CLIPBRIDGE-GAN: Bridging Prompt Semantics and Image Synthesis via CLIP-Guided Adversarial Training

Text-to-image synthesis remains a central challenge in generative modeling due to the complexity of aligning high-dimensional visual outputs with free-form natural language. While recent models based on diffusion or autoregressive techniques have achieved remarkable fidelity, they come at the cost of immense computational requirements, slow inference, and difficulty in controlling semantic consistency. In this paper, we propose CLIPBridge-GAN, a lightweight, efficient, and controllable generative adversarial framework that leverages the representational power of pretrained CLIP models in both the generator and discriminator. Our generator is guided by CLIP-based prompt embeddings fused through a learnable bridge module, enabling precise semantic conditioning from text. The discriminator uses CLIP to evaluate image-text alignment directly, eliminating the need for handcrafted auxiliary losses. This dual integration of CLIP enables CLIPBridge-GAN to operate with significantly fewer parameters and only a fraction of training data, while still producing competitive results compared to larger diffusion-based systems. Furthermore, the GAN architecture allows for real-time synthesis and interpretable latent space traversal. Qualitative and quantitative results demonstrate the effectiveness of our approach in generating coherent, high-quality images with robust prompt alignment, even under constrained computational settings.