GapFill: No Pixel Left Behind: Filling Gaps in Anime Colorization

GapFill is an interactive tool for helping professional anime colorists detect and fill small unpainted enclosed regions, or "gaps" (塗り残し) that are often left behind during digital manual colorization.

The system automatically detects gaps, highlights them, and suggests fill colors using a domain-specific deep-learning method that learns correspondences between image regions. GapFill also provides pop-up magnification, manual color correction, and sweep-to-apply interactions, reducing the repetitive work of finding gaps, zooming in, and selecting colors.

Quick Start

This repository contains two components. Refer to the README for each component for installation, data preparation, and usage instructions.

Component	Description	Setup and usage
Web application (web/)	Interactive GapFill interface for detecting, inspecting, and filling gaps	Web application README
Machine-learning pipeline (ml/)	Data preprocessing, model training, evaluation, and visualization	ML pipeline README

Notes on the Released Version

The model distributed with this repository was retrained after the paper was submitted and achieved a small improvement in prediction accuracy. As a result, its predictions may be slightly better than those produced by the version used in the user study, particularly for Task B (previous 'wrong prediction').

The codebase was also refactored in preparation for its public release. If you encounter any unexpected behavior or discrepancies, please do not hesitate to report them through GitHub Issues.

Citation

If you use GapFill in your research, please cite our paper:

@inproceedings{kono2026gapfill,
  author    = {Masahiro Kono and Akinobu Maejima and Yuki Koyama and
               Yotam Sechayk and Takeo Igarashi},
  title     = {No Pixel Left Behind: Filling Gaps in Anime Colorization},
  booktitle = {Proceedings of the 2026 CHI Conference on Human Factors in
               Computing Systems},
  year      = {2026},
  doi       = {10.1145/3772318.3790968}
}

Machine-readable citation metadata is available in CITATION.cff.

Image and Dataset Availability

All image materials used in this project were used with permission from ©IIS-P / Ponnomichi Production Committee.

The repository and hosted demo grant no license to extract or use these image materials independently. Without separate permission from the copyright holder, they may not be copied, modified, reused, or redistributed.

Task C uses full, unmodified source images. These assets are intentionally not distributed in this repository for copyright reasons and are ignored under web/public/preset-images/C/. The dataset used to train the released model is not distributed for the same reason.

Distributed preset images and documentation images contain embedded copyright metadata. Please do not remove this metadata when using, copying, or processing the images. See ASSET_LICENSE.md for the complete asset and training-data notice.

License

The source code is released under the MIT License. This license does not grant rights to the third-party image materials described above.

Updates

• June 21, 2026: Source code (refactored) for the web application is available.
• June 18, 2026: The trained GapFill model is available as trained_model.pth from GitHub Releases.
• June 10, 2026: Source code (refactored) for the GapFill model is available.
• March 10, 2026: The project website, paper, and demo web application are available.