变分扩散模型

1. VDM

1.1 VDM简介

VDM (Variational Diffusion Models) 基于 [[MHVAE]] 模型，但与 [[MHVAE]] 模型有3个不同：

• 对于所有时间步 $t$：隐变量 ${\mathbf{z}}_t$ 的维度和数据 $\mathbf{x}$ 的维度相等，即 ${\mathbf{z}}_t\in {\mathbb{R}}^d$, $x\in {\mathbb{R}}^d$；

• 对于所有时间步 $t$：隐变量 ${\mathbf{z}}_t$ 不是通过神经网络模型学习得到的，而是以前一个时间步 ${\mathbf{z}}_{t-1}$ 为均值的高斯分布。所以，Diffusion Models 不需要通过神经网络模型学习一个 Encoder；

• 随着时间步 $t$ 的增大，隐变量 ${\mathbf{z}}_t$ 逐渐逼近标准正态分布，最后在第 $T$ 步时 ${\mathbf{z}}_T\sim \mathcal{N}(0,\mathbf{I})$（T足够大）。

对照 [[MHVAE]] 的联合概率公式 (1) ，可得 VDM 的联合概率公式 (9)：

$$\begin{align}

\underbrace{p\left (x_{0:T}\right )}{\text{Joint Distribution}} = \underbrace{p\left (x_T\right )}{\text{Prior}} \prod_{t=1}^{T} \underbrace{p_{\theta}\left (x_{t-1}\mid x_t\right )}_{\text{Decoder}}

\end{align}$$

对照[[MHVAE]]的后验公式 (2)，可得VDM的后验公式 (10)：

$$\begin{align}

\underbrace{q\left (\boldsymbol{x}{1:T}\mid \boldsymbol{x}0\right )}{\text{Posterior Distribution}} = \prod{t=1}^{T} \underbrace{q\left (\boldsymbol{x}t\mid \boldsymbol{x}{t-1}\right )}_{\text{Encoder}}

\end{align}$$

注意[[MHVAE]]的公式和VDM的公式有以下区别：

• $q_{\varphi }$ 全部修改为 $q$ ，因为VDM模型的Encoder不需要用神经网络建模；

• ${\mathbf{z}}_t$ 全部修改为 ${\mathbf{x}}_t$，因为在VDM中 ${\mathbf{z}}_t$ 的维度和 ${\mathbf{x}}_t$ 的维度相等。

• $\mathbf{x}$ 全部修改为 ${\mathbf{x}}_0$

1.2 如何推导VDM的ELBo？

根据[[MHVAE]]的ELBo公式 (8) 将 $q_{\varphi }$ 改成 $q$，$z_t$ 改成 $x_t$ 即可得到VDM的ELBo：

$$
\begin{align}

& \log \underbrace{p(\boldsymbol{x})}_{\text{Evidence}} \

\geq & \underbrace{\mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p\left(\boldsymbol{x}{0: T}\right)}{q\left(\boldsymbol{x}_{1: T} \mid \boldsymbol{x}0\right)}\right]}{\text{ELBo of VDM}} \

