Overview of Diffusion Model Training and Sampling Process

扩散模型训练与采样过程梗概 Overview of Diffusion Model Training and Sampling Process Diffusion Models are inspired by non-equilibrium thermodynamics. They define a Markov chain of diffusion steps that gradually add random noise to the data, and then learn a counter-diffusion process to construct the desired data samples from the noise. In the training stage, noise is added to the picture, and the noise-added picture is input to the network. What the network needs to predict is the added noise. In the use stage, from randomly generated noise, the network is used to predict what noise has been added, and then the noise is gradually removed until it is restored. --- 1. Diffusion process - The process of diffusion is to continuously and gradually add noise to the image until the image completely becomes pure noise; - The added noise is Gaussian noise, and the next moment is obtained by adding noise to the image at the previous moment. Process of adding noise: Two parameters are defined here$t$的范围为0~$T$,$\betat$ 从$\beta1$;($\betaT$,是逐渐变大的，论文中是从0.0001等量的增加$T$The range is 0~$\alphat$与$\betat$an integer between,$\alphat = 1- \betat$ （1) $T$from$t$则代表的是$T$change to$xt=\sqrt{\alphat}x{t-1}+\sqrt{1-\alphat}zt$ (2) $xt$, is gradually getting larger. In the paper, it is an equal increase from 0.0001.$t$扩散步时，添加噪声后的图片，而$x{t-1}$times until 0.002) And$t-1$时刻所获得的图片；$zt$and$t$时刻所添加的噪声，该噪声采样自标准正态分布$N(0,1)$The relationship is:$x0$逐步添加噪声，扩散至$xT$(1)$x1=\sqrt{\alpha1}x{0}+\sqrt{1-\alpha1}z1$ $x2=\sqrt{\alpha2}x{1}+\sqrt{1-\alpha2}z2$It represents the…