Batik GAN for Generating Motif Synthesis using Multi-Discriminator and Self-Attention

This study proposes a novel approach for generating high-quality synthetic batik motifs by enhancing the Batik Generative Adversarial Network (Batik GAN) architecture with multi-discriminator and self-attention mechanisms. The proposed method addresses limitations in conventional GAN, particularly in capturing the intricate patterns, textures, and cultural elements inherent in traditional batik art. The Batik Nitik Sarimbit 120 dataset, comprising 120 images across 60 classes, serves as the foundation for training and evaluation, enriched with augmentation techniques to ensure robust learning. The multi discriminator framework enables comprehensive evaluation of synthetic motifs by focusing on distinct aspects such as global patterns, local details, and traditional elements, providing the generator with diverse and detailed feedback. Simultaneously, the self-attention layer allows the generator to dynamically prioritize critical motif elements, ensuring the preservation of spatial relationships and overall harmony in the patterns. Quantitative evaluation using FID, PSNR, and SSIM metrics demonstrates the superiority of the proposed method over the baseline Batik GAN SL and other tested variations. The proposed method achieved a significant reduction in FID, PSNR, and SSIM score. These results indicate enhanced visual fidelity, structural consistency, and noise reduction in the generated motifs. The findings highlight the effectiveness of integrating multi discriminator and self-attention mechanisms in Batik GAN, enabling the synthesis of realistic and culturally authentic batik motifs. This research contributes to the advancement of generative modeling in preserving and innovating traditional textile arts, with potential applications in digital heritage preservation and creative industries.