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.

👉 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).

This project is released under the Apache 2.0 license.

• [email protected]