---
title: 'Flow Marching: Integrating Learned ODEs & PDEs'
url: https://www.emergentmind.com/topics/flow-marching
type: topic
---

# Flow Marching: Integrating Learned ODEs & PDEs

to describe numerical integration along learned ODEs in flow matching and rectified flow models, as in FlowCast and StreamFlow. Also in "Transition Flow Matching" paper flow marching as concept. Need maybe mention "sometimes misheard as flow marching" from MAC-Flow and "informally called". Could address misconception that flow marching is not a standardized term in FM literature. Since data says "sometimes misheard as flow marching" and "The term 'flow marching' is not standard in literature but naturally refers to numerically marching". So can objectively mention that. Must be careful if concrete claim verbatim? Need quote maybe "The term 'flow marching' is not standard in the literature" from Blockwise and "sometimes informally called". That's in data, so okay.

Need start with intro paragraph no heading. Then sections numbered h2 titles. Maybe:

## 1. Term and scope
## 2. Flow marching in generative ODE models
## 3. Architectures and efficiency strategies
## 4. PDE and scientific computing interpretations
## 5. Environmental and hydrodynamic meanings
## 6. Conceptual issues, advantages, and open problems

Need number headings maybe "## 1. ..." etc.

Need citations within. Could cite introduction with main papers [2509.18611], [2511.09731], [2511.22009], [2603.15689]. Need cite single paper maybe one or two. Since intro can cite multiple.

Need include equations. Can include some central ones. Need not overdo. Since audience technical. Could include ODE, transition flow identity, CFM loss, Euler update, TMLT transformed PDE maybe in respective sections.

Need maybe include table կազմակերպ? Could use one small table to organize usages across literatures. Max 3 columns. But principles say tables brief. Could have columns: domain, meaning, representative paper. Might be useful. Detailed explanations outside. Let's do one small 3-column table in section 1 or 5. Need keep cells brief.

Need ensure all claims trace to data. No invented frameworks. Need mention exact metrics only if in data and relevant. Perhaps use them sparingly. But article task is comprehensive encyclopedia article, could mention concrete benchmarks. In generative section maybe cite efficiency and performance: FlowCast 10 NFEs vs DDIM etc. But if article general, maybe mention as exemplar. Since all numbers given. Need not overdo.

Need note "Editor's term" if coin shorthand. Might not need.

Need not greet or mention process.

Need same language as user's query English.

Let's craft with concise but comprehensive. Need cite at most once per paper per section. Can cite multiple papers in one section. Since many papers, careful not to repeat same ID too many times per section. Can repeat across sections.

Need maybe mention "Transition Flow Matching" directly learns transition flow as global quantity, enabling generation at arbitrary future time points [2603.15689]. "Stream-level flow matching with Gaussian processes" extends CFM by conditioning on latent streams modeled with GPs and remains simulation-free [2409.20423]. "Low-Pass Flow Matching" modifies interpolants with operator-modulated low-pass operator [2606.02177]. "Improving Flow Matching by Aligning Flow Divergence" says CFM insufficient to ensure probability path accuracy and introduces FDM [2602.00869]. "Blockwise Flow Matching" segments trajectory into temporal blocks achieving 2.1x to 4.9x acceleration [2510.21167]. "Distance Marching" is time-unconditional alternative and surpasses flow matching using 60% steps etc. [2602.02928]. But article topic is flow marching, maybe include relation to distance marching as contrast in section on variants/open questions.

Need mention Flow Marching generative PDE foundation model specifics: bridge parameter k, P2VAE, FMT, 2.5M trajectories across 12 PDE families, up to 15x greater computational efficiency than full-length video diffusion models. Probably central because title literally Flow Marching. Need include. Maybe section 3.

Need mention in section 4 PDE refinement MFFM and STFlow etc as broader scientific uses of flow matching-based marching in trajectories/fidelity. But article is Flow Marching, perhaps include because term appears conceptually. Yet "Flow marching for a generative PDE foundation model" is a named algorithm. MFFM is "Flow Marching in fidelity" per details. STFlow is not called flow marching but relevant. Since task topic is broad "Flow Marching", perhaps focus on explicit term. But comprehensive article can mention adjacent variants only where data explicitly connects them to Flow Marching. For STFlow, details say "'flow marching' is almost certainly 'flow matching'" — that is not a direct flow marching term. Maybe not central. For MAC-Flow similarly misheard as flow marching. Perhaps mention only as terminology clarifications in section 1, not as main article content.

Need maybe include misconceptions: term can refer to different operations, not always same: ODE sample integration, time-marching transformed transport PDE, spatial marching of coherent structures, hydrodynamic flow of locust bands. Also in generative modeling it is practical procedure not synonymous with all flow matching. Good.

