A Study of Effective Prompt Design Strategies for Generating Late Joseon Minhwa Images Using Generative AI - Exploring Keyword Mapping and Prompt Dictionary Construction Potential

This study proposes a systematic prompt design strategy that relies on generative AI to effectively reproduce the distinctive formative elements of Minhwa, Korean folk painting from the late Joseon period, and explores the potential of a keyword-based prompt dictionary. Although current text-to-image models such as Stable Diffusion demonstrate strong image generation capabilities, their predominantly Western-centric training data limits the accurate reproduction of Minhwa’s unique humor, exaggerated forms, symbolic motifs, colors, and compositions. To overcome these limitations, 482 Minhwa artworks were analyzed and visual keywords were systematically extracted, with then conducted categorization by form, composition, technique, and atmosphere. Extracted keywords were translated into AI-compatible terms and organized into a structured prompt dictionary, supported by detailed guidelines including synonym expansion, negative prompts, and weighting adjustments. Experiments conducted with Stable Diffusion on representative Minhwa subjects, specifically tiger and peony, compared general prompts against proposed keyword-based prompts structured as simple listings and descriptive sentences. Quantitative evaluation metrics, specifically FID, IS, and CLIP Score, demonstrated that descriptive, keyword-based prompts significantly enhanced alignment with the original Minhwa distributions and improved the reflection of prompt intentions. This research emphasizes the importance of systematic keyword analysis and structured prompt dictionaries for accurately reproducing cultural art styles using AI, providing foundational methods for future traditional cultural content creation.