Training-free style transfer via content-style image inversion

Abstract

Image style transfer aims to adapt a content image to a target style while preserving its structural information. Despite the strong generative capabilities of diffusion models, their application to style transfer faces two key challenges: (1) entangled content-style interplay during denoising, leading to suboptimal stylization, and (2) reliance on computationally expensive optimization (e.g., model fine-tuning or text inversion). To address these issues, we propose a training-free tri-path framework. The content and style paths separately leverage image inversion to extract content and style features, which are shared with the stylization path. Specifically, the content path preserves structure via residual connections and noised *h*-features, while the style path injects style through AdaIN-modulated self-attention features to avoid artifacts. Our method eliminates optimization overhead and ensures harmonious stylization by decoupling content-style control. Experiments demonstrate that our approach effectively retains content fidelity and style accuracy while minimizing artifacts.