How does it work ?

Brief description of how the paintings are generated

Principles

The paintings images presented on the site are generated with a model trained on more than 40,000 paintings thanks to Deep Learning. The model uses Junho Kim's excellent implementation of the StyleGAN algorithm developed by Nvidia. This implementation uses the TensorFlow machine learning library.

StyleGAN is a variation of GAN deep learning algorithms (Generative adversarial networks). It consists of training two models: a generator and a discriminator. The first one is trained to produce samples while the second one is trained to decide whether the produced sample is valid or not. Here, the samples produced are the images of the paintings. The discriminator therefore tries to detect whether an image of a painting is true or not while the generator produces images with the objective of cheating the discriminator.

The produced result is not intended to be perfect but will be improved over time, as the main purpose of this site is to develop Machine Learning skills. Model training requires high computing power and is therefore long and expensive.

Links

For more information, you can check the following links: