PEFT Strategies for Human–AI Co-Creation in Architectural Morphogenesis

Abstract

Generative AI is reshaping early-stage architectural design, yet the cognitive demands placed on designers navigating highly parameterised workflows remain underexamined. This study investigates how four parameter-efficient fine-tuning (PEFT) strategies in Stable Diffusion influence human–AI co-creation, interpretability, and designer control in biomimetic architectural morphogenesis. Four configurations—Baseline, Shared LoRA, IP-Adapter (w=0.8), and a custom fine-tuned LoRA—were evaluated using MorphEvalBench, a newly developed evaluation toolkit. Fidelity and diversity metrics were operationalised as proxies for designer cognition: biomimetic fidelity (DSS), distributional fidelity (FID/KID), and diversity (DIV). A visual benchmarking grid across representative token levels complemented quantitative analysis. Results reveal that IP-Adapter achieved the strongest quantitative profile across all dimensions, yet its visual conditioning mechanism introduces patterns independently of prompt specification at w≥0.6, reducing designer control and requiring management of multiple interacting parameters. Fine-tuned LoRA demonstrated a balanced profile with greater semantic stability, enabling prompt-driven control and emergent diversity driven by designer input. Notably, the fine-tuning process itself constitutes a design act, embedding the designer's intent into the model prior to generation. These findings demonstrate that quantitative superiority does not guarantee cognitive supportiveness, and propose four implications for designing cognitively supportive human–AI workflows in architectural practice.

Keywords: Generative AI, Human–AI Co-creation, Computational Design, Morphogenesis, Cognitive Load, Parameter-efficient Fine-tuning

DOI: 10.54941/ahfe1007968