Multi-Semantic Fusion Generative Adversarial Network for Text-to-Image Generation

Text-to-image synthesis always faces two challenges: image quality and image-text alignment. Existing methods mainly use a single sentence to synthesize images, which are challenging to extract adequate semantic features, resulting in the generated images being far apart from ground-truth images. In this paper, we propose a novel Multi-Semantic Fusion Generative Adversarial Network. Our model can fuse the same semantics from different sentences and preserve their unique semantics to generate accurate images. In addition, we have designed a multi-sentence joint discriminator to ensure that the generated images match all sentences. Experiments on CUB and MSCOCO datasets demonstrate that our model has significant advantages.