0、快速访问
论文阅读笔记:Denoising Diffusion Implicit Models (1)
论文阅读笔记:Denoising Diffusion Implicit Models (2)
论文阅读笔记:Denoising Diffusion Implicit Models (3)
论文阅读笔记:Denoising Diffusion Implicit Models (4)
4、DDPM与DDIM的相同点与不同点
4.1、 相同点
DDPM与DDIM的训练过程相同,因此DDPM训练的模型可以直接在DDIM当中使用,训练过程下所示
4.2、不同点
DDPM与DDIM在推理阶段是不同的。
DDPM在推理阶段的采样过程如下图所示。首先模型
ϵ
θ
\epsilon_\theta
ϵθ预测出
x
0
→
x
t
x_0\to x_t
x0→xt所添加的噪音
ϵ
t
\epsilon_t
ϵt,然后根据公式
(
x
t
−
1
−
α
t
1
−
α
ˉ
t
⋅
ϵ
t
)
\Big(x_t-\frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}_{t}}}\cdot \epsilon_t\Big)
(xt−1−αˉt1−αt⋅ϵt)得到
x
t
−
1
x_{t-1}
xt−1分布的均值,最后在均值上添加对应的噪音,得到
x
t
−
1
x_{t-1}
xt−1
接下来介绍DDIM的采样过程。根据上文论文阅读笔记:Denoising Diffusion Implicit Models (2)中公式(2)所示的前向加噪过程:在给定
x
0
x_0
x0和
x
t
x_t
xt的条件下,
x
t
−
1
x_{t-1}
xt−1的分布
q
σ
(
x
t
−
1
∣
x
t
,
x
0
)
=
N
(
x
t
−
1
∣
1
−
α
t
−
1
−
σ
t
2
1
−
α
t
⋅
x
t
+
[
α
t
−
1
−
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
]
⋅
x
0
,
σ
t
2
I
)
q_{\sigma}(x_{t-1}|x_t,x_0)=N\Bigg(x_{t-1}|\sqrt{\frac{1-\alpha_{t-1}-\sigma_t^2}{1-\alpha_{t}}}\cdot x_t+ \bigg[\sqrt{\alpha_{t-1}}- \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \bigg] \cdot x_0 ,\sigma_t^2 I\Bigg)
qσ(xt−1∣xt,x0)=N(xt−1∣1−αt1−αt−1−σt2⋅xt+[αt−1−1−αtαt⋅(1−αt−1−σt2)]⋅x0,σt2I),也就是
x
t
−
1
x_{t-1}
xt−1的计算过程如公式(1)所示。
x
t
−
1
=
α
t
−
1
⋅
x
0
+
1
−
α
t
−
1
−
σ
t
2
⋅
x
t
−
α
t
x
0
1
−
α
t
+
σ
t
2
ϵ
t
⏟
标准高斯分布
=
α
t
−
1
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
1
−
α
t
−
1
−
σ
t
2
⋅
1
1
−
α
t
⋅
(
x
t
−
α
t
⋅
(
x
t
−
1
−
α
t
⋅
z
t
α
t
)
)
+
σ
t
2
⋅
ϵ
t
=
α
t
−
1
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
1
−
α
t
−
1
−
σ
t
2
⋅
1
1
−
α
t
⋅
(
x
t
−
x
t
+
1
−
α
t
⋅
z
t
)
+
σ
t
2
⋅
ϵ
t
=
α
t
−
1
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
⏟
=
x
0
+
1
−
α
t
−
1
−
σ
t
2
⋅
z
t
+
σ
t
2
⋅
ϵ
t
\begin{equation} \begin{split} x_{t-1}&= \sqrt{\alpha_{t-1}}\cdot x_0+\sqrt{1-\alpha_{t-1}-\sigma_t^2}\cdot \frac{x_t-\sqrt{\alpha}_t x_0}{\sqrt{1-\alpha_t}} + \sigma_t^2 \underbrace{\epsilon_t}_{标准高斯分布} \\ &=\sqrt{\alpha_{t-1}}\cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\cdot \frac{1}{\sqrt{1-\alpha_t}}\cdot \bigg(x_t- \bcancel{\sqrt {\alpha_t}}\cdot \big(\frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\bcancel{\sqrt{\alpha_t}}} \big) \bigg) + \sigma_t^2 \cdot\epsilon_t\\ &=\sqrt{\alpha_{t-1}}\cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}}+ \sqrt{1-\alpha_{t-1}-\sigma_t^2}\cdot \frac{1}{\sqrt{1-\alpha_t}}\cdot (x_t - x_t + \sqrt{1-\alpha_t}\cdot z_t)+ \sigma_t^2 \cdot\epsilon_t\\ &=\sqrt{\alpha_{t-1}}\cdot \underbrace{ \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}}}_{=x_0}+ \sqrt{1-\alpha_{t-1}-\sigma_t^2}\cdot z_t + \sigma_t^2 \cdot\epsilon_t \end{split} \end{equation}
xt−1=αt−1⋅x0+1−αt−1−σt2⋅1−αtxt−αtx0+σt2标准高斯分布
ϵt=αt−1⋅αtxt−1−αt⋅zt+1−αt−1−σt2⋅1−αt1⋅(xt−αt
⋅(αt
xt−1−αt⋅zt))+σt2⋅ϵt=αt−1⋅αtxt−1−αt⋅zt+1−αt−1−σt2⋅1−αt1⋅(xt−xt+1−αt⋅zt)+σt2⋅ϵt=αt−1⋅=x0
