CSI-Net: revisit content style interaction in artistic style transfer

Artistic neural style transfer has garnered widespread attention for its remarkable ability to blend the structural content of a source image with the stylistic elements of a reference image. Existing methods frequently disentangle content and style in deeper network layers, often leading to artifacts at object boundaries and within high-frequency regions. In this paper, we revisit the intrinsic correlation between content and style and propose CSI-Net, a novel architecture inspired by Gated Recurrent Units (GRU). CSI-Net introduces a lightweight Content-Style Interaction (CSI) module that actively modulates how style features are activated or suppressed based on content. Furthermore, we devise a guidance loss in conjunction with a total variation loss to enhance edge coherence and reduce artifacts. Extensive experiments on multiple datasets demonstrate that CSI-Net improves stylization quality and yields fewer artifacts compared to state-of-the-art arbitrary style transfer methods.