An AI tool now designs motifs for Indonesia's Batak Ulos weaving, and actual weavers gave it a passing grade.
Researchers fine-tuned Stable Diffusion XL with LoRA and combined it with four separate ways to steer output - text prompts, reference images, a representation-based signal, and ControlNet-guided semantic maps - then tested combinations across shape, color, and high-complexity changes. Stacking more conditioning methods did not automatically help: text, image, and semantic-map conditioning produced the sharpest images (FID 270) but the least structurally consistent ones (SSIM 0.65), while text, image, and representation conditioning was the most balanced (SSIM 0.84, FID 280), and using all four at once scored worst overall (FID 330). The researchers attribute that last result to the conditioning signals working against each other. Nine professional weavers and thirty members of the public then rated the generated motifs, and both groups approved at statistically significant levels.
Ulos weaving is a centuries-old Batak textile tradition from North Sumatra, and like a lot of hand-woven heritage crafts, it is losing practitioners faster than it is gaining new motifs. This is a rare case of generative AI being pointed at helping a craft rather than replacing it - the tool is meant to hand weavers new design options, not cut them out of the process, and the fact that working weavers rated the output favorably matters more than the benchmark scores.
The paper also credits a multimodal language model it calls 'LLaMA 1.5-7B' for the text conditioning, but no such version exists - LLaMA releases run 1, 2, 3, not fractional numbers like 1.5, which is closer to LLaVA's naming scheme - so that detail is worth a raised eyebrow even if the rest of the study checks out.