Measuring Aesthetic Preferences of Neural Style Transfer: More Precision With the Two-Alternative-Forced-Choice Task

Abstract

The present work compares the two-alternative forced choice (2AFC) task to rating scales for measuring aesthetic perception of neural style transfer-generated images and investigates whether and to what extent the 2AFC task extracts clearer and more differentiated patterns of aesthetic preferences. To this aim, 8250 pairwise comparisons of 75 neural style transfer-generated images, varied in five parameter configurations, were measured by the 2AFC task and compared with rating scales. Statistical and qualitative results demonstrated higher precision of the 2AFC task over rating scales in detecting three different aesthetic preference patterns: (a) convergence (number of iterations), (b) an inverted U-shape (learning rate), and (c) a double peak (content-style ratio). Important for practitioners, finding such aesthetically optimal parameter configurations with the 2AFC task enables the reproducibility of aesthetic outcomes by the neural style transfer algorithm, which saves time and computational cost, and yields new insights about parameter-dependent aesthetic preferences.