| 논문명 |
LoRA와 ControlNet을 활용한 Stable Diffusion 기반 건축 입면 생성 방법과 평가에 관한 연구 / A Generation Method and Evaluation of Architectural Facade Design Using Stable Diffusion with LoRA and ControlNet |
| 저자명 |
박정민(Park, Jungmin) ; 홍순민(Hong, Soonmin) ; 추승연(Choo, Seungyeon) |
| 수록사항 |
대한건축학회논문집, Vol.41 No.8 (2025-08) |
| 주제어 |
스테이블 디퓨전; 건축 매스; 입면 디자인; 생성형 AI; 로라; 컨트롤넷 ; Stable Diffusion; Architecture massing; Facade Design; Generative AI; LoRA; ControlNet |
| 요약1 |
본 연구는 Stable Diffusion에 LoRA와 ControlNet을 결합하여 건축 입면 이미지를 생성하는 새로운 방법을 제안한다. 기존 Stable Diffusion은 건축 요소와 재료 표현에 한계가 있으며, 이를 해결하기 위해 LoRA를 활용한 건축 특화 학습과 ControlNet의 Canny Edge 및 Depth Map 제어를 적용하였다. 생성된 이미지는 CLIP 평가와 GPT-4V 기반 평가를 통해 분석되었다. 본 연구는 초기 건축 설계 과정에서의 효율성을 높이고, 입면 디자인 자동화에 기여할 수 있음을 시사한다. |
| 요약2 |
This study proposes a novel approach for generating architectural facade images by combining the Stable Diffusion model with Low-Rank
Adaptation (LoRA) and ControlNet. The standard Stable Diffusion model faces limitations in accurately reflecting architectural elements and
material characteristics, which are critical in the design process. To address these challenges, this research integrates domain-specific
fine-tuning using LoRA and precise shape control through ControlNet. LoRA allows the model to effectively learn architectural styles and
details, ensuring better representation of essential design elements such as windows, balconies, and facade materials. Meanwhile, ControlNet
utilizes Canny Edge and Depth Map information to enhance shape accuracy and spatial consistency, enabling more reliable image generation.
The generated images were evaluated through Contrastive Language-Image Pretraining (CLIP) scores for quantitative analysis and
GPT-4V-based qualitative evaluation, providing a more comprehensive understanding of architectural coherence and visual fidelity. The
GPT-4V assessment offered insights into spatial relationships, contextual relevance, and material expression that are not easily captured
through traditional metrics. This combined approach reduces the repetitive manual adjustments commonly required in text-prompt-based image
generation and facilitates a more intuitive and efficient design process during the early stages of architectural planning. By improving control
over detailed architectural features, the proposed method contributes to the automation of facade design, offering significant potential for
real-world applications in architectural design and visualization. Future research will focus on expanding the dataset to include diverse
architectural styles and validating its practical application in design and construction. |
