ArchDiff: Streamlining Architectural Design with Diffusion-Based Style Generation

The captivating appeal of architectural designs, with their ability to encapsulate both historical and contemporary aesthetics, is indisputable. However, current methods grapple with challenges such as constrained inspiration, lengthy production times, and complex adjustment procedures. To address these issues, we introduce the architectural style generation denoising Network (ArchDiff) in this paper. This novel approach, which utilizes diffusion models, aims to streamline the architectural design process. Specifically, we incorporate an Arch-Former module to distill fine-grained features from images. Our model then employs a denoising network to blend original architectural images with textual descriptions, facilitated by cross-attention layers. This integration offers users a broad spectrum of design references. Given the scarcity of publicly accessible datasets, we have curated the Arch-26 dataset, which comprises high-quality images across twenty-six classes, to train ArchDiff. We have also conducted classification experiments on this dataset, yielding promising results. Via extensive experimentation, we demonstrate that our method can generate pleasing architectural style images, thereby validating its efficacy as a tool for architectural design.