XLTransGAN: A Fusion of XLNet and GANs in Generative Models for Text-to-Image Synthesis

This paper demonstrates the working of an Artificial Intelligent System capable of generating realistic images from natural language descriptions, an area known as text-to-image generation. This system uses XLNET Transformer for language representation and Generative Adversarial Networks (GANs) to build a full text-to-image generation model. The Oxford 102 Flower dataset consisting textual descriptions linked with relevant photographs is used to train the model. The XLTransGAN comprises training a neural network with a paired dataset of images and text descriptions, with the goal of eventually achieving high-quality images that closely resemble input text description. As a result of XLNet’s better contextual comprehension than typical CNN-based text encoders, the XLTransGAN text-to-image model produced images with remarkable variety and quality, as evidenced by its Inception Score of 3.023. The fundamental goals of such systems are to generate coherent, unambiguous, and photorealistic visuals that correspond to the semantic context of the input text descriptions.