MeanFlow-Based One-Step Generative Corrector

• The paper's main contribution is the introduction of MeCo, which employs a one-step generative correction to transport separator outputs directly to the clean speech manifold.
• It integrates Data-Space Optimization that combines interval-level displacement and Endpoint SI-SDR objectives to enhance both perceptual quality and signal fidelity.
• Empirical evaluations demonstrate MeCo's competitive performance in multi-channel speech separation under various conditions with minimal additional computational cost.

Searching arXiv for MeCo and closely related MeanFlow papers to ground the article in current literature. MeanFlow-Based One-Step Generative Corrector (MeCo) is a conditional generative correction framework that uses a MeanFlow parameterization to map an imperfect estimate directly to a target sample in a single update. In its canonical form, MeCo was introduced for multi-channel speech separation, where it refines a discriminative separator output $\hat{\mathbf{s}}$ toward the clean speech manifold $\mathbf{s}$ by learning a conditional average velocity field over a finite time interval rather than an instantaneous velocity field that would require iterative ODE integration (Kim et al., 8 Jun 2026). Within the broader MeanFlow literature, MeCo belongs to a class of one-step generative methods that replace local slope estimation with interval-level displacement prediction, thereby turning generative inference from a multistep numerical solver into a direct transport or correction rule (Geng et al., 19 May 2025). Its distinctive contribution is to cast this one-step transport explicitly as a corrector operating on an existing estimate, and to optimize that corrector with Data-Space Optimization (DSO), which combines a long-interval data-space objective with an endpoint SI-SDR objective (Kim et al., 8 Jun 2026).

1. Conceptual position within the MeanFlow family

MeanFlow reframes flow matching around average velocity rather than instantaneous velocity. In standard flow matching, the learned object is a local field $v(z_t,t)$, and generation requires solving an ODE; in MeanFlow, the learned object is

$$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$

so that a full interval displacement can be recovered directly from one network evaluation (Geng et al., 19 May 2025). This shift is the mathematical basis for one-step generation.

MeCo adopts that MeanFlow construction but changes the operational role of the model. Rather than generating from a simple prior alone, it starts from a structured, already informative estimate and applies a single generative correction. In the multi-channel speech separation setting, the state at $t=1$ is the discriminative separator output $\hat{\mathbf{s}}$, and the endpoint at $t=0$ is clean speech $\mathbf{s}$ (Kim et al., 8 Jun 2026). This places MeCo between purely discriminative enhancement and fully generative reverse-time samplers: it preserves the efficiency of one-step MeanFlow while using the separator output as a strong initialization.

This role as a one-step corrector is consistent with related MeanFlow developments in other domains. MeanFlowSE describes a speech-enhancement instantiation in which the model predicts the average velocity over a finite time interval and performs a single backward displacement rather than iterative integration (Li et al., 18 Sep 2025). RMFlow, in image, molecule, and time-series generation, similarly treats baseline 1-NFE MeanFlow as a coarse transport and adds a lightweight refinement through analytical Gaussian noise injection (Huang et al., 31 Jan 2026). These works suggest that MeCo is not an isolated architecture but a domain-specific realization of a more general design pattern: one learned interval-scale transport step, optionally complemented by a minimal correction mechanism.

2. Mathematical formulation of the MeCo corrector

MeCo is built on the MeanFlow identity in conditional form. In standard conditional flow matching, one defines a path

$$x_t = \mu_t + \sigma_t z,\qquad z\sim\mathcal{N}(\mathbf{0},\mathbf{I}),$$

with ODE

$dx_t = v_t(x_t)\,dt,$

and learns the instantaneous field by minimizing

$\mathbf{s}$0

MeanFlow instead defines

$\mathbf{s}$1

which directly yields the displacement identity

$\mathbf{s}$2

For the full interval $\mathbf{s}$3, one evaluation produces the endpoint: $\mathbf{s}$4 This is the central one-step inference rule inherited by MeCo (Kim et al., 8 Jun 2026).

In MeCo, the conditional path is defined between clean speech and the separator estimate: $\mathbf{s}$5 and

$\mathbf{s}$6

Differentiation gives the analytical instantaneous target velocity

$\mathbf{s}$7

with

$\mathbf{s}$8

MeCo then learns the conditional average velocity field

$\mathbf{s}$9

conditioned on both the multichannel mixture $v(z_t,t)$0 and the discriminative estimate $v(z_t,t)$1 (Kim et al., 8 Jun 2026).

