AI for Cultural Heritage Textiles: Fine-Tuned Latent Diffusion for Novel Ulos Motif Synthesis

Technical Summary: AI for Cultural Heritage Textiles: Fine-Tuned Latent Diffusion for Novel Ulos Motif Synthesis

Problem Statement
The preservation and revitalization of Ulos, a traditional textile of the Batak ethnic group in North Sumatra, Indonesia, face a dual challenge: maintaining strict fidelity to cultural symbolism and design conventions while addressing the need for innovation to meet contemporary market demands. Traditional Ulos weaving is constrained by a narrow range of existing motifs and a time-intensive, skill-heavy design process. Previous attempts to automate motif generation using image quilting, SinGAN, and StyleGAN have demonstrated limitations. Image quilting often results in visible repetitions and unnatural seams, failing to generalize complex structural elements. While SinGAN and StyleGAN offer better style transfer, they struggle with preserving the semantic consistency and symbolic integrity of Ulos motifs, often producing structural artifacts, uncontrolled color shifts, or outputs that lack cultural coherence. Furthermore, these GAN-based approaches are computationally intensive and require massive datasets, which are often unavailable for specific cultural heritage contexts.

Methodology
To address these limitations, the authors propose a generative framework based on fine-tuned Latent Diffusion Models (LDMs). The methodology proceeds through five stages:

1. Data Collection and Preparation: The study utilizes a curated dataset of 88 unique Ulos motif fragments sourced from the DiTenun platform and field visits, representing Toba, Karo, and Simalungun sub-ethnic groups. To compensate for the small dataset size, the authors applied data augmentation (rotation, flipping, brightness variation) to expand the training set to 299 samples. All images were standardized to 512×512 pixels.
2. Model Selection and Fine-Tuning: Two pre-trained LDMs, Stable Diffusion v1.4 and Protogen v3.4, were fine-tuned on the Ulos dataset using an image-to-image strategy. The training incorporated text-conditioning via descriptive prompts that captured structural and symbolic attributes (e.g., "symmetrical vertical layout," "woven star motifs").
3. Training Configuration: The models were trained for 50 epochs with a batch size of 2, a learning rate of 1.0×10⁻⁵, and the Adam optimizer. The fine-tuning objective included a standard diffusion loss and an additional "Ulos Motif Loss" term to penalize deviations from the feature space of authentic motifs, ensuring cultural and geometric coherence.
4. Experimental Scenarios: The models were evaluated across three specific testing scenarios:
   • Scenario 1 (Structural Manipulation): Assessing the ability to alter motif shape and arrangement while preserving cultural essence.
   • Scenario 2 (Colour Variation): Testing precise chromatic transformations while strictly maintaining geometric integrity.
   • Scenario 3 (High Complexity): Evaluating robustness on dense, intricate patterns with layered symmetry.
5. Evaluation: Performance was measured quantitatively using Fréchet Inception Distance (FID) and Inception Score (IS) across varying strength (0.65–0.85) and guidance scale (5–19) parameters. Qualitative validation was conducted through interviews with three traditional weavers (assessing weaving feasibility and cultural fidelity) and a survey of 30 public participants (assessing aesthetic appeal).

Key Results
The study found that Protogen v3.4 consistently outperformed Stable Diffusion v1.4 across all metrics and scenarios:

• Quantitative Performance: Protogen v3.4 achieved a significantly lower FID (~10.5× lower than SD v1.4 in final convergence) and a higher Inception Score (2.0× higher), indicating superior visual fidelity, diversity, and alignment with the real Ulos motif distribution. While SD v1.4 showed some high IS values, its FID remained high, suggesting generated motifs were visually distinct but statistically divergent from authentic patterns.
• Parameter Sensitivity: Lower strength values (0.65) consistently yielded higher fidelity (lower FID), while higher strength values increased diversity at the cost of realism. A guidance scale of 5–9 was identified as the optimal operating range, providing the best balance between fidelity and diversity for both models.
• Qualitative Validation: Traditional weavers preferred motifs generated by Protogen v3.4 at moderate strength settings (0.65–0.75), citing better structural balance, clear thread paths, and adherence to traditional weaving orientation. In contrast, SD v1.4 outputs were frequently criticized for broken symmetry and "unweavable" vertical patterns. Public surveys also showed a strong preference for Protogen v3.4, with up to 96.67% of participants in complex scenarios favoring its outputs for their cultural representation and aesthetic appeal.

Significance and Claims
The paper claims that carefully fine-tuned generative AI, specifically Latent Diffusion Models, can effectively support the creative renewal of intangible cultural heritage without compromising its stylistic or symbolic integrity. The study demonstrates that:

1. Cultural Continuity: Diffusion models can generate novel Ulos motifs that extend beyond existing repositories while maintaining the structural rules and symbolic meanings inherent to the tradition.
2. Superiority of LDMs over GANs: Unlike previous GAN-based approaches, LDMs offer better semantic control through text conditioning and produce outputs with higher structural fidelity and fewer artifacts, making them more suitable for complex, symbolically rich textiles.
3. Human-AI Collaboration: The research positions AI not as a replacement for artisans but as a collaborative tool. The validation by traditional weavers confirms that these models can produce designs that are technically feasible for manual weaving, bridging the gap between computational creativity and traditional craftsmanship.

The authors conclude that this approach offers a sustainable paradigm for digital cultural preservation, allowing for the reinterpretation and advancement of Indonesian textile heritage in the era of intelligent design.