CoDeFont

Official implementation of CoDeFont

Project Website · GitHub Repository

---

Abstract

One-shot Font Generation (OFG) aims to synthesize a target glyph by transferring the stylistic characteristics of a single reference glyph while preserving the structure of the target character. However, accurately capturing font style from a single reference image remains challenging due to the limited stylistic information available in a single glyph, and this challenge is further amplified for complex scripts. We propose CoDeFont, a script-agnostic OFG framework that addresses this challenge through multi-level style-content factorization and progressive style-content reconciliation. CoDeFont independently aggregates complementary style and structural information across multiple spatial levels to obtain factor-specific representations, while a contrastive objective explicitly encourages separation between style and content embeddings. During reconstruction, the disentangled representations are progressively injected through multi-level style conditioning, enabling stylistic information to be incorporated at different decoding resolutions while preserving the underlying glyph structure. We evaluate CoDeFont on structurally distinct Bengali and Latin scripts across seen and unseen font styles. Extensive quantitative and qualitative analyses demonstrate consistent improvements over existing state-of-the-art methods, highlighting the effectiveness of the proposed framework for OFG across diverse writing systems.

Visual Results

Qualitative Comparison

Qualitative comparison of generated results. Replace this placeholder with the actual experimental figure.

Additional Results

Additional examples and comparisons.

Note: The image paths above are proposed locations for figures. Add the corresponding image files to docs/ before using these sections.

---

Repository Structure

CoDeFont/
├── dataloader/    # Data loading and preprocessing
├── docs/          # Documentation and figures
├── losses/        # Loss functions
├── models/        # Model and network components
├── scripts/       # Training and inference Python modules
├── train.sh       # Training entry point
├── infer.sh       # Inference entry point
├── LICENSE        # License information
└── README.md      # Project documentation

Which File Should I Run?

File	Purpose
train.sh	Start model training
infer.sh	Run inference using trained checkpoints
scripts/train.py	Training implementation invoked through python -m scripts.train
scripts/infer.py	Inference implementation invoked through python -m scripts.infer
dataloader/	Data-loading components used by the pipeline
models/	Model architecture and network components
losses/	Training objectives and loss implementations
docs/	Documentation and visual assets

The Python module paths shown above follow the launch commands in the shell scripts.

Installation

Clone the repository:

git clone https://github.com/rejoycs/codefont.git
cd codefont

Create and activate a Python environment, then install the dependencies required by the implementation.

conda create -n codefont python=3.10
conda activate codefont

Install a compatible PyTorch build for your CUDA environment, followed by the project's Python dependencies.

Dataset Preparation

The training and inference scripts expect language-specific directories.

For example, the launchers construct paths using:

Dataset_bengali/
results_bengali/
results_bengali_out/
ckpt/

Prepare the dataset in the format expected by the data-loading implementation before running the scripts. Check the dataset configuration and preprocessing code for the required directory structure and image format.

Training

The repository provides train.sh as the main training entry point.

bash train.sh

To specify another language:

bash train.sh english

The script invokes scripts.train with the following configured parameters:

• Batch size: 64
• Embedding dimension: 64
• Image size: 64 × 64
• Epochs: 500
• Device: cuda:0
• Test interval: 1

Training checkpoints are configured under ckpt/, with separate paths for the style encoder, content encoder, decoder, and discriminator.

Inference

Use infer.sh to generate results using trained model checkpoints.

bash infer.sh

To run inference for another language:

bash infer.sh english

The script invokes scripts.infer and reads the configured style encoder, content encoder, and decoder checkpoints.

By default, the output directory is:

results_bengali_out/

For another language, the output directory follows the pattern results_<language>_out/.

Important: Run inference only after the required model checkpoints are available at the paths expected by infer.sh. Verify that the dataset and checkpoint paths match your local setup.

Configuration

The current shell scripts define the primary execution parameters. To change the default language, batch size, embedding dimension, image size, checkpoint locations, or device, edit train.sh or infer.sh as appropriate.

For a different GPU, update the --DEVICE argument. For CPU execution, first confirm that the underlying implementation supports CPU inference and training.

Citation

If you use CoDeFont in your research, please cite the associated paper.

@misc{chakraborty2026codefontoneshotfontgeneration,
      title={CoDeFont: One-Shot Font Generation via Multi-Scale Factor-Specific Representation Learning}, 
      author={Rejoy Chakraborty and Prasun Roy and Saumik Bhattacharya and Umapada Pal},
      year={2026},
      eprint={2610.21364},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.21364}, 
}

Please refer to the official project website for project information.

License

This project is distributed under the Apache License 2.0. See LICENSE for details.