Dataset Card for Dataset Name

Official Dataset for Hanfu-Bench: A Multimodal Benchmark on Cross-Temporal Cultural Understanding and Transcreation

Dataset License Statement

This dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.

Core License Terms

Attribution (BY): Users must give appropriate credit to the original author(s) of the dataset.
NonCommercial (NC): This dataset is restricted to non-commercial use. Any use for commercial purposes is strictly prohibited.
ShareAlike (SA): Any derivative works based on this dataset must be distributed under the same CC BY-NC-SA 4.0 license.

Additional Restrictions

To further clarify usage, this dataset is strictly limited to academic research purposes, defined as:

Academic purposes: Activities related to education, academic research, and scholarly communication, including but not limited to writing academic papers, evaluating models, and conducting experiments.
Evaluation Only: The data may only be used for evaluation purposes and not for training models or systems.
Non-academic non-commercial activities, such as personal projects or non-academic nonprofit initiatives, are not permitted.
All users must ensure compliance with relevant institutional and national laws and regulations during usage.

Dataset Description

We introduce Hanfu-Bench, a novel, expert-curated multimodal dataset. In total, 496 sets of Hanfu are retained, comprising 1,192 images, with an average of 4.74 identified features per Hanfu. This dataset includes some figures (This dataset retains some images that were filtered out after annotation.)

Hanfu-Bench comprises two core tasks: cultural visual understanding and cultural image transcreation. The former task examines temporal-cultural feature recognition based on single- or multi-image inputs through multiple-choice visual question answering, while the latter focuses on transforming traditional attire into modern designs through cultural element inheritance and modern context adaptation. Our evaluation shows that closed VLMs perform comparably to non-experts on visual cutural understanding but fall short by 10% to human experts, while open VLMs lags further behind non-experts. For the transcreation task, multi-faceted human evaluation indicates that the best-performing model achieves a success rate of only 42%. Our benchmark provides an essential testbed, revealing significant challenges in this new direction of temporal cultural understanding and creative adaptation.

Overview of the dataset.

meta-info.json

Dataset({
  "1000": {
    "source": "TaoBao",
    "img_list": [
      "num1000_img1.jpg"
    ],
    "gender": "female",
    "type": "传统汉服形制",
    "period": "唐朝",
    "xiu": "unsure",
    "jin": "对襟",
    "ling": "直领",
    "bottoms": "unsure",
    "outerwear": "披帛"
  },
})

merged-sivqa.json

Dataset([
  {
    "question_id": "single_0",
    "question_type": "gender",
    "cloth_id": "1000",
    "img_list": [
      "num1000_img1.jpg"
    ],
    "base_question": "图片中的服饰通常适合什么性别？",
    "base_question_en": "Which gender is the clothing in this image typically suitable for, among the following options?",
    "choices": "A.男; B.女",
    "choices_en": "A.Male; B.Female",
    "answer": "B"
  },
])

merged-mivqa.json

Dataset([
  {
    "question_meta": {
      "answer": {
        "image": "num1080_img4.jpg",
        "meta": {
          "gender": "female",
          "type": "汉元素服饰",
          "style": "unsure",
          "period": "unsure",
          "xiu": "unsure",
          "jin": "对襟",
          "ling": "圆领",
          "bottoms": "unsure",
          "outerwear": ""
        }
      },
      "candidates": [
        {
          "image": "num166_img10.jpg",
          "meta": {
            "gender": "female",
            "type": "汉服改良版",
            "style": "TwoPiece",
            "period": "unsure",
            "xiu": "unsure",
            "jin": "对襟",
            "ling": "unsure",
            "bottoms": "马面裙",
            "outerwear": ""
          }
        },
        {
          "image": "num1169_img5.jpg",
          "meta": {
            "gender": "male",
            "type": "传统汉服形制",
            "style": "unsure",
            "period": "unsure",
            "xiu": "unsure",
            "jin": "大襟",
            "ling": "unsure",
            "bottoms": "裤",
            "outerwear": ""
          }
        },
        {
          "image": "num1310_img1.jpg",
          "meta": {
            "gender": "male",
            "type": "传统汉服形制",
            "style": "TwoPiece",
            "period": "unsure",
            "xiu": "unsure",
            "jin": "unsure",
            "ling": "unsure",
            "bottoms": "unsure",
            "outerwear": "甲胄"
          }
        }
      ],
      "answer_img": "num1080_img4.jpg",
      "candidate_imgs": [
        "num166_img10.jpg",
        "num1169_img5.jpg",
        "num1310_img1.jpg"
      ],
      "question": "以下图片中的服饰属于汉元素服饰的是？",
      "question_type": "type",
      "question_formular": "type_t1"
    },
    "question": "以下图片中的服饰属于汉元素服饰的是？",
    "options": [
      "num1080_img4.jpg",
      "num166_img10.jpg",
      "num1169_img5.jpg",
      "num1310_img1.jpg"
    ],
    "answer_idx": 0,
    "qid": "mivqa_0",
    "question_id": "multi_0",
    "question_type": "type"
  },
])

Citation

@misc{zhou2025hanfubenchmultimodalbenchmarkcrosstemporal,
      title={Hanfu-Bench: A Multimodal Benchmark on Cross-Temporal Cultural Understanding and Transcreation}, 
      author={Li Zhou and Lutong Yu and Dongchu Xie and Shaohuan Cheng and Wenyan Li and Haizhou Li},
      year={2025},
      eprint={2506.01565},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.01565}, 
}

Dataset Card Contact

If you have any questions regarding the dataset, please reach out to: lizhou21@cuhk.edu.cn

Datasets:

lizhou21
/

hanfu-bench

You need to agree to share your contact information to access this dataset