English | 简体中文

UniCalli is a groundbreaking unified diffusion framework that addresses column-level generation of Chinese calligraphy. Unlike existing methods that focus on isolated character generation or compromise calligraphic correctness for page-level synthesis, UniCalli integrates both recognition and generation tasks in a single model, achieving superior results in both stylistic fidelity and structural accuracy.

• Unified Architecture: First framework to unify column-level calligraphy generation and recognition
• Multi-Master Styles: Supports diverse calligraphic styles, including Wang Xizhi, Yan Zhenqing, Ouyang Xun, etc.
• Densely Annotated Data: Trained on large-scale calligraphy dataset with detailed annotations

For academic research and non-commercial use only. For commercial use, please contact the authors.

本模型仅供学术研究、非商业使用，商业使用请联系作者。

• Model Release - Base version without pred_box
• Inference Code
• Interactive Demo
• Dataset Release
• Training Code

git clone https://github.com/EnVision-Research/UniCalli.git
cd UniCalli
pip install -r requirements.txt

Download the pretrained model from Hugging Face:

# Using huggingface-cli
huggingface-cli download TSXu/UniCalli-base unicalli-base_cleaned.bin --local-dir ./checkpoints

Or from ModelScope:

# Using modelscope
pip install modelscope
python -c "from modelscope import snapshot_download; snapshot_download('tianshuo/UniCalli-base', local_dir='./checkpoints')"

Please note that you need to download additional components to ensure the model runs properly:

# InternVL3-1B:
https://huggingface.co/OpenGVLab/InternVL3-1B

# Fangzheng TTF:
https://www.fonts.net.cn/font-31659110985.html
MD5: 579e8932d773f5f58ebb2c643aa89ba9

You can also use the API directly:

from inference import CalligraphyGenerator

generator = CalligraphyGenerator(
    model_name="flux-dev",
    device="cuda",
    offload=False,
    intern_vlm_path="path/to/InternVL3-1B",
    checkpoint_path="unicalli-base_cleaned.bin",
    font_descriptions_path='dataset/chirography.json',
    author_descriptions_path='dataset/calligraphy_styles_en.json'
)

image, cond_img = generator.generate(
    text="生日快乐喵",  # Must be 5 characters
    font_style="楷",    # 楷(Regular)/草(Cursive)/行(Running)
    author="赵佶",    # Or None to use synthetic style
    save_path="output.png",
    num_steps=39,
    seed=1128293374,
)

For large models or limited GPU memory, you can use DeepSpeed ZeRO:

from inference import CalligraphyGenerator

generator = CalligraphyGenerator(
    model_name="flux-dev",
    device="cuda",
    offload=False,  # DeepSpeed manages memory
    intern_vlm_path="path/to/InternVL3-1B",
    checkpoint_path="unicalli-base_cleaned.bin",
    font_descriptions_path='dataset/chirography.json',
    author_descriptions_path='dataset/calligraphy_styles_en.json',
    use_deepspeed=True,
    deepspeed_config="ds_config_zero2.json"
)

image, cond_img = generator.generate(
    text="生日快乐喵",  # Must be 5 characters
    font_style="楷",    # 楷(Regular)/草(Cursive)/行(Running)
    author="赵佶",    # Or None to use synthetic style
    save_path="output.png",
    num_steps=39,
    seed=1128293374,
)

• 楷 (Regular Script / Kaishu): Standard, block-style characters
• 行 (Running Script / Xingshu): Semi-cursive, flowing style
• 草 (Cursive Script / Caoshu): Highly cursive, artistic style

The model supports various historical calligraphy masters including:

• 王羲之 (Wang Xizhi) - "Sage of Calligraphy"
• 颜真卿 (Yan Zhenqing) - Tang Dynasty master
• 欧阳询 (Ouyang Xun) - One of the Four Great Masters
• 赵佶 (Emperor Huizong) - Song Dynasty emperor and calligrapher
• And many more...

You can also use author=None to generate in a synthetic, averaged style.

• Base Architecture: FLUX diffusion model
• Model Size: ~23GB
• Input: Text (5 characters), font style, author style
• Output: Column-level calligraphy image
• Training Data: Large-scale Chinese calligraphy dataset with dense annotations

If you find UniCalli useful in your research, please consider citing:

@article{xu2025unicalli,
  title={UniCalli: A Unified Diffusion Framework for Column-Level Generation and Recognition of Chinese Calligraphy},
  author={Xu, Tianshuo and Wang, Kai and Chen, Zhifei and Wu, Leyi and Wen, Tianshui and Chao, Fei and Chen, Ying-Cong},
  journal={arXiv preprint arXiv:2025.13745},
  year={2025}
}

This work builds upon the FLUX architecture and benefits from the rich heritage of Chinese calligraphy. We thank the calligraphy masters whose works made this research possible.