An experimental Fang Song style Chinese font generated with skeleton-tracing and pix2pix

Overview

An experimental Fang Song style Chinese font generated with skeleton-tracing and pix2pix, with glyphs based on cwTeXFangSong. The font is optimised for vertical typesetting. Below is a sample:

The font contains roughly 13,000 glyphs, mostly for traditional Chinese.

I created the font for one of my own projects. The font is admittedly not perfect, but nevertheless have many ineteresting features; therefore I am sharing the font file and programs used to generate it.

Download

Download the font directly at dist/tkFangSong.ttf.

The name of the font is 剔骨仿宋 (thek-kwot-fang-song, So named because the algorithm that created it resembles "deboning"). It is licensed under SIL open font license. If you wish to credit the author, you might use my name 黃令東/黄令东, or the romanization "Lingdong Huang".

Features

The font builds on the elegant shapes from cwTeXFangSong to add more hand-made look and feel reminiscent of the aesthetics of old woodblock printed books.

The font has a wider proportion compared to the original cwTeXFangSong, and is further widened towards the bottom, to accentuate the finishing strokes. The "center of mass" is also moved downwards:

Above right is a visualization of the base function used to warp the skeleton.

The height of a glyph is additionally tweaked based on its vertical complexity, computed with Sobel operator and taking the max of each pixel row.

Many fonts are optimized for horizontal typesetting, and as such, when arranged vertically, the center of mass shifts left and right, giving a jagged look. This font attempts to solve the problem by computing centroids (via image moments) and aligning them.

The font has rich textures. Some of them are artifacts produced by pix2pix network; others are fine-tuned noises delibrately added.

It is to be noted that, as an automated process, it doesn't always produce optimal results; some characters might end up looking ugly, or use the wrong caligraphic movement for certain strokes; For some caligraphers, some strokes might appear too "weak" to their tastes.

Process

The medial axis (skeleton) is computed for each raster rendering of the glyphs in the original font. (The resultant hershey font can be found at ./dist/CWFS64J.HF.TXT)

Pairs of images: the original rendering vs the skeletons are sent to pix2pix for training. pix2pix learns the correspondance and becomes capable of turning skeletons to glyphs.

New skeletons are generated by warping the originals according to my (questionable) taste.

All the new skeletons are fed into the trained network to obtain the new glyphs. The new glyphs are warped in structure, but the weight and shape of the strokes still look legit.

Some post-processing is applied, and potrace is used to re-vectorize the glyphs. Finally, fontforge is used to create a TTF file.

Building the font

Note that to use the font, you can simply download it here. This sections is for reproducing the results from scratch.

The scripts used to build the font are included in the workflow/ folder. Note that making the font is a quite involved process (especially the part of training the neural net). You might also need to modify the scripts to fit your system/folder configuration, but here are some rough steps:

  • Get OpenCV, tensorflow<=1.13.1, numpy and friends
  • First install swig version of skeleton-tracing for python, then obtain cwTeXFangSong from here.
  • Run skel.py > CWFS64.HF.TXT, then join.py > CWFS64J.HF.TXT
  • Modify pairs.py to read CWFS64J.HF.TXT and output to a folder you will create.
  • Download pix2pix-tensorflow and train on the images created in the previous step. (Good luck getting a prehistoric tensorflow project running, you'll need it. Once you do it's pretty straightforward, follow their README)
  • I trained 40K steps on 1.3K images, afterwhich the quality didn't seem to improve much, your mileage might vary.
  • Run warp.py > CWFS64W3.HF.TXT. Modify pairs.py to read from it, and create an output folder like before. Run pairs.py.
  • Evaluate the new image pairs with pix2pix. Edit any of the images that have too much defect with Photoshop or software of your choice. Put the edited ones in a separated folder, say retouched/.
  • Modify refine.py to read from the images and retouched images folders, create an output folder fine/ for it, and run the script.
  • Download potrace, make it runnable from commandline, and run trace_all.py.
  • Run forgefont.py to create a TTF from SVGs generated in the previous step.
  • Done! You can also preview the glyphs with preview.py > index.html, or preview the skeleton with preview_hf.py > index.html.

A PDF containing all the glyphs can be found here. If you find this font not bad, you might also enjoy qiji-font, a more authentic reproduction of a historical typeface.

Owner
Lingdong Huang
Lingdong Huang
Add your new words to a text file and get them randomly.

