Backbone Choice Matters: Human and Perceptual Evaluation of CNN-Based Neural Style Transfer

Abstract

Neural style transfer (NST) commonly relies on pretrained convolutional neural networks (CNNs) to extract content and style representations, yet the effect of CNN backbone selection on the perceived quality of stylized images remains underexplored. This paper presents an empirical study of five CNN backbones for NST across multiple content images and artistic styles. We evaluate the generated images using both objective image quality assessment metrics and a human participant study with 96 respondents. Results show that images generated using higher-performing ImageNet backbones were more frequently preferred by participants and were often judged as better matching the target artistic style. However, standard image quality metrics only partially aligned with human preference and were weak predictors of perceived style imitation. These findings suggest that CNN backbone selection plays an important role in NST output quality and that human-centered evaluation remains necessary for assessing artistic style transfer systems.