---
title: Multi-Flow Matching Techniques
url: https://www.emergentmind.com/topics/multi-flow-matching
type: topic
---

# Multi-Flow Matching Techniques

Searching arXiv for relevant papers on multi-flow matching and closely related flow-matching variants.
Multi-Flow Matching denotes a family of flow-matching formulations that address settings where a single deterministic conditional velocity field is an inadequate representation of the supervision induced by standard flow matching. In the classical rectified or conditional flow-matching setup, a model learns a time-dependent vector field from source-target interpolation pairs, but at a fixed space-time location the supervision can be inherently ambiguous: multiple couplings or multiple observed marginals can induce different valid transport directions. Within the recent literature, “multi-flow” has therefore acquired several distinct but related meanings: a latent-conditioned mixture over local velocity fields, a training procedure over multisample couplings, a hierarchy of flows over flows, a cascade of conditional refinements across fidelities or scales, a blockwise decomposition of the time axis into specialized local flows, and a multi-marginal extension in which one global dynamics is constrained by more than two observed marginals. The common theme is that flow matching is generalized beyond a single unimodal conditional transport law by introducing additional latent structure, couplings, stages, scales, modalities, or time marginals [2502.09616], [2304.14772], [2508.04351], [2507.13350], [2605.16118], [2510.03569], [2510.01159].

## 1. Deterministic flow matching and the source of multi-flow ambiguity

Standard rectified flow matching starts from a source distribution \(p_0(x_0)\) and a target distribution \(p_1(x_1)\). At inference time, one samples \(x_0 \sim p_0\) and solves an ODE using a learned velocity field \(v_\theta(x_t,t)\). The likelihood relation is written via the instantaneous change of variables formula as
\[
\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,
\]
with transport PDE
\[
\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).
\]
In the rectified case, training uses linear interpolation
\[
x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,
\]
whose time derivative gives the target velocity
\[
v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.
\]
The standard objective is
\[
\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].
\]
This formulation is simulation-free in the usual FM sense, but it is only formally single-valued when the conditional target velocity at a given \((x_t,t)\) is unique [2502.09616].

The multi-flow issue appears because random source-target coupling can make the supervision at fixed \((x_t,t)\) ambiguous. Many distinct pairs can produce the same interpolated location \(x_t\) at the same time \(t\), while implying different target velocities \(x_1-x_0\). Under squared loss, the optimal predictor is the conditional mean,
\[
v^\ast(x_t,t) = \mathbb{E}_{\{(x_0,x_1,t) : \phi(x_0,x_1,t) = x_t\}\left[v(x_0,x_1,t)\right].
\]
The consequence is that classical rectified flow matching does not learn a multimodal conditional velocity distribution; it learns its average. The associated geometric intuition is that averaging causes trajectories to curve, make U-turns, or avoid intersections that the underlying coupled straight-line flows would naturally have [2502.09616].

A probabilistic reinterpretation makes this limitation explicit. If one assumes
\[
p(v|x_t,t) = {\cal N}(v; v_\theta(x_t,t),I),
\]
then maximizing the log-likelihood of empirical velocity targets yields
\[
\begin{split}
&\mathbb{E}_{t,x_0,x_1}\left[\log p(x_1-x_0|x_t,t)\right] \\
&\propto -\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - x_1 + x_0\|_2^2\right].
\end{split}
\]
Standard rectified flow matching is therefore maximum likelihood under a unimodal conditional velocity model. In this sense, the foundational motivation for Multi-Flow Matching is not merely that the data distribution is multimodal, but that the velocity distribution itself is often multimodal at fixed \((x_t,t)\) [2502.09616].

A closely related ambiguity argument appears in hierarchical rectified flow matching. There the target velocity distribution at fixed \((x_t,t)\) is denoted \(T_1(v;x_t,t)\), and the paper emphasizes that vanilla FM captures the mean conditional velocity rather than the full conditional velocity law. The distinction between multimodality of the data distribution and multimodality of the conditional velocity distribution is central: vanilla FM can handle the former, but collapses the latter under \(\ell_2\) regression [2507.13350].

