A Comprehensive Review on Noise Control of Diffusion Model

TL;DR

This paper addresses how the noise schedule in diffusion-based image synthesis critically impacts both training efficiency and sample quality. It provides a structured review of a wide range of schedules, from linear and exponential to cosine, sigmoid, Cauchy, Laplace, Fibonacci, logistic, and a learnable monotonic neural-network schedule, with explicit mathematical formulations for each. A key insight is that there is no one-size-fits-all schedule; different schedules excel under different data regimes and stages of the diffusion process, and learned schedules can reduce variance and improve stability. The work offers practical guidance for practitioners to tailor noise schedules to specific tasks and resolutions, ultimately improving convergence speed and image fidelity while acknowledging the importance of parameter tuning.

Abstract

Diffusion models have recently emerged as powerful generative frameworks for producing high-quality images. A pivotal component of these models is the noise schedule, which governs the rate of noise injection during the diffusion process. Since the noise schedule substantially influences sampling quality and training quality, understanding its design and implications is crucial. In this discussion, various noise schedules are examined, and their distinguishing features and performance characteristics are highlighted.

Figures (2)

• Figure 1:
• Figure 2: The noise control as time step increases under different types of noise schedule.