Generating Textual Art Using Hybrid Learning Techniques

Abstract

Textual Art holds a significant role in preserving the traditions and heritage of various historical cultures. Particularly in the Eastern culture, it has served as an embellishment for homes and mosques, meticulously crafted by skilled artisans endowed with a sense of aesthetics. Modern endeavors have been directed towards digitizing this art form, yet the resources available online for textual art, especially in Arabic and Urdu styles, remain quite limited. Consequently, utilizing pretrained models yields unsatisfactory outcomes. The essence of geometric patterns and textures holds paramount importance in visual style, particularly within Islamic art, where these patterns bear profound symbolic and artistic significance. Our approach involves a fusion of deformable style transfer (DST), an optimization-centric technique that harmonizes the texture and geometry of a given content image to closely align with a chosen style image. This methodology is coupled with diffusion model inpainting, which is applied to content images originating from Calliar: an online dataset featuring handwritten Arabic calligraphy. Diverging from previous generative art methodologies, our approach presents aesthetically pleasing results, all the while preserving the integrity and structure of the Arabic script. Our methodology is showcased across a diverse spectrum of Arabic calligraphy writing styles, encompassing even the most prominent types, including Diwani دیواني , Thuluth ثلث , Kufi كوفي , Farisi فارسي , Naskh ,نسخ and Rekaa رقعة .