Multi-Flow Matching Techniques

• Multi-Flow Matching is a framework that generalizes flow matching by introducing latent-conditioned mixtures, coupling strategies, and hierarchical flows to address inherent ambiguities in conditional velocity fields.
• It leverages specific techniques like VRFM, Latent-CFM, and multisample couplings to enable the representation of multimodal transport directions, avoiding the averaging pitfalls of traditional methods.
• This approach improves modeling accuracy and reduces gradient variance by constructing more coherent endpoint couplings and decomposing the transport problem across scales or modalities.

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 (Guo et al., 13 Feb 2025, Pooladian et al., 2023, Lee et al., 6 Aug 2025, Zhang et al., 17 Jul 2025, Chen et al., 15 May 2026, Islam et al., 3 Oct 2025, Kviman et al., 1 Oct 2025).

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 (Guo et al., 13 Feb 2025).

The multi-flow issue appears because random source-target coupling can make the supervision at fixed $p_1(x_1)$0 ambiguous. Many distinct pairs can produce the same interpolated location $p_1(x_1)$1 at the same time $p_1(x_1)$2, while implying different target velocities $p_1(x_1)$3. Under squared loss, the optimal predictor is the conditional mean,

$p_1(x_1)$4

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 (Guo et al., 13 Feb 2025).

A probabilistic reinterpretation makes this limitation explicit. If one assumes

$p_1(x_1)$5

then maximizing the log-likelihood of empirical velocity targets yields

$p_1(x_1)$6

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 $p_1(x_1)$7 (Guo et al., 13 Feb 2025).

A closely related ambiguity argument appears in hierarchical rectified flow matching. There the target velocity distribution at fixed $p_1(x_1)$8 is denoted $p_1(x_1)$9, 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 $x_0 \sim p_0$0 regression (Zhang et al., 17 Jul 2025).

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 $x_0 \sim p_0$1 so that the velocity model becomes conditionally unimodal but marginally multimodal: $x_0 \sim p_0$2 Marginalizing the latent gives

$x_0 \sim p_0$3

which is explicitly a Gaussian mixture. The method therefore replaces a single deterministic velocity field by a latent-conditioned family $x_0 \sim p_0$4, so that multiple valid transport directions can coexist at the same $x_0 \sim p_0$5 (Guo et al., 13 Feb 2025).

Because $x_0 \sim p_0$6 is unobserved, VRFM introduces a recognition model

$x_0 \sim p_0$7

with Gaussian posterior

$x_0 \sim p_0$8

The variational lower bound is

$x_0 \sim p_0$9

Substituting the Gaussian observation model yields the VRFM objective

$v_\theta(x_t,t)$0

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 (Guo et al., 13 Feb 2025).

At inference, VRFM samples $v_\theta(x_t,t)$1, samples $v_\theta(x_t,t)$2, and solves the ODE from $v_\theta(x_t,t)$3 to $v_\theta(x_t,t)$4 using $v_\theta(x_t,t)$5. The latent is sampled once before integration and then held fixed along the trajectory. Generation is therefore deterministic conditional on $v_\theta(x_t,t)$6, but diverse marginally because different $v_\theta(x_t,t)$7 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 (Guo et al., 13 Feb 2025).

A related but distinct latent-conditioned approach is Latent-CFM. It introduces a latent factor $v_\theta(x_t,t)$8 intended to capture latent structure of the target, and factorizes the endpoint coupling as

$v_\theta(x_t,t)$9

The flow network becomes

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$0

and the latent-conditioned training objective is

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$1

The paper further shows that if

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$2

then

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$3

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 (Samaddar et al., 7 May 2025).

The Variational Flow-Matching Policy (VFP) for robot manipulation extends the same logic to action generation. The policy is

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$4

with latent-conditioned velocity

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$5

The paper gives the ambiguity decomposition

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$6

implying

$\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$7

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 (Zhai et al., 3 Aug 2025).

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 $\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$8 over source and target endpoints, provided it has the correct marginals: $\log p_1(x_1) = \log p_0(x_0) + \int_1^0 \di v_\theta(x_t,t) dt,$9 Standard FM with independent pairings is the special case

$\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$0

With a nontrivial coupling, the same straight conditional interpolation

$\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$1

is retained, but the induced marginal vector field becomes less ambiguous because endpoint assignments are more coherent (Pooladian et al., 2023).

The central MFM objective is Joint Conditional Flow Matching: $\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$2 with $\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$3. In practice the coupling is constructed on minibatches via a doubly stochastic matrix $\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$4, giving an empirical joint

$\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$5

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 $\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$6 (Pooladian et al., 2023).

