Philip Jama

Artificial Impressionism

Training a deep neural network model to create machine-generated art by transferring painting styles to photography.

Project Overview

Artistic style transfer techniques combine the content of one image with the style of another through a novel use of a deep neural networks. In this project I explore Logan Engstrom's fast-style-transfer model, which is based on the research paper by Johnson et al, Perceptual Losses for Real-Time Style Transfer and Super-Resolution.

The model optimizes a loss function consisting of three terms: feature loss, style loss, and a regularization term,

  1. The 'Feature Reconstruction' loss term leverages a network that has been pre-trained for image classification (in this case VGG-19 based on this research paper). This encourages the 'content network' to have similar feature representation as the pre-trained net.
  2. The style loss is computed using a Gram matrix, which provides a 'good representation of our perception of style within images.' [1]
  3. Total Variation Regularization is used to encourage spatial smoothness.

This approach definitely produces some of the best results that I have seen so far. Enjoy the following renderings. (Click to enlarge.)

Project Details


August 2017


Model Training


Machine Learning, TensorFlow, Deep Learning, Python, Photography