Turning Quantum Noise on its Head: Using the Noise for Diffusion Models to Generate Images

In this work, we propose positively using noise from quantum computers, which is currently viewed as a hindrance for performing useful computation, instead of simulated noise to train generative image diffusion models, which have two primary advantages: True Noise Generation. A key quality of random quantum fluctuations is that they are independent of human input. Current means of generating noise in diffusion models are pseudo-random, meaning that randomness is simulated using human-defined algorithms. Parallel Noise Generation. In quantum computers, each quantum bit (or qubit) has random fluctuations, so we can use multiple quantum bits to generate randomness in parallel. To parallelize noise generation classically, we would need to use more resources, which is not needed in quantum computers. Our approach, Quantum Image Noise-based Generative Diffusion Models (referred to as QINGDM), utilizes these benefits of quantum noise.