ChipGAN-PyTorch

Visit the Original Code & Paper: This is the latest code for paper "ChipGAN: A Generative Adversarial Network for Chinese Ink Wash Painting Style Transfer". Original work by Bin He, Feng Gao etc.

Check the original code and paper at: CODE | PAPER

Description

In the paper above, He provides a new way to generate Chinese ink wash painting by using Generative Adversarial Network(GAN). The core modules of ChipGAN enforce three constraints – voids, brush strokes, and ink wash tone and diffusion – to address three key techniques commonly adopted in Chinese ink wash painting.

Example outputs

Dependencies

Tested with:

• OS: Windows 10/11 or Ubuntu 22.04 (Recommended)
• PyTorch 1.13
• Python 3.7

Create Environment

git clone https://github.com/Xzzit/ChipGAN-pytorch.git
cd ChipGAN-pytorch
conda create -n chipgan python=3.7
conda activate chipgan
pip install -r requirements.txt

Painting with Pre-trained Models

Download pre-trained models from ChipGAN Models

The downloaded file should be placed as following:

.
├── ...
├── saved_models                # Store all pre-trained models
│   ├── criticA.pth.tar             # Detect whether a image is a Landscape paintings or not
│   ├── criticB.pth.tar             # Detect whether a image is a Inkwash paintings or not
│   ├── criticINK.pth.tar           # Detect whether a image is a Blurred Inkwash painting or not
│   ├── genA.pth.tar                # Take inkwash painting as input and generate Landscape paintings
│   ├── genB.pth.tar                # Take Landscape painting as input and generate inkwash paintings
│   ├── hed-bsds500                 # Edge detection & blur
│   └── ...                         # Ohter models trained by user are also stored in here
└── saved_images                # Store all input and generated images

Run paint.py file to generate your ink wash painting images.

In this file, img_dir should be changed to the input image directory.

Open the file to read more details and I have written code comments for you. Have fun!

Training ChipGAN Models

Two datasets are needed to train ChipGAN. One is the landscape dataset(DatasetA), another is the ink wash paintings dataset(DatasetB).

For the DatasetA, I choose them from Kaggle|Landscape Pictures.

For the DatasetB, I choose them from Traditional Chinese Landscape Painting Dataset

The downloaded file should be placed as following:

.
├── ...
├── saved_models
│   ├── ...
├── saved_images
│   ├── ...
└── data                        # Store all training images
    ├── trainA                      # Dataset of landscape
    └── trainB                      # Dataset of ink wash painting

Run train.py to train ChipGAN models.

Issues

As shown in the example, the generated images has notable noisey and repeated pattern.

If you know what cause this issues, please contact me @[email protected] or post the solution in issues.