Abstract
Diffusion models typically inject isotropic Gaussian noise, disregardingstructure in the data. Motivated by the way quantum squeezed statesredistribute uncertainty according to the Heisenberg uncertainty principle, weintroduce Squeezed Diffusion Models (SDM), which scale noise anisotropicallyalong the principal component of the training distribution. As squeezingenhances the signal-to-noise ratio in physics, we hypothesize that scalingnoise in a data-dependent manner can better assist diffusion models in learningimportant data features. We study two configurations: (i) a Heisenbergdiffusion model that compensates the scaling on the principal axis with inversescaling on orthogonal directions and (ii) a standard SDM variant that scalesonly the principal axis. Counterintuitively, on CIFAR-10/100 and CelebA-64,mild antisqueezing - i.e. increasing variance on the principal axis -consistently improves FID by up to 15% and shifts the precision-recall frontiertoward higher recall. Our results demonstrate that simple, data-aware noiseshaping can deliver robust generative gains without architectural changes.