justbiubiu/tet-gan

TET-GAN: Text Effects Transfer via Stylization and Destylization. AAAI 2019

TET-GAN

This is a pytorch implementation of the paper.

Shuai Yang, Jiaying Liu, Wenjing Wang and Zongming Guo. TET-GAN: Text Effects Transfer via Stylization and Destylization, Accepted by AAAI Conference on Artificial Intelligence (AAAI), 2019.

[Project] | [Paper] | [Dataset]

It is provided for educational/research purpose only. Please consider citing our paper if you find the software useful for your work.

Usage:

Prerequisites

• Python 2.7
• Pytorch 0.4.1
• matplotlib
• scipy
• opencv-python
• Pillow

Install

• Clone this repo:

git clone https://github.com/williamyang1991/TET-GAN.git
cd TET-GAN/src

Testing Example

• Download a pre-trained model from [Google Drive] or [Baidu Cloud] to ../save/
• Style Transfer with default parameters
  • Results can be found in ../output/

python test.py

• Destylization with default parameters

python test.py --c2s 0

• Transfer the style of 26.jpg onto the text image 2.png and save the result as 26_2.png

python test.py --style_name ../data/style/26.png --content_name ../data/content/2.png --name 26_2.png

• For black and white text images, use option --content_type 1

python test.py --style_name ../data/style/1.png --content_name ../data/content/4.png --content_type 1

Training Examples

Full Training

• Prepare dataset in ../data/dataset/ where the images are arranged in this way (See [Dataset] for one example)

dataset/style1/train/1.png
dataset/style1/train/2.png
...
dataset/style2/train/1.png
dataset/style2/train/2.png
...

• The training image consists of a text image and a style image as follows

• Train TET-GAN with default parameters

python train.py 

Saved model can be found at ../save/

• Use --help to view more training options

python train.py --help

Oneshot Training

• Download a pre-trained model from [Google Drive] or [Baidu Cloud] to ../save/

  • Specify the pretrained model to load using the option --load_model_name

• Finetune TET-GAN on a new style/glyph image pair (supervised oneshot training)

python oneshotfinetune.py --style_name ../data/oneshotstyle/1-train.png

Saved model can be found at ../save/

• Finetune TET-GAN on a new style image without its glyph counterpart (unsupervised oneshot training)

python oneshotfinetune.py --style_name ../data/oneshotstyle/1-train.png --supervise 0

Saved model can be found at ../save/

• Use --help to view more finetuning options

python oneshotfinetune.py --help

Contact

Shuai Yang

[email protected]