Wing- Generative Adversarial Networks: Design of Diffusion Model for Text-to-Image Representation

Text-to-image generation aims to produce realistic images aligned with provided text. Current models face issues like weak resilience, repetitive outputs, and looping problems. Existing GAN-based approaches often rely on outdated designs, limiting their scalability. This paper presents a novel GAN method that leverages the parallelization power of transformer models. By generating multiple “wing sequences” from each training sample, our approach improves training stability and efficiency. This method enhances reward and loss signals for both the generator and discriminator, resulting in more stable training and better performance compared to traditional adversarial models.