Diffusion models demonstrate superior capabilities in tabular data generation tasks, yet challenges remain in modeling heterogeneous features and maintaining global distribution consistency. Moreover, existing research has limited exploration of conditional diffusion for enhancing minority-class samples. This paper proposes ConDTab, a conditional diffusion-based framework built on unified latent encoding and dual attention mechanisms for mixed-type tabular data generation. A Variational Autoencoder enhanced with dual row-column attention facilitates heterogeneous feature fusion and global pattern modeling, yielding an informative latent space for diffusion-based generation. To achieve fine-grained conditional control, a dynamic condition fusion module aligns condition information (e.g., class constraints) with sample features and injects the resulting signals into both input and intermediate layers, guiding the generative process to enhance minority-class fidelity while maintaining global distributional coherence. We evaluate ConDTab on eight public datasets. ConDTab maintains superior stability across distribution alignment and downstream performance when transitioning to imbalanced data, it also achieves the highest efficiency in utility–privacy trade-offs.