Visual heritage–informed generative modelling for Zhuang brocade motifs: Semantic control and multimodal evaluation with eye tracking and PAD

Generative AI offers novel computational avenues for the visualization and dissemination of intangible cultural heritage (ICH) motifs. However, a key challenge remains in computationally preserving cultural authenticity and evaluating how audiences visually and affectively perceive these AI-driven visual assets. To address this, we propose a culturally guided visual generation framework for Zhuang brocade patterns. This approach couples a structured three-level semantic annotation scheme—aimed at organizing visual heritage data—with LoRA fine-tuning of Stable Diffusion. We validate this framework through a multimodal human-perception study. Across two paired within-subject comparison blocks—same-colour/different-pattern and same-pattern/different-colour—54 participants viewed AI-generated and handcrafted/reference motifs while eye movements were recorded, alongside PAD affective ratings and interview feedback. When colour was held constant, the primary gaze measures did not differ robustly between the paired AI-generated and handcrafted conditions, suggesting comparable global gaze allocation rather than demonstrating equivalence. Conversely, when pattern was fixed, the AI-conditioned palette variants attracted longer and more frequent fixations, whereas PAD ratings showed that this attentional increase did not translate into a uniform increase in Pleasure. These results provide an evidence-based workflow for the generation and multimodal evaluation of visual heritage motifs, showing that palette manipulation can increase visual salience without a corresponding gain in affective acceptance.