TGC-Diff

KO | EN

gitlite — search

#python #hacktoberfest #react #ai #llm #go #rust #docker #macos #vue #awesome-list #security #ios #open-source

(★ 9)

ICDAR2025 Best Paper "Template-Guided Cascaded Diffusion for Stylized Handwritten Chinese Text-Line Generation"

TGC-Diff Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<a href=""><img src="https://img.shields.io/badge/Pytorch-2.5.1-green"></a>
<a href="https://github.com/newbie2niubility/TGC-Diff/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue"></a>
  
<h2 align="center">Template-Guided Cascaded Diffusion for Stylized Handwritten Chinese Text-Line Generation</a></h2>

<p align="center">
  🏆 <b>ICDAR 2025 Best Paper Award</b>
</p>

<div align="center">
  <a href="https://link.springer.com/chapter/10.1007/978-3-032-04614-7_9">Paper</a> 
</div>

## 🌟 Introduction
- We propose a novel cascaded diffusion framework designed to disentangle character-level and text-line-level styles. This framework comprises two models: Layout-Diff, which learns spatial styles, and Imitating-Diff, which mimics handwritten styles. These models are systematically integrated through a Template construct.
- We propose a latent aggregation module that integrates content as well as calligraphy style information into Imitating-Diff, effectively preserving the spatial layout and structural coherence of characters within the text-line while imitating the target writing style.
- We refine Imitating-Diff in the pixel space by incorporating a L1 loss weighted by a high-frequency mask, which enhances the clarity of fine-grained character details, thereby improving overall text-line readability.
- Extensive experiments on Chinese handwritten text-line datasets validate the effectiveness of our method, demonstrating robust performance in generating stylized handwriting at the line level.

<p align="center" style="margin-bottom: 10px;">
Overview of the proposed TGC-Diff
</p>

## 🔨 Requirements
```
conda create -n tgc_diff python=3.10 -y
conda activate tgc_diff
# install all dependencies
conda env create -f environment.yml
```
## ☀️ Datasets
We provide Chinese text-line datasets in [Baidu Netdisk](https://pan.baidu.com/s/1egncerfRPtaHNi3JYP43cg?pwd=qgd5) | [Google Drive](https://drive.google.com/file/d/1IAyfVImGdAPMAufQJHpBSdr-Afa2AAO_/view?usp=sharing). Please download these datasets, uzip them and move the extracted files to ./data.

We provide Chinese text-line template label datasets in [Baidu Netdisk](https://pan.baidu.com/s/1z6LXt00BEU3KKl0dpxLuiQ?pwd=knax) | [Google Drive](https://drive.google.com/file/d/14uHggU_ztBvxHGevld9jF1MCKQcnIYkd/view?usp=sharing). Please download these datasets, uzip them and move the extracted files to ./template/temp_data.

## 🐳 Model Zoo
|Model|Baidu Netdisk|Google Drive|
|---------------|--------------------------------|---------|
|Template Generation Model|[Baidu Netdisk](https://pan.baidu.com/s/102idFeSxBdrBMID-44gwhg?pwd=kdhi)|[Google Drive](https://drive.google.com/file/d/1rfH6ZFty-RQFtFf6_XwgXG8ZEx4njcvI/view?usp=sharing)|
|Pretrained VAE plus Resnet18|[Baidu Netdisk](https://pan.baidu.com/s/1-kIoYVE2qGsto94eRe3ANg?pwd=k5fi)|[Google Drive](https://drive.google.com/file/d/1MA0ndX0kdMik11klXBfbx3MFb0wXgyP7/view?usp=sharing)|
|Finetuned Imitating-Diff|[Baidu Netdisk](https://pan.baidu.com/s/1Bc8GpfcL1EBT2SYolRBdkw?pwd=wz4p)|[Google Drive](https://drive.google.com/file/d/1fjGTLNjXt7tVza1ORllxcuakFBWo5bZO/view?usp=sharing)|

**Note**:
Please download these weights. Move Template Generation Model to ./template/.  Move Pretrained VAE plus Resnet18 to ./.cache. Create a new folder ./weight and move finetuned Imitating-Diff to ./weight.
## 🏋️ Training
- **training on Chinese text-line dataset**
```Shell
sh train.sh
```
- **finetune on Chinese text-line dataset**
```Shell
sh finetune.sh
 ```
- **Inference on Chinese text-line dataset**
```Shell
sh test_ch.sh
```

- **training on Chinese template dataset**
```Shell
cd ./template
sh train_ch_temp.sh
 ```
 - **Inference on Chinese template dataset**
```Shell
cd ./template
sh sample.sh
 ```

**Note**:
Please modify ``--pre_train`` according to your own path.

## 📺 Exhibition
- **Comparisons with other methods on handwritten text-line generation**
<p align="center">
<img src="assets/comparison.png" style="width: 80%" align=center>
</p>

- **Style interpolation**
<p align="center">
<img src="assets/style_interpolation.png" style="width: 80%" align=center>
</p>

- **More visualization**
<p align="center">
<img src="assets/more_visual.png" style="width: 80%" align=center>
</p>

## ❤️ Citation
If you find our work inspiring or use our codebase in your research, please cite our work:
```
@InProceedings{TGC-Diff,
author="Wang, Honglie
and Ren, Minsi
and Zhang, Yan-Ming
and Yin, Fei
and Liu, Cheng-Lin",
title="Template-Guided Cascaded Diffusion for Stylized Handwritten Chinese Text-Line Generation",
booktitle="Document Analysis and Recognition -- ICDAR 2025",
year="2025",
publisher="Springer Nature Switzerland",
address="Cham",
pages="149--166",
}
```

TGC-Diff

TGC-Diff Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

TGC-Diff

TGC-Diff Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README