DA-Font: Few-Shot Font Generation via Dual-Attention Hybrid Integration (ACM MM 2025)

Official Pytorch Implementation of "DA-Font: Few-Shot Font Generation via Dual-Attention Hybrid Integration" by Weiran Chen, Guiqian Zhu, Ying Li, Yi Ji and Chunping Liu.

Paper can be found at Arxiv.

Abstract

Few-shot font generation aims to create new fonts with a limited number of glyph references. It can be used to significantly reduce the labor cost of manual font design. However, due to the variety and complexity of font styles, the results generated by existing methods often suffer from visible defects, such as stroke errors, artifacts and blurriness. To address these issues, we propose DA-Font, a novel framework which integrates a Dual-Attention Hybrid Module (DAHM). Specifically, we introduce two synergistic attention blocks: the component attention block that leverages component information from content images to guide the style transfer process, and the relation attention block that further refines spatial relationships through interacting the content feature with both original and stylized component-wise representations. These two blocks collaborate to preserve accurate character shapes and stylistic textures. Moreover, we also design a corner consistency loss and an elastic mesh feature loss to better improve geometric alignment. Extensive experiments show that our DA-Font outperforms the state-of-the-art methods across diverse font styles and characters, demonstrating its effectiveness in enhancing structural integrity and local fidelity.

Dependencies

python >= 3.8
torch >= 1.7.1
torchvision >= 0.8.2
sconf >= 0.2.5
lmdb >= 1.6.2
opencv-python >= 4.10.0.84
scipy >= 1.10.1

Data Preparation

Images and Characters

1. First, collect a set of font files in .ttf (TrueType Font) or .otf (OpenType Font) format. These fonts should be divided into three groups: content font, training fonts, and validation fonts. To ensure the model effectively learns style variations, it's recommended that the training fonts exhibit noticeable style diversity. In our experiments, the fonts were sourced from here.

2. Next, determine the target character set for both training and testing. For example, the first-level Chinese character set contains 3500 commonly used characters:

{梦、飞、龙、书、雪、月、花、山、海、云、星、雨、竹、茶、琴、棋、画、诗、酒、剑、...、etc}

3. Once the fonts and characters are prepared, use the script ./datasets/font2image.py to render character images from the fonts. The directory structure should be organized as follows:

Font Directory
|-- content
|   |-- content_font
|   |   |-- content_font_char1.png
|   |   |-- content_font_char2.png
|   |   |-- ...
|-- train
|   |-- train_font1
|   |-- train_font2
|   |-- train_font3
|   |   |-- train_font3_char1.png
|   |   |-- train_font3_char2.png
|   |   |-- ...
|   |-- ...
|-- val
|   |-- val_font1
|   |-- val_font2
|   |-- val_font3
|   |   |-- val_font3_char1.png
|   |   |-- val_font3_char2.png
|   |   |-- ...
|   |-- ...

Construct meta files and LMDB environment

1. The characters need to be split into training and validation sets, with each character saved in Unicode format. You can convert characters to Unicode using hex(ord(ch))[2:].upper():, examples can be found in ./meta/.

2. Execute the following command to generate the LMDB files:

  python3 ./build_dataset/build_meta4train.py \
  --saving_dir ./results/your_task_name/ \
  --content_font path\to\content \
  --train_font_dir path\to\training_font \
  --val_font_dir path\to\validation_font \
  --seen_unis_file path\to\train_unis.json \
  --unseen_unis_file path\to\val_unis.json 

Training Workflow

The overall training is divided into two stages:

1. Pre-training the content encoder and codebook based on VQ-VAE.
2. Training the few-shot font generation framework with GAN.

Pre-train the VQ-VAE

The VQ-VAE is trained using the content font. The relevant training code is provided at VQ-VAE.py.

Once VQ-VAE pre-training is complete, use the trained content encoder to compute similarity scores between each character in the training and validation sets. The similarity information is stored in a dictionary format, for example:

{'4E07': {'4E01': 0.2143, '4E03': 0.2374, ...}, '4E08': {'4E01': 0.1137, '4E03': 0.1020, ...}, ...}

Train the DA-Font

Adjust the configuration file as needed in the file ./cfgs/custom.yaml

Some Key Configuration Options:

• work_dir: Directory to store all output results. (should match the saving_dir used during dataset preparation)
• data_path: Path to the LMDB dataset. (saving_dir/lmdb)
• data_meta: Path to the meta information. (saving_dir/meta)
• content_font: Specify the source font to be used.
• all_content_char_json: JSON file listing all characters from both training and validation sets.
• other values are hyperparameters for training.

Launch Training

 python3 train.py task_name cfgs/custom.yaml
   #--resume \path\to\your\pretrain_model.pdparams

Infer the DA-Font

Run the script

  python3 inference.py ./cfgs/custom.yaml \
  --weight \path\to\saved_model.pdparams \
  --content_font \path\to\content_imgs \
  --img_path \path\to\test_imgs \
  --saving_root ./infer_res

Acknowledgements

Our project is inspired and modified by VQ-Font and FsFont. We would like to express our sincere gratitude to our collaborators for their valuable supports and to the reviewers for their insightful feedback and suggestions.

Citation

@InProceedings{Chen_2025_MM,
    author    = {Weiran Chen and Guiqian Zhu and Ying Li and Yi Ji and Chunping Liu},
    title     = {DA-Font: Few-Shot Font Generation via Dual-Attention Hybrid Integration},
    booktitle = {Proceedings of the {ACM} International Conference on Multimedia, {MM}},
    month     = {October},
    year      = {2025},
    pages     = {xxxx-xxxx}
}

Contact

If you have any questions, please feel free to contact wrchen2023@stu.suda.edu.cn or wrchen2023@outlook.com.