αtxt−1−αt⋅zt+1−αt−1−σt2⋅zt+σt2⋅ϵt
得到的公式(1)就是在推断时跳
1
1
1步的采样过程。
由前向加噪过程,可以推知
q
σ
(
x
t
−
2
∣
x
t
−
1
,
x
0
)
=
N
(
x
t
−
2
∣
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
x
t
−
1
+
[
α
t
−
2
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
]
⋅
x
0
,
σ
t
−
1
2
I
)
q_{\sigma}(x_{t-2}|x_{t-1},x_0)=N\bigg(x_{t-2}|\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot x_{t-1}+ \bigg[\sqrt{\alpha_{t-2}}- \frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \bigg] \cdot x_0 ,\sigma_{t-1}^2 I\bigg)
qσ(xt−2∣xt−1,x0)=N(xt−2∣1−αt−11−αt−2−σt−12⋅xt−1+[αt−2−1−αt−1αt−1⋅(1−αt−2−σt−12)]⋅x0,σt−12I)。接下来考虑,跳
2
2
2步时的采样过程,即在给定
x
0
x_0
x0和
x
t
x_t
xt时,
x
t
−
2
x_{t-2}
xt−2时的采样过程,即
q
σ
(
x
t
−
2
∣
x
0
,
x
t
)
q_\sigma(x_{t-2}|x_0,x_t)
qσ(xt−2∣x0,xt)的分布。
首先,我们可以确定
q
σ
(
x
t
−
2
∣
x
0
,
x
t
)
q_\sigma(x_{t-2}|x_0,x_t)
qσ(xt−2∣x0,xt)是高斯分布,假设其均值和方差分别为
μ
t
−
2
\mu_{t-2}
μt−2和
σ
t
−
2
2
\sigma_{t-2}^2
σt−22。由于
q
σ
(
x
t
−
2
∣
x
0
,
x
t
)
q_\sigma(x_{t-2}|x_0,x_t)
qσ(xt−2∣x0,xt)是
q
σ
(
x
t
−
2
,
x
t
−
1
∣
x
0
,
x
t
)
q_\sigma(x_{t-2},x_{t-1}|x_0,x_t)
qσ(xt−2,xt−1∣x0,xt) 的边缘分布。
q σ ( x t − 2 ∣ x 0 , x t ) = ∫ q σ ( x t − 2 , x t − 1 ∣ x 0 , x t ) ⋅ d x t − 1 = ∫ q σ ( x t − 2 ∣ x 0 , x t − 1 ) ⋅ q σ ( x t − 1 ∣ x 0 , x t ) ⋅ d x t − 1 \begin{equation} \begin{split} q_\sigma(x_{t-2}|x_0,x_t)&= \int q_\sigma(x_{t-2},x_{t-1}|x_0,x_t) \cdot dx_{t-1} \\ &=\int q_\sigma(x_{t-2}|x_0,x_{t-1}) \cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} \end{split} \end{equation} qσ(xt−2∣x0,xt)=∫qσ(xt−2,xt−1∣x0,xt)⋅dxt−1=∫qσ(xt−2∣x0,xt−1)⋅qσ(xt−1∣x0,xt)⋅dxt−1
因此
μ
t
−
2
=
E
(
q
σ
(
x
t
−
2
∣
x
0
,
x
t
)
)
=
∫
x
t
−
2
⋅
q
σ
(
x
t
−
2
∣
x
0
,
x
t
)
⋅
d
x
t
−
2
=
∫
x
t
−
2
⋅
(
∫
q
σ
(
x
t
−
2
,
∣
x
0
,
x
t
−
1
)
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
)
⋅
d
x
t
−
2
=
∫
∫
x
t
−
2
⋅
q
σ
(
x
t
−
2
,
∣
x
0
,
x
t
−
1
)
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
⋅
d
x
t
−
2
=
∫
(
∫
x
t
−
2
⋅
q
σ
(
x
t
−
2
∣
x
0
,
x
t
−
1
)
⋅
d
x
t
−
2
)
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
=
∫
(
E
(
q
σ
(
x
t
−
2
,
∣
x
0
,
x
t
−
1
)
)
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
=
∫
(
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
x
t
−
1
+
[
α
t
−
2
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
]
⋅
x
0
)
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
=
∫
(
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
x
t
−
1
)
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
+
∫
(
α
t
−
2
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
)
⋅
x
0
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
∫
x
t
−
1
⋅
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
+
(
α
t
−
2
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
)
⋅
x
0
∫
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
⋅
d
x
t
−
1
⏟
=
1
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
E
(
q
σ
(
x
t
−
