viewer: false
tags:
- deepfakes
- gen-ai
- text-to-video
pretty_name: DeepAction Dataset v1.0
size_categories:
- 1K<n<10K
task_categories:
- video-classification
The DeepAction dataset contains over 3,000 videos generated by six text-to-video AI models, along with real videos matched in terms of the action depicted. These videos show people performing ordinary actions such as walking, running, and cooking. The AI models used to generate these videos include, in alphabetic order, AnimateDiff, CogVideoX5B, Pexels, RunwayML, StableDiffusion, Veo (pre-release version), and VideoPoet. Refer to our our pre-print for details.
Getting Started
To get started, log into Hugging Face in your CLI environment, and run:
from datasets import load_dataset
dataset = load_dataset("faridlab/deepaction_v1", trust_remote_code=True)
Data
The data is structured into seven folders, with six folders corresponding to text-to-video AI models and one folder for real videos. Each of these folders has 100 subfolders corresponding to human action classes. All videos in a given subfolder were generated using the same prompt (see the list of prompts here).
Included below are example videos generated using the prompt "a person taking a selfie". Note that, since each text-to-video AI model generates videos with different ratios and resolutions, these videos were normalized 512x512.
Real |
AnimateDiff |
CogVideoX5B |
RunwayML |
StableDiffusion |
Veo (pre-release version) |
VideoPoet |
Licensing
The AI-generated videos (BDAnimateDiffLightning, CogVideoX5B, RunwayML, StableDiffusion, Veo, and VideoPoet folders) are released under the CC BY 4.0 license. The real videos (Pexels folder) are released under the Pexels license.
Misc
Please use the following citation when referring to this dataset:
@misc{bohacek2024human,
title={Human Action CLIPS: Detecting AI-generated Human Motion},
author={Matyas Bohacek and Hany Farid},
year={2024},
eprint={2412.00526},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2412.00526},
}
This work was done during the first author's (Matyas Bohacek) internship at Google.