## 2. Latent-conditioned multi-flow formulations

The most direct realization of Multi-Flow Matching as multimodal conditional transport is Variational Rectified Flow Matching (VRFM). VRFM introduces a latent variable \(z\) so that the velocity model becomes conditionally unimodal but marginally multimodal:
\[
p(v|x_t,t,z) = {\cal N}(v; v_\theta(x_t,t,z),I), \qquad p(z)={\cal N}(z;0,I).
\]
Marginalizing the latent gives
\[
p(v|x_t,t) = \int p(v|x_t,t,z)p(z)dz,
\]
which is explicitly a Gaussian mixture. The method therefore replaces a single deterministic velocity field by a latent-conditioned family \(v_\theta(x_t,t,z)\), so that multiple valid transport directions can coexist at the same \((x_t,t)\) [2502.09616].

Because \(z\) is unobserved, VRFM introduces a recognition model
\[
q_\phi(z|x_0,x_1,x_t,t),
\]
with Gaussian posterior
\[
q_\phi(z|x_0,x_1,x_t,t) = {\cal N}(z; \mu_\phi(x_0,x_1,x_t,t),\sigma_\phi(x_0,x_1,x_t,t)).
\]
The variational lower bound is
\[
\begin{split}
\log p(v|x_t,t) & \geq \mathbb{E}_{z\sim q_\phi}\left[\log p(v|x_t,t,z)\right] \\
&- D_\text{KL}(q_\phi(\cdot|x_0,x_1,x_t,t)|p(\cdot)).
\end{split}
\]
Substituting the Gaussian observation model yields the VRFM objective
\[
\begin{split}
&\mathbb{E}_{t,x_0,x_1}\left[\log p(x_1-x_0|x_t,t)\right] \geq \\
&\mathbb{E}_{t,x_0,x_1}[-\mathbb{E}_{z\sim q_\phi}\left[\|v_\theta(x_t,t,z) - x_1 + x_0\|_2^2\right] \\
& - D_\text{KL}(q_\phi(\cdot|x_0,x_1,x_t,t)|p(\cdot))].
\end{split}
\]
The conceptual advance is that the model no longer forces all valid velocities at a point into one mean vector; it represents multiple possible directions through latent-conditioned components [2502.09616].

At inference, VRFM samples \(x_0 \sim p_0\), samples \(z \sim p(z)\), and solves the ODE from \(t=0\) to \(t=1\) using \(v_\theta(x_t,t,z)\). The latent is sampled once before integration and then held fixed along the trajectory. Generation is therefore deterministic conditional on \((x_0,z)\), but diverse marginally because different \(z\) produce different trajectories. This gives a particularly clean form of multi-flow matching: one samples a latent that selects a coherent global flow direction or trajectory family, rather than injecting randomness at every time step [2502.09616].

A related but distinct latent-conditioned approach is Latent-CFM. It introduces a latent factor \(f\in \mathbb{R}^k\) intended to capture latent structure of the target, and factorizes the endpoint coupling as
\[
q(x_0,x_1) = \int q (f) q(x_0,x_1|f) df, \qquad q(x_0,x_1|f) = p_0(x_0) \times p_1(x_1|f).
\]
The flow network becomes
\[
v_\theta(x,f,t),
\]
and the latent-conditioned training objective is
\[
\mathcal{L}_{Latent-CFM} = \E_{t,q(f),q(x_0,x_1|f),p_t(x|x_0,x_1)}
 ||v_\theta (x,f,t) - u_t(x|x_0,x_1)||_2^2.
\]
The paper further shows that if
\[
v_\theta(x,t) = \E_{q(f|x_0,x_1)} v_\theta (x,f,t),
\]
then
\[
\mathcal{L}_{\text{CFM} \le \mathcal{L}_{Latent-CFM},
\]
with the difference equal to an expected squared deviation between latent-conditioned and averaged fields. This formalizes a mixture-of-flows interpretation: the single global flow can be viewed as the average of latent-conditioned flows, while Latent-CFM learns the finer decomposition directly [2505.04486].

