This dataset is strictly restricted to Academic and Research purposes only. Please ensure that the Primary email address in your Hugging Face account settings corresponds to your university or affiliated academic organization. Applications using personal email addresses will be rejected. Processing requests may take up to a week.

I understand and agree to the academic-only license for this dataset.

OmnigraphCanvas Dataset

⚠️ IMPORTANT NOTICE ⚠️

Academic and Research Purposes Only: The OmnigraphCanvas Dataset is distributed strictly for non-commercial, academic, and research purposes. By requesting access to this dataset, you agree that you will not distribute, reproduce, or use the contents of this repository for any commercial applications. All underlying manga artwork remains the intellectual property of its respective original creators and publishers. The dataset is provided under the principle of fair use for advancing scientific research in computer vision and natural language processing.

Dataset Overview

Manga understanding requires complex spatial and temporal reasoning. Models must interpret non-standard right-to-left reading order, varying panel layouts, and the continuity of character states across time. The OmnigraphCanvas Dataset is a high-fidelity collection of raw manga pages compiled specifically for multimodal machine learning, panel segmentation, temporal graph reasoning, and character re-identification.

OmnigraphCanvas provides tens of thousands of high-resolution pages across diverse art styles, functioning as the foundational infrastructure for training robust Vision-Language Models (VLMs) and OmniGraph architectures.

Character Re-identification: Tracking visually shifting characters across episodes.
Panel Extraction & Reading Order: Interpreting complex, overlapping manga panel grids.
Temporal Graph ML (OmniGraph): Mapping the chronological causal states (Flashbacks, Present, Next) of narrative sequences.

Dataset Structure

The dataset utilizes the standard WebDataset format optimized for scalable machine learning pipelines.

Data Instances

Each row in the dataset corresponds to a single, high-resolution manga page and its associated metadata.

{
  "png": "<PIL.Image.Image>",
  "__key__": "Gavactik Squad/Vol. 1 Ch. 0/page_0",
  "__url__": "hf://datasets/mastersubhajit/OmnigraphCanvas-Dataset"
}

Data Fields

png: The raw visual image of the manga page.
__key__: A unique string identifier encoding the original manga title, chapter, and index in the format {manga_title}/{chapter}/page_{index}.
__url__: The Hugging Face dataset repository URI.

Data Splits

The dataset consists of a single train split containing 43,533 images.

Use Cases & Supported Tasks

Unsupervised Panel Segmentation (Object Detection): Train models such as YOLO or Mask-RCNN to isolate irregular comic panels.
Vision-Language Alignment: Fine-tune foundational VLMs (e.g., Qwen-VL, LLaVA) on complex, high-contrast line art.
Temporal Dependency Modeling: Pre-train sequence models to predict chronological panel continuity based on dense visual context.

Curation & Processing

If you use this dataset in your research, please cite it as follows:

@misc{subhajit_ghosh_2026,
    author       = { Subhajit Ghosh },
    title        = { OmnigraphCanvas-Dataset (Revision 310c25c) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/mastersubhajit/OmnigraphCanvas-Dataset },
    doi          = { 10.57967/hf/10640 },
    publisher    = { Hugging Face }
}

Images were aggregated from raw digital manga scans and normalized for Hugging Face ingestion. All directory structures were systematically flattened and aligned with standard data streaming conventions to ensure maximum compatibility with the Hugging Face datasets library.

Affiliation & Credits

The OmnigraphCanvas Dataset was created and curated by Subhajit Ghosh. It is maintained exclusively for academic research and is proudly associated with the AI and Simulation Lab (ASL).

Downloads last month: 284

Total file size:

68.1 GB