Between Intuition and Instruction: Comparison of Generative Models in Symbolic Design Ideation Case Study of IKN Garuda Palace

Abstract

This study aims to map the comparison of various artificial intelligence (AI)-based generative models, particularly the shift from design intuition to text-based command instructions, as a potential source of ideas in symbolic architecture design. While generative AI is increasingly adopted in creative fields, there remains a critical research gap in the systematic evaluation of how different conversational platforms respond to abstract, symbol-based architectural narratives during early-stage ideation. To address it, this study employs a qualitative-comparative case study approach to evaluate three popular text-to-image generative models: Copilot, ChatGPT, and Gemini. The study was conducted by deploying identical command prompts based on the symbolic narrative of the Garuda National Capital (IKN) Palace and comparing the outputs across dimensions such as visual quality, contextual accuracy, diversity of ideas, generation speed, and adaptability to change. The concrete findings reveal distinct computational behaviours: Copilot excels at architectural integration through monumental abstraction, ChatGPT prioritises literal and explicit symbolic replication, and Gemini offers highly fluid, narrative-driven conceptual sketches. Based on these results, the concept of AI's potential is proposed through three main discussion frameworks: variation exploration, productivity efficiency, and communication visualisation. This study contributes a new evaluative framework for human-machine interaction in architectural pedagogy and practice, demonstrating how tailored prompt descriptions can transform AI from a passive visualisation tool into an active, reflective conceptual partner.