GryphOne

Official implementation of "GryphOne: Symbol-Aware Masked Diffusion for Structural Refinement in Offline Handwritten Mathematical Expression Recognition" (ECCV 2026).

Overview

GryphOne iteratively refines symbols and their relationships with symbol-aware masked diffusion:

Usage

Install

Install PyTorch 2.4+ and run the following command:

$ pip install -e .

Models

See releases.

Datasets

Download the following datasets:

• CROHME 2014
• CROHME 2016
• CROHME 2019
• MathWriting

Preprocess

Run preprocess.py as follows:

$ python3 preprocess.py datasets/crohme_test.yaml
$ python3 preprocess.py datasets/mathwriting.yaml

The datasets must be placed in data directory as follows:

$ ls ~/data
crohme/
  CROHME2014_data/
  CROHME2016_data/
  CROHME2019_data/
mathwriting/
  2024/
    test/
      000a4e8ca49c5a1c.inkml
      001083e26028da36.inkml
      0017bb5822bcba69.inkml
      002ae6d5dd4173e4.inkml
      00386113d577085b.inkml
    train/
      00001d1472a8709f.inkml
      00002504391b73b5.inkml
      00003037d3a6d0ba.inkml
      0000fe986018f92a.inkml
      00011d9f03970147.inkml
    valid/
      00000b332dcd6fe5.inkml
      000453d57c3d334d.inkml
      0005d0be8e507b24.inkml
      0005ea8e21185d36.inkml
      00061be0501a1fa8.inkml
pickle/
  gryph_crohme_test.pkl
  gryph_mathwriting.pkl

Training

Run train.py to start training for MathWriting using four GPUs:

$ torchrun --nproc-per-node=4 train.py configs/mathwriting.py --work-dir ~/work

Inference

Run infer.py to start inference using the trained model:

$ python3 infer.py --config ~/work/mathwriting.py --weight ~/work/epoch_60.pth --split test --store results.pkl

Use LgEval for strict evaluation.

License

This project is licensed under the MIT License. See LICENSE for more details.

Citation

@inproceedings{ECCV26KAT,
  author={Takaya Kawakatsu and Ryo Ishiyama},
  title={GryphOne: Symbol-Aware Masked Diffusion for Structural Refinement in Offline Handwritten Mathematical Expression Recognition},
  booktitle={Computer Vision -- ECCV 2026},
  publisher={Springer Nature Switzerland},
  year={2026},
  pages={609--624},
}

See Also

MuTabNet (ICDAR2024): our table recognition framework for ICDAR 2024 and 2026.