Fine-tuning GloVe Embedding with Contextual Information for Synthetic Batik Pattern Generation

Batik is an Indonesian cultural heritage that is rich in philosophical and symbolic meanings. To support preservation and innovation, this research examines the use of Generative Adversarial Networks (GAN) in generating synthetic batik patterns based on natural language description. The main challenge lies in the interpretation of semantically complex cultural texts. This research proposes a fine-tuning approach of the GloVe word insertion model with a batik domain-specific corpus. The dataset consists of 3,100 batik images of Parang and Kawung motifs, each accompanied by 10 textual descriptions. Two approaches were evaluated: GloVe generalized pre-training and GloVe enhanced. The GAN architecture combines multimodal input and up sampling techniques to generate images from text. Intrinsic evaluation results showed that the customized GloVe model improved the average cosine similarity value to 0.99. A paired t-test between the general model and the refined results yielded p < 0.01, indicating a statistically significant improvement. Extrinsic evaluation using Fréchet Inception Distance (FID) and Inception Score (IS) showed an improvement in visual quality: FID decreased from 64.5 to 48.1, and IS increased from 2.37 to 3.23. These findings demonstrate the effectiveness of semantic enhancement for improving the synthesis of culturally meaningful visuals. In addition to the technical contribution, this study demonstrates the potential of AI in the preservation of Indonesia's cultural heritage through.