Magic Videos
收藏资源简介:
该数据集由哈尔滨工业大学(深圳)联合鹏城实验室等机构构建,包含来自15种前沿开源和商业视频生成模型的14万条视频样本,涵盖240p-768p分辨率及1-10秒时长。数据通过VBench平台、MovieGen专有模型及定制提示库合成,重点采集景观、建筑和人际交互等高危伪造场景。其创新性在于保留原生分辨率时空特征,解决了传统下采样导致的高频伪影丢失问题,为AI生成视频检测领域提供了首个跨模型、多分辨率的基准测试平台。
This dataset was developed by Harbin Institute of Technology (Shenzhen) in partnership with institutions including Peng Cheng Laboratory. It comprises 140,000 video samples sourced from 15 state-of-the-art open-source and commercial video generation models, with resolutions spanning 240p to 768p and durations ranging from 1 to 10 seconds. The data was synthesized through the VBench platform, the proprietary MovieGen model, and a custom prompt library, with a focus on high-risk forgery scenarios such as landscapes, architectural scenes, and interpersonal interactions. Its core innovation lies in preserving native-resolution spatiotemporal features, which resolves the problem of high-frequency artifact loss caused by traditional downsampling, thereby providing the first cross-model, multi-resolution benchmark platform for the field of AI-generated video detection.
Qwen2.5-ViT For AI-generated video detection 数据集概述
一、 数据集来源与用途
该数据集用于支持ICLR26论文《Preserving Forgery Artifacts: AI-Generated Video Detection at Native Scale》的研究,旨在进行AI生成视频的检测。
二、 训练数据集
- VBench sampled videos
- 来源地址:https://github.com/Vchitect/VBench/tree/master/sampled_videos
- Kinetics
- 来源地址:https://github.com/cvdfoundation/kinetics-dataset
三、 评估数据集
- Magic Videos (Ours)
- 来源地址:https://huggingface.co/datasets/mgiant/magic_videos
- GenVideo-Val
- 来源地址:https://modelscope.cn/datasets/cccnju/Gen-Video/files
- DVF
- 来源地址:https://github.com/SparkleXFantasy/MM-Det
- DeepTraceReward
- 来源地址:https://huggingface.co/datasets/DeeptraceReward/RewardData




