Precise liver tumor segmentation from CT images is essential for diagnosis and treatment planning, yet the scarcity of labeled data impedes its accuracy. This study introduces a Conditional Generative Adversarial Network for modern synthetic data augmentation. The network features a U-Net generator and a convolutional discriminator, conditioned on liver tumor masks to produce anatomically plausible modern synthetic liver tumor images. We focused on tumor segmentation within pre-segmented liver regions, leveraging the high performance of existing liver segmentation methods. Modern synthetic images underwent rigorous selection, ensuring tumors were within liver mask boundaries. We compared state-of-the-art models across three augmentation strategies: no augmentation, traditional augmentation, and generated modern synthetic augmentation. Results demonstrated that the data augmentation technique significantly improved segmentation, achieving the highest Dice Per Case (DPC) and lowest Volume Overlap Error (VOE). Deep QRSA achieved a DPC of 0.889 and a VOE of 0.176 with generative augmentation. Consistent improvements across models highlight its robustness and generalizability. This study demonstrates the efficacy of generative data augmentation for enhancing liver tumor segmentation, offering a promising method for improving diagnostic accuracy in medical imaging.