| Title |
Comparative Analysis of AI Architectural Image Generation Using Prompts From Large Language Models |
| DOI |
https://doi.org/10.5659/JAIK.2025.41.10.113 |
| Keywords |
LLM; AI; ChatGPT; Claude; Midjourney; Image Generation; Prompt; Generative AI |
| Abstract |
This study compares the visual reproduction characteristics of architectural images generated by Midjourney using prompts from OpenAI GPT
and Anthropic Claude. Fifteen cases were evaluated across thirty items in three categories: form and color (FC), image atmosphere (IA), and
interpretation method (IM). A panel of five experts applied a three-point blind rating protocol. On average, Claude achieved a higher overall
reproduction score at 67.8 percent, compared to GPT at 62.5 percent, a difference of 5.3 percentage points. By category, Claude performed
better in color-related items such as primary color, color temperature, and light-dark contrast, while GPT showed relative strengths in
geometric composition and roof form. Claude also scored higher in the interpretation method category, particularly in scale interpretation,
degree of abstraction, and technical completeness. Both models received the lowest scores for image atmosphere, highlighting current
limitations in reproducing overall mood. These findings offer baseline guidance for prompt engineering in architectural applications, though
caution is needed when generalizing due to the small sample size and the inherent randomness of image generation. |