SGD-Font

---

This repo is the official implementation of paper, SGD-Font: Style and Glyph Decoupling for One-Shot Font Generation.

Dependencies

pytorch>=1.13.0
opencv-python
sklearn
pillow
tqdm
blobfile>=1.0.5
mpi4py
attrdict
yaml
lpips
pytorch-fid
fonttools

Dataset

We obtain fonts from the following font platforms under a personal non-commercial academic research license:

1. Foundertype (https://www.foundertype.com);
2. Font Meme (https://fontmeme.com/ziti/).

Example directory hierarchy

data_dir
    Seen_TRAIN
        |--- font1
        |--- font2
           |--- 0000.png
           |--- 0001.png
           |--- ...
        |--- ...
    Seen_TEST
    Unseen_TRAIN
    Unseen_TEST
    Source_TRAIN
    Source_TEST

Usage

Prepare dataset

python font2img.py --ttf_path ttf_folder --chara char.txt --save_path save_folder --img_size 80 --chara_size 50

Train

Modify the configuration file cfg/train_cfg.yaml, and then run

• single gpu

  python train.py --cfg_path cfg/train_cfg.yaml

• distributed training

  mpiexec -n $NUM_GPUS python train.py --cfg_path cfg/train_cfg.yaml

Test

Modify the configuration file cfg/test_cfg.yaml, and then run

bash sample.sh

Evaluate

Modify the path to your result, and run

bash evaluate.sh

Visual display

（1）Seen Font Unseen Content
$~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$ Reference $~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$ Result

（2）Unseen Font Unseen Content
$~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$ Reference $~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$ Result

Acknowledgements

---

This project is based on openai/guided-diffusion and CF-Font