README.md

---
license: cc-by-nc-4.0
task_categories:
  - image-to-image
language:
  - en
pretty_name: ChartStyle-100K
tags:
  - style-transfer
  - structured-visualization
  - image-editing
  - training-data
  - eccv-2026
configs:
  - config_name: preview
    default: true
    data_files:
      - split: preview
        path: preview/preview-*.parquet
  - config_name: train
    data_files:
      - split: train
        path: data/train-*.parquet
---

# ChartStyle-100K

**ChartStyle-100K** is a large-scale training dataset for **structured visualization style transfer**. It accompanies the ECCV 2026 paper **ChartStyle-100K: A Large-Scale Dataset for Structured Visualization Style Transfer**.

Each training example is a **triplet**: a style reference visualization, a content visualization, and the corresponding restyled target visualization. The goal is to train models that transfer visual style from the reference while faithfully preserving the content image's structure, text, and data-encoding geometry.

## Quick Facts

- **📄 Paper:** ChartStyle-100K: A Large-Scale Dataset for Structured Visualization Style Transfer
- **🏛️ Venue:** ECCV 2026
- **🎯 Task:** exemplar-guided structured visualization style transfer
- **🗂️ Split:** train
- **📊 Examples:** 100,744 style-transfer triplets
- **🖼️ Total images:** 302,232 (3 per triplet)

## Task Definition

Each training triplet consists of:

1. `style_reference`: a visualization image that defines the desired visual style;
2. `content_image`: a visualization image whose semantic content should be preserved;
3. `target_image`: the stylized visualization generated from `style_reference`, from which `content_image` is derived via restyling under a different visual family.

The triplets are constructed using a reverse-generation pipeline (ChartForge): the target is synthesized first from the style exemplar, and the content image is then produced by restyling the target, ensuring structural alignment between content and target at the data level.

## Dataset Structure

```text
ChartFoundation/ChartStyle-100k
├── README.md
├── preview/
│   └── preview-00000-of-00001.parquet   (100 samples for web preview)
└── data/
    ├── train-00000-of-00054.parquet
    ├── train-00001-of-00054.parquet
    ├── ...
    └── train-00053-of-00054.parquet
```

The `train` split contains **100,744** style-transfer triplets across 54 Parquet shards.

| Field | Type | Description |
| --- | --- | --- |
| `sample_id` | string | Sequential identifier from `000001` to `100744`. |
| `style_reference` | image | Reference visualization whose style should be transferred. |
| `content_image` | image | Input visualization whose content and structure should be preserved. |
| `target_image` | image | Pipeline-generated restyled visualization. |
| `content_type` | string | Coarse content family: `chart`, `flowchart`, `diagram`, or `table`. |
| `content_subject` | string | Thematic domain of the content visualization (e.g. `Finance`, `Biology`). |

The image columns are stored as Hugging Face `Image` features and decode to PIL images by default.

## Data Composition

### Content Type Distribution

| Content family | Count | Percentage |
| --- | ---: | ---: |
| `chart` | 76,122 | 75.6% |
| `diagram` | 11,244 | 11.2% |
| `flowchart` | 10,143 | 10.1% |
| `table` | 3,235 | 3.2% |
| **Total** | **100,744** | **100%** |

The `chart` category covers 36 fine-grained chart types including bar, pie, line, sankey, treemap, radar, violin, heatmap, and others. The `flowchart`, `diagram`, and `table` categories represent structural visualizations whose layout and topology must be preserved during style transfer.

### Content Subject Distribution

The content visualizations span **26 academic and professional domains** including Marketing, Psychology, Education, Biology, Finance, Physics, Engineering, Computer Science, and others, with a roughly uniform distribution across subjects.

### Style References

The style references are drawn from multiple sources to maximize visual diversity:

- real-world chart images from Chart-Galaxy-Real;
- Canva design templates with diverse professional styles;
- synthesized visualizations with diverse styles produced by the ChartForge pipeline.


## Loading

```python
from datasets import load_dataset

# Quick preview (100 samples, shown in the Dataset Viewer)
preview = load_dataset("ChartFoundation/ChartStyle-100k", "preview", split="preview")

# Load the full training dataset (100,744 triplets)
dataset = load_dataset("ChartFoundation/ChartStyle-100k", "train", split="train")

sample = dataset[0]
style_reference = sample["style_reference"]   # PIL Image
content_image   = sample["content_image"]      # PIL Image
target_image    = sample["target_image"]       # PIL Image
content_type    = sample["content_type"]       # str
content_subject = sample["content_subject"]    # str
```

Save images:

```python
style_reference.save("style_reference.png")
content_image.save("content_image.png")
target_image.save("target_image.png")
```

## Relationship to ChartStyleBench

ChartStyle-100K is the training dataset, while [ChartStyleBench](https://huggingface.co/datasets/ChartFoundation/ChartStyleBench) is the held-out evaluation benchmark. The benchmark images in ChartStyleBench are manually collected and have no overlap with the training data in ChartStyle-100K.

| Repository | Purpose | Size |
| --- | --- | --- |
| [ChartFoundation/ChartStyle-100k](https://huggingface.co/datasets/ChartFoundation/ChartStyle-100k) | Training data | 100,744 triplets |
| [ChartFoundation/ChartStyleBench](https://huggingface.co/datasets/ChartFoundation/ChartStyleBench) | Evaluation benchmark | 300 pairs |

## License

ChartStyle-100K is released under **CC BY-NC 4.0**.

## 📄 Citation

If you use ChartStyle-100K in your research or projects, please cite the following paper:

```bibtex
@inproceedings{yang2026chartstyle100k,
  title     = {ChartStyle-100K: A Large-Scale Dataset for Structured Visualization Style Transfer},
  author    = {Yang, Yuwei and Xie, Tianchi and Ni, Jinhong and Guo, Yukai and Zhang, Jing and Zheng, Liang and Bai, Yalong and Yuan, Yuhui},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```