earticle

영어

AI-driven image generation has transformed digital creativity but remains fundamentally limited in adhering to explicit design rules. Unlike traditional creative software, which incorporates structured validation and editing processes, AI-driven creative systems operate without real-time analytical refinement, leading to significant limitations in precision and accuracy. This study identifies a key issue: the disconnect between image analysis AI and image generation AI, resulting in persistent errors in numerical accuracy, spatial consistency, and stylistic adherence. Through a systematic evaluation of DALL·E 3, MidJourney, and Stable Diffusion 3.5, we demonstrate that even after 100 iterations, AI-generated outputs exhibited critical errors in fundamental concepts such as facial features (e.g., eyes, nose, and mouth) and numerical constraints (e.g., generating exactly three teeth), revealing AI’s inability to process structured and quantitative instructions reliably. To address these challenges, this study proposes a bidirectional feedback mechanism that integrates analytical validation layers into generative models, enabling real-time refinement. By establishing an AI interoperability framework, we bridge the gap between generative AI and analytical AI, allowing for real-time validation and iterative improvement of generated outputs. Additionally, we emphasize dataset optimization for geometric and minimalist structures and advocate for human-AI collaborative refinement to improve output accuracy. These findings highlight the necessity of structured analytical feedback in AI-driven creative systems, which we expect could contribute to the development of more adaptive, user-controlled design workflows.