| Title |
A Quantitative Study of Quality Degradation in Repeated Generative AI Architectural Imaging Using the Nano Banana Platform |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.8.179 |
| Keywords |
Generative AI; Architectural Design; Image Regeneration; Quality Degradation; Nano Banana |
| Abstract |
This study quantitatively analyzes quality changes during iterative regeneration (Redo) of architectural images on the Nano Banana generative
AI platform. Architectural form complexity is classified into three levels, and prompts at each level undergo consecutive Redo operations
within a single session. Six no-reference quality metrics, including sharpness, colorfulness, contrast, noise, edge density, and saturation, are
computed and analyzed using linear regression. The results reveal an over-sharpening phenomenon in which sharpness and noise increase
simultaneously as iterations progress. This pattern is driven by artificial edge enhancement and the accumulation of high-frequency noise
rather than genuine improvements in image quality. The effect becomes more pronounced at higher levels of geometric complexity, while
contrast consistently decreases across all levels, indicating a reduction in tonal range. Edge density increases in proportion to complexity,
whereas color-related metrics show inconsistent patterns, suggesting a stronger dependence on material and lighting conditions than on the
Redo process itself. These findings demonstrate that iterative regeneration does not produce substantive quality improvements and is instead
accompanied by facade proportion distortion and degradation of material texture representation. The results provide an analytical framework
for monitoring image quality in AI-assisted architectural design. |