Systematic Review of Arbitrary Neural Style Transfer: Eight Deep Learning Approaches Compared

Abstract

Neural style transfer has become an important computer vision technique for combining the semantic content of one image with the visual style of another. Arbitrary style transfer extends this capability by allowing a trained model to process previously unseen styles without retraining for each reference style. However, comparative evidence concerning architectural mechanisms, evaluation metrics, computational requirements, and reproducibility remains fragmented. This study presents a directed, systematic literature review of eight representative, arbitrary neural style transfer methods: AdaIN, SANet, MCCNet, ArtFlow, AdaAttN, IEContraAST, CSBNet, and RAST. The review protocol follows PRISMA 2020 principles and includes research questions, database searching, eligibility criteria, quality assessment, data extraction, and comparative synthesis. The findings show that normalization-based methods offer efficient inference, attention-based methods improve local style adaptation, flow-based and restoration-based methods address content leakage, contrastive learning strengthens style representation, and video-oriented methods emphasize temporal consistency. Recent transformer- and diffusion-based studies further improve global feature modeling and controllability but generally incur higher computational costs. This review contributes an architectural taxonomy, an evaluation metric map, a reproducibility analysis, and recommendations for standardized image and video benchmarks that integrate objective metrics, computational measurements, and human preference evaluation.