Neural Style Transfer in Artistic Rendering: A Quantitative Evaluation

Abstract

Neural style transfer enables the synthesis of images by combining the content of one image with theartistic style of another. While visually impressive results have been demonstrated, the quantitativeevaluation of such models remains limited. This paper presents a comprehensive evaluation of neuralstyle transfer techniques using both subjective and objective metrics. We implement several modelsincluding Gatys et al.'s optimization based method and real time feedforward models by Johnson etal. The eval uation is conducted using datasets of classical paintings and natural images, with metricssuch as Structural Similarity Index (SSIM), content loss, style loss, and user study ratings. Resultsindicate that feedforward models offer faster processing with r easonable stylization quality, thoughoften at the expense of finer details. Optimization based models produce higher quality outputs butare computationally intensive. User surveys confirm that different users prioritize style fidelity orcontent clarity differently, suggesting the need for tunable loss balancing. The paper also examines theimpact of resolution, color normalization, and feature extraction layers. We conclude that while neuralstyle transfer is an effective creative tool, its applications in professional design and photographyrequire further refinement in control mechanisms and interpretability. The findings guide researchersin improving model design and evaluation protocols.