Z*: Zero-shot Style Transfer via Attention Reweighting

InST CVPR '23 StyTr² CVPR '22 VCT ICCV'23 Ours Figure 1. Image style transfer results by the proposed Z * . Top: The stylized results by style/content references of different types. Our method can well balance the contents and styles in the results. Bottom: Comparisons with state-of-the-art methods, including diffusion-based models (VCT [9] and InST [51]), transformer-based model (StyTr 2 [13]), flow-based model (ArtFlow [2]), and CNN-based model (CAST [50]). Our method excels in generating stylized images with vivid style patterns and accurate content details.