Image Synthesis using Generative Adversarial Networks

Abstract

In the past few years, significant progress has been made in artificial intelligence transforming many industries. Recently, the emergence and rapid adoption of advanced large language models like OpenAI's GPT, Google's PaLM, and Meta's Llama have shown great potential and sparked considerable global interest. However, there is a critical knowledge gap in the construction sector when it comes to the opportunities and challenges of implementing Generative AI, specifically in image-to-image generation tasks using Generative Adversarial Networks (GANs). By processing massive datasets of pictures, GAN models learn the essence of visual elements like color, shape, and color composition in the image. Within the world of artistic exploration, the generation of visual content opens the limitless potential for creativity. It also has the potential to generate an image in the style of Van Gogh or in the style of Raja Ravi Varma. Among these, CycleGAN has emerged as a powerful model for image-to-image translation tasks without requiring paired data. The following study explores the application of CycleGAN for preserving the content of input images while experimenting with various hyperparameters to enhance the model’s performance. The objective is to assess the efficacy of CycleGAN in maintaining the integrity of the original content and compare its results against established tools such as NightCafe, Neural Style Transfer and Deep Dream Generator. The results of the following study indicate the that with careful tuning of hyperparameters, CycleGAN can achieve superior content preservation while generating visually appealing transformations. Furthermore, the comparative analysis reveals the strengths and limitations of each tool, highlighting CycleGAN’s potential as a robust alternative for image transformation tasks.