This study introduces Swin-MPGM, a novel scene text generation approach using GANs to separate text from complex backgrounds. It features a U-shaped architecture with a Swin-Transformer encoder for long-range dependencies and a CNN-decoder for image structure. The method addresses challenges like character recognition under wear, camera, and lighting variations. It innovatively generates image labels in real-time without pixel-level supervision, using text annotations and spatial info. It also includes an automatic multi-task loss balancing mechanism guided by homoscedastic uncertainty, optimizing training dynamically. Experiments on the Container Number datasets show improved generalization and robustness.