d3-deepfake-pipeline

D3: Training-Free AI-Generated Video Detection Using Second-Order Features

This repository contains the code for the paper "D3: Training-Free AI-Generated Video Detection Using Second-Order Features" (accepted at ICCV 2025) by Chende Zheng, Ruiqi Suo, Chenhao Lin, Zhengyu Zhao, Le Yang, Shuai Liu, Minghui Yang, Cong Wang, and Chao Shen.

Table of Contents

• Environment Setup
• Dataset Preparation and Preprocessing
• Inference
• Acknowledgement
• Citation

Environment Setup

Our codebase requires the following Python environment:

• Python >= 3.11.13

Follow these steps to set up the environment:

# Clone the repository
git clone https://github.com/Zig-HS/D3.git
cd D3

# Install required libraries
pip install -r requirements.txt

Dataset Preparation and Preprocessing

This section demonstrates dataset preparation using the GenVideo dataset as an example.

1. Create Dataset Directory Structure

mkdir GenVideo
cd GenVideo
mkdir video csv frames

2. Download and Organize Dataset

wget https://modelscope.cn/datasets/cccnju/Gen-Video/resolve/master/GenVideo-Val.zip
unzip GenVideo-Val.zip
mv GenVideo-Val/Real video/real_MSRVTT
mv GenVideo-Val/Fake/* video/
cd ..

3. Preprocessing Steps

Step 1: Frame Extraction Convert videos to frames using the video2frame utility:

python utils/video2frame.py --dataset-path GenVideo

Step 2: CSV Configuration Generation Generate CSV configuration files for the extracted frames:

python utils/folder2csv.py --is-real True --dataset-path GenVideo --folders real_MSRVTT
python utils/folder2csv.py --is-real False --dataset-path GenVideo --folders Crafter Gen2 HotShot Lavie ModelScope MoonValley MorphStudio Show_1 Sora WildScrape

Expected Dataset Structure

After proper processing, your dataset directory structure should look like this:

<dataset>/
├── video/
│   ├── <testsetA>/
│   │   ├── <video_id1>.mp4
│   │   ├── <video_id2>.mp4
│   │   └── ...
│   ├── <testsetB>/
│   │   ├── <video_idX>.mp4
│   │   └── ...
│   └── ... 
├── frames/
│   ├── <testsetA>/
│   │   ├── <video_id1>/
│   │   │   ├── 1.jpg
│   │   │   └── ...
│   │   └── ...
│   ├── <testsetB>/
│   │   └── ...
│   └── ... 
└── csv/
    ├── <testsetA>.csv
    ├── <testsetB>.csv
    └── ... 

For other datasets used in the paper (such as EvalCrafter, VideoPhy, and VidProM), you can download them from their official repositories and follow the same preprocessing steps described above.

Inference

After completing dataset preprocessing, run inference using eval.py:

python eval.py --gpu-id 0 --loss l2 --encoder XCLIP-16 --real-csv GenVideo/csv/real_MSRVTT.csv --fake-csv GenVideo/csv/Crafter.csv

Acknowledgement

Some of the design of our video processing and dataset structure is adopted from DeMamba. Thanks for their excellent work!

Citation

If you find this repository helpful, please consider citing it in your research:

@article{zheng2025d3,
  title={D3: Training-Free AI-Generated Video Detection Using Second-Order Features},
  author={Zheng, Chende and Lin, Chenhao and Zhao, Zhengyu and Yang, Le and Liu, Shuai and Yang, Minghui and Wang, Cong and Shen, Chao and others},
  journal={arXiv preprint arXiv:2508.00701},
  year={2025}
}

Updated

The original repo is updated for prediction task.

Ipynb Files

Upload the dataset(real and ai_generated) in the video folder. Run the "d3-evaluation-notebook-ipynb.ipynb" notebook directly for quick evaluation of the model and get the threshold for the prediction.