The Variational Flow-Matching Policy (VFP) for robot manipulation extends the same logic to action generation. The policy is
\[
p_{\theta,\psi}(a\mid s) = \int p_\theta(a\mid z,s)\,p_\psi(z\mid s)\,dz,
\]
with latent-conditioned velocity
\[
v_\theta(a_t,t,s,z).
\]
The paper gives the ambiguity decomposition
\[
\mathrm{Var}[v^*\mid a_t,t,s] = \mathbb{E}_{z\mid a_t,t,s}\!\left[ \mathrm{Var}[v^*\mid a_t,t,s,z] \right] + \mathrm{Var}_{z\mid a_t,t,s} \left( \mathbb{E}[v^*\mid a_t,t,s,z] \right),
\]
implying
\[
\mathcal{A}_\mathrm{VFP}\le \mathcal{A}_\mathrm{FM}.
\]
In precise terms, VFP treats multi-flow matching as mode factorization: a latent variable captures task-level and trajectory-level multi-modality, and a latent-conditioned decoder solves a lower-ambiguity conditional flow problem [2508.01622].

## 3. Coupling-based and hierarchical approaches

A second major line of Multi-Flow Matching modifies the pairing mechanism rather than introducing latent variables. Multisample Flow Matching (MFM) generalizes standard Flow Matching by allowing an arbitrary joint distribution \(q(x_0,x_1)\) over source and target endpoints, provided it has the correct marginals:
\[
\int q(x_0,x_1)\, dx_1 = q_0(x_0), \qquad \int q(x_0,x_1)\, dx_0 = q_1(x_1).
\]
Standard FM with independent pairings is the special case
\[
q(x_0,x_1)=q_0(x_0)q_1(x_1).
\]
With a nontrivial coupling, the same straight conditional interpolation
\[
x_t=(1-t)x_0+t x_1,\qquad u_t(x_t\mid x_1)=x_1-x_0
\]
is retained, but the induced marginal vector field becomes less ambiguous because endpoint assignments are more coherent [2304.14772].

The central MFM objective is Joint Conditional Flow Matching:
\[
\mathcal{L}_{\text{JCFM} = \mathbb{E}_{t,\, q(x_0,x_1)} \| v_t(x_t;\theta) - u_t(x_t \mid x_1)\|^2,
\]
with \(x_t=\psi_t(x_0\mid x_1)\). In practice the coupling is constructed on minibatches via a doubly stochastic matrix \(\pi(i,j)\), giving an empirical joint
\[
q^k(x_0,x_1)=\frac{1}{k}\sum_{i,j=1}^k \delta(x_0-x_0^{(i)})\delta(x_1-x_1^{(j)})\pi(i,j).
\]
This viewpoint supports an “ambiguity reduction” interpretation of Multi-Flow Matching: rather than explicitly representing multiple local flows, one alters the coupling so that fewer conflicting flows intersect at the same \((x_t,t)\) [2304.14772].

MFM further gives a variance argument. For fixed \(x,t\),
\[
\sigma^2_{t,x} = \mathrm{Tr}\!\left[ \mathrm{Cov}_{p_t(x_1\mid x)} \left( \nabla_\theta \|v_t(x;\theta)-u_t(x\mid x_1)\|^2 \right) \right]
\]
is bounded by
\[
\sigma^2_{t,x} \le \|\nabla_\theta v_t(x;\theta)\|^2 \, \mathbb{E}_{p_t(x_1\mid x)} \|u_t(x)-u_t(x\mid x_1)\|^2.
\]
A lower optimal JCFM value therefore implies a lower upper bound on gradient variance. This is one of the most explicit formal links between couplings and multi-flow ambiguity: better couplings make endpoint assignments more coherent, which reduces conditional ambiguity and variance [2304.14772].

