StrideDiffusion: Accelerating Diffusion Models for Time-series Generation

StrideDiffusion is a training-free, spectral-aware sampler for time-series diffusion models. It monitors band-level energy, log-power drift, and phase velocity during reverse diffusion, taking fine steps while high-frequency activity remains and larger jumps when mainly stable low-frequency structure is left. According to the abstract, it reduces sampling to 14–66 function evaluations across six unconditional generation benchmarks, versus 500 or 1,000 denoising steps, with up to 18.9x wall-clock speedup. Conditional imputation and forecasting reportedly achieve 5–14x average acceleration with comparable predictive accuracy.