Research on Text-to-Image Generation Model Based on Spatial Transformer Network Mechanism

The challenges of text-guided image generation technology mainly involve ensuring the semantic consistency of images and their visual realism, as well as enhancing the accuracy and trustworthiness of image generation. In view of the quality limitations exhibited by single - stage models during the image generation process, this paper presents a method to boost the semantic consistency between text descriptions and generated images. By incorporating the attention mechanism, the integration of cross - modal features between text and images is realized. Through the introduction of the spatial transform network model into the discriminator, the model only needs to concentrate on the image regions that are highly relevant to the text description, instead of all regions, thereby avoiding resource waste. By transforming and adjusting the local regions of the input image, it becomes possible to process and represent the spatial information of the image more precisely. Through experimental verification, the model proposed in this paper shows an improvement in the quality of generating complex scene images based on text descriptions. On the COCO dataset, the Inception Score (IS) has increased by 1.2%, and the Frechet Inception Distance (FID) has decreased by 0.61%. This result clearly demonstrates that the single - stage generation model integrated with the spatial transform network mechanism can generate more diverse images that are closer to real -world scenes.