The introduced XPCA Gen method is a novel approach for the generation of synthetic tabular data using the relevant information present in the original data. This is performed using the principal components obtained by the XPCA decomposition of the original data. XPCA is a probabilistic interpretation of standard PCA, which can handle mixtures of continuous and discrete variables. The new data points are generated by synthesizing the principal components, resulting in an accurate representation of real data that is noise-redundant and exhibits a good diversity in data points preserving the original distribution. The experimental results obtained on benchmark datasets (e.g. Credit, PID, Boston housing, etc.) show strong performance in machine learning utility metrics (accuracy, precision, recall) and similarity metrics like high Hausdorff distance, highlighting the method’s ability to capture inherent patterns in the dataset. Moreover, it achieves diversity in data points, without compromising statistical properties. The effect of variance scaling was investigated to regularize the variance of individual features, which increased due to redistribution of variances (since XPCA aims to capture maximum variance in the dataset). The comparative study showed that the regularized data resulted in higher ML utility performance compared to the generated data without regularization for the benchmark datasets. This makes the data suitable for several machine learning experiments due to the high data quality. In general, XPCA Gen and its regularized version emerge as a promising solution for data privacy preservation and robust model training with diverse samples.