A Unified Framework for Diffusion Bridge Problems: Flow Matching and Schrödinger Matching into One

Minyoung Kim · 2025

Read A Unified Framework for Diffusion Bridge Problems: Flow Matching and Schrödinger Matching into One by Minyoung Kim (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-21756.txt.

Introduction

Minyoung Kim unifies two families of methods for building paths between probability distributions in a 2025 machine learning study. The bridge problem asks for a stochastic or ordinary differential equation that carries one distribution into another. Generative modeling made the problem famous, and the century-old Schrödinger bridge is its best known special case.

Flow matching methods give solutions through ordinary equations, while iterative fitting methods target the Schrödinger version. Kim's framework folds both into one view, clarifying what each algorithm assumes and computes. A single framework should make it easier to borrow tricks between camps and to design bridges for new conditional generation tasks.

Catalog record

CREATORMinyoung Kim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-21756

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Unified Framework for Diffusion Bridge Problems: Flow Matching and Schrödinger Matching into One Minyoung Kim Abstract The bridge problem is to find an SDE (or sometimes an ODE) that bridges two given distributions. The application areas of the bridge problem are enormous, among which the recent generative modeling (e.g., conditional or unconditional image generation) is the most popular. Also the famous Schrödinger bridge problem, a widely known problem for a century, is a special instance of the bridge problem. Two most popular algorithms to tackle the bridge problems in the deep learning era are: (conditional) flow matching and iterative fitting algorithms, where the former confined to ODE solutions, and the latter specifically for the Schrödinger bridge problem. The main contribution of this article is in two folds: i) We provide concise reviews of these algorithms with technical details to some extent ; ii) We propose a novel unified perspective and framework that subsumes these seemingly unrelated algorithms (and their variants) into one . In particular, we show that our unified framework can instantiate the Flow Matching (FM) algorithm, the (mini-batch) optimal transport FM algorithm, the (mini-batch) Schrödinger bridge FM algorithm, and the deep Schrödinger bridge matching (DSBM) algorithm as its special cases. We believe that this unified framework will be useful for viewing the bridge problems in a more general and flexible perspective, and in turn can help researchers and practitioners to develop new bridge algorithms in their fields. 1 (Diffusion) Bridge Problems The diffusion bridge problem , or simply the bridge problem , can be defined as follows. \bullet Bridge problem. Given two distributions \pi_{0}(\cdot) and \pi_{1}(\cdot) in \mathbb{R}^{d} , find an SDE, more specifically, find the drift function u_{t}(x) where u:\mathbb{R}[0,1]\times\mathbb{R}^{d}\to\mathbb{R}^{d} with a specified diffusion coefficient \sigma , \displaystyle dx_{t}=u_{t}(x_{t})dt+\sigma dW_{t},\ x_{0}\!\sim\!\pi_{0}(\cdot) (1) that yields x_{1}\!\sim\!\pi_{1}(\cdot) . Here \{W_{t}\}_{t} is the Wiener process or the Brownian motion. Alternatively one can aim to find a reverse-time SDE (or both). That is, find \overline{u}_{t}(x) in \displaystyle\overleftarrow{d}x_{t}=\overline{u}_{t}(x_{t})dt+\sigma\overleftarrow{d}W_{t},\ x_{1}\!\sim\!\pi_{1}(\cdot) (2) that yields x_{0}\!\sim\!\pi_{0}(\cdot) . Note that if we specify \sigma\!=\!0 , then our goal is to find an ODE that bridges the two distributions \pi_{0}(\cdot) and \pi_{1}(\cdot) . Once solved, the solution to the bridge problem can give us the ability to sample from one of the \pi_{\{0,1\}} given the samples from the other, simply by integrating the learned SDE. The application areas of the bridge problem are enormous, among which the generative modeling (e.g., conditional or unconditional image generation) is the most popular. For instance, in typical generative modeling, \pi_{0} is usually a tractable density like Gaussian, while \pi_{1} is a target distribution that we want to sample from. In the bridge problem, however, \pi_{0} can also be an arbitrary distribution beyond tractable densities like Gaussians, and we do not make any particular assumption on \pi_{0} and \pi_{1} as long as we have samples from the two distributions. \bullet Two instances of the bridge problem. There are two interesting special instances of the bridge problem: the Schrödinger bridge problem and the ODE bridge problem . • ODE bridge problem. We strictly restrict ourselves to ODEs, i.e., \sigma=0 . • Schrödinger bridge problem. With \sigma\!