Hierarchical Rectified Flow Matching with Mini-Batch Couplings pushes the idea further by introducing a flow over the velocity distribution itself. Standard rectified flow uses
\[
x_t = (1-t)x_0 + tx_1,
\]
with target velocity \(x_1-x_0\). Hierarchical rectified flow introduces a source velocity \(v_0 \sim T_0\), a target velocity \(v_1(x_t,t)\sim T_1(v;x_t,t)\), and a second interpolation variable \(\tau\):
\[
u_\tau(x_t,t)=(1-\tau)v_0 + \tau v_1(x_t,t).
\]
The corresponding acceleration is
\[
a(x_t,t,u_\tau,\tau) = v_1(x_t,t)-v_0 = x_1-x_0-v_0.
\]
Generation then uses the coupled ODEs
\[
\begin{cases}
du_\tau(z_t,t)=a(z_t,t,u_\tau,\tau)\,d\tau,\qquad u_0\sim T_0,\\[3pt]
dz_t=u_1(z_t,t)\,dt,\qquad z_0\sim p_0.
\end{cases}
\]
This is a literal “flow over flows”: a position-space flow driven by a velocity sampled from a second flow in velocity space [2507.13350].

The paper’s central theorem connects the induced velocity distribution to the source-target coupling:
\[
T_1(v; x_t,t)= \frac{\gamma(x_t-tv,\;x_t+(1-t)v)}{p_t(x_t)},
\]
where \(\gamma(x_0,x_1)\) is any coupling with correct marginals. Thus the complexity of the velocity distribution depends directly on the coupling. Data coupling simplifies the velocity law; velocity coupling simplifies the acceleration law and straightens velocity trajectories. In this formulation, Multi-Flow Matching means not only modeling multiple possible local transport directions through a hierarchy, but also allocating their complexity across levels via couplings [2507.13350].

## 4. Multi-marginal and longitudinal extensions

In snapshot settings one does not observe paired source-target samples, but only independent samples from marginal distributions
\[
\rho_i = p_{t_i}, \qquad i=0,\dots,M,
\]
at multiple irregular times. Multi-Marginal Stochastic Flow Matching (MMSFM) extends simulation-free flow/score matching to this setting. The underlying state is modeled as an Itô diffusion
\[
dX(t) = u_t(X(t))\,dt + g(t)\,dW(t),
\]
with Fokker–Planck equation
\[
\frac{\partial p_t(x)}{\partial t} = -\nabla\cdot\bigl(p_t(x)u_t(x)\bigr) + \frac{g^2(t)}{2}\Delta p_t(x).
\]
The multi-marginal extension enters through a conditioning variable \(z\) built from aligned samples across multiple marginals using approximate MMOT and spline interpolation [2508.04351].

Instead of conditioning on an endpoint pair, MMSFM uses aligned tuples
\[
z=(x_{t_i}^\star,\dots,x_{t_{i+k}}^\star)
\]
over local windows and defines a Gaussian conditional path
\[
p_t(x\mid z) = \mathcal N(x\mid \mu_t,\sigma_t^2),
\]
where \(\mu_t\) is a spline through the aligned tuple. The deterministic flow target and score target are
\[
u_t^\circ(x\mid z) = \frac{\sigma_t'}{\sigma_t}(x-\mu_t)+\mu_t', \qquad
\nabla \log p_t(x\mid z) = \frac{\mu_t-x}{\sigma_t^2}.
\]
The method learns both a deterministic flow field \(v_t(x;\theta)\) and a score field \(s_t(x;\theta)\), combining them into a stochastic drift. In this setting, Multi-Flow Matching means learning one global stochastic dynamics from many overlapping local bridges across several observed marginals, rather than a single source-target bridge [2508.04351].

The paper’s preferred construction uses overlapping triplet windows (\(k=2\)) with monotonic cubic Hermite splines. The theorem-level contribution is that overlapping local FM losses combine into a regularized aggregate objective, and standard conditional flow matching is recovered as a special aligned case. This makes the method a direct extension of FM from two marginals to ordered multiple marginals [2508.04351].

