Abstract
The diffusion model has shown exceptional capabilities in controlled imagegeneration, which has further fueled interest in image style transfer. Existingworks mainly focus on training free-based methods (e.g., image inversion) dueto the scarcity of specific data. In this study, we present a data constructionpipeline for content-style-stylized image triplets that generates andautomatically cleanses stylized data triplets. Based on this pipeline, weconstruct a dataset IMAGStyle, the first large-scale style transfer datasetcontaining 210k image triplets, available for the community to explore andresearch. Equipped with IMAGStyle, we propose CSGO, a style transfer modelbased on end-to-end training, which explicitly decouples content and stylefeatures employing independent feature injection. The unified CSGO implementsimage-driven style transfer, text-driven stylized synthesis, and textediting-driven stylized synthesis. Extensive experiments demonstrate theeffectiveness of our approach in enhancing style control capabilities in imagegeneration. Additional visualization and access to the source code can belocated on the project page: \url{https://csgo-gen.github.io/}.