Core Concepts
Residual Denoising Diffusion Models (RDDM) propose a dual diffusion process to unify image generation and restoration by introducing residuals and noise diffusion.
Abstract
RDDM introduces a novel dual diffusion process that decouples residual diffusion and noise diffusion for image generation and restoration.
The forward process involves degrading the target image to the degraded input image while injecting noise, resulting in a dual diffusion framework.
The reverse generation process estimates residuals and noise injected in the forward process using two separate networks.
Empirical research suggests using SM-N for tasks requiring diversity and SM-Res for tasks demanding certainty.
An automatic objective selection algorithm is developed to choose between SM-N or SM-Res based on the task requirements.
A partially path-independent generation process is proposed, allowing modifications to the diffusion speed curve without affecting image semantics.
Stats
DDPM [17]とDDIM [51]によるサンプリングプロセスは、係数変換を通じてRDDMに適合する。
ノイズの速度曲線を変更することで画像生成プロセスが失敗する可能性がある。
Quotes
"Our RDDM enables a generic UNet, trained with only an L1 loss and a batch size of 1, to compete with state-of-the-art image restoration methods."
"We envision that our models can facilitate a unified and interpretable image-to-image distribution transformation methodology."