Modern deep learning systems are constrained by a persistent shortage of sufficiently large, high-quality training data. Synthesis of tabular training data using generative models such as conditional GANs offers a promising solution, but the utility of current models is still limited as they show problems in capturing complex relationships of categorical features. Therefore, we introduce the concept of multi-dependence conditional vectors (M-DCV), which allow for setting multiple conditions on the creation of categorical data directly during the training and sampling process of conditional GANs. We propose three different strategies to create M-DCV: first, based on Bayesian Networks (M-DCV Bayes), second on the association measurement Cramer’s V (M-DCV Cramer), and third one involving a cloning process (M-DCV RPC). We evaluate on two tabular GANs and three datasets using a unified pipeline for utility, fidelity, and privacy. Across all experiments, there is an overall average improvement in the quality of categorical fidelity by 33.2% with the RPC approach, 13.2% with the Bayes method, and 8.8% using the Cramer modification. Additionally, we can further enhance the machine learning utility with all M-DCV variants, although this comes at the cost of reduction in privacy scores.