InkFormer: From Characters to Signatures via Style-Transformation-Aware Online Handwriting Generation

This repository provides the training, inference, and evaluation code for InkFormer-Char, the character-level handwriting generation model. Additionally, the training code for both the content and style evaluators is available.

Updates

• [2026/10/23] Code and pretrained weights are released.
• [2026/10/15] InkFormer is accepted by TMLR 2026 🎉🎉🎉.

Repository Structure

InkFormer/
├── main.py                     # Training entry
├── inference.py                # Inference entry
├── dataset_char.py / utils.py  # Data loading and shared utilities
├── classifiers/                # Training code for the content/style classifiers (evaluators)
├── evaluation/                 # Evaluation scripts (nDTW / content / style)
├── modules/                    # Model architecture (Transformer_AR and encoders)
├── scripts/                    # train.sh / train_cont.sh / train_sty.sh / infer.sh / evaluate.sh
└── data_ckpt/
    ├── ckpt/                   # Model checkpoints
    └── datas/                  # Datasets

Environment Setup

A Linux machine with NVIDIA GPUs is recommended. Install the dependencies:

conda create -n inkformer python=3.9 -y
conda activate inkformer
pip install -r requirements.txt

Main dependencies: PyTorch, torchvision, accelerate, numpy, Pillow, matplotlib, tqdm, fastdtw, scipy, einops.

Data & Checkpoints

Download: Google Drive

After downloading, place the files under data_ckpt/ as follows:

data_ckpt/
├── ckpt/
│   ├── ar_generator_5000.pth    # Generator weights (for inference)
│   ├── cont_evaluator_2000.pth  # Content evaluator weights
│   ├── sty_evaluator_10000.pth  # Style evaluator weights
│   ├── lookup_char2int.npy      # Lookup table for content evaluation
│   └── lookup_idx2ori_on.npy    # Lookup table for style evaluation
└── datas/
    ├── train/                   # Training set (6 pools)
    │   ├── Pot1.0Train_pp1.npy
    │   ├── Pot1.1Train_pp1.npy
    │   ├── Pot1.2Train_pp1.npy
    │   ├── Pot1.0Test_pp1.npy
    │   ├── Pot1.1Test_pp1.npy
    │   └── Pot1.2Test_pp1.npy
    ├── test/                    # Test set
    │   └── competition_set_pp1.npy
    └── standard/                # Standard characters
        └── stand_zi.npy

Training

Generator

bash scripts/train.sh

• Uses 4 GPUs by default (CUDA_VISIBLE_DEVICES=0,1,2,3 in scripts/train.sh; adjust to your setup)
• Multi-GPU training is launched with accelerate; run accelerate config first if you have not configured it yet
• Key hyperparameters (batch-user, sample-per-user, n-reference, lr, epoch-end, etc.) can be edited directly in scripts/train.sh
• Checkpoints are saved every 500 epochs (save-epoch) to data_ckpt/ckpt/ar_generator_{epoch}.pth
• To resume training, set --epoch-begin to an existing epoch; data_ckpt/ckpt/ar_generator_{epoch_begin}.pth will be loaded automatically

Classifiers (Evaluators)

bash scripts/train_cont.sh   # Content classifier → data_ckpt/ckpt/cont_evaluator_{epoch}.pth
bash scripts/train_sty.sh    # Style classifier  → data_ckpt/ckpt/sty_evaluator_{epoch}.pth

If logs/infer/syn_saver_ref_15.npy (the generator's inference output) exists, classifier training additionally monitors the accuracy on synthesized characters (Acc_syn); otherwise this monitoring is skipped automatically, which does not affect training or checkpoint saving.

Inference

bash scripts/infer.sh

• Runs on a single GPU (GPU 0)
• Loads data_ckpt/ckpt/ar_generator_5000.pth by default; edit the --epoch argument in scripts/infer.sh to use another checkpoint
• Synthesized results are saved to logs/infer/syn_saver_ref_15.npy
• Side-by-side images of standard / synthesized / ground-truth characters are written to logs/infer/img_infer/

Evaluation

bash scripts/evaluate.sh

Runs three evaluations on the inference output logs/infer/syn_saver_ref_15.npy:

1. nDTW (evaluation/evaluate_ndtw_1.py): trajectory similarity
2. Content score (evaluation/evaluate_content_score.py): character recognizability, computed with the bundled content evaluator
3. Style score (evaluation/evaluate_style_score.py): style similarity, computed with the bundled style evaluator

Pipeline Overview

scripts/train.sh    →  data_ckpt/ckpt/ar_generator_*.pth (checkpoints)
scripts/infer.sh    →  logs/infer/syn_saver_ref_15.npy (synthesized results)
scripts/evaluate.sh →  nDTW / content score / style score

Acknowledgements

• CASIA-OLHWDB, ICDAR-2013 Competition, MSDS and Make Me a Hanzi are valuable resources for this work.

Citation

If you find this work useful or use this code in your research, please consider citing the following paper.

@article{
  pan2026inkformer,
  title={InkFormer: From Characters to Signatures via Style-Transformation-Aware Online Handwriting Generation},
  author={Wei Pan and Yilin Shi and Lianwen Jin},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2026},
  url={https://openreview.net/forum?id=Nbhw2Y5Nmh},
  note={}
}

License

Our code is released under the MIT License. The pretrained weights, which are trained using part of the CASIA-OLHWDB dataset, follow the original license of the dataset.