Aesthetic Preferences of Neural Style Transfer-Generated Portrait Images: An Exploratory Study with the Two-Alternative-Forced-Choice Task

Neural style transfer is a popular deep learning algorithm to generate images to mimic human artistry. This work applies the psychological method of the two-alternative forced choice (2afc) task to measure aesthetic preferences for neural style generated images. Portrait photos of three popular celebrities were generated by varying three parameters of neural style transfer in five configuration levels. Participants had to choose the image they preferred aesthetically from all pairwise combinations of configurations per style. The rate of being chosen was calculated for each neural style transfer configuration level. The findings show a differentiated picture of aesthetic preferences. On the one side, they indicate that people prefer images rendered with 500 iterations and a learning rate of 2e1, i.e. configurations that allow them to recognize the structure of the portrait image despite the stylization. On the other side, aesthetic preferences peak for two distinctly different content-tostyle weight ratios. Whereas the medium-high configuration (100:100) may be favored by people who like abstract arts, the high configuration (300:100) may be chosen by people who prefer realistic art. These results indicate that aesthetic preferences for neural style transfer-generated images can be characterized by unique patterns, and their optimal configuration levels can be captured by the 2afc task.