Experimental Exploration of Neural Style Transfer: Hyperparameter Impact and VGG Feature Dynamics in Batik Motif Generation

Abstract

Innovating traditional batik designs while preserving their cultural essence remains a significant challenge in the intersection of heritage conservation and computational creativity. This study addresses this challenge by optimizing Neural Style Transfer (NST), a deep learning technique to synthesize batik motifs that harmonize structural fidelity and stylistic authenticity. Focusing on hyperparameter adjustments tailored to batik’s abstract geometries, we systematically evaluate the impact of layer selection in VGG and pooling operations (max-pooling vs. average pooling) on style-content synthesis. Experiments reveal that shallow layers (e.g., conv2, conv4) preserve explicit motifs and edge details (SSIM 0.85), while deeper layers (conv16, conv24) generate abstract textures. Average pooling demonstrates superior stability, achieving smoother convergence (loss stabilized at 0.5 vs. 3.0 for max pooling) and higher structural coherence (SSIM 0.6963 vs. 0.6634), whereas max-pooling introduces fragmented artifacts due to gradient explosion. The optimized framework, validated through quantitative metrics (MSE Loss 0.02, SSIM 0.82) and qualitative artisan evaluations, successfully transferring the style image into the image, while maintaining the original content of the image, with adjusted hyperparameters. This work advances AI-driven tools for batik preservation, offering a scalable methodology to sustain Indonesia’s intangible heritage in the digital era.