StrideDiffusion implements a training-free sampler that adaptively adjusts denoising strides based on frequency band activity. By monitoring log-power drift and phase velocity, the method takes smaller steps during high-frequency transitions and larger jumps when low-frequency structures dominate.
HOW THIS AFFECTS YOU
●
builderThis provides a path to reducing latency in real-time time-series generation tasks.
●
researcherYou can optimize inference speed for time-series diffusion without retraining models.