| Title |
A Quantitative Analysis Approach for Biophilic Residential Design Using Generative AI |
| Authors |
김지연(Kim, Ji-Yeon) ; 박성준(Park, Sung-Jun) |
| DOI |
https://doi.org/10.14774/JKIID.2026.35.3.001 |
| Keywords |
Generative AI; Biophilic Design; Residential Space; Deep Learning; Quantitative Analysis |
| Abstract |
Previous studies on biophilic design have primarily treated natural elements as binary variables, limiting the precise evaluation of their psychological effects across varying levels of visual occupancy. This study proposes an integrated methodology to quantitatively control and visualize biophilic design attributes in residential spaces using Generative AI, establishing a systematic framework for evidence-based design. The spatial scope was limited to residential living rooms, with plants and wood selected as key biophilic attributes. These were tested across three target visual occupancy levels: L1 (15%), L2 (30%), and L3 (50%). The methodology integrated a five-step prompt structuring framework, LoRA fine-tuning, and Image-to-Image generation. To empirically validate the accuracy of the generated spatial images, YOLOv8-Seg was utilized for segmentation analysis. The study quantitatively compared the generation accuracy between a ‘Prompt Only’ condition and a ‘Prompt+LoRA’ condition. The analysis revealed that the ‘Prompt Only’ approach tended to consistently under-represent plant attributes, while inconsistently representing wood attributes ? over-representing at lower levels (L1?L2) and under-representing at the highest level (L3) ? due to pre-trained model biases. In contrast, the ‘Prompt+LoRA’ method demonstrated high precision, successfully maintaining error margins within ±3%p of the targeted visual occupancy ratios across all three levels for both attributes. These findings confirm that integrating LoRA fine-tuning with structured prompting provides reliable quantitative control over domain-specific spatial elements. Ultimately, this research advances biophilic design from a qualitative approach to a quantitative experimental paradigm, offering a robust data-driven visualization tool for early-stage architectural decision-making. |