Memorize-New-Words In this very very very little project, I've wrote a code to memorize new english words. Therefore you can add the words and their m

Mostafa 2 Jul 04, 2022
The Levenshtein Python C extension module contains functions for fast computation of Levenshtein distance and string similarity

Contents Maintainer wanted Introduction Installation Documentation License History Source code Authors Maintainer wanted I am looking for a new mainta

Antti Haapala 1.2k Dec 16, 2022
A program that looks through entered text and replaces certain commands with mathematical symbols

TextToSymbolConverter A program that looks through entered text and replaces certain commands with mathematical symbols Example: Syntax: Enter text in

1 Jan 02, 2022
A python tool one can extract the "hash" from a WINDOWS HELLO PIN

WINHELLO2hashcat About With this tool one can extract the "hash" from a WINDOWS HELLO PIN. This hash can be cracked with Hashcat, more precisely with

33 Dec 05, 2022
REST API for sentence tokenization and embedding using Multilingual Universal Sentence Encoder.

MUSE stands for Multilingual Universal Sentence Encoder - multilingual extension (supports 16 languages) of Universal Sentence Encoder (USE).

Dani El-Ayyass 47 Sep 05, 2022
Aml - anti-money laundering

Anti-money laundering Dedect relationship between A and E by tracing through payments with similar amounts and identifying payment chains. For example

3 Nov 21, 2022
Python port of Google's libphonenumber

phonenumbers Python Library This is a Python port of Google's libphonenumber library It supports Python 2.5-2.7 and Python 3.x (in the same codebase,

David Drysdale 3.1k Dec 29, 2022
Implementation of hashids (http://hashids.org) in Python. Compatible with Python 2 and Python 3

hashids for Python 2.7 & 3 A python port of the JavaScript hashids implementation. It generates YouTube-like hashes from one or many numbers. Use hash

David Aurelio 1.4k Jan 02, 2023
Amazing GitHub Template - Sane defaults for your next project!

🚀 Useful README.md, LICENSE, CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, GitHub Issues and Pull Requests and Actions templates to jumpstart your projects.

276 Jan 01, 2023
Auto translate Localizable.strings for multiple languages in Xcode

auto_localize Auto translate Localizable.strings for multiple languages in Xcode Usage put your origin Localizable.strings file in folder pip3 install

Wesley Zhang 13 Nov 22, 2022
🍋 A Python package to process food

Pyfood is a simple Python package to process food, in different languages. Pyfood's ambition is to be the go-to library to deal with food, recipes, on

Local Seasonal 8 Apr 04, 2022
Hotpotato is a recipe portfolio App that assists users to discover and comment new recipes.

Hotpotato Hotpotato is a recipe portfolio App that assists users to discover and comment new recipes. It is a fullstack React App made with a Redux st

Nico G Pierson 13 Nov 05, 2021
PyMultiDictionary is a Dictionary Module for Python 3+ to get meanings, translations, synonyms and antonyms of words in 20 different languages

PyMultiDictionary PyMultiDictionary is a Dictionary Module for Python 3+ to get meanings, translations, synonyms and antonyms of words in 20 different

Pablo Pizarro R. 19 Dec 26, 2022
LazyText is inspired b the idea of lazypredict, a library which helps build a lot of basic models without much code.

LazyText is inspired b the idea of lazypredict, a library which helps build a lot of basic models without much code. LazyText is for text what lazypredict is for numeric data.

Jay Vala 13 Nov 04, 2022
An extension to detect if the articles content match its title.

Clickbait Detector An extension to detect if the articles content match its title. This was developed in a period of 24-hours in a hackathon called 'H

Arvind Krishna 5 Jul 26, 2022
You can encode and decode base85, ascii85, base64, base32, and base16 with this tool.

You can encode and decode base85, ascii85, base64, base32, and base16 with this tool.

8 Dec 20, 2022
基于Pytex的数学建模工具,实现将md文件转换成pdf/tex文档的前后端

Pytex-for-MCM 基于Pytex的数学建模工具,实现将md文件转换成pdf/tex文档的前后端。

3 May 17, 2021
A python tool to convert Bangla Bijoy text to Unicode text.

Unicode Converter A python tool to convert Bangla Bijoy text to Unicode text. Installation Unicode Converter can be installed via PyPi. Make sure pip

Shahad Mahmud 10 Sep 29, 2022
pydantic-i18n is an extension to support an i18n for the pydantic error messages.

pydantic-i18n is an extension to support an i18n for the pydantic error messages

Boardpack 48 Dec 21, 2022
Answer some questions and get your brawler csvs ready!

BRAWL-STARS-V11-BRAWLER-MAKER-TOOL Answer some questions and get your brawler csvs ready! HOW TO RUN on android: Install pydroid3 from playstore, and

9 Jan 07, 2023