9th International Conference on ICT for Sustainable Development

Authors - Karan Chopra, Vatsal Mehta, Jayu Jain, Gayatri Joshi, Shanthi Therese
Abstract - The goal of this project is to increase the quality of the images generated using textual description by combining DF-GAN with ChatGPT. Initially, we experimented with GAN on MNIST dataset, and then we tried stackedGAN on Oxford-102 datasets. Unfortunately, these approaches had issues such as sub-par image quality and long training times. So, we moved onto DF-GAN with CUB & COCO dataset and saw the impact of better user prompts on improving image generation. A key development in this project is the integration of ChatGPT into the backend to improve prompt quality. By using ChatGPT, we can create more nuanced & contextually relevant prompts that significantly improve the expressiveness & accuracy of the images generated. The evaluation process includes metrics like sharpness & noise to provide an evaluation of the image quality that adds some value. In addition, the user-friendly interface using Streamlit improves accessibility, allowing a wider range of users to interact with our image generating model. This project develops as a systematic analysis of different GAN architectures and dataset combinations. It provides an extensive approach for advancing text to image generation, with an emphasis on practical usability.