The MeanFlow target is constructed with a first-order correction term: $v(z_t,t)$2 and the corresponding stop-gradient regression objective is

$v(z_t,t)$3

As in the broader MeanFlow literature, the derivative term is instantiated through a Jacobian-vector product, allowing local supervision of finite-interval motion without explicit trajectory integration (Geng et al., 19 May 2025).

3. Data-Space Optimization and the notion of a generative corrector

MeCo’s most specific technical addition to MeanFlow is Data-Space Optimization (DSO). The paper argues that matching average velocity alone is insufficient because errors in the velocity field matter more when they accumulate across long intervals. DSO therefore adds losses defined directly on displaced states and on the final one-step endpoint (Kim et al., 8 Jun 2026).

The first component is the $v(z_t,t)$4-loss. Using

$v(z_t,t)$5

MeCo defines

$v(z_t,t)$6

and minimizes

$v(z_t,t)$7

Because

$v(z_t,t)$8

prediction errors over longer displacement intervals are penalized more strongly. The paper interprets this as a generative objective for human listening quality, because one-step generation corresponds to the regime where $v(z_t,t)$9 (Kim et al., 8 Jun 2026).

The second component is the Endpoint SI-SDR loss. MeCo simulates inference during training by forming

$$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$0

then minimizes

$$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$1

where $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$2 is the optimal scale factor. This term directly targets terminal signal fidelity (Kim et al., 8 Jun 2026).

The resulting objective is

$$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$3

This is the sense in which MeCo is a generative corrector rather than merely a one-step generator. The corrector is not an auxiliary denoiser appended after sampling. It is the transport rule itself, optimized both for correct long-range trajectory displacement and for high-fidelity terminal reconstruction. A plausible implication is that DSO functions as an explicit bridge between generative manifold alignment and application-specific endpoint quality, rather than leaving that trade-off implicit in the MeanFlow identity alone.

4. Inference rule, conditioning, and architectural realization

At inference, MeCo performs exactly one correction step: $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$4 The initial state $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$5 corresponds to the separator output, and the update transports that estimate directly toward the clean speech manifold (Kim et al., 8 Jun 2026). This differs structurally from multistep flow or diffusion samplers, which repeatedly evaluate a neural field inside an ODE or SDE discretization.

The conditioning scheme is dual. The corrector receives the multichannel mixture $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$6, which carries spatial information from the microphones, and the discriminative estimate $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$7, which provides a task-specific coarse solution. In the reported implementation, MeCo operates on complex STFTs and uses channel-wise concatenation of the complex STFTs of $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$8 and $$u(z_t,r,t) \triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$9 (Kim et al., 8 Jun 2026). This design makes the corrector explicitly conditional on both acoustic context and the current estimate being corrected.

The backbone is NCSN++, following prior speech generative correctors, and for MeanFlow-based models the interval length $t=1$0 is embedded with Gaussian Fourier features and an MLP (Kim et al., 8 Jun 2026). The paper emphasizes that MeCo uses the same backbone as the generative baselines, so the performance gains are attributed primarily to the objective design rather than to increased model size.

This one-step displacement interpretation aligns closely with the MeanFlowSE formulation in speech enhancement, where the learned field is used as a direct backward-time displacement operator: $t=1$1 That parallel is important because it clarifies a possible misconception: in MeCo-style systems, the “corrector” is not necessarily a residual postprocessor layered onto a completed sample. It can be the generative update rule itself (Li et al., 18 Sep 2025).

5. Empirical profile in multi-channel speech separation

MeCo was evaluated in both in-domain and out-of-domain settings. The reported datasets are WSJ0 + WHAM! for in-domain evaluation, Librispeech + DEMAND for out-of-domain corpus/noise transfer, and six low-resource languages + DEMAND for out-of-domain language transfer (Kim et al., 8 Jun 2026).

The comparisons include a discriminative baseline, DeFTAN2, and generative correctors including Fast-GeCo and a plain MeanFlow model. The main efficiency claim is that MeCo achieves the best overall balance between reference-based fidelity and reference-free perceptual quality with only 1 additional NFE and a very small RTF increase of 0.0068 (Kim et al., 8 Jun 2026). The paper states that on WSJ0+WHAM! and Librispeech+DEMAND, MeCo improves or matches the best numbers across PESQ, ESTOI, SI-SDR, DNSMOS, UTMOS, and NISQA.

