CLIP for image style transfer : exploring text-image correlations

Abstract

Style transfer algorithms aim to transform images while preserving their content characteristics and enhancing stylistic similarities. This thesis evaluates the performance of two models, SigLIPstyler and CLIPstyler, trained with SigLIP and CLIP frameworks respectively, in achieving these objectives. SigLIPstyler demonstrates superior style capture, while CLIPstyler excels in content preservation. Additionally, content-style disentanglement is explored for its potential to enhance content preservation. The study also delves into Video Style Transfer, favoring random frame selection for training to first frame selection for training the style transfer model due to its effectiveness in maintaining temporal coherence. This research underscores the ongoing evolution and challenges in style transfer, emphasizing the need for continued innovation to enhance algorithmic capabilities and broaden their creative applications.