A decade ago, it seemed impossible that Artificial Intelligence (AI) might produce fine art: the human creativity could not be simulated effectively by a computer. Earlier this year, however, an AI painting (“Portrait of Edmond de Belamy”) was sold for a staggering £337,000.

How are we using AI software?

We want to tackle a centuries-old conundrum: what makes paintings aesthetically pleasing - their style or their semantic content? We can use convolutional neural networks (CNNs) to create images which combine any style with any content.

For example, here the same content is rendered in the style of (left) Cézanne and (right) Van Gogh:

 And here the content differs, but both images are in the style of Cézanne:

How can these images help us understand individual preferences for artistic styles?

We tracked people’s eye movements while they were looking at these sorts of paintings. Participants were shown two paintings, and picked the one they preferred. The two paintings shown had either the same style, or the same content. We calculated a single parameter from participants’ eye movements which reflected the strength of their preference for certain styles and contents.

Now we can predict future choices for paintings with similar style or portraying the same subject.

What is next?

We are working on a software algorithm that will provide automatic feedback about observers’ aesthetic preferences by looking at how people look at paintings. This is a crucial step toward personalised exhibition tailoring.

Would you like to see how this works?

Come and see us at the gallery. Support our research and subscribe to participate in the experiment: jasmina.stevanov@bristol.ac.uk  (07563 113943)

More details about how to participate can be found here.