DiffST: High-Level Style Feature Transfer Via Diffusion Models
Hongche Cheng · 2023
Image style transfer aims to transfer the style of one image to another to create a new image with a specific artistic style. Existing style transfer algorithms primarily focus on transferring low-level features, such as textures, colors, strokes, etc., while neglecting the transfer of high-level features like semantic elements and object shapes. To address this issue, we propose DiffST, a new style transfer architecture based on pretrained text-to-image diffusion models. Specifically, DiffST can fully extract the feature of the style images to guide the redrawing of the content image, thereby completing the style transfer. Moreover, DiffST introduces a new image redrawing method called two-stage inversion, which effectively preserves the information of the content image and enhances the quality of the generated image. We demonstrate the quality of our transfer using numerous different paintings and compare DiffST with state-of-the-art methods, proving the superiority of DiffST.