Generating New Ulos Motif with Generative AI Method in Digital Tenun Nusantara (DiTenun) Platform

Abstract

DiTenun is a startup developing a platform that utilizes artificial intelligence to create innovative digital textile patterns for woven fabrics. One of the woven motifs produced is the Ulos motif, a traditional weaving from the Batak tribe that consists of various types, patterns/motifs, and sizes. Currently, DiTenun platform applies two methods to generate Ulos motifs: image quilting and SinGAN. The image quilting method uses synthetic textures to form a new texture by combining blocks from the original texture. The SinGAN is a Generative Adversarial Network (GAN) method that accepts one image motif as input to generate a new motif that resembles the training motif. The new motifs generated by both methods are still repetitive and not diverse (less variation). Therefore, this paper focuses on improving the StyleGAN method, which utilizes two or more Ulos motif images as input to produce new innovative motifs by mixing regularization. Six experimental scenarios are carried out on the Ulos motif image dataset with different numbers of input motifs and hyperparameter tuning. The experiment results are new images with diverse patterns, colour combinations, and merge motif elements. The StyleGAN performance is measured with Frechet Inception Distance (FID) and Kernel Inception Distance (KID) to find the best-quality motif generated based on the six hyperparameter tuning scenarios. The results show that the fourth scenario on Ulos Batak Karo, Gundur Category (Min Max Resolution: 8 and 256, number images 4, on training iteration per resolution = 100000 and max iteration = 50000000) is the best motif generated, based on FID and KID score, are 91.32 and 0.04, respectively.