This paper exploits the relationship between DDPMs and probabilistic localization to overcome the inference bottleneck of denoising diffusion probabilistic models (DDPMs). By proving that the incrementality of DDPMs satisfies the exchangeability property, we demonstrate that various performance optimization techniques based on autoregressive models can be applied to the diffusion setting. Specifically, we propose "Automatic Predictive Decoding" (ASD), an extension of the predictive decoding algorithm widely used for DDPMs without requiring auxiliary models. We demonstrate through theoretical analysis that ASD achieves $\tilde{O}(K^{\frac{1}{3}})$ parallel execution speedups over K-stage sequential DDPMs, and experimentally demonstrate that it significantly accelerates DDPM inference in various applications.
