Framework GAN is designed with the two models, model G and model D constructed to compete against each other simultaneously in each training iteration. Both models are trained to determine whether a sample comes from a model distribution or data distribution. In this case the G model is designed to continuously replicate the original sample image so closely untill the D Model cannot tell the difference. After G model produces a fake image, then D Model uses these results to guess whether the incoming image is a fake image from G model or the original data from the original training data. In this competition, the two models continue to improve their performance so that the fake data cannot be distinguished from the original data.

Based on Figure 7a and 7b, it is apparent that the first model scheme produces several new batik images with quite unique contours and patterns. However, the scheme of the first model produces some images with an irregular and faded pattern as marked with the red outline. With the less apparent results, the first model almost succeeded in imitating a common batik pattern, which is the broken machete pattern as indicated by the green frame. As for the second scheme, the models are more able to generate batik patterns regularly with a fairly clear pattern and a sharper image. However, the second scheme produces irregular batik patterns of an opaque quality and is much more abundant than the first model scheme. The batik pattern of the broken parang, which was almost imitated by the first scheme, is better imitated by the second model scheme which results in a more apparent and sharper look as depicted in Figure 7b with a green frame marker. In addition, the green frame depicts batik pattern in which the generated results look sharp dominating the total output due to different learning rate ranges which also affect the training time of different models. Much lower learning rate ranges relatively takes longer training time than that in larger learning rate ranges. Conversely, a high learning rate value can also result in the model unable to learn well and optimally because changes in the learning rate value are too significant. With the results obtained in this study, it has been successful in depicting that DCGAN will succeed in generating a new, sharper and more diverse batik pattern if the dataset used to train the model is more and more diverse than previous studies [16]. In previous studies, the training dataset used solely includes 300 batik with 50 different types of batik less than the dataset used in this study. In addition, determining the right hyperparameter model is deemed influential in improving the performance of the DCGAN model. In this case, the application of two model schemes with different learning rates and weight initiations is proven to produce different synthetic batik patterns.