An explicit formulation of the learned noise predictor $ε_θ({\bf x}_t, t)$ via the forward-process noise $ε_{t}$ in denoising diffusion probabilistic models (DDPMs)

Abstract: In denoising diffusion probabilistic models (DDPMs), the learned noise predictor $ \epsilon_{\theta} ( {\bf x}t , t)$ is trained to approximate the forward-process noise $\epsilon_t$. The equality $\nabla{{\bf x}t} \log q({\bf x}_t) = -\frac 1 {\sqrt {1- {\bar \alpha}_t} } \epsilon{\theta} ( {\bf x}t , t)$ plays a fundamental role in both theoretical analyses and algorithmic design, and thus is frequently employed across diffusion-based generative models. In this paper, an explicit formulation of $ \epsilon{\theta} ( {\bf x}t , t)$ in terms of the forward-process noise $\epsilon_t$ is derived. This result show how the forward-process noise $\epsilon_t$ contributes to the learned predictor $ \epsilon{\theta} ( {\bf x}_t , t)$. Furthermore, based on this formulation, we present a novel and mathematically rigorous proof of the fundamental equality above, clarifying its origin and providing new theoretical insight into the structure of diffusion models.