The Use of Artificial Neural Networks in Architecture: Generating Depth-Aware and Explainable Building Facades from Floor Plans

This study aims to position the use of artificial intelligence in architectural design, not merely as a tool for visual generation, but as an exploratory decision-support framework in which selected design relationships can be represented in an interpretable and transparent manner. To this end, the study developed ArchAI 2.0, a five-stage hierarchical system that models semantic and spatial relationships between two-dimensional floor plans and generated building facades. The objective was to conditionally generate exterior building image alternatives using floor plan-derived structural information. Following a pair-integrity screening, 657 valid floor plan–exterior photograph project pairs sourced from the ArchDaily platform were retained and partitioned at the project level into training (n = 459), validation (n = 99), and independent test (n = 99) sets. Plan-derived geometric information was extracted using a deterministic computer-vision parsing pipeline and subsequently processed into structural tokens and a depth-aware parametric facade skeleton. The resulting skeleton representation was used as conditioning guidance for a ControlNet integrated with Stable Diffusion 1.5. A Plan-Edge ControlNet trained and evaluated under the same protocol was used as the primary conditioned baseline, while an unconditional text-only Stable Diffusion 1.5 model provided a secondary reference. On the independent test set, the Skeleton and Plan-Edge models showed comparable performance, with no statistically significant differences in Chamfer Distance, Edge IoU, SSIM, or CLIP cosine similarity after Holm correction. Relative to the text-only baseline, Skeleton conditioning produced a modest but statistically significant improvement in Edge IoU (0.0611 vs. 0.0574; Holm-adjusted p = 0.031), whereas the text-only model achieved significantly higher CLIP similarity (0.2749 vs. 0.2647; Holm-adjusted p < 0.001). FID was treated as a descriptive dataset-level measure because of the limited test set size. A bounded human conditioning adherence evaluation further showed that 58 of 98 evaluable cases reproduced the requested two-story organization (59.18%; 58.59% across the full 99-case test set), with a mean absolute story error of 0.4184 stories. Opening-column correspondence was more limited: across 87 evaluable cases, bipartite matching at the primary threshold of τ = 0.08 normalized facade width yielded a micro F1-score of 0.2970. An independent second-rater assessment on a predefined 25-case subset demonstrated high human annotation reliability, with 96.00% exact agreement and linearly weighted Cohen’s κw = 0.9351 for story-count ratings and an inter-rater opening-column matching F1-score of 0.7692. Rather than demonstrating uniform superiority or guaranteed realization of user-specified constraints, the findings indicate that the proposed depth-aware representation provides an alternative structured conditioning mechanism with a performance comparable to direct plan-edge guidance, while human evaluation reveals a clear distinction between representational editability and actual conditioning adherence. The deterministic rule engine complements this generative pipeline as an interpretable rule-based decision-support layer by expressing selected architectural relationships through explicit and reusable rules.