Light Wasserstein Generative Adversarial Network with Gradient Penalty (LWGAN-GP) for Batik Nitik Synthetic

The development of research in the field of image generation can now be one of the tools to introduce a country's artistic culture. Indonesia is known as one of the countries with cultural diversity. Batik is one of Indonesia's most iconic cultural arts. GAN studies on batik have produced good images but still have a long training process. Several GAN techniques have been developed to create synthetic batik, one of the pioneering ones is BATIK-GAN which proposes a double flow model. Several GAN techniques have been developed to create good images, one of the best known and most widely used as a model is DCGAN due to its lightweight training and ability to produce good synthetic images. In this research we propose the Light Wasserstein Generative Adversarial Network model with Gradient penalty (LWGAN-GP) to be able to generate batik motif images. This model is proposed to produce a model that can perform a faster training process but can still produce good images. The experimental results of LWGAN-GP compared with DCGAN show that the computation time of DCGAN is faster than that of LWGAN-GP but LWGAN-GP is superior to the variety of images produced and is able to remain stable (without collapse mode).