[ICLR26] UniCalli: A Unified Diffusion Framework for Column-Level Generation and Recognition of Chinese Calligraphy

English | 简体中文

Overview

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.

Key Features

• 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

Licence

For academic research and non-commercial use only.

本模型仅供学术研究、非商业使用。

TODO List

• Model Release - Base version without pred_box
• Inference Code (Generation)
• 4-bit Quantization - Only requires 18G GPU memory!
• Interactive Demo
• Dataset Release - UniCalli_dataset
• Inference Code (Recognition)
• Training Code

Getting Started

Installation

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

Download Model

Download the complete model package (includes checkpoint, InternVL embedding, and font) from Hugging Face:

# Using huggingface-cli (recommended)
huggingface-cli download TSXu/UniCalli-base --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')"

Usage

Run Demo (Gradio Interface)

python app.py

4-bit Quantization (GPU Memory < 18GB)

⚠️ Note: 4-bit quantization significantly reduces VRAM usage but may result in slightly degraded output quality.

pip install optimum-quanto

from inference import CalligraphyGenerator

generator = CalligraphyGenerator(
    model_name="flux-dev",
    device="cuda",
    offload=False,
    intern_vlm_path="./checkpoints/internvl_embedding",  # Path to downloaded embedding
    checkpoint_path="./checkpoints/unicalli-base_cleaned.bin",
    font_descriptions_path='dataset/chirography.json',
    author_descriptions_path='dataset/calligraphy_styles_en.json',
    use_deepspeed=False,
    use_4bit_quantization=True,  # Enable 4-bit quantization
)

image, cond_img = generator.generate(
    text="生日快乐喵",  # Must be 5 characters
    font_style="草",    # 楷(Regular)/草(Cursive)/行(Running)
    author="黄庭坚",    # Or None to use synthetic style
    save_path="output.png",
    num_steps=25,
    seed=42,
)

Using DeepSpeed for Memory Optimization (GPU Memory < 40G)

from inference import CalligraphyGenerator

generator = CalligraphyGenerator(
    model_name="flux-dev",
    device="cuda",
    offload=False,  # DeepSpeed manages memory
    intern_vlm_path="./checkpoints/internvl_embedding",  # Path to downloaded embedding
    checkpoint_path="./checkpoints/unicalli-base_cleaned.bin",
    font_descriptions_path='dataset/chirography.json',
    author_descriptions_path='dataset/calligraphy_styles_en.json',
    use_deepspeed=True,
    use_4bit_quantization=False,
    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,
)

Supported Font Styles

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

Supported Calligraphy Masters

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.

Model Details

• 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

Citation

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}
}

Acknowledgments

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.