Few shot font generation via transferring similarity guided global and quantization local styles（ICCV2023）

Official Pytorch Implementation of "Few shot font generation via transferring similarity guided global and quantization local styles" by Wei Pan, Anna Zhu*, Xinyu Zhou, Brian Kenji Iwana, and Shilin Li.

Our method is based on Vector Quantization, so we named our FFG method VQ-Font.

Paper can be found at ./Paper_IMG/ | Arxiv｜CVF.

Abstract

Automatic few-shot font generation (AFFG), aiming at generating new fonts with only a few glyph references, reduces the labor cost of manually designing fonts. However, the traditional AFFG paradigm of style-content disentanglement cannot capture the diverse local details of different fonts. So, many component-based approaches are proposed to tackle this problem. The issue with component-based approaches is that they usually require special pre-defined glyph components, e.g., strokes and radicals, which is infeasible for AFFG of different languages. In this paper, we present a novel font generation approach by aggregating styles from character similarity-guided global features and stylized component-level representations. We calculate the similarity scores of the target character and the referenced samples by measuring the distance along the corresponding channels from the content features, and assigning them as the weights for aggregating the global style features. To better capture the local styles, a cross-attention-based style transfer module is adopted to transfer the styles of reference glyphs to the components, where the components are self-learned discrete latent codes through vector quantization without manual definition. With these designs, our AFFG method could obtain a complete set of component-level style representations, and also control the global glyph characteristics. The experimental results reflect the effectiveness and generalization of the proposed method on different linguistic scripts, and also show its superiority when compared with other state-of-the-art methods.

The model receives several style reference characters (from the target style) and content characters (from the source font) to generate style-transformed characters.

Usage

Dependencies

python >= 3.7
torch >= 1.12.0
torchvision >= 0.13.0
sconf >= 0.2.5
lmdb >= 1.2.1

Data Preparation

Images and Characters

1) Collect a series of '.ttf'(TrueType) or '.otf'(OpenType) files to generate images for training models. and divide them into source content font and training set and test set. In order to better learn different styles, there should be differences and diversity in font styles in the training set. The fonts we used in our paper can be found in here.

2) Secondly, specify the characters to be generated (including training characters and test characters), eg the first-level Chinese character table contains 3500 Chinese characters.

trian_val_3500: {乙、十、丁、厂、七、卜、人、入、儿、匕、...、etc}
train_3000: {天、成、在、麻、...、etc}
val_500: {熊、湖、战、...、etc}

3) Convert the characters in the second step into unicode encoding and save them in json format, you can convert the utf8 format to unicode by using hex(ord(ch))[2:].upper():, examples can be found in ./meta/.

trian_val_all_characters: ["4E00", "4E01", "9576", "501F", ...]
train_unis: ["4E00", "4E01", ...]
val_unis: ["9576", "501F", ...]

4) After that, draw all font images via ./datasets/font2image.py. All images are named by 'characters + .png', such as ‘阿.png’. Organize directories structure as below, and train_3000.png means draw the image from train_unis: ["4E00", "4E01", ...].

Font Directory
|--| content
| --| kaiti4train_VAE
| --| train_3000.png
| --| ...
| --| kaiti4val_VAE
| --| val_500.png
| --| ...
| --| kaiti4train_FFG
| --| trian_val_3500.png
| --| ...
|--| train
| --| train_font1
| --| train_font2
| --| trian_val_3500.png
| --| ...
| --| ...
|--| val
| --| val_font1
| --| val_font2
| --| trian_val_3500.png
| --| ...
| --| ...

Build meta files and lmdb environment

Run script ./build_trainset.sh

  python3 ./build_dataset/build_meta4train.py \
  --saving_dir ./results/your_task_name/ \
  --content_font path\to\all_content \
  --train_font_dir path\to\training_font \
  --val_font_dir path\to\validation_font \
  --seen_unis_file path\to\train_unis.json \
  --unseen_unis_file path\to\val_unis.json

Training

The training process is divided into two stages: 1）Pre-training the content encoder and codebook via VQ-VAE, 2）Training the few shot font generation model via GAN.

Pre-train VQ-VAE

When pre-training VQ-VAE, the reconstructed character object comes from train_unis in the content font, The training process can be found at ./model/VQ-VAE.ipynb.

Then use the pre-trained content encoder to calculate a similarity between all training and test characters and store it as a dictionary.

{'4E07': {'4E01': 0.2143, '4E03': 0.2374, ...}, '4E08': {'4E01': 0.1137, '4E03': 0.1020, ...}, ...}