MVAD

# 版本问题已经解决，大家可以使用。训练集已上传完成，测试集还在上传中。
MVAD is the first general-purpose dataset specifically designed for detecting AI-generated multimodal video-audio content.
## :file_folder: Dataset Overview
### Statistics of multimodal video-audio data in the MVAD dataset:
| Video Source | Modality | Audio Generate | Length | Train |
|--------------------|----------|----------------------|----------|-------|
| HarmonySet | R-R | - | 10-60s | 20000 |
|Talkvid | R-R | - | 10-60s | 30000 |
| OpenVid-1M | R-F | FC&HY&MMA&AX | 1-10s | 8000 |
| InternVid-10M | R-F | FC&HY&MMA&AX | 1-10s | 8000 |
| MSR-VTT | R-F | FC&HY&MMA&AX | 1-10s | 8000 |
| JiMeng | F-F | FC&HY&MMA&AX | 5-10s | 4764 |
| KlingO1 | F-F | FC&HY&MMA&AX | 4s | 4400 |
| Sora2 | F-F | - | 5-10s | 5000 |
| Kling2.1 | F-F | - | 5-10s | 513 |
| Kling1.6 | F-F | - | 5-10s | 324 |
| Kling2.6 | F-F | - | 5-10s | 1902 |
| kling2.5Turbo | F-F | - | 5-10s | 297 |
| Humo | F-R | - | 3s | 8800 |
### MVAD Dataset Download
The previously released dataset has certain copyright issues, and we recommend discontinuing the use of the old version. We are currently accelerating the data screening and cleaning process, which is expected to be completed within two weeks. Once finished, we will release a new version of the dataset with higher quality and the most advanced audio and video content available.
### Demos of four multimodal video-audio data types
#### Fake Video - Fake Audio
[demo_animal](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_animal.mp4)
[demo_scene](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_scene.mp4)
[demo_object](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_object.mp4)
[demo_human](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_human.mp4)
#### Real Video - Fake Audio
[demo_scene](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Fake/real_fake_scene.mp4)
[demo_human](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Fake/real_fake_human.mp4)
#### Fake Video - Real Audio
[demo_animal](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_animal.mp4)
[demo_scene](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_scene.mp4)
[demo_object](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_object.mp4)
[demo_human](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_human.mp4)
#### Real Video - Real Audio
[demo_animal](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Real/real_real_animal.mp4)
[demo_human](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Real/real_real_human.mp4)

# MVAD：用于AIGC（Artificial Intelligence Generated Content）检测的多模态音视频综合数据集
MVAD是首个专为检测AI生成的多模态音视频内容打造的通用型数据集。
https://www.arxiv.org/abs/2512.00336

## 数据集概览
### MVAD数据集中的多模态音视频数据统计：
为贴合真实场景下多模态音视频内容的多样性，MVAD涵盖两类视觉领域（写实风格与动漫风格），并覆盖四大核心类别：人类、动物、物体与场景。本数据集包含三类音视频伪造类型与四种模态组合（假-假、假-真、真-假、真-真）。如表所示，MVAD共包含205,758条多模态音视频样本，由20余种不同生成方法制作而成，其中伪造样本104,578条，真实样本101,000条。遵循通用数据集设计规范，MVAD的伪造样本与真实样本比例保持1:1。各模态组合的样本分布如下：假-假（62,178条）、真-假（31,880条）、假-真（10,700条）、真-真（101,000条）。

| 视频来源 | 模态 | 音频生成方式 | 时长 | 训练集 | 测试集 | 单源样本数 | 总样本数 |
|-------------|----------|---------------|--------|-------|------|-------|-------------|
| Ugc-VideoCaptioner | R-R | - | 10-60s | - | 1,000 | | |
| HarmonySet | R-R | - | 10-60s | - | 30,000 | 31,000 | |
| TalkVid | R-R | - | 3s | 80,000 | - | 80,000 | 111,000 |
| Msvd | R-F | FC&HY&MMA&AX | 1-10s | - | 7,880 | | |
| OpenVid-1M | R-F | FC&HY&MMA&AX | 1-10s | - | 8,000 | 15,880 | |
| InternVid-10M | R-F | FC&HY&MMA&AX | 1-10s | 8,000 | - | | |
| MSR-VTT | R-F | FC&HY&MMA&AX | 1-10s | 8,000 | - | 16,000 | 31,880 |
| Sora | R-F | FC&HY&MMA&AX | 10-30s | - | 224 | | |
| Viva | F-F | FC&HY&MMA&AX | 2-5s | - | 3,980 | | |
| Vidu | F-F | FC&HY&MMA&AX | 3s | - | 2,908 | | |
| JiMeng | F-F | FC&HY&MMA&AX | 3s | - | 2,480 | | |
| Kling2.1 | F-F | - | 5s/10s | - | 513 | | |
| Kling2.5Turbo | F-F | - | 5s/10s | - | 249 | | |
| Sora2 | F-F | - | 10s | - | 996 | | |
| Voe3 | F-F | - | 8s-60s | - | 200 | 11,550 | |
| MoonValley | F-F | FC&HY&MMA&AX | 4s | 11,508 | - | | |
| pika | F-F | FC&HY&MMA&AX | 3s | 13,128 | - | | |
| Haiper | F-F | FC&HY&MMA&AX | 2s | 5,584 | - | | |
| Noisee | F-F | FC&HY&MMA&AX | 4s | 5,184 | - | | |
| Pixverse | F-F | FC&HY&MMA&AX | 4s | 8,628 | - | | |
| Emu3 | F-F | FC&HY&MMA&AX | 4s | 3,600 | - | | |
| Gen3 | F-F | FC&HY&MMA&AX | 10s | 2,668 | - | | |
| Kling1.6 | F-F | - | 5s/10s | 328 | - | 50,628 | 62,178 |
| Wan2.1 | F-R | - | 3s | - | 500 | 700 | |
| Kling-Avata | F-R | - | 3s | - | 200 | 700 | |
| Humo | F-R | - | 3s | 10,000 | - | 10,000 | 10,700 |
| **总样本数** | - | - | - | 176,628 | 59,130 | | 215,758 |

### MVAD数据集下载
本数据集即将正式发布。

### 四类多模态音视频数据样本演示
#### 假视频-假音频
[动物演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_animal.mp4)
[场景演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_scene.mp4)
[物体演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_object.mp4)
[人类演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Fake/fake_fake_human.mp4)
#### 真视频-假音频
[场景演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Fake/real_fake_scene.mp4)
[人类演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Fake/real_fake_human.mp4)
#### 假视频-真音频
[动物演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_animal.mp4)
[场景演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_scene.mp4)
[物体演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_object.mp4)
[人类演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Fake_Real/fake_real_human.mp4)
#### 真视频-真音频
[动物演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Real/real_real_animal.mp4)
[人类演示样例](https://github.com/HuMengXue0104/MVAD/blob/main/demos/Real_Real/real_real_human.mp4)