DDPM公式草稿1

推导1

1βt=αtxt=αtxt1+1αtztxt1=αt1xt2+1αt1zt1xt=αt(αt1xt2+1αt1zt1)+1αtzt=αtαt1xt2+αtαtαt1zt1+1αtztzt1N(0,I)ztN(0,I)αtαtαt1zt1N(0,αtαtαt1)1αtztN(0,1αt)(αtαtαt1zt1+1αtzt)N(0,1αtαt1)=αtαt1xt2+1αtαt1z,zN(0,I)xt=αtα1x0+1αtα1z,zN(0,I)xt=αtx0+1αtzαt=i=1Tαi\begin{aligned} 1-\beta_{t}=\alpha_t \quad \quad x_t&=\sqrt{\alpha_t}x_{t-1}+\sqrt{1-\alpha_t}z_t \\ x_{t-1}&=\sqrt{\alpha_{t-1}}x_{t-2}+\sqrt{1-\alpha_{t-1}}z_{t-1} \\ x_{t}&=\sqrt{\alpha_{t}}(\sqrt{\alpha_{t-1}}x_{t-2}+\sqrt{1-\alpha_{t-1}}z_{t-1})+\sqrt{1-\alpha_{t}}z_{t} \\ &=\sqrt{\alpha_t\alpha_{t-1}}x_{t-2}+\sqrt{\alpha_t-\alpha_t\alpha_{t-1}}z_{t-1}+\sqrt{1-\alpha_t}z_t \\ &\textcolor{red}{\because z_{t-1}\sim N(0,I) \quad z_t \sim N(0,I) }\\ &\textcolor{red}{\therefore \sqrt{\alpha_t-\alpha_t\alpha_{t-1}}z_{t-1} \sim N(0,\alpha_t-\alpha_t\alpha_{t-1}) \quad \sqrt{1-\alpha_t}z_t\sim N(0,1-\alpha_t) } \\ &\textcolor{red}{\quad (\sqrt{\alpha_t-\alpha_t\alpha_{t-1}}z_{t-1}+\sqrt{1-\alpha_t}z_t)\sim N(0,1-\alpha_t\alpha_{t-1})} \\ &=\sqrt{\alpha_t\alpha_{t-1}}x_{t-2}+\sqrt{1-\alpha_t\alpha_{t-1}}z,\quad z\sim N(0,I)\\ x_t&=\sqrt{\alpha_t\dots\alpha_1}x_0+\sqrt{1-\alpha_t\dots\alpha_1}z,\quad z\sim N(0,I)\\ x_t&=\sqrt{\overline{\alpha_t}}x_0+\sqrt{1-\overline{\alpha_t}}z \quad \quad \overline{\alpha_t}=\prod_{i = 1}^{T}\alpha_i \end{aligned}
xN(μ,σ2I)zN(0,I)x=μ+σzN(μ,σ2I)\begin{aligned} &\textcolor{red}{x\sim N(\mu,\sigma^2I)}\\ &\textcolor{red}{z\sim N(0,I)}\\ &\textcolor{red}{x=\mu+\sigma z\sim N(\mu,\sigma^2I)}\\ \end{aligned}
扩散模型学习 文章被收录于专栏

扩散模型学习

全部评论

相关推荐

狠赚笔第一人:学计算机自己不努力怪大环境?我大一就拿到了美团大厂的offer,好好看看自己有没有努力查看图片
点赞 评论 收藏
分享
点赞 收藏 评论
分享
牛客网
牛客企业服务