Generative Art: A DC-GAN Approach for Visualizing Descriptions

Text-to-Image Synthesis has gained significant attention due to its ability to generate realistic visuals from textual descriptions. Existing methods, such as Attention-based GAN s and Transformer models, often face challenges in maintaining semantic alignment between text and images, handling complex textual inputs, and ensuring high-quality outputs. To address these limitations, this paper proposes a Deep Convolutional Generative Adversarial Network (DC-GAN) framework leveraging GloVe embeddings for semantic representation. The Oxford 102 Flower dataset, comprising image-text pairs, is used to train and evaluate the model. The DC-GAN architecture includes a generator to synthesize images and a discriminator to validate authenticity, enabling iterative improvement. Evaluation metrics show an average Learned Perceptual Image Patch Similarity (LPIPS) score of 0.286 and a Mean Squared Error (MSE) of 1.455, demonstrating semantic coherence and low pixel-level deviations. Challenges such as refining fine-grained image details and enhancing robustness across diverse datasets persist. This study highlights the potential of DC-GAN for multimodal AI applications, including targeted advertising, virtual reality, and creative design. By advancing text-to-image synthesis, the research provides a robust framework for future exploration in generating semantically consistent and high-quality visuals from text.