MFM further gives a variance argument. For fixed $\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$7,

$\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$8

is bounded by

$\frac{\partial \log p_t(x_t)}{\partial t} = - \di v_\theta(x_t,t).$9

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 (Pooladian et al., 2023).

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 = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$0

with target velocity $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$1. Hierarchical rectified flow introduces a source velocity $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$2, a target velocity $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$3, and a second interpolation variable $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$4: $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$5 The corresponding acceleration is

$x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$6

Generation then uses the coupled ODEs

$x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$7

This is a literal “flow over flows”: a position-space flow driven by a velocity sampled from a second flow in velocity space (Zhang et al., 17 Jul 2025).

The paper’s central theorem connects the induced velocity distribution to the source-target coupling: $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$8 where $x_t = \phi(x_0,x_1,t) = (1-t)x_0 + tx_1,$9 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 (Zhang et al., 17 Jul 2025).

4. Multi-marginal and longitudinal extensions

In snapshot settings one does not observe paired source-target samples, but only independent samples from marginal distributions

$$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$0

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

$$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$1

with Fokker–Planck equation

$$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$2

The multi-marginal extension enters through a conditioning variable $$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$3 built from aligned samples across multiple marginals using approximate MMOT and spline interpolation (Lee et al., 6 Aug 2025).

Instead of conditioning on an endpoint pair, MMSFM uses aligned tuples

$$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$4

over local windows and defines a Gaussian conditional path

$$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$5

where $$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$6 is a spline through the aligned tuple. The deterministic flow target and score target are

$$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$7

The method learns both a deterministic flow field $$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$8 and a score field $$v(x_0,x_1,t)=\frac{\partial \phi(x_0,x_1,t)}{\partial t}=x_1-x_0.$$9, 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 (Lee et al., 6 Aug 2025).

The paper’s preferred construction uses overlapping triplet windows ($$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$0) 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 (Lee et al., 6 Aug 2025).

Interpolative Multi-Marginal Flow Matching (IMMFM) addresses longitudinal sparse trajectory modeling with a different path design. On each segment $$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$1, it uses a piecewise-quadratic conditional path

$$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$2

with

$$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$3

and Brownian-bridge-like variance

$$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$4

The induced drift target is

$$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$5

This is multi-marginal in the sense that each local path depends on more than one segment through $$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$6, so the current segment looks ahead to the following observation (Islam et al., 3 Oct 2025).

ALI-CFM constructs multi-marginal conditional paths in a different way. It learns neural interpolants

$$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$7

and requires their pushforward distributions at observed intermediate times to match the data marginals: $$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$8 The intermediate matching is enforced adversarially with

$$\mathbb{E}_{t,x_0,x_1}\left[\|v_\theta(x_t,t) - v(x_0,x_1,t)\|_2^2\right].$$9

plus regularization, and the learned interpolants are then marginalized by a CFM objective using

$(x_t,t)$0

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 (Kviman et al., 1 Oct 2025).

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

$(x_t,t)$1

or, across grids,

$(x_t,t)$2

using conditional flow matching with data-calibrated source noise: $(x_t,t)$3 The single-level objective is

$(x_t,t)$4

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 (Chen et al., 15 May 2026).

Blockwise Flow Matching (BFM) instead partitions the generative trajectory into temporal segments

$(x_t,t)$5

and assigns a distinct velocity block $(x_t,t)$6 to each interval. For segment $(x_t,t)$7,

$(x_t,t)$8

with local target velocity

$(x_t,t)$9

The blockwise loss is

$p_1(x_1)$00

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 (Park et al., 24 Oct 2025).

Laplacian Multi-scale Flow Matching (LapFlow) decomposes latent images into a Laplacian hierarchy

$p_1(x_1)$01

and defines scale-specific paths

$p_1(x_1)$02

A single joint transformer predicts velocities for the active subset of scales, with multiscale loss

$p_1(x_1)$03

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 (Zhao et al., 23 Feb 2026).

Heterogeneous variable-type settings motivate yet another form. FMIP for mixed-integer linear programming factorizes the conditional path as

$p_1(x_1)$04

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 (Li et al., 31 Jul 2025, Huang et al., 7 Jul 2025).

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 (Huang et al., 2024).

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 (Rouxel et al., 2024). 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 (Kruse et al., 23 Jun 2026). 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 (Lee et al., 7 Nov 2025).

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 (Ma et al., 19 May 2026).

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 (Guo et al., 13 Feb 2025, Pooladian et al., 2023, Lee et al., 6 Aug 2025, Zhang et al., 17 Jul 2025).