Home

Awesome

This is the official code of paper 'DeMamba: AI-Generated Video Detection on Million-Scale GenVideo Benchmark'.

:boom: News!!!

:dart: Todo

:file_folder: Dataset download

Data preparation process

Statistics of real and generated videos in the GenVideo dataset:

Video SourceTypeTaskTimeResolutionFPSLengthTraining SetTesting Set
Kinetics-400Real-17.05224-340-5-10s260,232-
Youku-mPLUGReal-23.07--10-120s953,279-
MSR-VTTReal-16.05--10-30s-10,000
ZeroScopeFakeT2V23.071024×57683s133,169-
I2VGen-XLFakeI2V23.121280×72082s61,975-
SVDFakeI2V23.121024×57684s149,026-
VideoCrafteFakeT2V24.011024×57682s39,485-
PikaFakeT2V&I2V24.021088×640243s98,377
DynamiCrafterFakeI2V24.031024×57683s46,205-
SDFakeT2V&I2V23-24512-102482-6s200,720-
SEINEFakeI2V24.041024×57682-4s24,737-
LatteFakeT2V24.03512×51282s149,979-
OpenSoraFakeT2V24.03512×51282s177,410-
ModelScopeFakeT2V23.03256×25684s-700
MorphStudioFakeT2V23.081280×72082s-700
MoonValleyFakeT2V24.011024×576163s-626
HotShotFakeT2V23.10672×38481s-700
Show_1FakeT2V23.10576×32084s-700
Gen2FakeI2V&T2V23.09896×512244s-1,380
CrafterFakeT2V23.04256×25684s-1,400
LavieFakeT2V23.091280×204882s-1,400
SoraFakeT2V24.02---60s-56
WildScrapeFakeT2V&I2V24512-10248-162-6s-926
Total Count------2,294,59419,588

:snake: Detail Mamba (DeMamba)

<p align="center"><em>In memory of Kobe Bryant (generated by GPT-4o)</em></p>

"Determination wins games, but Detail wins championships." — Kobe Bryant, in his Show Detail, 2018

<p align="center"><em>The overall framework of our Detail Mamba (DeMamba)</em></p>

:space_invader: Citing GenVideo&DeMamba

If you use GenVideo or DeMamba in your research or use the codebase here, please use the following BibTeX entry.

@article{DeMamba,
      title={DeMamba: AI-Generated Video Detection on Million-Scale GenVideo Benchmark},
      author={Haoxing Chen and Yan Hong and Zizheng Huang and Zhuoer Xu and Zhangxuan Gu and Yaohui Li and Jun Lan and Huijia Zhu and Jianfu Zhang and Weiqiang Wang and Huaxiong Li},
      journal={arXiv preprint arXiv:2405.19707},
      year={2024}
}

Acknowledgement

Many thanks to the nice work of STIL, CLIP, XCLIP, NPR and VideoMAE.

:email: Contact

If you have any questions, feel free to contact us: hx.chen@hotmail.com.