Text-to-Image GAN for Realistic Flower Image Synthesis using Textual Input

This research investigates a unique method of using GANs to generate pictures, primarily for the purpose of creating lifelike floral images from textual input of flower names. The GAN model was trained to capture the distinctive traits, colors, and textures that are exclusive to each flower category by utilizing a large dataset that includes a variety of photos of different flower varieties. In order to reduce artifacts and improve realism, this technique uses a dual-model structure (Generator and Discriminator) that allows the model to repeatedly revise and enhance the produced pictures. The final model produced accurate findings, effectively capturing the minor differences between flower varieties, as demonstrated by the qualitative analysis of the photos that were produced. Our approach included meticulous data pretreatment, model architecture optimization, and hyper parameter tweaking to guarantee precise color representation and structural clarity in the pictures that were produced. High-quality photos that capture the distinctive characteristics linked to each flower name are produced by optimizing the GAN's capacity to distinguish between different flower varieties through the model architecture and training procedure. This study offers a framework for producing particular picture kinds from descriptive inputs, which may find use in augmented reality, educational resources, and creative design. The produced photos were thoroughly assessed using a sophisticated accuracy criteria for five flower categories: rose, jasmine, hibiscus, periwinkle, and crossandra. With an average accuracy of 87%, the model constantly performed well, demonstrating its strong capacity to create realistic and varied floral pictures. These findings highlight the effectiveness of GANs in tasks involving the production of text to images as well as its potential for wider use in visual synthesis.