HAVIC

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 33 Open GitHub ↗

[CVPR'26 Findings] Leave No Stone Unturned: Uncovering Holistic Audio-Visual Intrinsic Coherence for Deepfake Detection

Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<div align="center">
  <h2><b> Leave No Stone Unturned: Uncovering Holistic Audio-Visual Intrinsic Coherence for Deepfake Detection </b></h2>
</div>

</div>

## News

- **[Dec 2025]** Pre-training, finetuning, evaluation and inference code are released.
- **[Feb 2026]** Our paper is accepted at CVPR 2026 Findings.
- **[Mar 2026]** Our paper is available on arXiv: [https://arxiv.org/abs/2603.23960](https://arxiv.org/abs/2603.23960).
- **[Apr 2026]** Model weights are provided [here](https://huggingface.co/JielunPeng/HAVIC/).
- **[May 2026]** HiFi-AVDF dataset is provided [here](https://huggingface.co/datasets/JielunPeng/HiFi-AVDF).

## Overview
This repository contains the official implementation of [Leave No Stone Unturned: Uncovering Holistic Audio-Visual Intrinsic Coherence for Deepfake Detection](https://arxiv.org/abs/2603.23960). The proposed **HiFi-AVDF** dataset is available [here](https://huggingface.co/datasets/JielunPeng/HiFi-AVDF).

![alt text](https://raw.githubusercontent.com/tuffy-studio/HAVIC/main/assets/poster.png)

## 1 Requirements

### Create a conda environment and activate it

```bash
conda create -n HAVIC python=3.10 ffmpeg
conda activate HAVIC
```
### Install Python dependencies

```bash
pip install -r requirements.txt
```

<!-- > ❗**Note:** If you are in mainland China, you can use the following commands to speed up Python package installation:

```bash
pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/
``` -->

## 2 Dataset Download

We pre-train the HAVIC on the **LRS2** dataset, which contains only real videos. Please download LRS2 [here](https://www.robots.ox.ac.uk/~vgg/data/lip_reading/lrs2.html). Due to copyright we cannot release the data.

We use the **FakeAVCeleb** dataset to finetune the model and evaluate intra-dataset performance. Please follow the instructions on their [official site](https://sites.google.com/view/fakeavcelebdash-lab/) to request and download the dataset.

To further evaluate the cross-dataset generalization of the model, we use **KoDF** and **HiFi-AVDF** dataset:

* **KoDF official page**: [KoDF](https://deepbrainai-research.github.io/kodf/)
* **HiFi-AVDF download page**: [HiFi-AVDF](https://huggingface.co/datasets/JielunPeng/HiFi-AVDF)

## 3 Training

### 3.1 Pre-training

#### Step 1: Data Preprocessing
We perform preprocessing on LRS2 dataset to crop the face regions:
```bash
# Make sure you are in the project root directory.
cd ./video_data_engine/
python preprocess_pt_dataset.py \
    --train_set_dir <path to the training set root> \
    --test_set_dir <path to the test set root> \
    --save_dir <the root directory to save processed videos and csvs>
```
"This step generates two CSV files with preprocessed training and testing data, saved under the specified `save_dir`. Each file contains five columns: `video_path, face_crop_folder, audio_label, visual_label, overall_label`. The labels will not be used.

#### Step 2: Initialize Weights for Pretraining

Following [AVFF](https://openaccess.thecvf.com/content/CVPR2024/html/Oorloff_AVFF_Audio-Visual_Feature_Fusion_for_Video_Deepfake_Detection_CVPR_2024_paper.html), We initialize the audio and visual encoder–decoders separately with the pretrained weights of **AudioMAE** and **MARLIN**. 
Please download the official pretrained [AudioMAE](https://drive.google.com/file/d/1ni_DV4dRf7GxM8k-Eirx71WP9Gg89wwu/view?usp=share_link) and [MARLIN](https://huggingface.co/ControlNet/MARLIN/blob/main/marlin_vit_base_ytf.full.pt) model weights and place them in the `weights/` folder. Then Run the following to get init weights:

```bash
# Make sure you are in the project root directory.
cd ./weights/
python initialize_pretrain_weights.py
```

This step will generate a `model_to_be_pt.pth` file, which contains the initialized model weights for pre-training.