=& \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) \prod{t=1}^T p{\boldsymbol{\theta}}\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}t\right)}{\prod{t=1}^T q\left(\boldsymbol{x}t \mid \boldsymbol{x}{t-1}\right)}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}0 \mid \boldsymbol{x}1\right) \prod{t=2}^T p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t\right)}{q\left(\boldsymbol{x}_1 \mid \boldsymbol{x}0\right) \prod{t=2}^T q\left(\boldsymbol{x}t \mid \boldsymbol{x}{t-1}\right)}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}0 \mid \boldsymbol{x}1\right) \prod{t=2}^T p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t\right)}{q\left(\boldsymbol{x}_1 \mid \boldsymbol{x}0\right) \prod{t=2}^T q\left(\boldsymbol{x}t \mid \boldsymbol{x}{t-1}, \boldsymbol{x}_0\right)}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}_0 \mid \boldsymbol{x}1\right)}{q\left(\boldsymbol{x}1 \mid \boldsymbol{x}0\right)}+\log \prod{t=2}^T \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t\right)}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}{t-1}, \boldsymbol{x}_0\right)}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}_0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}_0 \mid \boldsymbol{x}1\right)}{q\left(\boldsymbol{x}1 \mid \boldsymbol{x}0\right)}+\log \prod{t=2}^T \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{\frac{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0\right) q\left(\boldsymbol{x}_t \mid \boldsymbol{x}0\right)}{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_0\right)}}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}_0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}_0 \mid \boldsymbol{x}1\right)}{q\left(\boldsymbol{x}1 \mid \boldsymbol{x}0\right)}+\log \prod{t=2}^T \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{\frac{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0\right) q\left(\boldsymbol{x}_t \mid \boldsymbol{x}0\right)}{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_0\right)}}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}_0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}_0 \mid \boldsymbol{x}_1\right)}{q\left(\boldsymbol{x}_1 \mid \boldsymbol{x}_0\right)}+\log \frac{q\left(\boldsymbol{x}_1 \mid \boldsymbol{x}0\right)}{q\left(\boldsymbol{x}T \mid \boldsymbol{x}0\right)}+\log \prod{t=2}^T \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0\right)}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}_0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right) p{\boldsymbol{\theta}}\left(\boldsymbol{x}_0 \mid \boldsymbol{x}1\right)}{q\left(\boldsymbol{x}T \mid \boldsymbol{x}0\right)}+\sum{t=2}^T \log \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0\right)}\right] \

= & \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log p{\boldsymbol{\theta}}\left(\boldsymbol{x}0 \mid \boldsymbol{x}1\right)\right]+\mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right)}{q\left(\boldsymbol{x}T \mid \boldsymbol{x}0\right)}\right]+\sum{t=2}^T \mathbb{E}{q\left(\boldsymbol{x}{1: T} \mid \boldsymbol{x}0\right)}\left[\log \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0\right)}\right] \

= & \mathbb{E}_{q\left(\boldsymbol{x}_1 \mid \boldsymbol{x}0\right)}\left[\log p{\boldsymbol{\theta}}\left(\boldsymbol{x}_0 \mid \boldsymbol{x}1\right)\right]+\mathbb{E}{q\left(\boldsymbol{x}_T \mid \boldsymbol{x}_0\right)}\left[\log \frac{p\left(\boldsymbol{x}T\right)}{q\left(\boldsymbol{x}T \mid \boldsymbol{x}0\right)}\right]+\sum{t=2}^T \mathbb{E}{q\left(\boldsymbol{x}t, \boldsymbol{x}{t-1} \mid \boldsymbol{x}0\right)}\left[\log \frac{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0\right)}\right] \

= & \underbrace{\underbrace{\mathbb{E}{q\left(\boldsymbol{x}1 \mid \boldsymbol{x}0\right)}\left[\log p{\boldsymbol{\theta}}\left(\boldsymbol{x}0 \mid \boldsymbol{x}1\right)\right]}{x_0\approx{x1}}}{\text{reconstruction term} \color{red}{\approx 0}}-\underbrace{D{\mathrm{KL}}\left(\underbrace{q\left(\boldsymbol{x}T \mid \boldsymbol{x}0\right)}{\approx N(0, I)} \parallel \underbrace{p\left(\boldsymbol{x}T\right)}{=N(0, I)}\right)}{\text{prior matching term}\color{red}{\approx 0}} -\underbrace{\sum{t=2}^T \underbrace{\mathbb{E}_{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0\right)}\left[

D_{\mathrm{KL}}\left(\underbrace{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t, \boldsymbol{x}0\right)}{\color{red}{\text {complexity posterior}}} \parallel \underbrace{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{\textcolor{green}{\text{Decoder of VDM}}}\right)

\right]}{\text {denoising matching term }}}{\textbf{Objective function to optimize}}

\end{align}
$$

1.3 如何从ELBo推导VDM的目标函数？

1.3.1 重建项 (reconstruction term)

结论：可以通过蒙特卡罗估计计算，但是真实情况是当 $T$ 比较大时 ${\mathbf{x}}_0\approx {\mathbf{x}}_1$，可忽略不计。

$$\begin{align}

\underbrace{\underbrace{\mathbb{E}_{q\left(\boldsymbol{x}_1 \mid \boldsymbol{x}0\right)}\left[\log p{\boldsymbol{\theta}}\left(\boldsymbol{x}0 \mid \boldsymbol{x}1\right)\right]}{x_0\approx{x1}}}{\text{reconstruction term}} \approx 0

