Deep learning-based generative models for residential layout fusion and their performance evaluation: a case study of Beijing Siheyuan

초록

With the rise of generative AI in architecture, integrating traditional spatial logic with modern residential functionality has become a key challenge. This study proposes a method to generate hybrid floor plans by fusing Beijing Siheyuan and modern layouts using VAE + Pix2Pix and CycleGAN models. A dataset of 170 function-labelled image pairs was constructed, with RGB and grayscale variants and unpaired samples. A fusion loss factor (lambda_fuse) was introduced to control the balance between traditional structure and modern flexibility. Objective and expert evaluations based on spatial symmetry and enclosure showed that the VAE + Pix2Pix model with lambda_fuse = 1 achieved the best performance in structural clarity and cultural coherence. This research confirms the potential of deep generative models for culturally informed spatial design.

키워드