos/fake/video125_038.mp4';
outDir    = 'results/';

pythonPath   = 'path/to/anaconda/envs/matlab_env/python';
pyScriptsDir = 'path/to/vfs/python/metrics';

metrics       = {'BRISQUE','LIQE','MANIQA','MUSIQ','DBCNN','TRES'};
functionTypes = {'matlab','python','python','python','python','python'};

pair_run(realVideo, fakeVideo, outDir, pythonPath, pyScriptsDir, metrics, functionTypes);

The output folder will contain real_iqm.csv, fake_iqm.csv, and vfs_summary.csv.

The provided demo pair (real and fake videos) is a subset of FaceForensics++ dataset.

⚠️ Note: The example data in example_videos/ originates from FaceForensics++ and is subject to its dataset license.

Figure 2: Facial regions from a sample frame of a real video125 and its manipulated versions, with VFS scores computed via MANIQA.

Code License: MIT License
Dataset License: FaceForensics++ dataset license (see link above)