Improved Single-Stage Text-to-Image Method

In recent years, text-to-image as an important research direction of multimodal technology, are widely used in design, education and other fields. Generative Adversarial Network (GAN) has become a mainstream method due to its powerful generative capability, but still faces problems such as mode collapse, semantic consistency and image quality. In this paper, we propose an improved DF-GAN model that introduces a word-level attention mechanism to enhance textual information utilization, and adds diversity loss to the generator loss function to mitigate mode collapse. Experimental results show that the improved model significantly improves image generation quality and semantic consistency on the CUB and COCO datasets, and the FID is reduced by 7.29% and 4.55% respectively compared to the baseline model. The ablation experiments validate the effectiveness of the diversity loss and word-level attention mechanisms. Future research can further explore more efficient attention mechanisms and more stable training methods.