PROJECT TITLE :

Improved ArtGAN for Conditional Synthesis of Natural Image and Artwork

ABSTRACT:

This research offers a number of innovative ways to improve the generative adversarial network (GAN) for conditional picture synthesis, and we call the suggested model 'ArtGAN'. î As an important novelty in ArtGAN, back-propagation of the loss function gradient from the categorical discriminator to the generator has been implemented. The generator is able to learn more quickly and produce images of a higher quality thanks to the label information. Incorporating an autoencoder into the categorical discriminator was inspired by current research. Finally, we introduce a new method for improving image quality. ArtGAN is tested on CIFAR-10 and STL-10 using ablation investigations in the experiments. Experiments on CIFAR-10 reveal that the Inception score of our proposed model beats the current state of the art results. On Oxford-102 and CUB-200, we show that ArtGAN can generate plausible-looking images, and can draw realistic artworks based on style, artist, and genre. ArtGAN is a github repository where you can find the source code and model files.