An explicit formulation of the learned noise predictor $\hat{I}{\mu }_{\hat{I}}¸({\mathbf{x}}_t,t)$ via the forward-process noise $\hat{I}{\mu }_t$ in denoising diffusion probabilistic models (DDPMs)

KiHyun Yun

In denoising diffusion probabilistic models (DDPMs), the learned noise predictor $\hat{I}{\mu }_{\hat{I}}¸({\mathbf{x}}_t,t)$ is trained to approximate the forward-process noise $\hat{I}{\mu }_t$. The equality…