<p>We propose a data-generation framework that consists of deep conditional quantile function estimation and sequential high-density data sampling. To address the quantile crossing problem, the proposed conditional quantile representation integrates I-spline regression and neural networks for varying coefficients. The proposed conditional quantile estimates also effectively represent Rosenblatt’s transformation by sequentially arranging the estimators. In addition, we propose a data sampling method that sequentially generates data on the quantile of the highest density regions. The proposed model can be used to determine the highest-density regions because it can locally estimate the slope of the conditional quantile functions for each knot interval. This process assumes that high-quality data are related to high-density data. We demonstrated the weak consistency of the proposed model in terms of its connectivity with the proposed sampling method. An attractive feature of our proposed framework is its unified approach, which generates diverse types of data, such as image, text, and mixed-type data, as well as its efficiency, which can reduce cherry-picking behavior during data generation. The performance of the proposed framework is demonstrated by numerical experiments and real data examples.</p>