File size: 5,459 Bytes
10a0f39 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 |
---
task_categories:
- video-text-to-text
license: cc-by-4.0
language:
- en
tags:
- video-detection
- ai-generated-content
- explainable-ai
- multimodal
---
# ViF-CoT-4K Dataset
This repository hosts the **ViF-CoT-4K** dataset, a key component of the [Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning](https://huggingface.co/papers/2512.15693) paper. Skyra is a specialized multimodal large language model (MLLM) designed to identify human-perceivable visual artifacts in AI-generated videos, leveraging them as grounded evidence for both detection and explanation.
- **Paper**: [Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning](https://huggingface.co/papers/2512.15693)
- **Project Page**: https://joeleelyf.github.io/Skyra/
- **Code**: https://github.com/JoeLeelyf/Skyra
## Introduction
The misuse of AI-driven video generation technologies has raised serious social concerns, highlighting the urgent need for reliable AI-generated video detectors. Most existing methods are limited to binary classification and lack the necessary explanations for human interpretation. **ViF-CoT-4K** addresses this by providing a specialized dataset to train multimodal large language models (MLLMs) to identify human-perceivable visual artifacts in AI-generated videos and leverage them as grounded evidence for both detection and explanation.
ViF-CoT-4K represents the first large-scale AI-generated video artifact dataset with fine-grained human annotations, supporting the development of models capable of spatio-temporal artifact perception, explanation capability, and detection accuracy.
### Hierarchical Artifact Taxonomy
The dataset defines a comprehensive taxonomy to categorize AI generation errors, dividing them into **Low-level Forgery** (e.g., texture/color anomalies) and **Violation of Laws** (e.g., physical inconsistencies).
<p align="center">
<img src="https://github.com/JoeLeelyf/Skyra/raw/main/static/images/taxonomy.png" alt="Taxonomy of Artifacts" width="60%">
</p>
## Dataset: ViF-CoT-4K
**ViF-CoT-4K** is constructed to address the lack of detailed artifact annotations in existing datasets.
- **Scale**: ~4,000 videos, including high-quality samples from **Sora-2, Wan2.1, Kling**, and more.
- **Annotation**: Fine-grained labels including artifact type, textual explanation, timestamps, and bounding boxes.
- **Real-Fake Pairs**: Generated videos are semantically aligned with real counterparts to prevent shortcut learning.
<p align="center">
<img src="https://github.com/JoeLeelyf/Skyra/raw/main/static/images/statistics.png" alt="Dataset Statistics" width="90%">
</p>
## Usage
### Requirements
- **SFT Stage**: follow [LlaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) for environment setup.
- **RL Stage**: follow [verl](https://github.com/volcengine/verl) for environment setup.
- **Inference**: follow [Qwen-2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) for quick start and [vLLM](https://github.com/vllm-project/vllm) for deployment.
### Data Preparation
- Training data: Download and prepare the **ViF-CoT-4K** dataset from [here](https://huggingface.co/datasets/JoeLeelyf/ViF-CoT-4K).
- Evaluation data: Download evaluation datasets (e.g., **ViF-Bench**) from [here](https://huggingface.co/datasets/JoeLeelyf/ViF-Bench). And modify the path to your local directory in `test_index.json`.
The `test_index.json` file should contain the following format:
```json
{
"Real": [
"path_to_parsed_frames_dir/Real/gdymHI9S6gM-0",
...
],
"LTX-Video-13B-T": [
"path_to_parsed_frames_dir/Fake/LTX-Video-13B-T/gdymHI9S6gM-0",
...
],
...
}
```
### Supervised Fine-Tuning (SFT)
We use LLaMA-Factory for SFT. You can start training after setup the dataset config following the instructions in the LLaMA-Factory repository.
```bash
cd train/LLaMA-Factory
bash train.sh
```
### Reinforcement Learning (RL)
We use verl for RL training with GRPO, with adapted reward design provided in `train/verl/verl/utils/reward_score/ladm.py`.
### Evaluation
Evaluate scripts are provided in the `eval/` directory. You can run the evaluation script as follows:
- inference: Run inference to get model predictions and explanations, save the results in a JSON file.
```bash
cd eval
bash scripts/Skyra/inference.sh
# or
python inference.py \
--index_json /path_to/test_index.json \
--model_path /path_to/Skyra-SFT \
--model_name Skyra-SFT \
--save_dir results/Skyra
```
- evaluation: Evaluate the model predictions against ground truth and compute metrics.
```bash
cd eval
bash scripts/Skyra/eval.sh
# or
python eval.py \
--json_file_path results/Skyra/Skyra-SFT_predictions.json
```
## License
The **ViF-CoT-4K** dataset and **Skyra** model weights are released under the **CC BY 4.0** license. Users must adhere to the terms of source datasets (Kinetics-400, Panda-70M, HD-VILA-100M).
## Citation
If you find Skyra or ViF-CoT-4K useful, please cite our paper:
```bibtex
@misc{li2025skyraaigeneratedvideodetection,
title={Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning},
author={Yifei Li and Wenzhao Zheng and Yanran Zhang and Runze Sun and Yu Zheng and Lei Chen and Jie Zhou and Jiwen Lu},
year={2025},
eprint={2512.15693},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2512.15693},
}
``` |