Style Transfer with Keras

All art is at once surface and symbol.
All art is quite useless.
From Wilde's preface to The Picture of Dorian Gray

The slides from a talk I gave on neural networks and style transfer are available here.

Without me telling you engaging stories while you look at them, I hope that something is lost. But some of them are pretty, and I hope they give a feel for the effect that the parameters have on the final image.

I've copied some of the images to this page to give an idea - for more, see the slides.

The input image for many of the examples is this view of Nottingham, the location of the conference:

Input image

In the style of Starry Night

Big Bold Colours

Since everyone knows what a Starry Night looks like, here's something by Matisse with big bold colours:

The Snail

More Bold

Picasso's Guitar

Thomas Bewick's Bear

I deliberately showed examples where the algorithm failed. Here's one - the slides show one way of removing those white erasures.

A Breakout Image

This style was used because of an earlier presentation about teaching neural networks to play Atari games, also using Keras.


Not bad


The dappling is art, not compression artifact

Voyager Animation

The beautiful animation of Voyager below flickers because the starting image for each frame is a different random set of values. It shows that the outcome is sensitive to the starting condition, but not critically so. In the slides there is another version where I used the same initial condition for each. The flickering is replaced with twinkling, because the optimizer is non-deterministic.