Generating Images from Text Using GANs vs Transformers

This paper compares Generative Adversarial Networks (GANs) and transformer-based models for text-to-image generation, examining each approach's ability to convert textual descriptions into realistic images. GANs, leveraging a generator-discriminator framework, have made strides with models like DF-GAN [1] and StackGAN [2], which improve image fidelity through multi-stage generation and attention. However, GANs face challenges with complex datasets. Transformer-based models, initially used in natural language processing, offer robust handling of dependencies, enhancing the capture of image features from text. Emerging hybrid models combine GAN and transformer strengths for improved stability and image quality. This study aims to identify optimal techniques by analyzing their architectures, training, and output quality, contributing to advancements in multi-modal generative modeling.