NO MORE BLAH-BLAH: EMBRACING REAL TEXT IN THE IMAGE SYNTHESIS WORLD

Aref Tabatabaei, Zahra Dehghanian, Negar Movaghatian, Maryam Amirmazlaghani

Keywords: Text-to-Image, Stable Diffusion, ControlNet, Image Synthesis

TL;DR: This paper introduces a novel method to better integrate text into images, improving the appearance of text on objects.

Abstract: The integration of text onto objects within an image frequently results in an unnatural appearance, where the text appears overlaid rather than seamlessly embedded. Existing text-to-image models often encounter challenges in accurately incorporating text within an image context. This paper introduces a novel conditional inpainting technique aimed at overcoming these limitations. Our proposed model showcases exceptional effectiveness in addressing this issue, achieving compelling results by integrating text into images, thus significantly enhancing the naturalness and coherence of the final compositions.

Submission Number: 200