The architectural circuit above is a modified BatikGAN SL series with the addition of content loss. The first step in the model is to enter a collection of batik patches on the generator. Each patchA and patchB batik will be entered into generator B1 and generator B2. Both generators will produce a series of matrices called feature maps. The result of feature maps from the two generators will be combined into a 128x128 image form. In the formation of the image there is a local loss l2 function which makes the generated image have a batik pattern. The resulting image from the generator will be compared with the original batik image using several loss functions, including style loss, adversarial loss or global discriminator loss, and local loss (local adversarial loss + local style loss). In addition, the image generated by the generator will be compared with the batik patch input using the content loss function.

The picture above is an image of batik produced by each scenario model. The resulting models include BatikGAN SL with Augmentation, BatikGAN SL without Augmentation, ContentLoss Weight 1, ContentLoss Weight 10, and ContentLoss Weight 100. The results obtained by each model will be evaluated by FID (Frechet Inception Distance). The smaller the FID value, the better. To try to see other results and see the FID evaluation value, you can try the application demo listed.