\end{align}$$

1.3.2 先验匹配项 (prior matching term)

在 VDM 中，我们有如下假设：

$$\begin{align}

\underbrace{q\left(\boldsymbol{x}t \mid \boldsymbol{x}{t-1} \right)}{\textcolor{red}{\text{Encoder of VDM}}} & =\mathcal{N}\left(\boldsymbol{x}t ; \sqrt{\alpha_t} \boldsymbol{x}{t-1},\left(1-\alpha_t\right) \mathbf{I}\right) \quad \text{ with } \boldsymbol{\alpha{t} \in (0, 1)}

\end{align}$$

利用重参数化技巧，可得：

$$\begin{align}

\boldsymbol{x}t & =\sqrt{\alpha_t} \boldsymbol{x}{t-1}+\sqrt{1-\alpha_t} \boldsymbol{\epsilon} \quad \text { with } \boldsymbol{\epsilon} \sim \mathcal{N}(\boldsymbol{\epsilon} ; \mathbf{0}, \mathbf{I}) \text{, }\boldsymbol{\alpha_{t} \in (0, 1)}

\end{align}$$
基于重参数化技巧继续推导，可得：

$$

\begin{align}

\boldsymbol{x}t & =\sqrt{\alpha_t} \boldsymbol{x}{t-1}+\sqrt{1-\alpha_t} \boldsymbol{\epsilon}_{t-1}^* \

& =\sqrt{\alpha_t}\left(\sqrt{\alpha_{t-1}} \boldsymbol{x}{t-2}+\sqrt{1-\alpha{t-1}} \epsilon_{t-2}^\right)+\sqrt{1-\alpha_t} \boldsymbol{\epsilon}_{t-1}^ \

& =\sqrt{\alpha_t \alpha_{t-1}} \boldsymbol{x}{t-2}+\sqrt{\alpha_t-\alpha_t \alpha{t-1}} \boldsymbol{\epsilon}{t-2}^*+\sqrt{1-\alpha_t} \boldsymbol{\epsilon}{t-1}^* \

& =\sqrt{\alpha_t \alpha_{t-1}} \boldsymbol{x}{t-2}+\sqrt{{\sqrt{\alpha_t-\alpha_t \alpha{t-1}}}^2+\sqrt{1-\alpha_t}} \boldsymbol{\epsilon}{t-2} \quad \text{(apply } \boldsymbol{\lbrace \epsilon_t^*,\epsilon_t \rbrace{t=0}^{T}\overset{iid}{\sim}\mathcal{N}\boldsymbol{(\epsilon; \mathbf{0}, \mathbf{I})}}\text{)}\

& =\sqrt{\alpha_t \alpha_{t-1}} \boldsymbol{x}{t-2}+\sqrt{\alpha_t-\alpha_t \alpha{t-1}+1-\alpha_t} \boldsymbol{\epsilon}_{t-2} \

& =\sqrt{\alpha_t \alpha_{t-1}} \boldsymbol{x}{t-2}+\sqrt{1-\alpha_t \alpha{t-1}} \boldsymbol{\epsilon}_{t-2} \

& =\ldots \

& =\sqrt{\prod_{i=1}^t \alpha_i} \boldsymbol{x}0+\sqrt{1-\prod{i=1}^t \alpha_i \boldsymbol{\epsilon}0} \quad \text{ with } \boldsymbol{\epsilon}{0} \sim \mathcal{N}(\boldsymbol{\epsilon}; \mathbf{0}, \mathbf{I}) \

& =\sqrt{\bar{\alpha}_t} \boldsymbol{x}_0+\sqrt{1-\bar{\alpha}_t} \boldsymbol{\epsilon}0 \quad \text{ with } \boldsymbol{\epsilon}{0} \sim \mathcal{N}(\boldsymbol{\epsilon}; \mathbf{0}, \mathbf{I}) \

& \sim \mathcal{N}\left(\boldsymbol{x}_t ; \sqrt{\bar{\alpha}_t} \boldsymbol{x}_0,\left(1-\bar{\alpha}_t\right) \mathbf{I}\right)

