Parameter-efficient Quantum Denoising Diffusion Probabilistic Models with temporal encoding

Abstract

Quantum generative models have attracted growing interest for their potential to transform generative learning through the principles of quantum computing. The recently proposed Quantum Denoising Diffusion Probabilistic Models (QuDDPM) represent a significant advancement by integrating classical diffusion mechanisms with quantum computation. However, QuDDPM suffers from a key scalability bottleneck: its parameter count grows linearly with the number of denoising steps, as each step requires independent optimization. To overcome this limitation, we propose a Temporal-aware Quantum Denoising Diffusion Probabilistic Model (TQuDDPM), a parameter-sharing framework that incorporates temporal encoding into the denoising process. Our numerical simulations show that TQuDDPM significantly reduces parameter requirements by up to 94% and training time by up to 90%, all while preserving or even improving generative performance. This work introduces a novel approach to timestep representation in quantum generative learning and demonstrates that TQuDDPM achieves substantial computational efficiency alongside high-fidelity generation.