QRNG-DDPM: Enhancing Diffusion Models Through Fitting Mixture Noise with Quantum Random Number

Abstract: Recent research has demonstrated that the denoising diffusion probabilistic model (DDPM) can generate high-quality images in artificial intelligence (AI), showing its distinctive capabilities. However, despite this, the diversity of generated images is often limited by the predictability of traditional pseudo-random number generators in the stochastic process. To address this problem, this paper proposes a new mixed noise model based on quantum random numbers QRNG-DDPM. By operating on single qubits, we generate quantum random numbers and apply a self-developed encoding scheme to convert the quantum random numbers into distributions suitable for noise models. Due to the inherent unpredictability of quantum phenomena, quantum random numbers offer a higher level of randomness and diversity compared to traditional pseudo-random numbers. Our experimental results demonstrate that the proposed method significantly enhances the diversity and unpredictability of the generated images, achieving a \(5.4\\%\) reduction in the Fréchet Inception Distance (FID) score on the CIFAR-10 dataset.

Recommended citation: Y. Peng, X. Li and Y. Wang, "QRNG-DDPM: Enhancing Diffusion Models Through Fitting Mixture Noise with Quantum Random Number," 2024 IEEE International Conference on Quantum Computing and Engineering (QCE), Montreal, QC, Canada, 2024, pp. 92-96, doi: 10.1109/QCE60285.2024.10259.
Download Paper