Abstract Diffusion probabilistic models (DPMs) have demonstrated remarkable performance in high-resolution image synthesis. However, their sampling efficiency remains limited due to the large number of sampling steps required. Recent advancements in high-order numerical solvers have reduced the sampling steps, yet the discretization schedules employed are often optimized solely based on pixel-space Euclidean distances. Since natural images exhibit a power-law spectral distribution where low-frequency components dominate the signal energy, this optimization inherently exhibits a spectral bias. Consequently, the solver fails to fully leverage the distinct recovery dynamics of different frequency components, often leading to the loss of critical high-frequency details in few-step regimes. We propose a novel frequency-aware optimization framework for designing discretization schedules. By decomposing the approximation error of the numerical solver in the Fourier domain, we reformulate the time-step search as a constrained optimization problem. This objective incorporates a frequency-dependent weighting function, guiding the optimization algorithm to allocate denser time steps during the critical intervals. Our method is training-free and compatible with various solvers. Extensive experiments on CIFAR and ImageNet demonstrate that the optimized schedules consistently improve generation quality in the few-step regime, achieving competitive results with only 5–15 neural function evaluations. These results highlight the effectiveness of frequency-aware optimization for efficient diffusion sampling.
Zheng et al. (Fri,) studied this question.