Amidst the rapid evolution of modern visual design and human-computer interaction, the automatic synthesis of high-fidelity images from natural language descriptions has emerged as a pivotal research frontier in artificial intelligence within the domain of visual communication. Conventional text-to-image generation techniques continue to grapple with limitations in semantic-visual alignment, image fidelity, and model controllability, rendering them insufficient for complex, multi-scene, and multi-semantic generation tasks. In this study, we introduce Variational-Contrastive Learning Generative Adversarial Network (VCL-GAN). This novel image generation framework synergistically integrates a variational autoencoder with a contrastive learning paradigm to address the core challenges of language-driven image synthesis in visual communication optimization. Built upon the GAN architecture, the proposed model employs a dual-path semantic encoding strategy within the generator, wherein text representations derived via Bidirectional Encoder Representations from Transformers (BERT) are simultaneously infused into both a canonical semantic path and a structured latent path facilitated by the variational autoencoder (VAE). To further enrich semantic granularity, we incorporate cross-modal and intra-modal contrastive losses, thereby enhancing semantic coherence and perceptual fidelity in the generated outputs. Extensive experiments conducted on two benchmark datasets, Caltech–UCSD Birds (CUB) and Microsoft Common Objects in Context (MS-COCO), demonstrate that VCL-GAN surpasses state-of-the-art models such as AttnGAN in terms of Inception Score (IS) and Fréchet Inception Distance (FID), while also exhibiting superior robustness and generalizability in subjective user evaluations and ablation studies. The proposed approach not only advances the precision, clarity, and semantic richness of language-conditioned image generation but also presents a compelling paradigm for intelligent visual communication systems, holding significant promise for applications in Artificial Intelligence (AI)-assisted design, cultural content creation, and multimodal human-computer interaction.
Cai et al. (Thu,) studied this question.