Interpolative Multi-Marginal Flow Matching (IMMFM) addresses longitudinal sparse trajectory modeling with a different path design. On each segment \([t_i,t_{i+1}]\), it uses a piecewise-quadratic conditional path
\[
p_t(x\mid z) = \mathcal{N}\!\left(x\bigm|\mu_t(z),\;\sigma^2(t)I\right),
\]
with
\[
\mu_t = x_i + v_i(t - t_i) + \frac{1}{2}(v_i - v_{i+1})\frac{(t - t_i)(t_{i+1} - t)}{t_{i+1} - t_i},
\]
and Brownian-bridge-like variance
\[
\sigma(t)=\sigma_0\frac{(t-t_i)(t_{i+1}-t)}{t_{i+1}-t_i}.
\]
The induced drift target is
\[
u^\circ_t(x \mid z)
=
v_i + \frac{1}{2}(v_i - v_{i+1})(2\alpha_t - 1)
+ \frac{\sigma'(t)}{\sigma(t)}\bigl(x - \mu_t(z)\bigr).
\]
This is multi-marginal in the sense that each local path depends on more than one segment through \(v_{i+1}\), so the current segment looks ahead to the following observation [2510.03569].

ALI-CFM constructs multi-marginal conditional paths in a different way. It learns neural interpolants
\[
G_\phi(x_0, x_1, t) = (1-t)x_0 + tx_1 + t(1-t) f_\phi(x_0, x_1, t),
\]
and requires their pushforward distributions at observed intermediate times to match the data marginals:
\[
(G_\phi(\cdot,\cdot,t_i))_\#\pi \approx q_{t_i}.
\]
The intermediate matching is enforced adversarially with
\[
L_\text{GAN}(G_\phi,D_\gamma;t_i)
\]
plus regularization, and the learned interpolants are then marginalized by a CFM objective using
\[
\frac{d}{dt}G_\phi(x_0,x_1,t)
\]
as the target conditional velocity. Here Multi-Flow Matching means learning a single smooth conditional path across more than two observed marginals and then plugging that path into conditional flow matching [2510.01159].

## 5. Multi-scale, cascaded, blockwise, and heterogeneous multi-flow systems

Another interpretation of Multi-Flow Matching organizes transport across scales, fidelities, or modality types rather than latent modes or time marginals.

Multi-Fidelity Flow Matching (MFFM) is a conditional, residual, cascaded form of flow matching for PDE solution refinement. At a single level it models residuals
\[
\delta = u_H - u_L
\]
or, across grids,
\[
\delta_\ell = u_{\ell+1} - I_{\ell\to\ell+1}(u_\ell),
\]
using conditional flow matching with data-calibrated source noise:
\[
\delta_t = t\delta + (1-t)\varepsilon,\qquad \varepsilon\sim q_\delta.
\]
The single-level objective is
\[
\mathcal{J}_{\mathrm{MFFM}(\theta) = \mathbb{E}_{t\sim \mathcal{U}[0,1]} \mathbb{E}_{(u_L,\delta)\sim \mathcal{D} \mathbb{E}_{\varepsilon\sim q_\delta} \left\| v_\theta(\delta_t,t,u_L) - (\delta-\varepsilon) \right\|^2.
\]
A cascade then composes one such conditional flow per adjacent fidelity transition. In this sense, MFFM is multi-flow matching as a composition of multiple conditional transports, one per fidelity gap, rather than as a mixture or multimodal velocity field [2605.16118].

Blockwise Flow Matching (BFM) instead partitions the generative trajectory into temporal segments
\[
0=t_0 < t_1 < \cdots < t_M=1,
\]
and assigns a distinct velocity block \(v_\theta^{(m)}\) to each interval. For segment \(m\),
\[
x_t = (1-a_m(t))x_{t_{m-1}} + a_m(t)x_{t_m},\qquad
a_m(t)= \frac{t-t_{m-1}{t_m-t_{m-1},
\]
with local target velocity
\[
v_t^{(m)} = \frac{x_{t_m} - x_{t_{m-1}}}{t_m - t_{m-1}}.
\]
The blockwise loss is
\[
\mathcal{L}_{\text{BFM}(\theta)
=
\mathbb{E}_{x_0, x_1, t}
\left[
\| v_{\theta}^{(m)}(x_t, c) - v_t^{(m)} \|^2
\right].
\]
BFM is therefore a piecewise, temporally routed, multi-velocity-field formulation of flow matching: one global transport is decomposed into multiple local flows over adjacent time intervals, each with its own parameters and local regression target [2510.21167].

