Exploring Semantic Consistency in Generative Artificial Intelligence via Text-to-Image and Image-to-Text Transformation

Authors : Onur Doğan, Almila Altıntaş, Buse Yücetürk, Doğa Aydın, Fatih Soygazi, Yılmaz Kılıçaslan

Pages : 53-62

View : 101 | Download : 66

Publication Date : 2025-06-27

Article Type : Research Paper

Abstract :Recent advancements in artificial intelligence (AI) have brought Generative AI models dealing with Text-to-Image and Image-to-Text transformation to the forefront. While these models offer significant potential, their effectiveness hinges on the proper utilization of prompts – user-provided instructions guiding the generation process. It is crucial to ascertain the success of images generated through prompt input. In this context, following the text-to-image generation process, the creation of descriptive text for the produced image is also of significant importance. Thus, by comparing the input prompt and the output text, it becomes possible to determine the degree of success of the generatively produced image. This study discusses the consistency between natural language used by humans and prompt language used by Generative AI models. We propose a novel approach: a natural language text-to-image model generates an image, which is then described in text by an image-to-text model, and this text is subsequently used as a prompt. A comparison module then identifies the prompt and corresponding image pair from a pre-built database that has the highest similarity to a human-generated description. This approach aims to maximize the benefit from existing AI models and promote explainability – a crucial principle in AI. This solution addresses the problem of improving the AI model\\\'s ability to generate human-like descriptions and enhancing the process of evaluating the accuracy of these descriptions.
Keywords : Üretken yapay zeka, görüntüden-metine modeller, doğal dil işleme, prompt mühendisliği, metinden-görüntüye modeller

ORIGINAL ARTICLE URL