\end{align}

$$

可得：

$$

q(\boldsymbol{x}_t) = \mathcal{N}\left(\boldsymbol{x}_t ; \sqrt{\bar{\alpha}_t} \boldsymbol{x}_0,\left(1-\bar{\alpha}_t\right) \mathbf{I}\right)

$$

由马尔可夫性：

$$\begin{align}

q(\boldsymbol{x}_t \mid \boldsymbol{x}_0) = q(\boldsymbol{x}_t) = \mathcal{N}\left(\boldsymbol{x}_t ; \sqrt{\bar{\alpha}_t} \boldsymbol{x}_0,\left(1-\bar{\alpha}_t\right) \mathbf{I}\right)

\end{align}$$
当 $T$ 足够大，比如 $T=1000$ ：

$$

\alpha_t \in (0, 1) \implies \bar{\alpha}_T \approx 0

$$

所以：

$$

\boldsymbol{q}(\boldsymbol{x}_T \mid \boldsymbol{x}_0) \approx \mathcal{N}\left(\mathbf{0}, \mathbf{I}\right)

$$

结论：先验匹配项可以忽略不计：

$$

\underbrace{D_{\mathrm{KL}}\left(\underbrace{q\left(\boldsymbol{x}T \mid \boldsymbol{x}0\right)}{\approx N(0, I)} \parallel \underbrace{p\left(\boldsymbol{x}T\right)}{=N(0, I)}\right)}{\text{prior matching term}} \approx 0

$$

1.3.3 降噪匹配项 (denoising matching term)

下面对降噪匹配项中的子式分别进行推导：

$$

\begin{align}

& \underbrace{q\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}0\right) }{\text{complexity posterior}} \

= & \frac{q\left(\boldsymbol{x}t \mid \boldsymbol{x}{t-1}, \boldsymbol{x}0\right) q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}_0\right)}{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0\right)} \

= & \frac{\overbrace{\mathcal{N}\left(\boldsymbol{x}t ; \sqrt{\alpha_t} \boldsymbol{x}{t-1},\left(1-\alpha_t\right) \mathbf{I}\right)}^{\text{apply Markov Property in Eq.(25)}} \overbrace{\mathcal{N}\left(\boldsymbol{x}{t-1} ; \sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0,\left(1-\bar{\alpha}{t-1}\right) \mathbf{I}\right)}^{\text{apply Eq.(37)}}}{\underbrace{\mathcal{N}\left(\boldsymbol{x}_t ; \sqrt{\bar{\alpha}_t} \boldsymbol{x}_0,\left(1-\bar{\alpha}t\right) \mathbf{I}\right)}{\text{apply Eq.(37)}}} \

\propto & \exp \left{-\left[\frac{\left(\boldsymbol{x}t-\sqrt{\alpha_t} \boldsymbol{x}{t-1}\right)^2}{2\left(1-\alpha_t\right)}+\frac{\left(\boldsymbol{x}{t-1}-\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0\right)^2}{2\left(1-\bar{\alpha}{t-1}\right)}-\frac{\left(\boldsymbol{x}_t-\sqrt{\bar{\alpha}_t} \boldsymbol{x}_0\right)^2}{2\left(1-\bar{\alpha}_t\right)}\right]\right} \

= & \exp \left{-\frac{1}{2}\left[\frac{\left(\boldsymbol{x}t-\sqrt{\alpha_t} \boldsymbol{x}{t-1}\right)^2}{1-\alpha_t}+\frac{\left(\boldsymbol{x}{t-1}-\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0\right)^2}{1-\bar{\alpha}{t-1}}-\frac{\left(\boldsymbol{x}_t-\sqrt{\bar{\alpha}_t} \boldsymbol{x}_0\right)^2}{1-\bar{\alpha}_t}\right]\right} \

= & \exp \left{-\frac{1}{2}\left[\frac{\left(-2 \sqrt{\alpha_t} \boldsymbol{x}t \boldsymbol{x}{t-1}+\alpha_t \boldsymbol{x}{t-1}^{2\right)}{1-\alpha_t}+\frac{\left(\boldsymbol{x}_{t-1}}2-2 \sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}_{t-1} \boldsymbol{x}0\right)}{1-\bar{\alpha}{t-1}}+C\left(\boldsymbol{x}_t, \boldsymbol{x}_0\right)\right]\right} \

