DDPM公式草稿3

A=αtβt+11αt1B=2(αtxtβt+αt1x01αt1)μ=B2Aσ2=1Aσ2=1αt11αtβtμ=αt1αt11αtxt+αt1βt1αtx0xt=αtxt1+1αtztx0=1αt(xt1αtz~)μ=1αt(xtβt1αtz~)q(xt1xt)N(1αt(xtβt1αtz~),1αt11αtβt)xt1=1αt(xtβt1αtz~)+1αt11αtβtzz~=UNet(xt,t)zN(0,I) A=\frac{\alpha_t}{\beta_t}+\frac{1}{1-\overline{\alpha_{t-1}}} \quad B=-2(\frac{\sqrt{\alpha_t}x_t}{\beta_t}+\frac{\sqrt{\overline{\alpha_{t-1}}}x_0}{1-\overline{\alpha_{t-1}}}) \\ \quad \\ \because \mu=-\frac{B}{2A} \quad \sigma^2=\frac{1}{A}\\ \therefore \sigma^2=\frac{1-\overline{\alpha_{t-1}}}{1-\overline{\alpha_t}}\beta_t \quad \mu=\sqrt{\alpha_t}\frac{1-\overline{\alpha_{t-1}}}{1-\overline{\alpha_t}}x_t+\frac{\sqrt{\overline{\alpha_{t-1}}}\beta_t}{1-\overline{\alpha_t}}x_0\\ \quad \\ \begin{aligned} \because x_t&=\sqrt{\alpha_t}x_{t-1}+\sqrt{1-\alpha_t}z_t\\ x_0&=\frac{1}{\sqrt{\overline{\alpha_t}}}(x_t-\sqrt{1-\overline{\alpha_t}}\widetilde{z})\\ \therefore \mu&=\frac{1}{\sqrt{\alpha_t}}(x_t-\frac{\beta_t}{\sqrt{1-\overline{\alpha_t}}}\widetilde{z}) \end{aligned}\\ 即:q(x_{t-1}|x_t) \sim N(\frac{1}{\sqrt{\alpha_t}}(x_t-\frac{\beta_t}{\sqrt{1-\overline{\alpha_t}}}\widetilde{z}),\frac{1-\overline{\alpha_{t-1}}}{1-\overline{\alpha_t}}\beta_t)\\ x_{t-1}=\frac{1}{\sqrt{\alpha_t}}(x_t-\frac{\beta_t}{\sqrt{1-\overline{\alpha_t}}}\widetilde{z})+\sqrt{\frac{1-\overline{\alpha_{t-1}}}{1-\overline{\alpha_t}}\beta_t}z\\ \widetilde{z}=UNet(x_t,t) \qquad z \sim N(0,I)
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

孤寡孤寡的牛牛很热情:为什么我2本9硕投了很多,都是简历或者挂,难道那个恶心人的测评真的得认真做吗
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