AI Driven Text To Image (Flowers)

Abstract

This project explores the use of advanced Generative Adversarial Networks (GANs) for text-to-image(flower) synthesis, aiming to generate high-quality, realistic images from textual descriptions. The project leverages various state-of-the-art GAN models, including StackGAN, AttnGAN, DMGAN, and DFGAN, to compare and evaluate their performance in generating images that accurately reflect the input text. The models utilize techniques such as attention mechanisms and memory networks to improve text-image alignment and image quality. The Adam optimizer is employed for efficient and stable training of the generator and discriminator networks, ensuring fast convergence and high performance. Additionally, an Affine Transformer is used to enhance spatial transformations and preserve consistency between text and image features. The project evaluates the models using quantita tive metrics like Fr�echet Inception Distance (FID), which measures the similarity between real and generated images, and qualitative assessments based on human judgment. The expected outcomes include generating high-resolution, diverse images that align well with the given text, providing valuable insights into the effec tiveness of different GAN architectures for text-to-image synthesis. This research has potential applications in fields such as AI-driven content creation, digital art, and virtual design.