1
∣
x
0
,
x
t
)
)
+
(
α
t
−
2
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
)
⋅
x
0
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
(
1
−
α
t
−
1
−
σ
t
2
1
−
α
t
⋅
x
t
+
[
α
t
−
1
−
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
]
⋅
x
0
)
+
(
α
t
−
2
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
)
⋅
x
0
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
1
−
α
t
−
1
−
σ
t
2
1
−
α
t
⋅
x
t
+
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
−
1
⋅
x
0
−
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
⋅
x
0
+
α
t
−
2
⋅
x
0
−
α
t
−
1
⋅
(
1
−
α
t
−
2
−
σ
t
−
1
2
)
1
−
α
t
−
1
⋅
x
0
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
1
−
α
t
−
1
−
σ
t
2
1
−
α
t
⋅
x
t
−
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
⋅
x
0
+
α
t
−
2
⋅
x
0
⏟
x
0
=
x
t
−
1
−
α
t
⋅
z
t
α
t
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
1
−
α
t
−
1
−
σ
t
2
1
−
α
t
⋅
x
t
−
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
α
t
−
2
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
1
−
α
t
−
1
−
σ
t
2
1
−
α
t
⋅
x
t
−
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
⋅
x
t
α
t
+
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
⋅
1
−
α
t
⋅
z
t
α
t
+
α
t
−
2
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
=
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
α
t
⋅
⋅
(
1
−
α
t
−
1
−
σ
t
2
)
1
−
α
t
⋅
1
−
α
t
⋅
z
t
α
t
+
α
t
−
2
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
=
α
t
−
2
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
⏟
=
x
0
+
1
−
α
t
−
2
−
σ
t
−
1
2
1
−
α
t
−
1
⋅
1
−
α
t
−
1
−
σ
t
2
⋅
z
t
=
=
=
=
=
=
令所有的
σ
=
0
α
t
−
2
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
1
−
α
t
−
2
⋅
z
t
\begin{equation} \begin{split} \mu_{t-2}&=E\big(q_\sigma(x_{t-2}|x_0,x_t)\big) \\ &=\int x_{t-2} \cdot q_\sigma(x_{t-2}|x_0,x_t)\cdot dx_{t-2} \\ &=\int x_{t-2} \cdot \bigg(\int q_\sigma(x_{t-2},|x_0,x_{t-1}) \cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} \bigg) \cdot dx_{t-2} \\ &=\int \int x_{t-2} \cdot q_\sigma(x_{t-2},|x_0,x_{t-1}) \cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} \cdot dx_{t-2} \\ &=\int \bigg( \int x_{t-2} \cdot q_\sigma(x_{t-2}|x_0,x_{t-1}) \cdot dx_{t-2} \bigg)\cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} \\ &=\int \bigg(E(q_\sigma(x_{t-2},|x_0,x_{t-1}) \bigg)\cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} \\ &=\int \bigg(\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot x_{t-1}+ \bigg[\sqrt{\alpha_{t-2}}- \frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \bigg] \cdot x_0 \bigg)\cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} \\ &=\int \bigg(\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot x_{t-1} \bigg)\cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} + \int \bigg(\sqrt{\alpha_{t-2}}- \frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \bigg) \cdot x_0 \cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1}\\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \int x_{t-1}\cdot q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1} +\bigg(\sqrt{\alpha_{t-2}}- \frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \bigg) \cdot x_0 \underbrace{ \int q_\sigma(x_{t-1}|x_0,x_{t}) \cdot dx_{t-1}}_{=1}\\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot E\bigg(q_\sigma(x_{t-1}|x_0,x_{t})\bigg) +\bigg(\sqrt{\alpha_{t-2}}- \frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \bigg) \cdot x_0 \\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \bigg(\sqrt{\frac{1-\alpha_{t-1}-\sigma_t^2}{1-\alpha_{t}}}\cdot x_t+ \bigg[\sqrt{\alpha_{t-1}}- \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \bigg] \cdot x_0 \bigg) +\bigg(\sqrt{\alpha_{t-2}}- \frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \bigg) \cdot x_0 \\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \sqrt{\frac{1-\alpha_{t-1}-\sigma_t^2}{1-\alpha_{t}}}\cdot x_t + \bcancel{\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}} \cdot \sqrt{\alpha_{t-1}}\cdot x_0} -\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \cdot x_0 + \sqrt{\alpha_{t-2}} \cdot x_0 - \bcancel{\frac{\sqrt{ \alpha_{t-1}\cdot (1-\alpha_{t-2}-\sigma_{t-1}^2} )}{\sqrt{1-\alpha_{t-1}}} \cdot x_0}\\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \sqrt{\frac{1-\alpha_{t-1}-\sigma_t^2}{1-\alpha_{t}}}\cdot x_t -\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \cdot x_0 + \sqrt{\alpha_{t-2}} \cdot \underbrace{x_0}_{x_0=\frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}}} \\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \sqrt{\frac{1-\alpha_{t-1}-\sigma_t^2}{1-\alpha_{t}}}\cdot x_t -\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{\alpha_{t-2}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} \\ &=\bcancel{\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \sqrt{\frac{1-\alpha_{t-1}-\sigma_t^2}{1-\alpha_{t}}}\cdot x_t} -\bcancel{\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \cdot \frac{x_t}{\sqrt{\alpha_t}}} +\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \frac{\sqrt{ \alpha_t\cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\sqrt{1-\alpha_t}} \cdot \frac{{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{\alpha_{t-2}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} \\ &=\sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \frac{\bcancel{\sqrt{\alpha_t}}\cdot \sqrt{ \cdot (1-\alpha_{t-1}-\sigma_t^2} )}{\bcancel{\sqrt{1-\alpha_t}}} \cdot \frac{{\bcancel{\sqrt{1-\alpha_t}}\cdot z_t}}{\bcancel{\sqrt{\alpha_t}}} + \sqrt{\alpha_{t-2}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} \\ &=\sqrt{\alpha_{t-2}} \cdot \underbrace{ \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}}}_{=x_0}+ \sqrt{\frac{1-\alpha_{t-2}-\sigma_{t-1}^2}{1-\alpha_{t-1}}}\cdot \sqrt{ 1-\alpha_{t-1}-\sigma_t^2} \cdot z_t \\ &\stackrel{\mathrm{令所有的\sigma=0}}{======}\sqrt{\alpha_{t-2}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{1-\alpha_{t-2}}\cdot z_t \end{split} \end{equation}