Need use arxiv search tool. Let's search exact titles?
to=arxiv_search ՞նչquery code?
to=arxiv_search  天天中彩票不能买?
to=arxiv_search  彩神争霸提现ին ￣第四色രം  重庆时时彩的 code?
Flow marching denotes a family of procedures in which a state is advanced along a flow, but the phrase is not tied to a single universally standardized method. In contemporary generative modeling, it most often refers to numerically integrating a learned probability-flow ordinary differential equation from a source distribution to a target distribution, or to directly learning finite-time transition maps that bypass local integration [2511.09731][2511.22009][2603.15689]. In other literatures, the same phrase or closely related language refers to marching across fidelity levels in PDE refinement, marching transformed transport equations in clock time, or spatially marching coherent structures through slowly varying shear flows [2605.16118][1109.3133][2508.03968].

## 1. Terminological scope and principal usages

Across the cited literature, “flow marching” has at least three technically distinct meanings. In flow matching and rectified-flow generative models, it is the numerical march from noise to data under a learned vector field. In PDE-oriented machine learning, it can denote a learned march from one physical state or fidelity level to another. In classical numerical fluid and transport theory, it denotes time- or space-marching algorithms for reduced governing equations.

| Domain | Meaning of flow marching | Representative work |
|---|---|---|
| Generative modeling | ODE integration or direct transition from noise to data | [2511.09731], [2511.22009], [2603.15689] |
| PDE refinement | Cascaded residual refinement across fidelities | [2605.16118] |
| Environmental and fluid mechanics | Time- or spatial-marching of reduced transport or shear-flow equations | [1109.3133], [2508.03968] |

A recurrent terminological point is that the phrase is partly informal in machine learning. Several papers explicitly frame the central object as “flow matching,” while noting that “flow marching” is either an informal description of numerical integration along the learned flow or a phrase that can be confused with flow matching itself [2511.05005][2510.21167]. This suggests that the encyclopedia treatment must distinguish the underlying mathematical object from the particular marching algorithm used to realize it.

## 2. Flow marching in conditional flow matching and rectified-flow models

In the most common machine-learning usage, flow marching is the numerical realization of a learned transport ODE. FlowCast formulates precipitation nowcasting in latent space with a time-dependent vector field
$$
\frac{dZ(t)}{dt} = v_\theta(Z(t), t, Z_{\text{past}}), \quad t\in[0,1],
$$
where a VAE encodes radar frames into latent sequences and the generative model transports a noise latent sequence at \(t=0\) to a forecast latent sequence at \(t=1\) [2511.09731]. FlowCast uses Independent Conditional Flow Matching with Gaussian interpolation
$$
p_t(Z_t\mid Z_0,Z_1)=\mathcal{N}((1-t)Z_0+tZ_1,\sigma^2 I),
$$
target velocity
$$
u_t(Z_0,Z_1)=Z_1-Z_0,
$$
and an MSE loss that regresses the learned velocity field to \(Z_{\text{future}}-Z_P\). At inference, the model performs explicit Euler marching,
$$
Z_{k+1}=Z_k+\Delta t\,v_\theta(Z_k,t_k,Z_{\text{past}}),
$$
with \(K=10\) steps. In direct comparison on the same architecture, CFM with 10 steps achieved CRPS \(=0.0168\) and CSI-M \(=0.455\), while DDIM with 10 steps gave CRPS \(=0.0262\) and CSI-M \(=0.395\), and DDIM with 100 steps still lagged at CRPS \(=0.0208\) and CSI-M \(=0.398\) [2511.09731].

StreamFlow generalizes the same basic notion to high-throughput rectified-flow sampling. It defines flow marching as numerically integrating
$$
\frac{dx_t}{dt}=v_\theta(x_t,t)
$$
from noise to data, then redesigns the inference pipeline around batched velocity-field evaluation, heterogeneous timesteps, and dynamic TensorRT execution. The reported gain is up to \(611\%\) acceleration for \(512\times512\) image generation, with the central operation still being the march of samples through the learned ODE field [2511.22009].

A more probabilistic extension appears in stream-level flow matching with Gaussian processes, where the conditional paths are not determined only by endpoints but by latent stochastic streams \(s\). There the marginal vector field is written as
$$
u_t(x)=\mathbb{E}[\dot{x}_t(s)\mid x_t(s)=x],
$$
and training remains simulation-free because \((x_t,\dot{x}_t)\) are sampled from conditional Gaussian-process laws rather than from the learned ODE itself [2409.20423]. In that setting, flow marching still refers to integrating the learned CNF at test time, but the training geometry is widened by stochastic path ensembles.

