<p>Generating semantically consistent and high-quality complex images from text remains a challenging task. While large-scale diffusion and autoregressive models have made progress, they suffer from three key limitations: high computational cost, slow generation speed, and weak fine-grained alignment between text and image. To address these issues, we propose Cross-modal Memory GAN (CM-GAN), a novel GAN-based framework for efficient and controllable text-to-image synthesis. CM-GAN incorporates a MALIP-based discriminator, which leverages the complex scene understanding ability of Mamba CLIP (MALIP) to improve image quality assessment. We further design a MALIP-empowered generator, which bridges the semantic gap between text and image by inducing visual concepts from MALIP through predicted joint features and text-conditioned prompts. To enhance fine-grained alignment without additional supervision, we introduce an Implicit Semantic Reasoning (ISR) module that implicitly captures local interactions. We redesign the original CLIP-ViT with Mamba Vision Blocks to form MALIP, enabling efficient processing of long and complex text sequences. Experiments on the CUB dataset, COCO dataset and CC3M dataset show that CM-GAN outperforms several state-of-the-art methods in terms of image fidelity, semantic alignment, and generation diversity. These results highlight the effectiveness of combining understanding and generation, offering new insights for building unified and lightweight multimodal generative models.</p>