OmniDocLayout-1M

OmniDocLayout-1M is a large-scale dataset for document layout generation, featuring about one million document pages across six contemporary document types. The dataset is designed to support research on layout generation, layout representation learning, and coarse-to-fine LLM learning for Document AI.

The dataset focuses on diverse and realistic document layouts, including document types that are underrepresented in prior layout generation datasets (e.g. PubLayNet and DocBank), such as newspapers, textbooks, magazines, exam papers, academic papers, and slides.

Dataset Summary

OmniDocLayout-1M contains page-level document layout annotations. Each sample describes the canvas size of a document page and a sequence of layout elements represented by category labels and bounding boxes.

The current released files are organized by document types:

OmniDocLayout-1M/
├── README.md
└── data/
    ├── academic.json
    ├── exam.json
    ├── magazine.json
    ├── newspaper.json
    ├── slide.json
    └── textbook.json

Dataset Statistics

The dataset contains about 1M pages in total.

Covered Document Types

OmniDocLayout-1M covers six common document types:

• Magazine: multi-column magazine pages with rich combinations of text blocks, titles, images, and other visual elements.
• Textbook: educational pages with dense text, figures, tables, equations, captions, and structured sections.
• Academic paper: research paper pages with scholarly layouts, including text, titles, figures, tables, equations, and captions.
• Exam paper: exam-style pages containing questions, options, tables, formulas, and structured educational content.
• Newspaper: complex and dense newspaper layouts with multi-column text, headlines, figures, and irregular spatial arrangements.
• Slide: presentation-style pages with larger visual regions, text blocks, figures, tables, and equation-heavy educational or academic content.

Data Format

Each JSON file contains a list of page-level samples. Each sample has the following structure:

{
  "Document Type": "document_type",
  "Canvas Width": W,
  "Canvas Height": H,
  "Bbox Number": N,
  "Layout Info": [
    ["category_1", x_1, y_1, w_1, h_1],
    ["category_2", x_2, y_2, w_2, h_2],
    ["category_3", x_3, y_3, w_3, h_3],
    ...
    ["category_N", x_N, y_N, w_N, h_N]
  ]
}

Field Descriptions

Category Labels

The category label is stored as a string in each layout element. The released JSON files use layout-related categories such as:

text, title, image, table, image_caption, table_caption, image_footnote, table_footnote, equation

Different document types may contain different distributions of categories.

Bounding Box Format

Each bounding box in Layout Info follows the absolute xywh pixel format:

[category, x, y, width, height]

where:

• category is the semantic category of the layout element.
• x and y denote the element position on the canvas.
• width and height denote the element size.
• All coordinates and sizes are absolute pixel values on the corresponding canvas.
• The coordinate system follows the page canvas defined by Canvas Width and Canvas Height.

Loading the Dataset

Since each document type is stored as a standalone JSON file, the dataset can be loaded directly with Python:

import json

json_path = "magazine.json"

with open(json_path, "r", encoding="utf-8") as f:
    data = json.load(f)

print(type(data))
print(len(data))
print(data[0].keys())
print(data[0]["Layout Info"][0])

To load all types:

import json
from pathlib import Path

root = Path("OmniDocLayout-1M/data")

document_files = {
    "magazine": "magazine.json",
    "textbook": "textbook.json",
    "academic": "academic.json",
    "exam": "exam.json",
    "newspaper": "newspaper.json",
    "slide": "slide.json",
}

all_data = {}

for document, filename in document_files.items():
    with open(root / filename, "r", encoding="utf-8") as f:
        all_data[document] = json.load(f)

for document, samples in all_data.items():
    print(document, len(samples))

Dataset Construction

OmniDocLayout-1M was constructed to provide a large-scale and diverse source of document layouts. The dataset was collected from 36 public and copyright-clean sources and processed through an automated pipeline.

The general construction process includes:

1. Collecting documents from diverse public sources.
2. Standardizing input document formats.
3. Rendering document pages when necessary.
4. Removing duplicates and low-quality pages.
5. Parsing pages into layout elements.
6. Filtering and cleaning annotations.
7. Exporting page-level layouts into JSON format.

The dataset covers multiple contemporary domains, including education, academia, news, publishing, and slides.

Annotation

Fully automatic using MinerU toolkit. Newspaper layouts with dense/complex structure are additionally refined via fine-tuned DocLayout-YOLO.

Quality check

Human evaluation on 1,200 pages shows ≥92% perceived quality consistency between automatic annotations and manual labels.

Citation

If you find OmniDocLayout-1M useful in your research, please cite:

@inproceedings{kang2026omnidoclayout,
  title={OmniDocLayout: Towards Diverse Document Layout Generation via Coarse-to-Fine LLM Learning},
  author={Kang, Hengrui and Gu, Zhuangcheng and Zhao, Zhiyuan and Wen, Zichen and Wang, Bin and Li, Weijia and He, Conghui},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3208--3218},
  year={2026}
}