A Study of Visual Representation Methods of Generative AI Models

Images generated by Generative AI are actively utilized in visual content including design, advertising, video, and social media, and their importance and influence are continuously expanding. These generated images are highly realistic and detailed, making it difficult to distinguish whether they were created by a human or generated by AI. Generative AI learns from vast datasets and generates new images by analyzing their statistical characteristics and patterns. These datasets may consist of various types of images, such as photographs, drawings, and graphics, reflecting diverse aspects of reality, including shapes, colors, textures, and structures. This suggests that images created by generative AI inherently possess the ability to represent reality. This study aims to explore how Generative AI represents the real world, moving beyond simple visual expression. To this end, it examines the gradual development of the concept of representation from traditional to modern times and its application in art. It also analyzes representative image-generation AI models to investigate the technical principles and methods by which Generative AI represents the real world. To contemplate the image-generating principle and procedure from the perspective of representation, the study utilizes three viewpoints; the originals and datasets, the representation of essence, and indeterminacy and diversity during image generation. The study suggests the need to revisit the traditional concept of the original in the context of technological advancements and proposes an expanded perspective on representation. It emphasizes that Generative AI does not simply replicate original data but creates diverse representations through its indeterministic processes.