Answers are grounded in this paper and cite their sources.
Observational analysis measures perceptual alignment in architectural imagery across geographical and affective descriptors, suggesting bias in AI outputs.
Generative AI is increasingly applied in architectural research, from automated ideation and reshaping design workflows to design education. Despite the increasing realism of synthetic imagery, several research gaps remain including alignment, plausibility, explainability, and control. This study focuses on alignment with human perceptions, specifically examining how synthetic architectural façade imagery aligns with geographical, objective, and affective text descriptors. We propose a pipeline that applies a Latent Diffusion Model to generate façade images and then evaluate this alignment through both AI-based and human-based evaluations. The results reveal that while images generated with geographical prompts are notably aligned, they also show clear biases. The results also reveal that images synthesised from objective descriptors (e.g., angular/curvy) are more aligned with human perceptions than affective descriptors (e.g., utopian/dystopian). These initial results highlight the opportunities and limits of current generative AI models, hinting at data biases and the potential lack of embodied understanding to grasp the complexity in experiencing architecture. Limitations of the study remain. Future work can expand on exploring cultural biases and semantic overlaps, and in testing more advanced embodied AI models and methods.
No takes yet. Share an insight, caveat, or question.
Law et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: