An inference stage taxonomy-guided prompt framework for culturally aware generation of indigenous motifs

Traditional ethnic patterns face severe threats due to rapid urbanisation, declining intergenerational craft transmission, and the loss of indigenous knowledge systems. Although generative artificial intelligence (AI), specifically diffusion-based text-to-image algorithms, offers potential workflows for digital preservation, existing models frequently generate culturally stereotypical or homogenised patterns that lack regional symbolic specificity. To address this limitation, this study introduces a taxonomy-guided prompt framework applied entirely during inference. The framework operationalises structured vernacular keywords derived from ethnographic fieldwork in Sarawak, Malaysia, using indigenous Dayak motifs as a foundational case study. The system was validated by evaluating 10 distinct Dayak motifs across two inference settings: a Simple Prompt Condition (SPC) and an Engineered Prompt Condition (EPC). Comparative validation was conducted by a blind panel of six experts (two cultural scholars and four master artisans) using a 5-point Likert scale to assess morphological accuracy, material authenticity, semiotic alignment, and cultural integrity. Statistical analysis demonstrates that the taxonomy-guided EPC framework significantly outperformed baseline prompts, achieving higher mean scores in both morphological accuracy (4.6 vs. 2.9) and cultural integrity (4.8 vs. 2.7) compared to SPC outputs ( p < 0.001). The results demonstrate that a structured vernacular prompt architecture can serve effectively as an inference-level computational governance layer for culturally sensitive generative AI, offering a training-free, scalable framework for responsible digital heritage conservation.