Getting real-world data into quantum computers is a non-trivial task and the first step to many relevant applications. The encoding choice may jeopardize potential quantum speedups, and therefore is a fundamental part of quantum algorithm design. Out of NISQ-suitable embeddings, the Qsample encoding stands out for its qubit-efficiency and because qGANs can prepare it using a polynomial number of gates. However, literature only exemplifies hybrid qGANs up to three qubits.

In this work, we achieve state-of-the-art accuracy for Qsample encoding on registers up to five qubits. Our high-resolution data loading benefits algorithms of practical interest that accept approximate state preparation, such as quantum amplitude estimation and HHL. Our contribution consists of the novel UniqGAN architecture, an offline optimizer based on manifold line search, and a progressive growth training scheme, whose applicability is shown in a profit estimation use case in energy economics. When training with progressive growth, the UniqGAN fits the test distribution 63.69% more accurately on average and is 80.30% shallower than the best fitting benchmark qGAN. Additionally, the quantum option pricing algorithm using said qGAN produces results up to 3.07% close to classical Monte Carlo in our experimental setup.