creative-graphic-design/AesEvalBench

Dataset Card for AesEval-Bench

Dataset Summary

AesEval-Bench is a benchmark for evaluating whether vision-language models can assess graphic design aesthetics. The benchmark contains perturbed graphic design samples with preview images, highlighted previews, element-level metadata, and labels for aesthetic issues.

The benchmark spans four dimensions and twelve indicators:

Dimension	Indicators
Layout	balance, layering, whitespace, alignment
Typography	legibility, hierarchy
Graphics	quality, relevance
Color	harmony, contrast, appeal, psychology

The original evaluation code supports binary aesthetic judgment, problematic-region choice, and bounding-box localization. This Hugging Face loader exposes the benchmark samples and labels in a single test split.

Languages

The metadata text is in English.

Dataset Structure

Data Instances

Each row corresponds to one {sample_id}-perturbs_new directory from the released benchmark_data/ archive.

{
  "sample_name": "5888a55a95a7a863ddcc1d1d_3-perturbs_new",
  "source_id": "5888a55a95a7a863ddcc1d1d",
  "perturbation_id": 3,
  "canvas_width": 1200,
  "canvas_height": 600,
  "title": "Citation about volunteer work",
  "preview": "<image>",
  "preview_highlight": "<image>",
  "task_labels": [
    {
      "dimension": "layout",
      "task": "balance",
      "key": "layout-balance",
      "has_issue": false
    }
  ],
  "gt_annotations": [
    {
      "dimension": "graphic",
      "task": "quality",
      "key": "graphic-quality",
      "element_index": 5,
      "attribute": "Shape",
      "left": 458.0,
      "top": 410.0,
      "width": 282.0,
      "height": 65.0
    }
  ]
}

Data Fields

• sample_name: Directory name in the original benchmark release.
• source_id: Base design identifier parsed from sample_name.
• perturbation_id: Perturbation index parsed from sample_name.
• canvas_width, canvas_height: Canvas size.
• title, category, keywords, industries: Design metadata from meta_info.json.
• preview: Perturbed preview image.
• preview_highlight: Preview image with highlighted regions.
• element_images: Variable-length list of per-element PNG assets.
• elements: Simplified element metadata from simplified_meta_info.json.
• task_labels: One binary label for each of the 12 aesthetic indicators.
• gt_annotations: Positive issue annotations with element index, attribute, and bounding box.
• gt_json: Original GT.json serialized as JSON text.
• changes_json: Original changes.json serialized as JSON text.
• meta_info_json: Original meta_info.json serialized as JSON text.
• simplified_meta_info_json: Original simplified_meta_info.json serialized as JSON text.

Data Splits

Split	Rows
test	1,198

Dataset Creation

Source Data

The original repository distributes the benchmark as a Google Drive archive. The expected extracted layout is:

benchmark_data/
`-- {sample_id}-perturbs_new/
    |-- preview.png
    |-- preview_highlight.png
    |-- simplified_meta_info.json
    |-- GT.json
    |-- changes.json
    |-- 0.png, 1.png, ...
    `-- meta_info.json

This loader downloads the archive with gdown when no local data_dir is provided. For local testing, pass either the downloaded ZIP file or an extracted benchmark_data/ directory through data_dir.

Additional Information

Licensing Information

The source repository did not declare a license at the time this dataset loader was added, so the dataset license is marked as unknown.

Citation Information

@misc{an2026canvisionlanguagemodelsassess,
  title={Can Vision Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective},
  author={An, Arctanx and Sun, Shizhao and Huang, Danqing and Cheng, Mingxi and Gao, Yan and Li, Ji and Qiao, Yu and Bian, Jiang},
  year={2026},
  eprint={2603.01083},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.01083}
}

Contributions

Thanks to @arctanxarc and the AesEval-Bench authors for creating this dataset.