Abstract:Dynamic artistic text style transfer aims to migrate the style in terms of both the appearance and motion patterns from a reference style video to the target text to create artistic text animation. Recent researches have improved the usability of transfer models by introducing texture control. However, it remains an important open challenge to investigate the control of the stylistic degree with respect to shape deformation. In this paper, we explore a new problem of dynamic artistic text style transfer with glyph stylistic degree control. The key idea is to build multi-scale glyph-style shape mappings through a novel bidirectional shape matching framework. Following this idea, we first introduce a scale-ware Shape-Matching GAN to learn such mappings to simultaneously model the style shape features at multiple scales and transfer them onto the target glyph. Furthermore, an advanced Shape-Matching GAN++ is proposed to animate a static text image based on the reference style video. Our Shape-Matching GAN++ characterizes the short-term consistency of motion patterns via shape matchings within consecutive frames, which are propagated to achieve effective long-term consistency. Experiments show that the proposed method outperforms previous state-of-the-arts both qualitatively and quantitatively, and generate high-quality and controllable artistic text.

Shape-Matching GAN++: Scale Controllable Dynamic Artistic Text Style Transfer