Mapping Architectural Precedents through Visual Similarity Graphs: A SigLIP-Based Analysis of the Serpentine Pavilion Archive

Electronic Materials Letters · Springer Nature · 2026-10-19 · DOI: 10.1007/s00004-026-00921-8

Аннотация

This study presents a computable methodology for examining architectural precedents through machine-derived visual similarity. Using the Serpentine Pavilion archive (2000–2025), the SigLIP vision encoder maps 210 images from 24 pavilion/year groups into a multidimensional latent space. Cosine similarity and graph-theoretical modeling generate two complementary networks: an image-level graph and a pavilion-centroid graph based on averaged, L2-normalized embeddings. The centroid graph provides the primary basis for pavilion-level interpretation by controlling for multiple photographs. Network metrics identify central, bridge, and peripheral cases. Comparison across the two analytical levels shows that network-role assignments are sensitive to image aggregation and the unit of analysis, while bridge roles exhibit greater cross-level consistency. The workflow is reproducible and inspectable without treating AI embeddings as neutral or bias-free representations of architectural morphology.

Читать у издателя →

Похожие статьи