shahdsaf / weirdgan

Artistic Generative Neural Network (GAN), so-called WeirdGAN, generating alien and weird fascinating images.

Artistic GAN with TensorFlow and Keras

A novel artistic GAN that generates alien interestingly-looking objects as a result of the input dataset we've created to train our GAN on and bias the generator to create such fascinating novel mixed colors as well.

This work has been shown in the exhibition of Future Proof at Galerie Postel in Hamburg!

Requirement

To run the code, you need

• keras
• TensorFlow

WeirdGAN components

Dataset:

The dataset is a collection of weirdly-looking alien-like objects including space rocks from different point of views, abstract artworks etc. The dataset was fully prepared by Ben Nurgenc, an artist that initiated this project's idea.

Models:

• 64x64 GAN: This model uses a GAN model for generating images of size 64x64.
• 256x256 GAN: For generating images of size 256x256.

Training

To train the network, run python <model_version>_version/training.py.

To load the model and create test images in inference phase, see the /256x256_scaled_version/test_notebook.ipynb notebook.

Results

Please refer to the project samples PDF file under docs/ for more insight into our dataset and training observations.

Dataset samples:

Generated samples:

Acknowledgement

Our project relied highly on the following repo:

• https://github.com/javiermzll/GAN-Dog-Generator

Authors

Shahd Safarani and Ben Nurgenc