Balance of Unity and Variety in Fine Art Paintings: A Computational Study

A balanced composition is essential for achieving high aesthetics in visual artworks. Seemingly endowed with a natural sense of balance, viewers can often instantly feel the lack of balance in a composition. Yet balance is a rich and abstract notion hard to articulate in quantitative terms. Existing studies of balance in pictures have focused on the spatial arrangement of visual elements, e.g., color and texture. Inspired by the early twentieth century painter Harold Speed’s seminal book “The Practice and Science of Drawing”, I investigate in this chapter the balance between two opposing qualities of a picture: unity and variety. An overarching principle implicitly assumed by Speed is that a high level of variety in one visual aspect, e.g., tone, should be compensated by a low level in other visual aspects, e.g., shape, to maintain a proper extent of unity. This work aims at developing computational methods to facilitate the application of Speed’s principle in automatic systems. In particular, I have studied the interplay between the variety levels in shape, tone, and color in a painting. First, features to measure variety in shape, tone, and color, were derived using machine learning methods. Then the relationships between these features were examined by linear regression and quantile regression. Because the unity level of a good design varies in an acceptable range, only when the unity level risks being too low do we start to observe a clear tradeoff between the variety levels in different visual aspects. I thus employed quantile regression to model the extreme cases, for instance, paintings with the top variety levels in shape at any given level of variety in tone or color. Experiments on more than 4000 fine art paintings of 34 artists yielded findings that support the design principle of maintaining enough unity by limiting the overall amount of variety in different visual aspects. The techniques developed here enable us to transform the notion of balance advocated by Speed into well-defined quantities. They have potential applications in multimedia information systems, such as providing automatic critiques on the composition of photos or making suggestions for altering an artwork.