A Case Study of a Machine Learning Usage in Design: Conceptual Models from Graph-based GANs

Authors : Mustafa Koç, İmdat As

Pages : 621-639

Doi:10.37246/grid.1591809

View : 105 | Download : 280

Publication Date : 2025-10-28

Article Type : Research Paper

Abstract :This paper presents a novel machine learning (ML) pipeline that transforms architectural graph representations into fully rendered three-dimensional (3D) conceptual massing models. Unlike previous ML approaches that focus primarily on 2D floorplan generation, our method integrates multiple components into a single workflow: (1) graph-based input using HouseGAN++, (2) image-based shape extraction via custom MATLAB processing, (3) 3D model construction with FloorplanToBlender, and (4) diffusion model–based style transfer for visual enhancement. This end-to-end approach is distinctive in its combination of automated plan-to-volume conversion and aesthetic exploration through generative image synthesis. The results show that our pipeline enables efficient, multi-stage architectural ideation while significantly reducing manual effort. The proposed method contributes to early-stage design processes by accelerating concept development and offering stylistically diverse outputs from abstract spatial inputs.
Keywords : makine öğrenmesi, kütle modelleri, GAN, konsept tasarım, Grafik-tabanlı Derin Öğrenme

ORIGINAL ARTICLE URL