DuyHieu63/SpiS_GAN

SpiS-GAN: Spiral-Modulated Handwriting Synthesis with Star Operation

SpiS-GAN is a GAN-based handwriting synthesis framework designed to generate realistic, legible, and writer-consistent handwriting for improving downstream handwritten text recognition (HTR) systems.

Introduction

This repository contains the reference code and dataset for the paper:

SpiS-GAN: Spiral-Modulated Handwriting Synthesis with Star Operation

Nguyen Duy Hieu, Dang Hoai Nam, Pham Hoang Giap, Quang Huu Hieu, Vo Nguyen Le Duy

Overview

Installation

Create a Python environment and install PyTorch for your CUDA version first. Then install the remaining dependencies:

# Install torch/torchvision following the official PyTorch command for your system.
# Example: https://pytorch.org/get-started/locally/
pip install -r requirements.txt

Code

We released in on GitHub : https://github.com/DAIR-Group/SpiS-GAN

Current release status:

Resource	Status
32px datasets/checkpoints	Available
64px datasets/checkpoints	To be updated

Data Preparation

The dataset loader expects HDF5 files under ./data/. Current path settings are defined in lib/path_config.py.

Expected files:

data/
|-- train_32.hdf5         # IAM train/validation split
|-- test_32.hdf5          # IAM test split
|-- train_vn.h5           # Vietnamese train/validation split
|-- test_vn.h5            # Vietnamese test split
|-- english_words.txt     # English lexicon
`-- vietnamese_words.txt  # Vietnamese lexicon

Training

Train on English handwriting data:

python train.py --config configs/SpiS_gan_iam_32.yml

Train on Vietnamese handwriting data:

python train.py --config configs/SpiS_gan_vn_32.yml

64px configurations are also available:

python train.py --config configs/SpiS_gan_iam_64.yml
python train.py --config configs/SpiS_gan_vn_64.yml

The 64px configuration files are included for reproducibility and future use. The public 64px datasets/checkpoints will be updated later.

Training outputs are written to runs/<config-name>-<timestamp>/, including generated samples and checkpoints according to each YAML configuration.

Generation

Generate handwriting samples from a config file:

python generate.py --config configs/SpiS_gan_iam_32.yml

Use random lexicon sampling:

python generate.py --config configs/SpiS_gan_vn_32.yml --random_lexicon

Set the ckpt field in the YAML config to the trained checkpoint path before running generation.

Configuration

Main configuration files:

Config	Dataset	Resolution
configs/SpiS_gan_iam_32.yml	IAM English handwriting	32px
configs/SpiS_gan_iam_64.yml	IAM English handwriting	64px
configs/SpiS_gan_vn_32.yml	Vietnamese handwriting	32px
configs/SpiS_gan_vn_64.yml	Vietnamese handwriting	64px

Handwriting synthesis and reconstruction results on IAM dataset

Handwriting synthesis results on HANDS-VNOnDB dataset

Repository Structure

.
|-- configs/              # Training and generation configs for IAM and Vietnamese data
|-- docs/               # README figures and result images
|-- data/                 # Lexicons and expected dataset/checkpoint location
|-- fid_kid/              # FID/KID evaluation utilities
|-- font/                 # Font assets used by the pipeline
|-- lib/                  # Dataset, alphabet, path, and utility code
|-- networks/             # Generator, discriminator, recognizer, and model modules
|-- generate.py           # Generate handwriting samples from a trained checkpoint
|-- train.py              # Train SpiS-GAN from a config file
`-- README.md

Citation

If you use this repository, please cite:

@misc{hieu2026spisganspiralmodulatedhandwritingsynthesis,
  title         = {SpiS-GAN: Spiral-Modulated Handwriting Synthesis with Star Operation},
  author        = {Nguyen Duy Hieu and Dang Hoai Nam and Pham Hoang Giap and Quang Huu Hieu and Vo Nguyen Le Duy},
  year          = {2026},
  eprint        = {2607.06949},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.06949}
}