/ShapeMatchingGAN

[ICCV 2019, Oral] Controllable Artistic Text Style Transfer via Shape-Matching GAN

Primary LanguageJupyter NotebookMIT LicenseMIT

ShapeMatchingGAN

sourceadjustable stylistic degree of glyphstylized textapplication
liquid artistic text renderingsmoke artistic text rendering

This is a pytorch implementation of the paper.

Shuai Yang, Zhangyang Wang, Zhaowen Wang, Ning Xu, Jiaying Liu and Zongming Guo. Controllable Artistic Text Style Transfer via Shape-Matching GAN, accepted by International Conference on Computer Vision (ICCV), 2019.

[Project] | [Paper] | More about artistic text style transfer [Link]

Please consider citing our paper if you find the software useful for your work.

Usage:

Prerequisites

  • Python 2.7
  • Pytorch 1.1.0
  • matplotlib
  • scipy
  • Pillow

Install

  • Clone this repo:
git clone https://github.com/TAMU-VITA/ShapeMatchingGAN.git
cd ShapeMatchingGAN/src

Testing Example

  • Download pre-trained models from [Google Drive] or [Baidu Cloud](code:rjpi) to ../save/
  • Artisic text style transfer using fire style with scale 0.0
    • Results can be found in ../output/

python test.py \
--scale 0.0
--structure_model ../save/fire-GS-iccv.ckpt \
--texture_model ../save/fire-GT-iccv.ckpt \
--gpu
  • Artisic text style transfer with specified parameters
    • setting scale to -1 means testing with multiple scales in [0,1] with step of scale_step
    • specify the input text name, output image path and name with text_name, result_dir and name, respectively
python test.py \
--text_name ../data/rawtext/yaheiB/val/0801.png \
--scale -1 --scale_step 0.2 \
--structure_model ../save/fire-GS-iccv.ckpt \
--texture_model ../save/fire-GT-iccv.ckpt \
--result_dir ../output --name fire-0801 \
--gpu

or just modifying and running

sh ../script/launch_test.sh
  • For black and white text images, use option --text_type 1
    • utils.text_image_preprocessing will transform BW images into distance-based images
    • distance-based images make the network better deal with the saturated regions

Training Examples

Training Sketch Module G_B

  • Download text dataset from [Google Drive] or [Baidu Cloud](code:rjpi) to ../data/

  • Train G_B with default parameters

    • Adding augmented images to the training set can make G_B more robust
python trainSketchModule.py \
--text_path ../data/rawtext/yaheiB/train --text_datasize 708 \
--augment_text_path ../data/rawtext/augment --augment_text_datasize 5 \
--batchsize 16 --Btraining_num 12800 \
--save_GB_name ../save/GB.ckpt \
--gpu

or just modifying and running

sh ../script/launch_SketchModule.sh

Saved model can be found at ../save/

  • Use --help to view more training options
python trainSketchModule.py --help

Training Structure Transfer G_S

  • Train G_S with default parameters
    • step1: G_S is first trained with a fixed l = 1 to learn the greatest deformation
    • step2: we then use l ∈ {0, 1} to learn two extremes
    • step3: G_S is tuned on l ∈ {i/K}, i=0,...,K where K = 3 (i.e. --scale_num 4)
    • for structure with directional patterns, training without --Sanglejitter will be a good option
python trainStructureTransfer.py \
--style_name ../data/style/fire.png \
--batchsize 16 --Straining_num 2560 \
--step1_epochs 30 --step2_epochs 40 --step3_epochs 80 \
--scale_num 4 \
--Sanglejitter \
--save_path ../save --save_name fire \
--gpu

or just modifying and running

sh ../script/launch_ShapeMGAN_structure.sh

Saved model can be found at ../save/

  • To preserve the glyph legibility (Eq. (7) in the paper), use option --glyph_preserve
    • need to specify the text dataset --text_path ../data/rawtext/yaheiB/train and --text_datasize 708
    • need to load pre-trained G_B model --load_GB_name ../save/GB-iccv.ckpt
    • in most cases, --glyph_preserve is not necessary, since one can alternatively use a smaller l
  • Use --help to view more training options
python trainStructureTransfer.py --help

Training Texture Transfer G_T

  • Train G_T with default parameters
    • for complicated style or style with directional patterns, training without --Tanglejitter will be a good option
python trainTextureTransfer.py \
--style_name ../data/style/fire.png \
--batchsize 4 --Ttraining_num 800 \
--texture_step1_epochs 40 \
--Tanglejitter \
--save_path ../save --save_name fire \
--gpu

or just modifying and running

sh ../script/launch_ShapeMGAN_texture.sh

Saved model can be found at ../save/

  • To train with style loss, use option --style_loss
    • need to specify the text dataset --text_path ../data/rawtext/yaheiB/train and --text_datasize 708
    • need to load pre-trained G_S model --load_GS_name ../save/fire-GS.ckpt
    • adding --style_loss can slightly improve the texture details
  • Use --help to view more training options
python trainTextureTransfer.py --help

More

Three training examples are in the IPythonNotebook ShapeMatchingGAN.ipynb

Have fun :-)

Try with your own style images

  • Style image preparation
    • Applicable style types: To make the stylized text easy to recognize, it is desirable to have a certain distinction between the text and the background. If the texture has no distinct shape, the generated stylized text will be mixed with the background. Therefore, textures with distinct shapes as the reference style are recommended.
    • Prepare (X,Y): Use Image Matting Algorithm or the Quick Selection Tool in Photoshop to obtain the black and white structure map X (i.e. foreground mask) of the style image Y.
    • Prepare distance-based structure map: Use utils.text_image_preprocessing to transform black and white X into distance-based X.
    • Concatenate distance-based X with Y as the format of images in ../data/style/ and copy the result to ../data/style/.

Contact

Shuai Yang

williamyang@pku.edu.cn