Antithetic Noise in Diffusion Models

Abstract: We 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, we find that pairing each initial noise with its negation consistently yields strongly negatively correlated samples. To explain this phenomenon, we combine experiments and theoretical analysis, leading to a symmetry conjecture that the learned score function is approximately affine antisymmetric (odd symmetry up to a constant shift), and provide evidence supporting it. Leveraging this negative correlation, we enable two applications: (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, we extend the two-point pairing to a randomized quasi-Monte Carlo estimator, which further improves estimation accuracy. Our framework is training-free, model-agnostic, and adds no runtime overhead.

• The paper demonstrates that pairing Gaussian noise with its negation produces strongly negatively correlated outputs in diffusion models.
• It introduces a symmetry conjecture suggesting that the learned score function behaves approximately as an affine antisymmetric function.
• Practical benefits include enhanced image diversity and a variance-reducing estimator that improves uncertainty quantification without additional computational cost.

Antithetic Noise in Diffusion Models: An Analytical Perspective

The study "Antithetic Noise in Diffusion Models" presents a novel investigation into the impact of antithetic noise on the outputs of diffusion models, a category of generative models that have demonstrated substantial efficacy in producing high-quality synthetic media such as images, audio, and video. The research conducted by Jing Jia et al. systematically assesses how antithetic sampling, a method where Gaussian noise is paired with its negation, affects diffusion model outputs. Notably, this approach is shown to yield samples that are strongly negatively correlated, a finding that remains consistent across various architectures and data types.

Core Findings

The research initiative elucidated a noteworthy phenomenon: when every initial noise vector used in the sampling process of a pretrained diffusion model is paired with its negative counterpart, the resulting samples demonstrate significant negative correlation. This finding holds irrespective of the diffusion model type, dataset, or sampling conditions. The authors propose a "symmetry conjecture" to explain this phenomenon, asserting that the learned score function of these models likely exhibits affine antisymmetry. This assumption aligns with experimental data, which shows that the function is approximately odd-symmetric modulo a constant shift.

Theoretical Contributions and Analysis

The theoretical contribution is pivotal in understanding why negating the Gaussian noise leads to samples with strong negative correlation. The authors support their symmetry conjecture with empirical evidence and a formal analysis, suggesting that the score network behaves almost like a linear, affine antisymmetric function with respect to its inputs. This insights into the inner mechanics of the score function extend the theoretical understanding of diffusion models and potentially unlock new avenues for further refinements.

Practical Implications

Two practical applications of this theoretical investigation are articulated. First, the use of antithetic noise can enhance the diversity of generated images while preserving their quality. This occurs because antithetic pairs push the reverse-diffusion process into distinct regions of the image space, fostering a diversely populated output. Second, antithetic sampling facilitates sharper uncertainty quantification. By leveraging the negative correlation inherent in antithetic pairs, the authors devise a variance-reducing estimator that significantly narrows confidence intervals for downstream statistical estimations—by up to 90% in certain metrics—without incurring additional computational overhead.

Extensions and Future Directions

Building on the antithetic pair framework, the authors extended their methodology to a randomized quasi-Monte Carlo estimator, demonstrating a further improvement in estimation accuracy due to the more optimal sampling of the probabilistic space. The simplicity, model-agnostic nature, and computational efficiency of the presented framework indicate potential applicability in various fields relying on generative models, including computer graphics and Bayesian inference.

Conclusion

This paper presents a compelling case for the systematic exploration of initial noise manipulation in diffusion models. The clear empirical and theoretical evidence provided supports the utility of antithetic noise as a tool for improving both diversity and statistical estimation accuracy in the context of generative modeling. Future research will benefit from verifying and extending the symmetry conjecture and further exploring the effective dimensionality exploited by quasi-Monte Carlo methods in high-dimensional model spaces. Such investigations could significantly influence the design and optimization of generative models across multiple applications.