ChartGalaxyPP/ChartGalaxyPlusPlus

ChartGalaxy++

A Richly Annotated Dataset for Chart Understanding and Generation

Project   ·   Image2SceneGraph model   ·   Application data   ·   Loading guide

ChartGalaxy++ connects what is in an infographic chart with how it is organized. Each chart is paired with a scene graph: visual elements and semantic groups form the nodes, while hierarchical and spatial relationships connect them. These annotations support chart structure prediction, visual question answering, and evaluation of image generation.

Infographic charts	Annotated nodes	Relationships
217,195	18.90 million	39.99 million

Explore the dataset

Selected synthetic examples from the released PNGs, shown in full. The gallery spans compact comparisons, layered circular charts, dense radial marks, and illustrated narratives. View at full resolution · Sample identities

What is annotated?

The pasta-production example from the paper: each country groups a value, a pasta image, a flag, and a country label. The scene graph records these groups, element attributes, and hierarchical and spatial relationships. Enlarge

Layer	Annotation content
Visual elements	Text, images, and shapes with bounding boxes, semantic roles, text content, and appearance attributes
Semantic groups	Charts, axes, legends, legend items, data items, series, and panels
Hierarchy	Parent–child links connecting elements to groups and the chart composition
Spatial relationships	Relative position, alignment, overlap, and Boolean proximity (is_near)

See the annotation guide, JSON schema, and data format for all fields and coordinate conventions.

Dataset at a glance

Split	Real charts	Synthetic charts	Total
Train	56,244	159,951	216,195
Test	500	500	1,000
Total	56,744	160,451	217,195

Real charts: URLs + annotation JSON. Synthetic charts: PNG + annotation JSON. The main dataset is distributed in 926 independently extractable tar.gz shards. sample_index.jsonl.gz maps each sample ID to its shard and member files. The separate QA package provides image URLs without image files; unavailable URLs are left empty.

Get started

Install the download client with pip install huggingface_hub. Start with one standalone example before downloading shards:

import json
from huggingface_hub import HfApi, hf_hub_download

repo = "ChartGalaxyPP/ChartGalaxyPlusPlus"
revision = HfApi().dataset_info(repo).sha
def download(name):
    return hf_hub_download(repo, name, repo_type="dataset", revision=revision)

image_path = download("examples/01-layout/image.png")
graph_path = download("examples/01-layout/scene_graph.json")
spatial_path = download("examples/01-layout/spatial_relations.json")
with open(graph_path, encoding="utf-8") as f:
    graph = json.load(f)
nodes = graph["compositional_deconstruction"]["nodes"]
print(f"{len(nodes)} explicit nodes", image_path)

The example is an existing training record, not an additional sample. To load any chart, use the sample index and the shard-loading example. Bounding boxes in dataset annotations use [y0, x0, y1, x1], normalized to 0–1000.

Download the complete dataset

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id=repo,
    repo_type="dataset",
    revision=revision,
    local_dir="chartgalaxy-plus-plus",
)

The data shards total approximately 105 GB. The release includes shard hashes in manifest.json. See USAGE.md for extraction and integrity checks.

What can you do with it?

Application	What the scene graph enables	Released resource
Image-to-scene-graph prediction	Recover elements, semantic groups, and chart hierarchy	Image2SceneGraph model; 1,000-chart benchmark
Infographic question answering	Associate text and marks through explicit groups and relationships	1,266 questions, reference answers, and image URLs
Scene graph preservation	Evaluate structural fidelity in generated infographic charts	11-model benchmark and generated outputs

See the project page for qualitative results and links to all three application packages. Under the paper's evaluation protocol, the released Image2SceneGraph model achieves 89.4% node F1, 85.1% hierarchy F1, and 88.0% spatial F1.

Annotation and release notes

Exact annotation counts and release notes

• Nodes: 15,591,656 visual elements + 3,088,632 explicit groups + 217,195 implicit roots = 18,897,483.
• Relationships: 18,680,288 parent links + 21,308,862 stored spatial records = 39,989,150. Unrecorded pairs are not negative labels.
• is_near is Boolean. Positive-area overlaps and containment are excluded; both absolute and relative distance thresholds must hold. The format guide specifies the rule.
• The 1,000-chart test set, comprising 500 real and 500 synthetic charts, has been manually verified. human_gold is true for the test split and false for the training split.
• Generated replacement illustrations and their affected annotations are included in the main dataset. Historical benchmark packages retain the identities of their evaluated inputs; use their supplied references when inspecting reported results.
• Real-image references are supplied without checking their current availability. Source pages, direct image URLs, and archive references are distinguished in the format guide.

This release contains images or image references, scene graph annotations, and spatial records. Underlying data-table files and annotation/training/evaluation pipeline code are not included. The gallery is a curated visual preview, not a random sample or annotation-quality evaluation.

License

Contributed annotations are licensed under CC BY-NC 4.0. Third-party chart content and required upstream attributions retain their respective rights; see LICENSE and license and sources.