Architectural plan generation via pix2pix series algorithms faces dual challenges:the absence of domain-specific evaluation metrics and a lack of systematic insights into the joint impact of training configurations.To...Architectural plan generation via pix2pix series algorithms faces dual challenges:the absence of domain-specific evaluation metrics and a lack of systematic insights into the joint impact of training configurations.To address the limitations of pix2pix-based models adaptation to architectural design,we designed a training regimen involving 12 experiments with varying training set sizes,dataset characteristics,and algorithms.These experiments utilized our self-built,high-quality,large-volume synthetic dataset of architectural-like plans.By saving intermediate models,we obtained 240 generative models for evaluation on a fixed test set.To quantify model performance,we developed a dual-aspect evaluation method that assesses predictions through pixel similarity(principle adherence)and segmentation line continuity(vectorization quality).Analysis revealed algorithm choice and training set size as primary factors,with larger sets enhancing the benefits of high-resolution and enhancedannotation datasets.The optimal model achieved high-quality predictions,demonstrating strict adherence to predefined principles(0.81 similarity)and effective vectorization(0.86 segmentation line continuity).Testing on 7695 samples of varying complexity confirmed the model’s robustness,strong generative capability,and controlled innovation within defined principles,validated through 3D model conversion.This work provides a domain-adapted framework for training and evaluating pix2pix-based architectural generators,bridging generative research and practical applications.展开更多
文摘Architectural plan generation via pix2pix series algorithms faces dual challenges:the absence of domain-specific evaluation metrics and a lack of systematic insights into the joint impact of training configurations.To address the limitations of pix2pix-based models adaptation to architectural design,we designed a training regimen involving 12 experiments with varying training set sizes,dataset characteristics,and algorithms.These experiments utilized our self-built,high-quality,large-volume synthetic dataset of architectural-like plans.By saving intermediate models,we obtained 240 generative models for evaluation on a fixed test set.To quantify model performance,we developed a dual-aspect evaluation method that assesses predictions through pixel similarity(principle adherence)and segmentation line continuity(vectorization quality).Analysis revealed algorithm choice and training set size as primary factors,with larger sets enhancing the benefits of high-resolution and enhancedannotation datasets.The optimal model achieved high-quality predictions,demonstrating strict adherence to predefined principles(0.81 similarity)and effective vectorization(0.86 segmentation line continuity).Testing on 7695 samples of varying complexity confirmed the model’s robustness,strong generative capability,and controlled innovation within defined principles,validated through 3D model conversion.This work provides a domain-adapted framework for training and evaluating pix2pix-based architectural generators,bridging generative research and practical applications.