The reported trend relative to baselines is structured rather than uniform. Compared with DeFTAN2, MeCo improves both perceptual quality and fidelity. Compared with Fast-GeCo, MeCo is usually better on human-listening metrics and competitive or better on signal-fidelity metrics. Compared with plain MeanFlow, MeCo is consistently stronger, which the paper treats as evidence that DSO is materially important (Kim et al., 8 Jun 2026).

The ablation study gives the internal logic of the method. MeanFlow alone improves over the discriminative separator. Adding $t=1$2-loss yields a modest gain in both fidelity and perceptual quality. Adding Endpoint SI-SDR gives another gain, especially in fidelity. Using both together gives the best overall result (Kim et al., 8 Jun 2026). This supports the intended division of labor within DSO: $t=1$3-loss acts as a trajectory-level generative regularizer, while Endpoint SI-SDR sharpens the terminal output.

Out-of-domain behavior is emphasized as a major property. On unseen languages, MeCo remains the strongest method overall. The paper argues that this follows from learning the clean-speech distribution rather than a purely deterministic correction mapping (Kim et al., 8 Jun 2026). This suggests that MeCo’s generative prior is functioning not merely as a perceptual polish but as a robustness mechanism under distribution shift.

6. Related variants, extensions, and open technical questions

MeCo sits within a rapidly expanding MeanFlow ecosystem, and several adjacent results clarify both its promise and its limitations. The original MeanFlow paper established that one-step generation can be trained from scratch with no pre-training, distillation, or curriculum learning, and reported FID 3.43 at 1-NFE on ImageNet 256x256 (Geng et al., 19 May 2025). Subsequent work showed that one-step quality depends strongly on training dynamics. In particular, analysis of MeanFlow training found that a well-established instantaneous velocity is a prerequisite for learning large-gap average velocity, and a progressive curriculum improved 1-NFE ImageNet $t=1$4 performance from 3.43 to 2.87 with the same DiT-XL backbone (Kim et al., 24 Nov 2025). For MeCo-like systems, this suggests that one-step correction quality may depend critically on how local and long-range correction regimes are staged during training.

Domain-specific MeanFlow adaptations also sharpen the notion of a corrector. MeanFlowSE in speech enhancement demonstrates that a conditional average-velocity field can replace multistep ODE solvers while maintaining strong intelligibility, fidelity, and perceptual quality at low computational cost (Li et al., 18 Sep 2025). RMFlow, by contrast, starts from the observation that baseline 1-NFE MeanFlow is often too crude for multimodal generation and adds a “free” analytical Gaussian refinement step, together with a likelihood term $t=1$5, to improve 1-NFE quality at essentially unchanged cost (Huang et al., 31 Jan 2026). These results imply two distinct correction paradigms inside the MeanFlow literature: MeCo integrates the correction objective into the learned transport itself, whereas RMFlow applies a stochastic refinement after coarse one-step transport.

There are also structural variations that modify MeanFlow supervision rather than inference. Rectified MeanFlow trains the mean-velocity field on rectified couplings produced by a 1-rectified flow, arguing that standard MeanFlow suffers from noisy supervision on curved trajectories (Zhang et al., 28 Nov 2025). OT-based Mean Flows replace independent pairings with minibatch optimal-transport couplings, improving one-step fidelity by making training trajectories more geometry-consistent (Akbari et al., 26 Sep 2025). On manifolds, Riemannian MeanFlow extends the average-velocity construction to location-dependent tangent spaces and retains 1-NFE generation through an intrinsic identity based on covariant differentiation (Zhong et al., 11 Mar 2026). These developments indicate that MeCo’s core idea is portable, but also that one-step correction depends sensitively on trajectory geometry, target parameterization, and derivative-consistency constraints.

A common misconception is that MeanFlow-based correctors are merely heuristic shortcuts to diffusion-style sampling. The cited literature does not support that characterization. MeanFlow methods derive supervision from an explicit identity linking average and instantaneous velocities, and MeCo preserves that structure while adding objectives targeted at long-range displacement and endpoint fidelity (Geng et al., 19 May 2025). Another misconception is that one-step correction necessarily requires teacher distillation. MeanFlow, MeanFlowSE, and MeCo are all described as self-contained formulations that do not rely on knowledge distillation or external teachers (Geng et al., 19 May 2025). The unresolved issue is not whether one-step correction is principled, but how much residual mismatch remains after compressing a full trajectory into one interval-level prediction. The continuing appearance of refinements, curricula, and geometry-aware coupling schemes suggests that this compression remains the central technical challenge for MeanFlow-based one-step generative correction.