## 3. Flow Marching as a named algorithm for PDE foundation models

A more specialized usage is the named algorithm “Flow Marching” for generative PDE foundation models. That work targets conditional transitions
$$
p(\mathbf{x}_{s+1}\mid \mathbf{x}_{0:s}),
$$
and replaces a deterministic neural-operator time-stepper with a learned transport from a noisy current state toward the next clean state [2509.18611]. For consecutive states \((\mathbf{x}_0,\mathbf{x}_1)\), it defines a bridge
$$
\mathbf{x}_t^k = \mu_t + \sigma_t \mathbf{z},\qquad
\mu_t=t\mathbf{x}_1+k(1-t)\mathbf{x}_0,\qquad
\sigma_t=(1-t)(1-k),
$$
with \(k\in[0,1]\). The limiting cases are explicit: \(k=0\) recovers the standard rectified-flow path from noise to data, while \(k=1\) yields deterministic interpolation between adjacent PDE states [2509.18611].

The training target is the frame-interpolation velocity
$$
\mathbf{u}_t^k=\frac{\mathbf{x}_1-\mathbf{x}_t^k}{1-t},
$$
and the practical objective uses the preconditioned form
$$
\mathcal{L}_{\text{FM}}
=
\frac{1}{2}\mathbb{E}\Big[\|(1-t)\mathbf{g}_\theta(\mathbf{x}_t^k,t)-(\mathbf{x}_1-\mathbf{x}_t^k)\|^2\Big].
$$
The paper couples this objective to a Physics-Pretrained Variational Autoencoder and a Flow Marching Transformer with diffusion forcing and latent temporal pyramids. It curates a corpus of approximately \(2.5\)M trajectories across \(12\) PDE families and reports up to \(15\times\) greater computational efficiency than full-length video diffusion models, few-shot adaptation on unseen Kolmogorov turbulence, and improved long-term rollout stability over deterministic counterparts [2509.18611].

In this usage, flow marching is not merely Euler stepping on a static generative ODE. It is a generative time-stepping mechanism for PDE evolution, with explicit control over initial-condition uncertainty through the bridge parameter \(k\) and separate aleatoric sampling through a reverse-time SDE construction [2509.18611].

## 4. From local velocities to global transitions, and other reformulations

A major recent development is the shift from learning local velocity fields to learning global finite-time objects. Transition Flow Matching directly learns a transition flow
$$
X^\theta(x_t,t,r),
$$
which maps a state at time \(t\) to a state at arbitrary future time \(r\), rather than learning only the local velocity \(v(x_t,t)\) [2603.15689]. The paper defines the average velocity
$$
(r-t)\,u(x_t,t,r)=x_{t\to r}-x_t=\int_t^r v(x_\tau,\tau)\,d\tau,
$$
and then the transition flow identity
$$
X(x_t,t,r)=x_{t\to r}+(r-t)\frac{d}{dt}X(x_t,t,r).
$$
This yields single-step generation by \(x_1=X^\theta(x_0,0,1)\) and arbitrary-time marching by \(x_{t_{k+1}}=X^\theta(x_{t_k},t_k,t_{k+1})\) [2603.15689]. A plausible implication is that “flow marching” in generative modeling increasingly denotes not only numerical integration, but also evaluation of learned global flow maps.

Several other works modify the geometry, solver, or decomposition of marching. Blockwise Flow Matching partitions the generative trajectory into temporal segments, assigns a specialized velocity block to each segment, and reports \(2.1\times\) to \(4.9\times\) accelerations in inference complexity on ImageNet \(256\times256\) at comparable generation quality [2510.21167]. Low-Pass Flow Matching replaces scalar noise scaling with an operator-modulated interpolant
$$
x_t = t x_1 + \mathcal{L}_t x_0,
$$
thereby inducing a time-varying spectral bias; on Galaxy10 DECaLS with adaptive solvers, RLP-CFM improved FID from \(11.23\) to \(10.03\) while reducing NFE from \(92.0\) to \(56.0\) relative to CFM/FM [2606.02177].

Other papers refine the theoretical fidelity of the learned path. “Improving Flow Matching by Aligning Flow Divergence” proves that conditional flow matching alone is insufficient to ensure accuracy of the learned probability path and bounds the total-variation gap in terms of a divergence mismatch. Its combined flow-and-divergence matching objective improves CIFAR-10 from NLL \(2.99\) and FID \(6.35\) under FM (OT) to NLL \(2.85\) and FID \(5.62\) under FDM (OT) [2602.00869]. Distance Marching moves in a different direction: it removes explicit time conditioning, learns a distance-like scalar field and denoising direction, and reports lower FID than flow matching’s final performance using \(60\%\) of the sampling steps, together with \(13.6\%\) lower FID on average across backbone sizes on class-conditional ImageNet [2602.02928].