D3: Training-Free AI-Generated Video Detection Using Second-Order Features

This repository contains the code for the paper "D3: Training-Free AI-Generated Video Detection Using Second-Order Features" (accepted at ICCV 2025) by Chende Zheng, Ruiqi Suo, Chenhao Lin, Zhengyu Zhao, Le Yang, Shuai Liu, Minghui Yang, Cong Wang, and Chao Shen.

Table of Contents

• Environment Setup
• Dataset Preparation and Preprocessing
• Inference
• Acknowledgement
• Citation

Environment Setup

Our codebase requires the following Python environment:

• Python >= 3.11.13

Follow these steps to set up the environment:

# Clone the repository
git clone https://github.com/Zig-HS/D3.git
cd D3

# Install required libraries
pip install -r requirements.txt

Dataset Preparation and Preprocessing

This section demonstrates dataset preparation using the GenVideo dataset as an example.

1. Create Dataset Directory Structure

mkdir GenVideo
cd GenVideo
mkdir video csv frames

2. Download and Organize Dataset

wget https://modelscope.cn/datasets/cccnju/Gen-Video/resolve/master/GenVideo-Val.zip
unzip GenVideo-Val.zip
mv GenVideo-Val/Real video/real_MSRVTT
mv GenVideo-Val/Fake/* video/
cd ..

3. Preprocessing Steps

Step 1: Frame Extraction Convert videos to frames using the video2frame utility:

python utils/video2frame.py --dataset-path GenVideo

Step 2: CSV Configuration Generation Generate CSV configuration files for the extracted frames:

python utils/folder2csv.py --is-real True --dataset-path GenVideo --folders real_MSRVTT
python utils/folder2csv.py --is-real False --dataset-path GenVideo --folders Crafter Gen2 HotShot Lavie ModelScope MoonValley MorphStudio Show_1 Sora WildScrape

Expected Dataset Structure

After proper processing, your dataset directory structure should look like this:

<dataset>/
├── video/
│   ├── <testsetA>/
│   │   ├── <video_id1>.mp4
│   │   ├── <video_id2>.mp4
│   │   └── ...
│   ├── <testsetB>/
│   │   ├── <video_idX>.mp4
│   │   └── ...
│   └── ... 
├── frames/
│   ├── <testsetA>/
│   │   ├── <video_id1>/
│   │   │   ├── 1.jpg
│   │   │   └── ...
│   │   └── ...
│   ├── <testsetB>/
│   │   └── ...
│   └── ... 
└── csv/
    ├── <testsetA>.csv
    ├── <testsetB>.csv
    └── ... 

For other datasets used in the paper (such as EvalCrafter, VideoPhy, and VidProM), you can download them from their official repositories and follow the same preprocessing steps described above.

Inference

After completing dataset preprocessing, run inference using eval.py:

python eval.py --gpu-id 0 --loss l2 --encoder XCLIP-16 --real-csv GenVideo/csv/real_MSRVTT.csv --fake-csv GenVideo/csv/Crafter.csv

Acknowledgement

Some of the design of our video processing and dataset structure is adopted from DeMamba. Thanks for their excellent work!

Citation

If you find this repository helpful, please consider citing it in your research:

@article{zheng2025d3,
  title={D3: Training-Free AI-Generated Video Detection Using Second-Order Features},
  author={Zheng, Chende and Lin, Chenhao and Zhao, Zhengyu and Yang, Le and Liu, Shuai and Yang, Minghui and Wang, Cong and Shen, Chao and others},
  journal={arXiv preprint arXiv:2508.00701},
  year={2025}
}

Updated

The original repo is updated for prediction task.

Ipynb Files

Upload the dataset(real and ai_generated) in the video folder. Run the "d3-evaluation-notebook-ipynb.ipynb" notebook directly for quick evaluation of the model and get the threshold for the prediction. Run the "d3-prediction-notebook-ipynb.ipynb" notebook for prediction.