Paper: De-rendering Stylized Texts

Wataru Shimoda¹, Daichi Haraguchi², Seiichi Uchida², Kota Yamaguchi^1
1CyberAgent.Inc, ² Kyushu University
Accepted to ICCV2021. [Publication] [Arxiv] [project-page]

Introduction

This repository contains the codes for "De-rendering stylized texts".

Concept

We propose to parse rendering parameters of stylized texts utilizing a neural net.

Demo

The proposed model parses rendering parameters based on famous 2d graphic engine[Skia.org|python implementation], which has compatibility with CSS in the Web. We can export the estimated rendering parameters and edit texts by an off-the-shelf rendering engine.

Installation

Requirements

• Python >= 3.7
• Pytorch >= 1.8.1
• torchvision >= 0.9.1

pip install -r requirements.txt

Font data

• The proposed model is trained with google fonts.
• Download google fonts and locate in data/fonts/ as gfonts.
  • Note: the organization of font files in the google fonts is updated from our environment.
• Download font files from this link(ofl) and locate in data/fonts/gfonts/.

- cd　data/fonts
- git clone https://github.com/google/fonts.git gfonts
+ mkdir data/fonts/gfonts; cd data/fonts/gfonts
+ tar xvzf ofl.tar.gz

Pre-rendered alpha maps

• The proposed model parses rendering parameters and refines them through the differentiable rendering model, which uses pre-rendered alpha maps.
• Generate pre-rendered alpha maps.

python -m util_lib.gen_pams

Pre-rendered alpha maps would be generated in data/fonts/prerendered_alpha.

Usage

Test

• Download the pre-trained weight from this link (weight).
• Locate the weight file in weights/font100_unified.pth.

Example usage.

python test.py --imgfile=example/sample.jpg

Note

• imgfile option: path of an input image
• results would be generated in res/

Text image editing

The proposed model generates a reconstructed image and a pickle file for the parsed rendering parameters.
Here, we prepare a notebook file:text_edit.ipynb for the guide of the processings to edit text images using the parsed rendering parameters.

Some examples from text_edit.ipynb:

Background editing

Text editing

Border effect editing

Shadow effect editing

Text offsets editing

Font editing

Data generation

Quick start.

python gen.py --bgtype=load --bg_dir=src/modules/generator/example/bg --mask_dir=src/modules/generator/example/mask

The generated text images would be located in gen_data/.

For the detail, see generator.

Train text parser model

Quick start. Generate training data using simple background dataset.

python gen.py --bgtype=color

Train text parser model with the generated simple background data.

python train.py

For the detail, see trainer.

Attribute details

Todo

• Testing codes
• Codes for the text image generator
• Notebook for text editing
• Training codes for text paraser model
• Training codes for inpainting model
• Demo app

Reference

@InProceedings{Shimoda_2021_ICCV,
    author    = {Shimoda, Wataru and Haraguchi, Daichi and Uchida, Seiichi and Yamaguchi, Kota},
    title     = {De-Rendering Stylized Texts},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {1076-1085}
}

Contact

This repository is maintained by Wataru shimoda(wataru_shimoda[at]cyberagent.co.jp).