AI-Based Approach to Visualization of Therapeutic Garden Design† - Exploring the Availability of Stable Diffusion Models -

This study aims to visualize subject gardens such as the healing gardens that requires specific design criteria. The existing visualization of healing gardens has a limitation: it is difficult to objectively verify because it relies on the designer’s subjective judgment. To compensate for this, the main functions of Stable Diffusion (Text-to-Image, Image-to-Image, Checkpoint, and LoRA) were compared under the same prompt and parameter conditions. Quantitative evaluation using CLIP Score and qualitative evaluation based on aesthetics (shape and shaping, spatial composition and harmony, color/material, arrangement and scale) and function (clarity of entrance and predictability, movement path simplicity and predictability, barrier-free elements, planting/resting/work space compartment, and planting of five senses) were performed. As a result of quantitative evaluation, LoRA was the highest at 62.76, followed by Checkpoint 62.29, Image-to-Image 61.92, and Text-to-Image 60.79. The final visualization combining LoRA and Checkpoint recorded 63.48, securing the balance between reality and formability. In the qualitative evaluation, the average aesthetics score was high at 4.19 (SD=0.13), and the average functionality score was relatively low at 3.25 (SD = 0.25). The recognition and circulation of the entrance were improved, but the specific expression of barrier-free details, work spaces, and five-sensory planting was limited. By presenting an evaluation procedure that combines a design-guideline-based prompt with a CLIP Score, this study explored the possibility of using generative artificial intelligence in landscape design visualization, suggesting that it is necessary to improve precision through the parallel application of control modules such as ControlNet and the and construction of specialized healing gardens in the future.