GANs and Style Transfer to the Next Level

Style Transfer

Style Transfer

Previous Articles:
Generative Adversarial Networks, what are they and how do they work?
Style Transfer Using a Pre-Trained Convolutional Neural Network

In previous articles, I talked about GANs and CNNs in relation to Style Transfer techniques. Typically during Style Transfer, you are taking a target picture and applying a style to it so that it still resembles or is inspired by the original target picture. For example you could take a photo and apply an art style based on artworks of Van Gogh to get a new picture that looks very similar to “Starry Night”. A recent paper published in 2018 by three NVIDIA AI engineers, Tero Karras, Samuli Laine and Timo Aila, discuss their architecture using a GAN to recreate photo realistic pictures of people’s face with extreme granular control over style. They demonstrated with a source image, you could tweak the style to different attributes such as pose, identity, facial features, eyes and color scheme. The results were impossible to distinguish between AI generated photos and real source photos (in my opinion). Be your own judge and check out the video below.


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.