A Review of Deep Learning-Based Image Style Transfer Methods: From Gatys to Diffusion Models

Abstract

This paper systematically reviews how the image style conversion technology based on deep learning has developed step by step, from the classic Gatys method to the current diffusion models. We propose a classification framework, which includes four technical types: iterative optimization method, feedforward network method, video timing method and diffusion model method. In this framework, we compare the representative methods, and reveal the fundamental differences and performance trade-offs between different technology types from the perspectives of content expression, style modeling, semantic controllability and computational efficiency. We also pointed out several key challenges, including the incomplete separation of style and content, the limited ability to deal with complex style conversion, the difficulty in balancing the consistency of video timing and computational efficiency, the insufficient controllability and efficiency of diffusion model, and the lack of standardized evaluation system. We also look forward to the future direction, such as enhancing interpretability, expanding to 3D scenes, realizing lightweight deployment and building ethical framework. This review hopes to provide a clear technical roadmap and method reference for researchers.