New Design Ideas

Artificial intelligence (AI) is increasingly used in architecture and interior design. Previous studies have mainly focused on image generation, while the systematic evaluation of visual outputs remains limited. Another gap is the lack of datasets built on the stylistic language of specific design offices, which makes it difficult to test whether AI can reproduce a design office’s established identity. This study aimed to develop a customized text-to-image AI model for hotel interiors and to evaluate its outputs through a structured framework. The scope covers hotel rooms, using a dataset of 5,964 renderings from 17 projects prepared by Yeşim Kozanlı Architecture & Interior Design (YKA). The methodology was organized into two phases. Phase 1 (initial experiment) involved five steps: (i) data collection (1,252 renderings from 9 hotels), (ii) data preparation, (iii) caption-guided model training, (iv) image generation and (v) expert review. Phase 2 (main experiment) involved six steps: (i) expanded data collection (5,964 renderings from 17 projects), (ii) categorization into styles, (iii) caption-free model training, (iv) defining evaluation criteria through the Analytic Hierarchy Process (AHP) with five weighted dimensions - style (30%), material/texture (20.9%), furnishing and fixture organization (22.5%), furnishing and fixture details (13.3%) and atmosphere/lighting (13.3%), (v) image generation and (vi) survey-based evaluation (n = 100, using 1-9 Likert scale). Initial results show that caption-free models produced more scalable and consistent outputs. Atmosphere/lighting achieved the highest mean score (M = 6.45), followed by style (M = 6.33), while material/texture (M = 6.06), furnishing and fixture organization (M = 6.00) and furnishing and fixture details (M = 5.94) scored lower, indicating challenges in technical precision. The AHP-based framework provided a structured assessment that confirmed relative strengths and weaknesses. The study proposes a replicable generation-evaluation cycle linking customized dataset-based training with systematic assessment, demonstrating tailored text-to-image workflows for design offices.