Tabular data are fundamental to critical decision-making across healthcare, finance, and manufacturing domains. However, generating high-quality synthetic tabular data remains challenging because of heterogeneous feature types, severe class imbalance, and stringent privacy requirements. To address these challenges, we propose TabCL, a generative framework that combines a variational autoencoder (VAE) with a denoising diffusion model, reinforced by contrastive learning and consistency regularization. The contrastive learning component sharpens class boundaries in the VAE's latent space, improving class separation and representation quality. Further, consistency regularization ensures that latent codes perturbed with different noise levels reconstruct to identical outputs, which enhances both sample diversity and model robustness without increasing computational complexity. Extensive experiments on six public benchmarks, four classification and two regression datasets, demonstrate that the proposed TabCL outperforms the existing methods including SMOTE, CTGAN, and TVAE across all standard quantitative metrics. Distributional analyses further reveal that TabCL more accurately reproduces rare categorical levels, heavy-tailed numerical outliers, and complex cross-feature correlation, resulting in synthetic data whose statistical properties closely align with those of the real data even under severe class imbalance and limited-sample sizes. By simultaneously improving latent-space structure and diffusion-based generation, TabCL produces high-fidelity, privacy-respecting synthetic tabular data suitable for downstream modelling and data sharing. Future extensions will target longitudinal datasets and incorporate formal differential privacy guarantees to enable broader deployment in privacy-sensitive industrial environments.