Generative artificial intelligence has a habit of flattening the world's visual cultures into a single, glossy aesthetic, and indigenous design traditions are among the most frequent casualties. A new study from Sarawak, Malaysia, reports that a carefully structured prompting method can pull diffusion-based image generators back toward cultural accuracy, without retraining a single model parameter. The research, published in Discover Artificial Intelligence, shows that when ethnographic knowledge about indigenous Dayak motifs is embedded directly into the text prompts fed to a pre-trained image generator, expert evaluators rated the resulting images dramatically higher on morphological accuracy, material authenticity, semiotic alignment and cultural integrity than images produced with ordinary descriptive prompts.
The problem the researchers set out to address is well documented. Diffusion models such as Stable Diffusion XL are trained on vast, globally aggregated image datasets that over-represent dominant visual styles. When users type generic phrases like
The study's methodological foundation rests on a deliberate pairing of two intellectual traditions that rarely meet in machine learning research. The authors constructed their motif taxonomy using Erwin Panofsky's three-tier model of iconographic analysis, a framework developed in art history that moves from raw visual description through recognisable subject matter to deep ideological interpretation. They then aligned these tiers with structural semiotics, drawing on Ferdinand de Saussure's distinction between signifier and signified and Roland Barthes' theory of cultural connotation.
In practical terms, this means a single Dayak motif is documented at three levels simultaneously: its objective geometry, such as curvilinear spirals or radial symmetry; the specific animal, plant, or natural phenomenon it represents; and the cosmological narrative it carries, such as protection, status, or spiritual statehood. Each of these levels was then converted into a machine-readable prompt variable, so that the generative model received not just a name for a pattern but a layered semantic specification of what the pattern means and how it must be rendered. The ethnographic groundwork behind the taxonomy was substantial.
Fieldwork conducted in Kuching and Miri documented visual structures, materials, and techniques across five craft media: basket weaving, beadwork, woodcarving, textiles, and painting. The researchers compiled a pool of 32 ethnographically documented motifs, from which 10 were selected using stratified purposive sampling to ensure balanced representation across material media and across the three core semiotic domains of zoomorphic, phytomorphic, and abstract-geometric forms. The selected motifs included two zoomorphic examples, the kapit ulut and the tiger; two phytomorphic motifs, lelingkok betayok and terjat tepun; and six abstract-geometric designs.
This stratification matters because a validation focused on a single craft medium could easily produce results that fail to generalise, whereas the balanced sample spans rattan basketry, traditional weaving, deep-relief woodcarving, and multi-coloured glass seed beadwork. The evaluation panel itself reflects a design choice aimed at epistemological breadth. Six experts, two indigenous knowledge scholars and four native Sarawak artisans, assessed anonymised outputs on a five-point Likert scale and also participated in semi-structured interviews covering native terminology, regional dialects, cultural connotations, functional values, and customary taboos.
Inter-rater reliability, measured with Fleiss' Kappa, reached 0.82, a coefficient the authors characterise as substantial to near-perfect agreement. This level of consensus across evaluators drawn from both academic and practitioner traditions is notable, since artisans and scholars might plausibly attend to different features of a generated image. The kappa figure suggests that the improvements were recognised uniformly rather than driven by any single evaluator's perspective.
The statistical design also deserves attention. Because each of the 10 motifs was generated under both the simple prompt condition and the engineered prompt condition, the researchers used a paired-samples t-test, with each motif serving as its own control. This pairing neutralises confounds such as motif complexity or geometric density, isolating prompt architecture as the variable driving performance differences.
With nine degrees of freedom, the critical t-value at the 0.001 level is 4.781, and the observed t-statistics ranged from 7.85 for semiotic alignment to 11.02 for cultural integrity, all comfortably exceeding that threshold. The largest gaps appeared in material authenticity, where the engineered prompts scored 4.7 against 2.6 for baseline prompts, and cultural integrity, at 4.8 versus 2.7. Perhaps the most methodologically significant element of the study is its ablation experiment, which addresses an obvious alternative explanation: that the engineered prompts simply worked better because they were longer.
The baseline prompts averaged 36.2 words, while the framework prompts averaged 145.8 words, so any performance gain could in principle reflect sheer descriptive volume rather than structure. To test this, the researchers introduced a length-matched control condition, restricted to a mean of 144.5 words, which matched the framework's token density but stripped away all taxonomic classifications, hierarchical semiotic variables, and material boundary tags, replacing them with unorganised decorative descriptors. The result was decisive.
The length-matched condition produced only marginal gains over the baseline, lifting morphological accuracy from 2.9 to 3.1 and cultural integrity from 2.7 to 2.9, while the full framework scored at least 4.5 on every dimension. Semantic structure, not word count, is the causal factor. The qualitative expert commentary illuminates why unstructured length fails.
Evaluators observed that baseline images frequently treated cultural patterns as flat surface decorations that appeared digitally printed or sprayed onto objects, lacking the internal construction logic native to the craft medium. One artisan evaluator noted that the engineered output portrayed real beading logic and physical hand-tension, while the baseline image could be seen as decorative but lacked material truth. Another recurring failure mode was superficial literalism: when prompted with plant-based terms, baseline generations sometimes produced anatomically realistic three-dimensional trees rather than the stylised phytomorphic conventions of authentic Dayak design.
The engineered prompts, by contrast, captured subtle technical traits such as uniform strip alignments in rattan basketry, natural dye absorption variations across organic fibres, and consistent weave tensions, including the dye-bleeding characteristics along the warp threads of Pua Kumbu weaves. The prompt architecture itself combines positive formatting variables with rigid negative constraints, a technique aimed at preventing what the authors call cultural hallucination. Negative prompts explicitly exclude neon colours, plastic textures, cyberpunk aesthetics, modern geometric abstraction, 3D renders, and glossy finishes, forcing the model toward traditional, organic rendering.
