Text-to-Image Generation of Bird Images Using GAN-CLS and MS-GAN

This research describes a text-to-image creation system that generates realistic bird images from textual descriptions using the CUB-200-2011 dataset. The technique combines Conditional Generative Adversarial Networks (CGAN) and Mode-Seeking GANs to improve image quality and variety. The pipeline includes Word2Vec embeddings for text preprocessing, a GAN framework trained over 1000 epochs to generate images, and a Gradio-based interface for real-time user interaction. The quantitative evaluation results show an Inception Score of 1.0568, a standard deviation of 0.0044, and a Frechet Inception Distance of $\mathbf{2 9 9. 0 6 8 5}$. Classifier metrics indicate 76.3% accuracy and a 6.5% F1-score. Despite minor mode collapse concerns, the system generates visually coherent images with various colors and shapes, highlighting the effectiveness of integrating GAN-CLS with MS-GAN for fine-grained synthesis tasks. This work lays a foundation for future advancements in multimodal learning and text-to-image generation, with potential applications in automated design, education, and accessibility.