>\!0 , there is an additional constraint that the path measure of the SDE ( 1 ), denoted by P^{u} , is closest to a given reference SDE path measure P^{ref} . That is, \displaystyle\min_{u}\ \textrm{KL}(P^{u}||P^{ref})\ \ \textrm{s.t.}\ P^{u}_{0}(x_{0})=\pi_{0}(x_{0}),P^{u}_{1}(x_{1})=\pi_{1}(x_{1}) (3) where P_{t} of a path measure P indicates the marginal distribution at time t . In this case to have finite KL divergence, \sigma of P^{u} has to be set equal to \sigma_{ref} of P^{ref} , i.e., \sigma=\sigma_{ref} , (from the Girsanov theorem). 2 Flow Matching and Schrödinger Bridge Matching Algorithms Among several existing algorithms that aim to solve the bridge problems, in this article we focus on two recent matching algorithms: (conditional) flow matching ( Lipman et al, 2023 ; Tong et al, 2023 ; Albergo and Vanden-Eijnden, 2023 ; Liu et al, 2023 ) and Schrödinger bridge matching algorithms ( De Bortoli et al, 2021 ; Vargas et al, 2021 ; Shi et al, 2023 ) . These algorithms were developed independently: the former aimed to solve the ODE bridge problem while the latter the Schrödinger bridge problem. In this section we review the algorithms focusing mainly on the key ideas with some technical details, but being mathematically less rigorous for better readability. In Sec. 3 , we propose a novel unified framework that subsumes these two seemingly unrelated algorithms and their variants into one. 2.1 (Conditional) Flow Matching for ODE Bridge Problems The Flow Matching (FM) ( Lipman et al, 2023 ) or its extension Conditional Flow Matching (CFM) ( Tong et al, 2023 ) is one promising way to solve the ODE bridge problem. The key idea of the FM is quite intuitive. We first design some marginal distribution path \{P_{t}(x_{t})\}_{t} with the boundary conditions P_{0}\!=\!\pi_{0} , P_{1}\!=\!\pi_{1} . We then derive the ODE dx_{t}=u_{t}(x_{t})dt that yields \{P_{t}(x_{t})\}_{t} as its marginal distributions. The drift u_{t}(x_{t}) is approximated by a neural network v_{\theta}(t,x_{t}) with parameters \theta by solving: \displaystyle\min_{\theta}\ \mathbb{E}_{t,x_{t}\sim P_{t}}||u_{t}(x_{t})-v_{\theta}(t,x_{t})||^{2} (4) Once solved, we can generate samples from \pi_{1} (or \pi_{0} ) approximately by simulating: dx_{t}=v_{\theta}(t,x_{t})dt , x_{0}\!\sim\!\pi_{0} (resp., x_{1}\!\sim\!\pi_{1} ). However, one of the main limitations of this strategy is that designing the marginal path \{P_{t}(x_{t})\}_{t} satisfying the boundary condition is often difficult. And it is this issue that motivated the CFM. \bullet Conditional Flow Matching (CFM). To make the path design easier, we introduce some latent random variable z to condition x_{t} . Although CFM derivations hold regardless of the choice of z , it is typically chosen as the terminal random variates z=(x_{0},x_{1}) , and we will follow this practice and notation. Specifically, in CFM we design the so-called pinned marginal path \{P(x_{t}|x_{0},x_{1})\}_{t} and the coupling distribution Q(x_{0},x_{1}) subject to the condition P_{0}(x_{0})=\pi_{0}(x_{0}) , P_{1}(x_{1})=\pi_{1}(x_{1}) where P_{t}(x_{t}) is defined as \displaystyle P_{t}(x_{t}):=\int P_{t}(x_{t}|x_{0},x_{1})Q(x_{0},x_{1})d(x_{0},x_{1}) (5) Then we derive the ODE dx_{t}=u_{t}(x_{t}|x_{0},x_{1})dt that yields \{P(x_{t}|x_{0},x_{1})\}_{t} as its marginal distributions for each (x_{0},x_{1}) , which admits a closed form if \{P(x_{t}|x_{0},x_{1})\}_{t} are Gaussians ( Lipman et al, 2023 ) . We then approximate \mathbb{E}[u_{t}(x_{t}|x_{0},x_{1})|x_{t}] , the conditional expectation derived from the joint P(x_{t}|x_{0},x_{1})Q(x_{0},x_{1}) , by a neural network v_{\theta}(t,x_{t}) by solving the following optimization: \displaystyle\min_{\theta}\ \mathbb{E}\ ||u_{t}(x_{t}|x_{0},x_{1})-v_{\theta}(t,x_{t})||^{2} (6) where the expectation is taken with respect to the joint P(x_{t}|x_{0},x_{1})Q(x_{0},x_{1}) and uniform t . Surprisingly, it can be shown ( Tong et al, 2023 ) that the gradient of the objective in ( 6 ) coincides with that in ( 4 ) for u_{t}(x_{t}) defined as: \displaystyle u_{t}(x)=\frac{1}{P_{t}(x_{t})}\mathbb{E}_{Q(x_{0},x_{1})}[u_{t}(x_{t}|x_{0},x_{1})P_{t}(x_{t}|x_{0},x_{1})] (7) hence sharing the same training dynamics as ( 4 ). Therefore the optimal v_{\theta}(t,x_{t}) of ( 6 ) is a good estimate for u_{t}(x_{t}) . Since the ODE dx_{t}=u_{t}(x_{t})dt admits \{P_{t}(x_{t})\}_{t} as its marginal distributions, so does dx_{t}=v_{\theta}(t,x_{t})dt approximately. Many existing flow matching variants including FM ( Lipman et al, 2023 ) , Stochastic Interpolation ( Albergo and Vanden-Eijnden, 2023 ) , and Rectified FM ( Liu et al, 2023 ) can be viewed as special instances of this CFM framework. For instance, these models can be realized by having a straight line (linear interpolation) pinned marginal path ( 8 ) with vanishing variance while

Read the complete text