arXiv 2508.14871
Squeezed Diffusion Models
By Jyotirmai Singh, Samar Khanna, et al.
Published 2025-08-20
Wiki summary
Explore the paper's summary, context, and related research on Papiers.
Diffusion models typically inject isotropic Gaussian noise, disregarding structure in the data. Motivated by the way quantum squeezed states redistribute uncertainty according to the Heisenberg uncertainty principle, we introduce Squeezed Diffusion Models (SDM), which scale noise anisotropically along the principal component of the training distribution. As squeezing enhances the signal-to-noise ratio in physics, we…