Quantum Circuits in Diffusion Models: Fair Study and Failure Analysis

Researchers conducted a fair-comparison study on integrating variational quantum circuits (VQCs) into diffusion models, using a squeeze-and-excitation (SE) channel-modulation scaffold to isolate the quantum contribution. Across various models like DDPM, latent diffusion, and score-based NCSN on datasets like MNIST and CIFAR-10, quantum cores achieved mean FID scores comparable to their classical counterparts. Despite using significantly fewer core parameters (4.5 to 9 times less), the study did not establish a quantum parameter-efficiency advantage, as parameter-matched classical controls also attained comparable mean FID. This suggests that while quantum circuits can perform similarly, they don't inherently offer a parameter-efficiency edge in these specific configurations. A critical finding was a structural failure in score-based NCSN models: the unbounded score target, which is proportional to 1/sigma, drives angle-embedding inputs far beyond the 2pi period of rotation gates. This leads to phase aliasing and a collapse of the quantum modulator. The researchers successfully remedied this by applying a bounding transformation, theta = pi * tanh(.), which maps inputs to a non-aliasing domain, substantially improving the performance of both quantum cores. The study emphasizes a fair-comparison protocol for quantum-enhanced generative models and provides a mechanistic understanding of angle embedding failures.