#### Step 3: Start Pre-training

After the above steps, you can start the pre-training process by running the following command. Note that you need to configure several settings in the shell script `pretrain.sh` in advance, including the path of  **model saving directory**, etc.
```bash
# Make sure you are in the project root directory.
cd ./scripts/
bash pretrain.sh
```

**The pre-trained model weights are provided [here](https://huggingface.co/JielunPeng/HAVIC/).**

Additionally, since our pre-training method incorporates the **Masked Autoencoders (MAE)** to learn self-supervised representations, we implemented visualizations of the masked regions and the reconstructed outputs in [MAE_visualization.py](./MAE_visualization.py). Before running, make sure to configure the necessary settings in the shell script [mae_visual.sh](./mae_visual.sh), then you can run the shell script [mae_visual.sh](./mae_visual.sh) to perform visualization.

```bash
# Make sure you are in the project root directory.
bash mae_visual.sh
```

---

### 3.2 Finetuning

#### Step 1: Data Preprocessing

First, we randomly split the FakeAVCeleb dataset into **70% training** and **30% test** sets by running the following command:

```bash
# Make sure you are in the project root directory.
cd ./video_data_engine/
python split_FakeAVCeleb_dataset.py \
    --dataset_root <path to dataset root, e.g., /data/FakeAVCeleb_v1.2>
```
This will produce two CSV files under the `video_data_engine` directory: `training_set.csv` for the training split and `test_set.csv` for the test split. Each CSV file contains four columns:
`
video_path, audio_label, visual_label, overall_label
`
. The labels are automatically inferred from the video path.

Then, we perform preprocessing on the two splits to crop the face regions:

```bash
# Make sure you are in the project root directory.
cd ./video_data_engine/
python preprocess_ft_dataset.py \
    --training_set_csv <path to the training_set.csv file> \
    --test_set_csv <path to the test_set.csv file>
```
This step produces two new CSV files containing the preprocessed data for training and testing: `processed_training_set.csv` and `processed_test_set.csv`. Each file contains five columns: `video_path, face_crop_folder, audio_label, visual_label, overall_label`.

#### Step 2: Initialize Weights for Finetuning

After pre-training, the pre-trained weights need to be transferred to be loaded into the model for finetuning. Please put the pre-trained weights or the weights downloaded from our release into the `./weights` directory, then run the following command to obtain the initial weights for finetuning:

```bash
# Make sure you are in the project root directory.
python ./weights/initialize_finetune_weights.py
```
This step will generate a `model_to_be_ft.pth` file, which contains the initialized model weights for finetuning, and a `newly_added_modules.txt` file, which is a list recording the newly added modules that will be trained at a larger learning rate during the finetuning stage.

#### Step 3: Start Finetuning

After the above steps, you can start the finetuning process by running the following command. Note that you need to configure several settings in the shell script `finetune.sh` in advance, including the path of  **model saving directory**, etc.
```bash
# Make sure you are in the project root directory.
cd ./scripts/
bash finetune.sh
```

**The finetuned model weights are provided [here](https://huggingface.co/JielunPeng/HAVIC/).**

## 4 Evaluation and Inference
Before evaluation or inference, please prepare your fine-tuned model, or download the model provided by us.

To evaluate or run inference on videos, please first organize the input videos into a CSV file. You may use [our provided code](evaluation/make_label_csv.py) or prepare the CSV using your own method.

For evaluation, the CSV file should contain two columns: `video_path, overall_label`, where `video_path` is the absolute path to the video file, and `overall_label` indicates the ground-truth label of the sample. For inference, the CSV file should contain a single column: `video_path`.

>❗**Note:** No additional video pre-processing is required, as the entire video will be automatically processed using a sliding-window strategy during evaluation and inference, and the face detection module from FaceX-Zoo is integrated into the pipeline. During execution, a temporary directory named sliding_window_inference_tmp will be created in the current working directory to store intermediate files.

Then you can run evaluation or inference using the following commands:

```bash
# Make sure you are in the project root directory.
cd ./evaluation/
python sliding_window_infer.py \
    --csv_file_path <path_to_input_csv> \
    --save_csv_path <path_to_output_csv> \
    --finetune_path <path_to_finetune_weight> \
    --mode <evalution or inference>
```

For each input video, the model outputs a deepfake probability score, indicating the likelihood that the video is manipulated (1:fake; 0:real). The prediction results will be saved to  `save_csv_path`, where each row contains a video path, the ground-truth label (if in evaluation mode), and predicted probability.

## 5 Acknowledgement

We appreciate the following github repos for their valuable code and contributions:

- MARLIN (https://github.com/ControlNet/MARLIN)
- AudioMAE (https://github.com/facebookresearch/AudioMAE)
- OpenAVFF (https://github.com/JoeLeelyf/OpenAVFF)
- FaceX-Zoo (https://github.com/JDAI-CV/faceX-Zoo)

## 6 Contact

If you have any questions or concerns, please contact:

📧 **[jielunpeng.hit@gmail.com](mailto:jielunpeng.hit@gmail.com)**

📧 **[25s003052@stu.hit.edu.cn](25s003052@stu.hit.edu.cn)**

or feel free to submit an issue in this repository.

HAVIC

Download

Today's picks

Was this content helpful?

Recent Feedback

HAVIC

Download

Today's picks

Was this content helpful?

Recent Feedback

Download README