Image Generation from Random Noise using Generative Adversarial Networks

Abstract

Generative Adversarial Networks (GAN) are an exciting and rapidly changing field that promises to generate realistic samples across a range of domains. They generate new and unique images that have never been seen before, opening a wide range of creative applications, such as in art and design, where they can be used to create novel and mind-blowing visual content. In this research paper, we test a GAN model that intakes random noise and creates realistic output images. The generator generates new images from random noise while updating its weights on the basis of whether the discriminator can be fooled to predict the generated images as real ones. We modeled the architectures for the generator and discriminator and evaluated the model on two datasets, the Simpson faces dataset and the Abstract paintings dataset. We study the results generated from the two datasets to understand the type of data suitable for quality image generation. The datasets vary in terms of feature simplicity and size. Learning curves for both datasets showing model losses have been plotted. Batches of images produced by training both datasets have been analyzed to state which one gives the best results. It was found that images with simple features perform better when trained on GANs compared to images containing randomness and difficult-to-learn features. Overall, the research demonstrates the versatility and effectiveness of GANs for image generation from random noise and also provides insights into the current situation of the state-of-the-art in this field.