Generative image models based on ordinary differential equations can be seen as forms of variational auto-encoders with a partially deterministic inference network. \(\newcommand{\coloneqq}{\mathrel{\vcenter{:}}=}\)
Variational Auto-encoders
Given a dataset of samples \(y\) from an unknown distribution \(\pi_1\), generative models attempt to find parameters \(\theta\) for a fixed family that maximizes the likelihood \(p_\theta(y)\). When the family implicitly marginalizes over a latent variable \(z\) so that \(\log p_\theta(y) = \log E_{z \sim p} p_\theta(y \vert z)\), maximum likelihood estimation becomes intractable in general. Variational auto-encoders address this problem by maximizing a lower bound on the likelihood: \[ \begin{align*} \log p_\theta(y) &= \log E_{z \sim p} p_\theta(y \vert z) \\ &=\log E_{q(z \vert y)} \frac{p(y\vert z) p(z)}{q(z\vert y)} &\text{ (importance sampling)}\\ & \geq E_{q(z \vert y)} \log \frac{p(y\vert z) p(z)}{q(z)} &\text{ (Jensen's inequality)} \end{align*} \]
This quantity \(\mathcal{L} \coloneqq E_{q(z \vert y)} \log p(y,z) - \log q(z)\) is known as the evidence lower bound or ELBO. The distribution \(q(z\vert y)\) used for importance sampling is usually parameterized by a neural net and is known as the inference network. In what follows, we will implicitly assume that \(q\) is a function of neural net parameters \(\phi\).
From Variational Auto-encoders to Rectified Flows
Assume that the model’s prior over \(z\) factors as \(p(z) = p(z_0)\prod_t p(z_{t+\Delta} \vert z_t)\), where \(t\) ranges from \(0\) to \(1\) in increments of \(\Delta\) and \(z_1\) is observed (taking the place of \(y\) in the discussion so far). Instead of fixing the model \(p\) and learning parameters for a distribution \(q\) to approximate its posterior, with diffusion-based approaches, we do the opposite.. Fix the posterior approximation \(q(z_t \vert z_1, z_0)\) to be a Delta distribution at \(tz_1 + (1 - t)z_0\) (linearly interpolating between observations). Let \(p(z_{t+\Delta} \vert z_t)\) be distributed as \(\mathcal{N}(z_t + \Delta f_\theta(z_t, t), \sigma^2/\Delta^2)\), where \(f_\theta\) is a neural network parameterized by \(\theta\). Under this linear interpolation scheme, \(z_{t + \Delta} - z_t = \Delta (z_1 - z_0)\). This means that \[ \begin{align*} &E\left[\log p((z_{t+\Delta} \vert z_t)\right] \\ &= E\left[-\frac{1}{2} \frac{(z_{t + \Delta} - [z_t + \Delta f_\theta (z_t, t)])^2}{\sigma^2 \Delta^2}\right] \\ &= E\left[\frac{-1}{2\sigma^2 \Delta} ((z_1 - z_0) - f_\theta (z_t, t))^2\right] \end{align*} \]
We can plug this simplification into the full ELBO, noting that the sum over time can be written as an expectation.
\[ E_{z_0} \left[ \log p(z_0) - \log q(z_0\vert z_1) + \sum_t \log p(z_{t+\Delta} \vert z_t) \right] = \\ -KL[q(z_0\vert z_1) \vert p(z_0)] + \frac{-1}{2\sigma^2} E_{t, z_0} \left[ ((z_1 - z_0) - f_\theta (z_t, t))^2 \right] \]
It breaks into two parts: a KL divergence keeping \(q(z_0\vert z_1)\) close to \(p(z_0)\), and a stochastic squared loss in which both time and latent code are sampled. As \(\Delta \to 0\), the distribution over times approaches uniform. This is precisely the loss for the Rectified Flow model for the case of a fixed \(q(z_0 \vert z_1)\). Originally, the authors interpreted this loss as a way to learn an ODE \(\frac{\partial z}{\partial t} = f(z, t)\) which transports samples \(z_0\) to observations \(z_1\). By casting the loss in the language of variational auto-encoders, however, we can learn a posterior distribution for \(z_0\) at the same time.