The trained GAN can create data similar to the training data, and various expressive abilities are acquired inside the GAN to generate data. We utilized this expressive ability to develop a new expressive method. The gif image below is an example of a new representation image using GAN. Stained glass images can be obtained by inputting special noises instead of the GAN intermediate representation learned from CelebA images.

Details will be added later.

• 64-bit Python 3.6 installation with numpy 1.13.3 or newer. We recommend Anaconda3.
• NVIDIA driver 385 or newer
• Tensorflow 1.2 or newer
• numpy>=1.13.3
• Pillow>=3.1.1

A minimal example of using a pre-trained Generative_Art_with_GAN generator is given in generate_art_minimal.py.

Before executed, you must download a learned model from Google Drive: https://drive.google.com/open?id=1N9AvyKx6atBy-Ewcmcd3EyVZwXT3xQ-t

Put the downloaded checkpoints in the same directory as generate_art_minimal.py.

The execution command is python generate_art_minimal.py

Once the datasets are set up, you can train your own Generative_Art_with_GAN networks as follows:

1. Edit train_celeba_wgan_gp_pos.py to specify the dataset directory configuration by editing specific lines.
2. Run the training script with python train_celeba_wgan_gp_pos.py.
3. The results are written to a newly created directory ./sample_images_while_training/<image name>.
4. The training may take several hours to complete.