μt−2=E(qσ(xt−2∣x0,xt))=∫xt−2⋅qσ(xt−2∣x0,xt)⋅dxt−2=∫xt−2⋅(∫qσ(xt−2,∣x0,xt−1)⋅qσ(xt−1∣x0,xt)⋅dxt−1)⋅dxt−2=∫∫xt−2⋅qσ(xt−2,∣x0,xt−1)⋅qσ(xt−1∣x0,xt)⋅dxt−1⋅dxt−2=∫(∫xt−2⋅qσ(xt−2∣x0,xt−1)⋅dxt−2)⋅qσ(xt−1∣x0,xt)⋅dxt−1=∫(E(qσ(xt−2,∣x0,xt−1))⋅qσ(xt−1∣x0,xt)⋅dxt−1=∫(1−αt−11−αt−2−σt−12⋅xt−1+[αt−2−1−αt−1αt−1⋅(1−αt−2−σt−12)]⋅x0)⋅qσ(xt−1∣x0,xt)⋅dxt−1=∫(1−αt−11−αt−2−σt−12⋅xt−1)⋅qσ(xt−1∣x0,xt)⋅dxt−1+∫(αt−2−1−αt−1αt−1⋅(1−αt−2−σt−12))⋅x0⋅qσ(xt−1∣x0,xt)⋅dxt−1=1−αt−11−αt−2−σt−12⋅∫xt−1⋅qσ(xt−1∣x0,xt)⋅dxt−1+(αt−2−1−αt−1αt−1⋅(1−αt−2−σt−12))⋅x0=1
∫qσ(xt−1∣x0,xt)⋅dxt−1=1−αt−11−αt−2−σt−12⋅E(qσ(xt−1∣x0,xt))+(αt−2−1−αt−1αt−1⋅(1−αt−2−σt−12))⋅x0=1−αt−11−αt−2−σt−12⋅(1−αt1−αt−1−σt2⋅xt+[αt−1−1−αtαt⋅(1−αt−1−σt2)]⋅x0)+(αt−2−1−αt−1αt−1⋅(1−αt−2−σt−12))⋅x0=1−αt−11−αt−2−σt−12⋅1−αt1−αt−1−σt2⋅xt+1−αt−11−αt−2−σt−12⋅αt−1⋅x0
−1−αt−11−αt−2−σt−12⋅1−αtαt⋅(1−αt−1−σt2)⋅x0+αt−2⋅x0−1−αt−1αt−1⋅(1−αt−2−σt−12)⋅x0
=1−αt−11−αt−2−σt−12⋅1−αt1−αt−1−σt2⋅xt−1−αt−11−αt−2−σt−12⋅1−αtαt⋅(1−αt−1−σt2)⋅x0+αt−2⋅x0=αtxt−1−αt⋅zt
x0=1−αt−11−αt−2−σt−12⋅1−αt1−αt−1−σt2⋅xt−1−αt−11−αt−2−σt−12⋅1−αtαt⋅(1−αt−1−σt2)⋅αtxt−1−αt⋅zt+αt−2⋅αtxt−1−αt⋅zt=1−αt−11−αt−2−σt−12⋅1−αt1−αt−1−σt2⋅xt
−1−αt−11−αt−2−σt−12⋅1−αtαt⋅(1−αt−1−σt2)⋅αtxt
+1−αt−11−αt−2−σt−12⋅1−αtαt⋅(1−αt−1−σt2)⋅αt1−αt⋅zt+αt−2⋅αtxt−1−αt⋅zt=1−αt−11−αt−2−σt−12⋅1−αt
αt
⋅⋅(1−αt−1−σt2)⋅αt
1−αt
⋅zt+αt−2⋅αtxt−1−αt⋅zt=αt−2⋅=x0
αtxt−1−αt⋅zt+1−αt−11−αt−2−σt−12⋅1−αt−1−σt2⋅zt======令所有的σ=0αt−2⋅αtxt−1−αt⋅zt+1−αt−2⋅zt
论文中跳步过程如公式(4)所示,结果貌似与论文中公式略有不同。但是,公式中的
σ
t
\sigma_t
σt是自定义的,因此如果我们令
σ
t
=
0
\sigma_t=0
σt=0,也就是
x
t
x_t
xt的取值都是确定的,不存在方差了,那么就可以实现跳
2
2
2步了。这个只是均值,至于方差,既然已经令
σ
t
=
0
\sigma_t=0
σt=0了,那分布
q
σ
(
x
t
−
2
∣
x
0
,
x
t
)
q_\sigma(x_{t-2}|x_0,x_t)
qσ(xt−2∣x0,xt)的方差
σ
t
−
2
2
\sigma_{t-2}^2
σt−22肯定等于
0
0
0了。跳
1
1
1步和跳
2
2
2步的分布如公式(4)所示。
跳
1
步时:
x
t
−
1
=
α
t
−
1
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
1
−
α
t
−
1
⋅
z
t
跳
2
步时:
x
t
−
2
=
α
t
−
2
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
1
−
α
t
−
2
⋅
z
t
\begin{equation} \begin{split} 跳1步时:x_{t-1}&=\sqrt{\alpha_{t-1}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{1-\alpha_{t-1}}\cdot z_t\\ 跳2步时:x_{t-2}&=\sqrt{\alpha_{t-2}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{1-\alpha_{t-2}}\cdot z_t \end{split} \end{equation}
跳1步时:xt−1跳2步时:xt−2=αt−1⋅αtxt−1−αt⋅zt+1−αt−1⋅zt=αt−2⋅αtxt−1−αt⋅zt+1−αt−2⋅zt
结合跳
1
1
1步和
2
2
2步,我们可以推知:跳
n
n
n步的公式就是
q
σ
(
x
t
−
n
∣
x
t
,
x
0
)
=
α
t
−
n
⋅
x
t
−
1
−
α
t
⋅
z
t
α
t
+
1
−
α
t
−
n
⋅
z
t
q_\sigma(x_{t-n}|x_t,x_0)=\sqrt{\alpha_{t-n}} \cdot \frac{x_t-{\sqrt{1-\alpha_t}\cdot z_t}}{\sqrt{\alpha_t}} + \sqrt{1-\alpha_{t-n}}\cdot z_t
qσ(xt−n∣xt,x0)=αt−n⋅αtxt−1−αt⋅zt+1−αt−n⋅zt,接下来用数学归纳法证明,转下文论文阅读笔记:Denoising Diffusion Implicit Models (4)