TY - GEN
T1 - Adjacency-Driven Interactive Planning Meets Pix2Pix: A Hybrid Workflow for Early Architectural Layouts
AU - Rezve, Md Fahim Hasan
AU - Chowdhury, Shuva
PY - 2025
Y1 - 2025
N2 - The demand for flexible and rapid residential plan solutions has intensified, necessitating innovative tools that bridge user intent, spatial logic, and architectural design automation. This research introduces an interactive generative design platform that enables users - including stakeholders, and individuals without formal architectural training - to define, manipulate, and visualize spatial relationships for housing layouts using an adjacency matrix framework. Through an intuitive interface, users establish spatial logic by specifying room connectivity and proximity, represented as a dynamically generated spatial relationship diagram (SRD). A key contribution of this system is the integration of a Pix2Pix conditional GAN model that translates SRDs into architectural floor plans. Unlike traditional procedural methods, the GAN-based model learns to generate spatially valid, architecturally coherent plans directly from user-defined logic. The platform enforces adjacency and privacy constraints while offering automated evaluation through scoring metrics such as spatial accuracy, adjacency matching, privacy consistency, and circulation efficiency. Suggestions for improvement guide users toward more optimized configurations while preserving their design intent.The research adopts a design-thinking approach and combines symbolic reasoning with deep image-to-image learning. A structured user study involving eight participants-grouped as architecture students, general users, and professional architects—was conducted to evaluate system performance. Participants completed two design attempts: one unguided and another using system feedback. Results showed measurable improvement in adjacency scores, circulation efficiency, and final layout quality across all user groups, validating the platform’s impact on iterative learning and spatial decision-making. This work contributes a hybrid methodology to computational design, merging user-driven logic, rule-based reasoning, and generative AI. It demonstrates how mass customization and participatory input can be integrated into early-stage planning for residential floor plan generation and improve user performance particularly in the affordable housing context.
AB - The demand for flexible and rapid residential plan solutions has intensified, necessitating innovative tools that bridge user intent, spatial logic, and architectural design automation. This research introduces an interactive generative design platform that enables users - including stakeholders, and individuals without formal architectural training - to define, manipulate, and visualize spatial relationships for housing layouts using an adjacency matrix framework. Through an intuitive interface, users establish spatial logic by specifying room connectivity and proximity, represented as a dynamically generated spatial relationship diagram (SRD). A key contribution of this system is the integration of a Pix2Pix conditional GAN model that translates SRDs into architectural floor plans. Unlike traditional procedural methods, the GAN-based model learns to generate spatially valid, architecturally coherent plans directly from user-defined logic. The platform enforces adjacency and privacy constraints while offering automated evaluation through scoring metrics such as spatial accuracy, adjacency matching, privacy consistency, and circulation efficiency. Suggestions for improvement guide users toward more optimized configurations while preserving their design intent.The research adopts a design-thinking approach and combines symbolic reasoning with deep image-to-image learning. A structured user study involving eight participants-grouped as architecture students, general users, and professional architects—was conducted to evaluate system performance. Participants completed two design attempts: one unguided and another using system feedback. Results showed measurable improvement in adjacency scores, circulation efficiency, and final layout quality across all user groups, validating the platform’s impact on iterative learning and spatial decision-making. This work contributes a hybrid methodology to computational design, merging user-driven logic, rule-based reasoning, and generative AI. It demonstrates how mass customization and participatory input can be integrated into early-stage planning for residential floor plan generation and improve user performance particularly in the affordable housing context.
M3 - Other contribution
VL - May
T3 - Supervised Graduate Thesis
ER -