Fine-tuned AI generates culturally authentic Batak Ulos textile motifs

Researchers from Indonesia (Simanjuntak et al.) fine-tuned two pretrained latent diffusion models—Protogen v3.4 and Stable Diffusion v1.4—on a curated dataset of high-resolution Ulos motifs, traditional textiles of the Batak ethnic group in North Sumatra. The goal was to overcome historical limitations of a narrow motif range and time-intensive design processes. The framework generates novel, culturally consistent patterns. Quantitative evaluation used Frechet Inception Distance (FID) and Inception Score (IS), while qualitative assessments involved traditional weavers and the public. Protogen v3.4 consistently outperformed Stable Diffusion v1.4, achieving ~10.5x lower FID and 2.0x higher IS, indicating better visual fidelity, diversity, and alignment with real Ulos motif distributions.

Further analysis of generation parameters revealed a clear fidelity-diversity tradeoff. Lower strength values produced higher fidelity (lower FID), while higher strength increased diversity at the cost of realism. Across all configurations, a guidance scale of 5–9 provided the most effective balance, stabilizing FID, KID, and IS. The study demonstrates that carefully fine-tuned generative AI can support creative renewal of intangible cultural heritage while preserving its stylistic and symbolic integrity. The paper is under review at the 2026 International Conference on Data, Information and Computing Science.