One of the big challenges in the use of copula in many problem is its estimation, as a proposition. In this work, we propose a simple yet effective estimation approach: we reformulate copula-density learning as a classification task that separates dependent samples from independent samples obtained by breaking dependence. The classifier’s score yields a density-ratio estimate, from which we directly recover the copula density. The method is flexible, and avoids restrictive parametric assumptions. Experiments on synthetic and real data show strong performance and improved stability compared to standard baselines, highlighting the effectiveness of the proposed approach.