\propto & \exp \left{-\frac{1}{2}\left[-\frac{2 \sqrt{\alpha_t} \boldsymbol{x}t \boldsymbol{x}{t-1}}{1-\alpha_t}+\frac{\alpha_t \boldsymbol{x}{t-1}^{2}{1-\alpha_t}+\frac{\boldsymbol{x}_{t-1}}2}{1-\bar{\alpha}{t-1}}-\frac{2 \sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}{t-1} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right]\right} \

= & \exp \left{-\frac{1}{2}\left[\left(\frac{\alpha_t}{1-\alpha_t}+\frac{1}{1-\bar{\alpha}{t-1}}\right) \boldsymbol{x}{t-1}^2-2\left(\frac{\sqrt{\alpha_t} \boldsymbol{x}t}{1-\alpha_t}+\frac{\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right) \boldsymbol{x}_{t-1}\right]\right} \

= & \exp \left{-\frac{1}{2}\left[\frac{\alpha_t\left(1-\bar{\alpha}{t-1}\right)+1-\alpha_t}{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)} \boldsymbol{x}{t-1}^2-2\left(\frac{\sqrt{\alpha_t} \boldsymbol{x}t}{1-\alpha_t}+\frac{\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right) \boldsymbol{x}{t-1}\right]\right} \

= & \exp \left{-\frac{1}{2}\left[\frac{\alpha_t-\bar{\alpha}t+1-\alpha_t}{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)} \boldsymbol{x}{t-1}^2-2\left(\frac{\sqrt{\alpha_t} \boldsymbol{x}t}{1-\alpha_t}+\frac{\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right) \boldsymbol{x}{t-1}\right]\right} \

= & \exp \left{-\frac{1}{2}\left[\frac{1-\bar{\alpha}t}{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)} \boldsymbol{x}{t-1}^2-2\left(\frac{\sqrt{\alpha_t} \boldsymbol{x}t}{1-\alpha_t}+\frac{\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right) \boldsymbol{x}{t-1}\right]\right} \

= & \exp \left{-\frac{1}{2}\left(\frac{1-\bar{\alpha}t}{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)}\right)\left[\boldsymbol{x}{t-1}^2-2 \frac{\left(\frac{\sqrt{\alpha_t} \boldsymbol{x}t}{1-\alpha_t}+\frac{\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right)}{\frac{1-\bar{\alpha}t}{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)}} \boldsymbol{x}{t-1}\right]\right} \

= & \exp \left{-\frac{1}{2}\left(\frac{1-\bar{\alpha}t}{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)}\right)\left[\boldsymbol{x}{t-1}^2-2 \frac{\left(\frac{\sqrt{\alpha_t} \boldsymbol{x}t}{1-\alpha_t}+\frac{\sqrt{\bar{\alpha}{t-1}} \boldsymbol{x}0}{1-\bar{\alpha}{t-1}}\right)\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)}{1-\bar{\alpha}t} \boldsymbol{x}{t-1}\right]\right} \

= & \exp \left{-\frac{1}{2}\left(\frac{1}{\frac{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)}{1-\bar{\alpha}t}}\right)\left[\boldsymbol{x}{t-1}^2-2 \frac{\sqrt{\alpha_t}\left(1-\bar{\alpha}{t-1}\right) \boldsymbol{x}t+\sqrt{\bar{\alpha}{t-1}}\left(1-\alpha_t\right) \boldsymbol{x}_0}{1-\bar{\alpha}t} \boldsymbol{x}{t-1}\right]\right} \

\propto & \mathcal{N}(\boldsymbol{x}{t-1} ; \underbrace{\frac{\sqrt{\alpha_t}\left(1-\bar{\alpha}{t-1}\right) \boldsymbol{x}t+\sqrt{\bar{\alpha}{t-1}}\left(1-\alpha_t\right) \boldsymbol{x}_0}{1-\bar{\alpha}t}}{\mu_q\left(\boldsymbol{x}_t, \boldsymbol{x}0\right)}, \underbrace{\left.\frac{\left(1-\alpha_t\right)\left(1-\bar{\alpha}{t-1}\right)}{1-\bar{\alpha}t} \mathbf{I}\right)}{\boldsymbol{\Sigma}_q(t)}