Laplacian Multi-scale Flow Matching (LapFlow) decomposes latent images into a Laplacian hierarchy
\[
x_1^{(2)} = \mathrm{Down}(\mathrm{Down}(x_1)), \qquad
x_1^{(1)} = \mathrm{Down}(x_1) - \mathrm{Up}(x_1^{(2)}), \qquad
x_1^{(0)} = x_1 - \mathrm{Up}(\mathrm{Down}(x_1)),
\]
and defines scale-specific paths
\[
x_t^{(k)} = \alpha_t^{(k)} x_1^{(k)} + \sigma_t^{(k)} x_0^{(k)}.
\]
A single joint transformer predicts velocities for the active subset of scales, with multiscale loss
\[
\mathcal{L}_{\mathrm{mv} = \sum_{k=s}^{2} w_k\, \mathbb{E}
\left\| v_t^{(k)} - u_t^{(k)}(x_t^{(k)} \mid x_1^{(k)}) \right\|^2.
\]
This is multi-flow matching in a specifically multiscale sense: separate conditional flows per scale, jointly modeled by one network, with staggered temporal supports and causal cross-scale attention [2602.19461].

Heterogeneous variable-type settings motivate yet another form. FMIP for mixed-integer linear programming factorizes the conditional path as
\[
p_{t\mid1}(\boldsymbol{G}_t\mid\boldsymbol{G}_1)
=
p_{t\mid1}(\boldsymbol{d}_t \mid\boldsymbol{d}_1, \boldsymbol{G})
\;
p_{t\mid1}(\boldsymbol{c}_t \mid\boldsymbol{c}_1, \boldsymbol{G}),
\]
and combines a continuous flow for continuous variables with a discrete flow for integer variables. MolFORM for structure-based drug design likewise combines conditional flow matching for coordinates with discrete flow matching for atom types. In both cases, “multi-flow” means modality-specific transport laws within a shared conditional model [2507.23390], [2507.05503].

## 6. Scope, distinctions, and non-examples

The literature uses similar language for conceptually different ideas, and several papers are explicit about what they are not. Flow Generator Matching (FGM) compresses a pretrained continuous-time flow-matching teacher into a one-step generator. It is relevant mainly because it collapses a multi-step ODE trajectory into a one-step map, but it does not propose multiple learned flows, a mixture of vector fields, multiple probability paths, or a multi-branch flow architecture. Formally it is a single-flow, single-path distillation method rather than a multi-flow formulation [2410.19310].

A similar distinction appears in application papers. “Flow Matching Imitation Learning for Multi-Support Manipulation” uses a single conditional flow field over a multimodal trajectory distribution of whole-body/contact strategies. The “multi” is in support/contact strategies and trajectory modes, not in multiple explicit flow models [2407.12381]. MATCH for multi-view anomaly detection uses one shared OT-CFM latent flow conditioned on view index, with view-wise max aggregation. Its multi-view nature does not amount to multiple coupled flows [2606.24375]. MAC-Flow for multi-agent coordination uses a joint conditional flow over concatenated multi-agent actions, followed by decentralized policy distillation; it is a multi-agent application of standard flow matching, not a new multi-flow objective [2511.05005].

By contrast, heat-based or blur-based multiscale path constructions such as Heat Dissipation Flow Matching are best described as multi-scale single-flow matching: one ODE and one vector field, but a path explicitly structured by a continuum of blur scales. This is highly relevant to multiscale FM, but not to multiple interacting flows in the strict sense [2605.19371].

These distinctions suggest that “Multi-Flow Matching” is best treated as an umbrella term rather than a single formalism. In current usage it can denote at least five technically distinct patterns: latent-conditioned mixtures over vector fields, coupling-based ambiguity reduction, hierarchical flows over velocities, multi-marginal path constructions over several observed marginals, and staged or modality-specific compositions of several local flows. A plausible implication is that the term is stabilizing around the broader problem—how to generalize flow matching when a single unimodal conditional transport law is structurally mismatched to the data—rather than around one canonical algorithm [2502.09616], [2304.14772], [2508.04351], [2507.13350].

Source: https://www.emergentmind.com/topics/multi-flow-matching