## 5. Marching across fidelity levels and trajectory spaces

The same transport language has been adapted to scientific surrogate modeling in ways that make “marching” refer to progression across resolution or trajectory space rather than solely across diffusion time. Multi-Fidelity Flow Matching models the residual
$$
\delta = u_{\text{HF}} - u_{\text{LF (prolongated)}}
$$
between adjacent discretization levels and learns a conditional residual flow on each level of a grid hierarchy. Its source distribution is calibrated to empirical residual scale and local Gaussian-blur correlation, and after level-wise pretraining it fine-tunes the composed cascade end-to-end with a deterministic one-step rollout. The resulting inference path reaches the finest grid in \(L\) deterministic network evaluations per query, one per cascade level, and the paper describes this as a learned analog of multigrid refinement validated on eight benchmarks [2605.16118].

Flow matching for geometric trajectory simulation uses the same broad philosophy in another form. STFlow models full trajectories \(\mathbf{x}\in\mathbb{R}^{T\times N\times d}\) with conditional flow matching and data-dependent couplings, using physics-informed priors rather than uninformed noise. The prior preserves observed frames and extrapolates the future by a random walk fitted to observed mean velocity and variance. The reported outcome is significantly lower prediction errors and more efficient inference on N-body systems, molecular dynamics, and pedestrian dynamics [2505.18647]. Although that work primarily speaks in the language of flow matching rather than flow marching, it exemplifies the same shift from local denoising to structured transport over physically meaningful trajectory objects.

A related but conceptually separate case is multi-agent coordination. MAC-Flow learns a continuous-time flow over joint actions and then distills it into decentralized one-step policies, achieving about \(\times 14.5\) faster inference than diffusion-based MARL methods while maintaining good performance [2511.05005]. Here the phrase “flow marching” is best understood as a loose analogy for moving through a learned action-space flow, not as a standard term of art.

## 6. Classical numerical and fluid-mechanical meanings

Outside generative modeling, “marching” retains its classical numerical meaning. In environmental transport, the Time-Marching Laplace Transform technique treats water age as an extra coordinate, applies a Laplace transform in that coordinate,
$$
\hat{g}(x,t,s)=\int_0^\infty e^{-s\tau}g(x,t,\tau)\,d\tau,
$$
and then marches the transformed problem forward in physical time using standard time-marching schemes. The transformed equation has the form of a transient advection-dispersion equation with a first-order decay term, so full transient age distributions can be computed without explicitly discretizing age [1109.3133].

In high-Reynolds-number shear-flow theory, spatial marching means downstream propagation of reduced asymptotic equations. The multiple-scale method for non-uniformly curved channel flow introduces slow variables \(X=R^{-1}x\) and \(T=R^{-1}t\), couples Boundary Region Equations for the slowly varying mean flow to local exact coherent structures for fast fluctuations, and then spatially marches the coupled system in \(X\). The framework is consistent with high-Reynolds-number asymptotic theory and is used to estimate momentum transport and the downstream persistence of subcritical self-sustained states [2508.03968].

A looser hydrodynamic use appears in the study of marching locust bands. There, “flow” and “marching” are literal rather than algorithmic: the band is treated as an active fluid described by an effective one-dimensional Toner–Tu equation, and the observed effective pressure increases as a linear function of density in highly polarized segments [2308.02589]. This usage shares the language of flow and marching but not the transport-learning formalism of flow matching.

A final boundary case is Quantum Flow Matching, which extends the flow-matching idea from classical probability densities to quantum density matrices by applying stepwise quantum circuits to ensembles of pure states. The resulting discrete quantum flow transports an initial ensemble approximating \(\rho_0\) to one approximating a target density matrix, enabling tasks such as thermal-state preparation and transport analysis without repeated circuit redesigns [2508.12413]. This suggests that the modern “flow marching” family is broadening from numerical integration of classical ODE transports to operator-level transport in quantum state space.

The accumulated evidence indicates that flow marching is best understood as a family resemblance term. In its dominant modern sense it denotes the realization of learned transports—either by integrating local velocity fields or by evaluating global transition maps. In adjacent scientific literatures it denotes time-marching or spatial-marching of reduced flow equations. What unifies these uses is not a single algorithm, but the idea of advancing a state along a mathematically specified flow, with the precise meaning determined by the governing representation: probability path, fidelity hierarchy, transformed transport equation, or asymptotic fluid model.

Source: https://www.emergentmind.com/topics/flow-marching