First Creating Backgrounds Then Rendering Texts: A New Paradigm for Visual Text Blending

<a href='https://arxiv.org/abs/2410.10168'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a> <a href='https://github.com/Zhenhang-Li/GlyphOnly/'><img src='https://img.shields.io/badge/Code-Github-green'></a>

TODOs

• [x] Release SynthGlyph10K and Better_Background
• [ ] Release inference code and checkpoints
• [x] Release training code
• [x] Release training dataset and test dataset

Datasets

A better background image for generating scene text images is available at Google Drive.

SynthGlyph10K is available at Google Drive.

Instance categories suitable for text generation on surfaces can be obtained at Google Drive.

Train dataset is available at Google Drive.

Test dataset is available at Google Drive.

Checkpoint is available at Google Drive

Installation

Clone this repo:

git clone https://github.com/Zhenhang-Li/GlyphOnly.git
cd GlyphOnly

Build up a new environment and install packages as follows:

conda create -n glyphonly python=3.8
conda activate glyphonly
pip install -r requirements.txt
cd diffusers
pip install -e .

Inference

Training

Download all subfolders under the following folder to the current directory. Google Drive

bash train.sh

Evaluation

Citation

@incollection{li2024first,
  title={First Creating Backgrounds Then Rendering Texts: A New Paradigm for Visual Text Blending},
  author={Li, Zhenhang and Shu, Yan and Zeng, Weichao and Yang, Dongbao and Zhou, Yu},
  booktitle={ECAI 2024},
  pages={346--353},
  year={2024},
  publisher={IOS Press}
}