More strikingly, a second category of negative constraints blocks cross-cultural visual contamination by explicitly forbidding Polynesian tribal patterns, Maori tattoo styles, Native American dreamcatchers, Aztec motifs, and generic tribal aesthetics. This addresses a documented failure mode of diffusion models, which tend to blend visual elements from unrelated indigenous traditions into pan-tribal hybrids. The ablation results showed that without such structural boundary tags, even long descriptive prompts led the model to hallucinate hybridised styles, mixing Iban patterns with Polynesian line placements, and to generate physically implausible textures, such as attempting to render warp-ikat patterns onto solid bamboo basket weaves.
The choice of Stable Diffusion XL as the generative backbone was justified on technical grounds. Compared with generative adversarial networks, which are prone to mode collapse and instabilities under complex text conditioning, diffusion models offer greater stylistic variation and more precise structural alignment with textual instructions. This stability matters when prompts carry dense, multi-clause semantic specifications.
The authors also quantified the lexical profile of their prompt corpus using token counts from standard tokenisers, along with the Type-Token Ratio and the Flesch-Kincaid Grade Level, ensuring that differences between conditions could be mapped precisely to sub-word density and vocabulary structure. Beyond the technical results, the study engages seriously with the ethical risks of applying generative AI to indigenous cultural material. The authors catalogue documented pitfalls including unauthorised reproduction of sacred symbols, decontextualisation of cultural artefacts, dataset bias leading to visual homogenisation, and extractive data pipelines aligned with Western epistemologies.
They note that traditional copyright frameworks fail to protect communal cultural property, which motivates a proactive, rights-oriented development strategy. Their proposed mitigations include data sovereignty mechanisms, taxonomy-guided prompt filtering, and collaborative co-creation with indigenous knowledge holders. The framework's cultural knowledge repository was validated by community elders, and the authors emphasise that indigenous communities must act not merely as data sources but as active knowledge holders and co-designers of AI systems.
The semiotic framing of Dayak motifs provides essential context for why superficial imitation is inadequate. Prior field research in Borneo has established that exclusive motifs such as the aso, a dragon-dog figure, and stylised human figures are strictly reserved for the aristocracy and function as formal visual indicators of social rank. Textile designs have been shown to register communal protection and spiritual statehood rather than arbitrary geometry, and traditional banners incorporate specific environmental motifs, colour hierarchies, and textile textures that project ethnic pride during major cultural festivals.
In such a system, a generated image that gets the geometry roughly right but the social or cosmological semantics wrong is not a minor approximation but a substantive misrepresentation, which is why the study treats semiotic alignment and cultural integrity as distinct evaluation dimensions. The research also situates itself within a broader landscape of AI applications to folk art. Convolutional neural networks have achieved high precision in classifying Indonesian batik motifs and have been applied to Indian folk art and the Inaul fabrics of the Philippines, supporting digital documentation of knowledge held by ageing artisan populations.
Generative models, including GANs and Stable Diffusion, have been used to repair damaged artworks and synthesise new works in traditional styles, and hybrid GAN-diffusion pipelines have restored historical paintings and temple murals. What has been missing, the authors argue, is any standardised vernacular prompt system grounded in ethnographic taxonomy, and any quantitative validation of cultural accuracy in generated outputs. This study addresses both gaps simultaneously.
Several limitations temper the findings. The evaluator panel of six is small, and the authors themselves describe the validation as exploratory rather than statistically generalisable. The framework has been tested only on Sarawak Dayak motifs, and its transferability to other indigenous systems remains unverified.
The ablation study addresses prompt length but not the full space of possible confounds, and further testing across different generative architectures and larger evaluator groups would strengthen the evidence base. The authors also flag systemic hurdles facing digital preservation in Sarawak more broadly, including a shortage of local technical personnel, minimal regional research investment, rapid software obsolescence, and the absence of data-safety frameworks, all of which create long-term storage vulnerabilities for any digital archive. The conceptual contribution with the widest implications may be the reframing of taxonomy itself.
Rather than serving as a passive classification scheme, the taxonomy operates as an active semantic boundary mechanism that governs how a diffusion model navigates its latent visual space. Because large-scale models are trained on globally aggregated data, they reproduce dominant visual patterns and generalised tribal aesthetics by default; structured vernacular descriptors counteract this bias at inference time. Future directions identified by the authors include expanding the taxonomy to additional indigenous communities, developing multilingual vernacular prompt systems that accommodate regional dialects, building community-led interfaces that let artisans directly control and evaluate generative outputs, and exploring hybrid approaches that combine structured prompting with carefully controlled model adaptation under strict community consent protocols.
The overarching lesson is that cultural fidelity in generative AI is not a function of model scale alone, but of whether indigenous knowledge structures are deliberately embedded in the generative process itself.
Subject of Research: Taxonomy-guided prompt engineering for culturally accurate AI generation of indigenous Dayak motifs
Article Title: An inference stage taxonomy-guided prompt framework for culturally aware generation of indigenous motifs
Article References: Hatef Jalil, M., & Abdullah, J. (2026). An inference stage taxonomy-guided prompt framework for culturally aware generation of indigenous motifs. Discover Artificial Intelligence, 6(1), Article 959. https://doi.org/10.1007/s44163-026-01944-3
Image Credits: AI Generated
DOI: 10.1007/s44163-026-01944-3
Keywords: generative AI, diffusion models, indigenous heritage, Dayak motifs, prompt engineering, Sarawak, cultural preservation, semiotics, digital heritage, Stable Diffusion XL, ethnography, inference
News Source: Denise Maddox. (October 28, 2026). Structured prompts steer AI toward authentic indigenous Dayak motifs, study finds. Scienmag.
