DL-Siheyuan: deep generative and intelligent evaluation framework for culturally adaptive housing layouts — a case study of Beijing Courtyards

This study proposes a deep generative framework integrating ResNet encoder, VAE latent space, and Pix2Pix image translation for layout generation between traditional Beijing Siheyuan courtyards and modern residences under small-sample conditions, achieving both cultural semantic preservation and functional performance.Under identical data and hyperparameters, the proposed ResNet+VAE+Pix2Pix surpasses VAE+Pix2Pix in visual metrics (peak SSIM 0.9015 vs. 0.9000; peak PSNR 15.274 dB vs. 15.181 dB) and shows more stable late-stage convergence, while the DDGAN baseline exhibits numerical instability and structural collapse with limited data. To validate generalization and scalability, learning curves on a fixed validation set across 25%/50%/75%/100% training sizes show SSIM improving by 4.77% and PSNR by 6.70%, with isotonic regression confirming monotonic, unsaturated growth. A multidimensional evaluation combining expert scoring and environmental simulation (Rhino/CFD and Radiance) indicates high ratings across five spatial quality dimensions and consistent satisfaction of daylighting and ventilation thresholds. Moreover, CNNs predicting expert scores and performance metrics achieve mean squared errors < 0.2, and GAN-predicted performance maps deviate < 5% from simulations, supporting engineering applicability. The framework balances cultural continuity with functional performance, establishing a generalizable, data-efficient pathway for small-sample architectural layout generation.