DisenStyler: Text-driven fast image stylization using content disentanglement and style adaptive matching

Abstract

The emergence of the CLIP(Contrastive Language-Image Pre-Training) model has drawn widespread attention to text-driven image style transfer. However, existing methods are prone to content distortion when generating images and the transfer process is time-consuming. In this paper, we present DisenStyler, a novel Text-Driven Fast Image Stylization using Content Disentanglement and Style Adaptive Matching. The Global-Local Feature Disentanglement and Fusion (GLFDF) to fuse the content features extracted from the frequency and the spatial, enabling the detail information of the generated images can be well preserved. Furthermore, the Style Adaptive Matching Module (SAMM) is designed to map text features into the image space and conduct style adaptive matching by utilizing the means and variances of text and images. This not only significantly improves the speed of style transfer but also optimizes the local stylization effect of the generated images. Qualitative and quantitative experimental results show that the DisenStyler can better balance the content and style of the generated images while achieving fast image stylization.