\end{align}

$$

参考式 (52) ，试图将 $p_{\theta }({\mathbf{x}}_{t-1}\mid {\mathbf{x}}_t)$ 也改写为正态分布的形式，令方差与式 (52) 相等：

$$\begin{align}

p_{\theta}(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t) = \mathcal{N}(\boldsymbol{x}{t-1}; \underbrace{\mu{\theta}(\boldsymbol{x}t, t)}{\text{learned by model}}, \Sigma_{q}(t))

\end{align}$$

参考式 (52) 中的 ${\mu }_q({\mathbf{x}}_t,{\mathbf{x}}_0)$，可得：

$$\begin{align}

\boldsymbol{\mu}_{\boldsymbol{\theta}}\left(\boldsymbol{x}_t, t\right)

=\frac{\sqrt{\alpha_t}\left(1-\bar{\alpha}{t-1}\right) \boldsymbol{x}t+\sqrt{\bar{\alpha}{t-1}}\left(1-\alpha_t\right) \overbrace{\hat{\boldsymbol{x}}{\boldsymbol{\theta}}\left(\boldsymbol{x}_t, t\right)}^{\text{learned by model}}}{1-\bar{\alpha}_t}

\end{align}$$

2.3.4 VDM的目标函数

公式 (23) 中的重建项和先验匹配项小到可以忽略不计，可得目标函数：

$$\begin{align}

&\operatorname{arg}\max \log{p(\boldsymbol{x})} \

\propto &\operatorname{arg}\min_{\theta} \underbrace{\sum_{t=2}^T \underbrace{\mathbb{E}_{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0\right)}\left[

D_{\mathrm{KL}}\left(\underbrace{q\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t, \boldsymbol{x}0\right)}{\color{red}{\text {complexity posterior}}} \parallel \underbrace{p{\boldsymbol{\theta}}\left(\boldsymbol{x}{t-1} \mid \boldsymbol{x}t\right)}{\textcolor{red}{\text{Decoder of VDM}}}\right)

\right]}{\text {denoising matching term }}}{\textbf{Objective function to optimize}}

\end{align}$$根据公式(56)，VDM的目标函数为：$$\begin{align}

&\operatorname{arg}\min \sum_{t=2}^{T}\mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)}\left[ D{\text{KL}}(q(\boldsymbol{x}{t-1}|\boldsymbol{x}{t}, \boldsymbol{x}{0})\parallel p{\theta}(\boldsymbol{x}{t-1}|\boldsymbol{x}{t}))\right] \

=& \operatorname{arg}\min\sum_{t=2}^{T}\mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)}\left[ D{\operatorname{KL}}(\mathcal{N}(\boldsymbol{x}{t-1}; \boldsymbol{\mu}{q}(\boldsymbol{x}t, \boldsymbol{x}0), \boldsymbol{\Sigma}{q}(t)) \parallel \mathcal{N}(\boldsymbol{x}{t-1}; \boldsymbol{\mu}_{\theta}(\boldsymbol{x}t,\boldsymbol{t}), \boldsymbol{\Sigma}{q}(t)))\right] \

=& \operatorname{arg}\min\sum_{t=2}^{T}\mathbb{E}_{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0\right)}\left[ \dfrac{1}{2}[\log\dfrac{|\boldsymbol{\Sigma}_q(t)|}{|\boldsymbol{\Sigma}_q(t)|}-d+\operatorname{tr}(\boldsymbol{\Sigma}_q(t)^{-1}\boldsymbol{\Sigma}q(t))+(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}0))^{T\boldsymbol{\Sigma}_q(t)}{-1}(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}_0))]\right] \quad \text{(apply )}\

=& \operatorname{arg}\min\sum_{t=2}^{T}\mathbb{E}_{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}0\right)}\left[ \dfrac{1}{2}\left[\log1-d+d+(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}0))^T\sum_q(t){-1}(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}_0))\right]\right] \

=& \operatorname{arg}\min_{\theta} \sum_{t=2}^{T}\mathbb{E}_{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}0\right)}\left[ \frac{1}{2}\left[\left(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}0)\right)^{{T}\sum_{q}(t)}{-1}\left(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}_0)\right)\right]\right] \

