A1427
Title: Antithetic noise in diffusion models
Authors: Guanyang Wang - Rutgers University (United States) [presenting]
Abstract: The purpose is to initiate a systematic study of antithetic initial noise in diffusion models. Across unconditional models trained on diverse datasets, text-conditioned latent-diffusion models, and diffusion-posterior samplers, it is found that pairing each initial noise with its negation consistently yields strongly negatively correlated samples. To explain this phenomenon, experiments and theoretical analysis are combined, leading to a symmetry conjecture that the learned score function is approximately affine antisymmetric (odd symmetry up to a constant shift), and evidence is provided supporting it. Leveraging this negative correlation, two applications are enabled: (1) enhancing image diversity in models like stable diffusion without quality loss, and (2) sharpening uncertainty quantification (e.g., up to 90\% narrower confidence intervals) when estimating downstream statistics. Building on these gains, the two-point pairing is extended to a randomized quasi-Monte Carlo estimator, which further improves estimation accuracy. The framework is training-free, model-agnostic, and adds no runtime overhead.
