Example of Creativity Scores assigned by the algorithm to paintings between 1850 and 1950. The horizontal axis is the year of the painting and the vertical axis is the creativity score.

An Algorithm for Assessing Creativity of Art

Can we develop a computer algorithm that assesses the creativity of a painting, given its context within art history? Is creativity in art something that can be quantified? What would be the definition of creativity and how to measure it?

At Rutgers’ Art and Artificial Intelligence Laboratory we investigated this problem. We proposed a novel algorithm for assessing the creativity of artistic products, such as paintings and sculptures. We use the most common definition of creativity, which emphasizes the originality of the work and its influential value. The algorithm analyzes paintings using only visual analysis and the date of creation, with no prior knowledge about art history, and generates a creativity score for each painting.

We applied the proposed algorithm to the task of quantifying creativity of paintings and sculptures. We experimented on collections of over 62K paintings, ranging from 1400 till 2010 to understand the behavior of the proposed algorithm.

The most important conclusion of this work is that, when introduced with a large collection of images of paintings and sculptures, the algorithm can successfully highlight paintings that are considered creative (original and influential) by art historians. For example the algorithm pointed out to “the scream” by Edvard Munch’s (1893) as scoring very high relative to other paintings in the late 19th century. This painting is considered as the second iconic figure after Leonardo’s Mona Lisa in the history of art, and it is known to be the most-reproduced painting in the twentieth century. It is also one of the most outstanding Expressionist paintings. The algorithm also gave Picasso’s Young Ladies of Avignon (1907) the highest creativity score above all the paintings it analyzed in the persiod between 1904 and 1911. Art historians indicate that the flat picture plane and the application of primitivism in this painting made it an innovative work of art, which lead to Picasso’s cubism.

The algorithm pointed out Picasso’s Young Ladies of Avignon (1907) as the most creative painting between 1904 and 1911.

The algorithm favored Picasso’s Maquette for Guitar (1912) above all other cubism paintings and sculptures it analyzed. The algorithm pointed out to several of Kasimir Malevich’s first Suprematism paintings that appeared in 1915 (for example the “Red Square”) as highly creative above the paintings of that period. For the period between 1916 and 1945, the majority of the top-scoring paintings were by Piet Mondrian and Georgia O’Keeffee.

The algorithm achieved these conclusions without any knowledge about art or art history encoded in its input. The algorithm achieved this assessment only by visual analysis of paintings and considering the dates of the paintings.

Of course the outcome of the algorithm does not necessarily coincide with what art historian would point out. For example the algorithm gave a much higher score for Domenico Ghirlandaio’s 1476 Last Supper over Leonardo da Vinci’s famous fresco, which appeared later. The algorithm favored da Vinci’s St. John the Baptist (1515) over his other paintings that it analyzed.

Leonardo da Vinci’s St. John the Baptist (1515) scored the highest among all da Vinci’s religious paintings that were analyzed by the algorithm.

How can we validate the assessment of creativity given by the proposed algorithm, or any algorithm that do this task? Besides the qualitative evidence, we proposed a validation methodology, which we call “time machine experiments”, where we change the date of an artwork to some point in the past or in the future, relative to its correct time of creation. Then we compute the creativity scores using the wrong date, by running the algorithm on the whole data. We then compute the gain (or loss) in the creativity score of that artwork compared to its score using correct dating. What should we expect from an algorithm that assigns creativity scores in a sensible way? Moving a creative painting back in history would increase its creativity score, while moving a painting forward would decrease its score. When we performed these experiments we found that paintings from Impressionist, Post-Impressionist, Expressionist, and Cubism movements have significant gain in their creativity scores when moved back to around 1600 AD. In contrast, Neoclassicism paintings did not gain much when moved back to 1600. This makes sense, because Neoclassicism can be considered as revival to Renaissance. On the other hand, paintings from Renaissance and Baroque styles had losses in their creativity scores when moved forward to 1900 AD.

The outcome of this research will be presented to in the 6th International Conference on Computational Creativity, Park City Utah, June 29-July 2nd, 2015. The paper can be accessed at http://arxiv.org/abs/1506.00711