DCGAN based Pre-trained model for Image Reconstruction using ImageNet

Despite recent achievements in generative image modeling, generating better quality image samples from complex datasets such as ImageNet remains an illusory goal. The objective of this paper is to train Deep Convolutional Generative Adversarial Network at the well-known CIFAR10 dataset and study the instabilities specific to such scale and then test the large-scale ImageNet dataset for establishment of the proposed DCGAN. We find that applying a pre-trained DCGAN can remove the complexity and also can learn prior details of images and improve the quality of generated image. Our modifications on DCGAN lead to models which set the new cutting edge in class-contingent image reconstruction on pre-trained GAN's. When tested on ImageNet at 128 × 128 resolution, our model (DCGAN) lowers the loss between the generated and real image samples which shows that the proposed DCGAN model works well with both the datasets.