Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
147
You are an expert evaluator for text-to-video alignment. Your task is to judge how well a generated video matches a given text description.
### Step 1: Analyze
First, carefully read the provided text description and identify its key visual, temporal, and emotional elements. Focus on details such as:
- Scene (location, background, lighting)
- Characters (gender, appearance, clothing, expressions)
- Actions or movements
- Atmosphere or emotions
### Step 2: Compare
Then, analyze the content of the provided video (as described or observed) and reason step-by-step whether it aligns with each element in the text.
### Step 3: Rate
Based on your reasoning, give a score from **1 to 5**, where:
- **5** = Perfectly matches the text description in all major aspects.
- **4** = Mostly matches, with only minor deviations or missing details.
- **3** = Partially matches, some elements missing or inaccurate.
- **2** = Weak match, only a few aspects are correct.
- **1** = Does not match the description at all.
### Step 4: Output
Summarize your reasoning and provide the final score with format: "Score: <score>"
---
**Text Description:**
{video_text}
**Your Evaluation:**
(1) Key elements identified from text
(2) Step-by-step comparison reasoning
(3) Final alignment score (1–5)

Soul


🤓 Project    | 📑 Paper    | 🤖 Online Experience    | 🤖 API Documentation    | 🤗 Soul Model | Eval Suite   | 🤗 Soul-Bench | Results  


Soul: Breathe Life into Digital Human for High-fidelity Long-term Multimodal Animation

🎋 Click ↓ to watch brief introduction for Soul, Soul-1M, and Soul-Bench

TODO

  • Release evaluation tool for Soul-Bench.
  • Release inference code.
  • Release training code.

Inference (Soul Model)

It will be released soon.

Soul Bench

pip install "huggingface_hub[cli]"
huggingface-cli download --repo-type dataset APRIL-AIGC/Soul-Bench Soul-Bench/ --local-dir ./Soul-Bench --resume-download [--token hf_xxx]

SOTA Results on Soul Bench

pip install "huggingface_hub[cli]"
huggingface-cli download --repo-type dataset APRIL-AIGC/Soul-Bench Soul_Results/ --local-dir ./Soul_Results --resume-download [--token hf_xxx]

Evaluate your model on Soul Bench

We provide evaluation tools; please refer to subfolder evaluation tool.

License Agreement

Commercialization: Note that our license is non-commercial. If commercialization is required, please use Tencent Cloud Video Creation Large Model: Online Experience / API Documentation

Acknowledgements

We would like to thank the contributors to the Wan2.1, Wan2.2, Qwen, umt5-xxl, diffusers and HuggingFace repositories, for their open researches.

Citation

If you find our work helpful, please cite us.

@misc{soul,
    title={Soul: Breathe Life into Digital Human for High-fidelity Long-term Multimodal Animation}, 
    author={Jiangning Zhang and Junwei Zhu and Zhenye Gan and Donghao Luo and Chuming Lin and Feifan Xu and Xu Peng and Jianlong Hu and Yuansen Liu and Yijia Hong and Weijian Cao and Han Feng and Xu Chen and Chencan Fu and Keke He and Xiaobin Hu and Chengjie Wang},
    year={2025},
    eprint={2512.13495},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2512.13495}, 
}
Downloads last month
297