ChartGalaxy is a million-scale dataset of synthetic and real infographic charts with data tables, supporting applications in infographic chart understanding, code generation, and chart generation. The dataset addresses the challenge that existing datasets are mostly limited to plain charts, failing to capture the diverse range of design styles and layouts that are key characteristics of infographic charts.

2026.02.05: 🎉🎉 A new batch of data has been added, comprising 108,208 infographic charts. This update features broader diversity in title designs and more polished layouts, improving overall readability.

2025.12.15: 🎉🎉 A new batch of data has been added on Hugging Face 🤗, comprising 19,458 infographic charts. This update includes more diverse title designs and fixes for layout issues that might cause ambiguity.

👉 Access the full ChartGalaxy dataset on Hugging Face 🤗! 👈

Due to storage limitations, we only include a collection of representative examples below (and more in examples).

• Size: 1,763,189 infographic charts (1,701,356 synthetic + 61,833 real)
• Content: Each infographic chart is paired with the tabular data used to create it
• Chart Types: 75 chart types with 440 chart variations
• Layout Templates: 68 layout templates

ChartGalaxy was constructed through:

1. Real Infographic Chart Collection: Charts were collected from 18 reputable chart-rich websites, such as Visual Capitalist and Statista.

2. Synthetic Infographic Chart Creation: Following an inductive structuring process that:

   • Identifies 75 chart types (e.g., bar charts) and 440 chart variations reflecting different visual element styles
   • Extracts 68 layout templates defining spatial relationships among elements
   • Programmatically generates synthetic charts based on these patterns

The utility of ChartGalaxy is demonstrated through three representative applications:

Fine-tuning on ChartGalaxy improves the performance of foundation models on infographic chart understanding.

We provide the code and data for the evaluation introduced in our paper (code, data).

A benchmark for assessing LVLMs' code generation for infographic charts. The benchmark evaluates the similarity between charts rendered by the generated D3.js code and ground-truth ones at two levels of granularity: high-level (overall visual similarity) and low-level (average similarity across fine-grained SVG elements).

We provide the code and data for the constructed benchmark introduced in our paper (benchmark).

An example-based method that transforms user-provided tabular data into an infographic chart, aligning with the layout and visual style of a given example chart. User studies show this method outperforms GPT-Image-1 on fidelity, aesthetics, and creativity.

We provide the code for the example-based infographic chart generation introduced in our paper (code).

ChartGalaxy can also support a broad range of related tasks, including:

• infographic chart visual question answering
• chart-to-table / data extraction
• infographic chart generation
• layout-aware generation and editing
• retrieval-based infographic chart generation

If you find ChartGalaxy useful in your research, please cite our paper:

@misc{li2025chartgalaxy,
      title={ChartGalaxy: A Dataset for Infographic Chart Understanding and Generation}, 
      author={Zhen Li and Duan Li and Yukai Guo and Xinyuan Guo and Bowen Li and Lanxi Xiao and Shenyu Qiao and Jiashu Chen and Zijian Wu and Hui Zhang and Xinhuan Shu and Shixia Liu},
      year={2025},
      eprint={2505.18668},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.18668},
}

This project is released under the Apache 2.0 license.

• chartgalaxy@163.com