=& \operatorname{arg}\min\limits_{\theta} \sum_{t=2}^{T}\mathbb{E}_{q\left(\boldsymbol{x}_t \mid \boldsymbol{x}0\right)}\left[ \dfrac{1}{2}\left[\left(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t,\boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}0)\right)^{{T}\left(\sigma_{q}}{2}(t)I\right)^{-1}\left(\boldsymbol{\mu}{\theta}(\boldsymbol{x}t,\boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}_0)\right)\right]\right] \

=& \operatorname{arg}\min_{\theta}\sum_{t=2}^{T}\mathbb{E}_{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)}\left[ \frac{1}{2\sigma{q}^{2}(t)}\left[\left|\boldsymbol{\mu}{\theta}(\boldsymbol{x}t, \boldsymbol{t})-\boldsymbol{\mu}{q}(\boldsymbol{x}_t, \boldsymbol{x}0)\right|{2}^{2}\right] \right] \

=& \operatorname{arg}\min_{\theta} \sum_{t=2}^{T}\mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)}\left[ \frac{1}{2\sigma{q}^{2}(t)}\left[\left| \underbrace{\frac{\sqrt{\alpha_t}\left(1-\bar{\alpha}{t-1}\right) \boldsymbol{x}t+\sqrt{\bar{\alpha}{t-1}}\left(1-\alpha_t\right) \overbrace{\hat{\boldsymbol{x}}{\boldsymbol{\theta}}\left(\boldsymbol{x}t, t\right)}^{\text{learned by model}}}{1-\bar{\alpha}t}}{\mu{\theta} \text{ apply Eq.(55)}} - \underbrace{\frac{\sqrt{\alpha_t}\left(1-\bar{\alpha}{t-1}\right) \boldsymbol{x}t+\sqrt{\bar{\alpha}{t-1}}\left(1-\alpha_t\right) \boldsymbol{x}0}{1-\bar{\alpha}t}}{\mu{q} \text{ apply Eq.(53)}} \right|{2}^{2}\right] \right] \

=& \operatorname{arg}\min_{\theta} \sum_{t=2}^{T}\mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)} \left[ \frac{1}{2\sigma{q}^{2}(t)} \left[ \left| \frac{\sqrt{\bar{\alpha}{t-1}}(1-\alpha{t})\hat{x}{\theta}(x_t,t)}{1-\bar{\alpha}t} - \frac{\sqrt{\bar{\alpha}{t-1}}(1-\alpha{t})x_0}{1-\bar{\alpha}t} \right|{2}^{2} \right] \right] \

=& \operatorname{arg}\min_{\theta} \sum_{t=2}^{T}\mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)} \left[ \frac{1}{2\sigma{q}^{2}(t)} \left[ \left| \frac{\sqrt{\bar{\alpha}{t-1}}(1-\alpha{t})}{1-\bar{\alpha}t} (\hat{x}{\theta}(x_t,t) - x_0) \right|_{2}^{2} \right] \right] \

=& \underbrace{\operatorname{arg}\min_{\theta} \sum_{t=2}^{T}\mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)} \left[ \frac{1}{2\sigma{q}^{2}(t)} \frac{\bar{\alpha}{t-1}(1-\alpha{t})^{2}{(1-\bar{\alpha}_t)}2} \left[ \left| \hat{x}{\theta}(x_t,t) - x_0 \right|{2}^{2} \right] \right]}_{\text{Objective function of Diffusion Model}}

\end{align}$$
式 (67) 采用蒙特卡洛估计，可得：

$$\begin{align}

\operatorname{arg}\min_{\theta} \mathbb{E}{t\sim \mathbf{U}(2,T)} \left [ \mathbb{E}{q\left(\boldsymbol{x}t \mid \boldsymbol{x}0\right)} \left[ \frac{1}{2\sigma{q}^{2}(t)} \frac{\bar{\alpha}{t-1}(1-\alpha_{t})^{2}{(1-\bar{\alpha}_t)}2} \left[ \left| \hat{x}{\theta}(x_t,t) - x_0 \right|{2}^{2} \right] \right] \right]

\end{align}$$

综上所述，VDM模型学习的预测原图 ${\mathbf{x}}_0$，因为目标函数中的 $\Vert {\hat{\mathbf{x}}}_{\theta }({\mathbf{x}}_t,t)-{\mathbf{x}}_0{\Vert }_2^2$ 。