StrideDiffusion Accelerates Time-Series Diffusion Models Significantly.

Diffusion models have emerged as powerful tools for generating time-series data, but their practical application is often hindered by the extensive number of sequential denoising steps required during inference. Traditional fast samplers typically employ fixed or generic schedules, failing to account for the unique characteristic of time-series diffusion where different spectral bands evolve at varying rates. Introducing StrideDiffusion, a novel training-free sampler that intelligently adapts its denoising stride based on the activity of spectral bands. This method monitors relative band energy, log-power drift, and phase velocity at each step to determine whether high-frequency dynamics are still active or if the trajectory is dominated by stable low-frequency structures. It then takes finer steps when rapid changes occur and larger jumps when only coarse components remain. Evaluations across six time-series generation benchmarks demonstrate StrideDiffusion's effectiveness, reducing the number of function evaluations from hundreds to as few as 14-66. This translates to up to an 18.9x wall-clock speedup for unconditional generation and 5-14x acceleration for conditional imputation and forecasting, all while preserving or enhancing generation quality. The research highlights that spectral evolution provides a practical and principled signal for significantly faster time-series diffusion sampling.