Decoupling Layout from Glyph in Online Chinese Handwriting Generation

🎉🎉🎉 This work has been accepted by ICLR2025 🎉🎉🎉

Data and Environment

The data can be download at "https://huggingface.co/datasets/Immortalman12/OLHWD/tree/main".

The description of each data file can be found there.

The required environment is particularly simple, only basic libraries such as torch, numpy, matplotlib, and math are needed.

How to build a framework for Online Chinese Handwriting Generation

It consists of two steps:

1. Single Character Generation
2. Layout Generation

Single Character Generation

File name	Main functions
basemodel.py	Network Architecture (1D-Unet)
unet_ddpm.py	Diffusion Model Implementation
train_ddpm.py	Train and test the diffusion model
style_classifier.py	Style Encoder
evaluate_char.py	Evaluate the character generator
./evaluate_char/char_classifier.py	Train the content classifier for evaluation
./evaluate_char/style_classifier.py	Train the style classifier for evaluation

1. Download the data into ./datas
2. Train the diffusion model train_ddpm.py
3. For evaluation, train the content and style classifiers
4. Evaluate and Finished evaluate_char.py

Layout Generation

File name	Main functions
./bonibox_gen/count.py	Count the border information of each character type in the dataset
./bonibox_gen/char_boxes.npy	The border information of each character type （Use Gaussian Distribution
./bonibox_gen/train_box_generator.py	Train the layout planner LSTM module
./bonibox_gen/simpebox.py	Generate layout utilizing Gaussian distribution for each character
./bonibox_gen/generatebox.py	Generate layout utilizing LSTM network
./bonibox_gen/evaluate_box.py	Evaluate the generated layout

1. Get the data ./bonibox_gen/char_boxes.npy
2. Train the layout model ./bonibox_gen/train_box_generator.py
3. Evaluate and Finished ./bonibox_gen/evaluate_box.py

Full Text Line Generation

Utilizing generate_line.py

Notification

The code is a bit messy. Please be understanding! Due to graduation, some files were lost. If you have any questions, feel free to communicate with the author at any time. Wish you all smooth progress in your scientific research!

If you find our work helpful, please cite us using:

@article{ren2024decoupling,
  title={Decoupling Layout from Glyph in Online Chinese Handwriting Generation},
  author={Ren, Min-Si and Zhang, Yan-Ming and Chen, Yi},
  journal={arXiv preprint arXiv:2410.02309},
  year={2024}
}

@inproceedings{ren2025decoupling,
  title={Decoupling Layout from Glyph in Online Chinese Handwriting Generation},
  author={Minsi Ren and Yan-Ming Zhang and Yi Chen},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=DhHIw9Nbl1}
}