MeanFlow Parametrizations

Updated 24 June 2026

MeanFlow parametrizations are techniques that replace instantaneous dynamics with interval-averaged quantities to enhance stability and enable efficient one-step or few-step sampling across diverse domains.
They employ neural network architectures with stop-gradient training, specialized time embeddings, and consistent mathematical identities to ensure reliability in Euclidean, manifold, and discrete settings.
Empirical studies demonstrate that MeanFlow methods reduce inference steps and computation time while achieving competitive performance on benchmarks in generative modeling and control tasks.

MeanFlow parametrizations define the mechanism by which deep generative models, flow-based policies, or control laws use interval-averaged dynamical objects—such as average velocities or conditional kernels—rather than instantaneous ones. This approach is designed to enable scalable, stable, and extremely efficient one-step or few-step sampling and control across a wide spectrum of domains, including diffusion generative models, sampled-data control, Riemannian manifolds, discrete state spaces, multimodal data, and reinforcement learning. The precise method of parameterization, training, and deployment is domain-dependent but shares unifying mathematical structures and consistency identities that enforce correctness and sample efficiency across these contexts.

1. Mathematical Foundations of MeanFlow Parametrization

MeanFlow parametrizations replace the modeling of instantaneous velocity fields $v(z_t, t)$ (or analogous quantities, e.g., policies, transition rates) with the prediction of an interval-averaged quantity over $[r, t]$ . In the continuous Euclidean case, the average velocity is defined as

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

where $z_\tau$ is the system state following the underlying dynamics from time $r$ to $t$ .

This averaged object satisfies a differential identity—termed the MeanFlow identity:

$u(z_t, r, t) = v(z_t, t) - (t-r) \Big[ v(z_t, t) \cdot \nabla_z u(z_t, r, t) + \partial_t u(z_t, r, t) \Big]$

or, equivalently,

$u(z_t, r, t) = v(z_t, t) - (t-r) \frac{d}{dt}u(z_t, r, t)$

where the total derivative accounts for both explicit and pathwise variations. This identity underpins the correctness and unambiguous regression target for neural parametrizations (Geng et al., 19 May 2025, Kim et al., 24 Nov 2025).

On Riemannian manifolds, the interval-averaged velocity is defined intrinsically via parallel transport and geodesic log maps. On discrete spaces, the averaged object becomes a mean rate or transition kernel, satisfying an identity derived from the Kolmogorov forward equation rather than a spatial chain rule (Khan et al., 12 May 2026, Woo et al., 8 Feb 2026, Zhong et al., 11 Mar 2026).

2. Neural Network Parameterizations

2.1. Continuous Euclidean Generative Models

For imaging or general vector data, the interval-averaged velocity $u_\theta(z_t, r, t)$ is parameterized by neural networks (U-Nets, Vision Transformers, DiT backbones) that take as input the noisy/intermediate data $z_t$ and time variables $[r, t]$ 0, typically embedded via sinusoidal or MLP-based modules. The same network (or head) is often shared for both instantaneous and averaged velocity (with $[r, t]$ 1), enforcing the identity at $[r, t]$ 2 and maintaining consistency (Geng et al., 19 May 2025, Kim et al., 24 Nov 2025, Hu et al., 17 Nov 2025, Kim et al., 24 Nov 2025).

For conditional and multimodal tasks, relevant context or conditioning is concatenated or cross-attended into the network (Kim et al., 8 Jun 2026, Dai et al., 16 Jun 2026).

2.2. Sampled-Data Control

In sampled-data control of linear systems, the key object is a finite-horizon minimum-energy control coefficient $[r, t]$ 3, which encapsulates the necessary effect of inputs over an interval $[r, t]$ 4. The coefficient admits both analytic and integral representations as well as a local differential identity:

$[r, t]$ 5

where $[r, t]$ 6 is the controllability Gramian and $[r, t]$ 7 the state-transition matrix. $[r, t]$ 8 is recast as a neural network and trained to satisfy this identity via a stop-gradient regression, making all updates model-consistent (Dong et al., 20 Mar 2026).

2.3. Manifolds and Discrete State Spaces

On Riemannian manifolds, MeanFlow-like quantities are predicted either in tangent spaces ("v-prediction"), via endpoint prediction with log maps, or via direct $[r, t]$ 9 point-prediction parameterizations, often with stabilization by loss reweighting or careful time embeddings (Woo et al., 8 Feb 2026, Zhong et al., 11 Mar 2026).

In the discrete (finite-state) setting, the key object becomes the observed kernel or transition matrix $u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 0, parametrized via a "boundary-by-construction" mixture-of-delta form:

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

with $u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 2 a softmax neural output, so that kernel correctness and boundary validity are guaranteed. The mean rate is defined and regressed according to the discrete MeanFlow identity, using conditional transition probabilities as the target (Khan et al., 12 May 2026, Xu et al., 9 Apr 2026).

3. Training Techniques and Objectives

The prevailing objective is regression against a stop-gradient target derived from the MeanFlow identity, typically:

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

or discrete analogs using kernels (Geng et al., 19 May 2025, Hu et al., 17 Nov 2025, Dong et al., 20 Mar 2026, Khan et al., 12 May 2026).

Several stability and performance advancements have been developed:

Curriculum/Progressive Weighting: Early emphasis on short interval (small $u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 4) velocity regression to accelerate instantaneous velocity formation, then progressive increase in supervision for large-gap average velocities (Kim et al., 24 Nov 2025).
Stop-Gradient and Warmup: To prevent instability early in training, MeanFlow targets are sometimes approximated discretely (via finite differences or ODE endpoint evaluations) and only switched to the differential version after partial convergence (He et al., 18 May 2026, Khan et al., 12 May 2026).
Gradient Modulation: Interpolation between stop-gradient and full-backprop for higher-order terms is used to balance stability and expressivity (You et al., 24 Aug 2025). Warmup schedules for the modulation parameter further facilitate smooth convergence.
Objective Decomposition: For geometric and Riemannian settings, multi-term losses (centered on both "Eulerian" and "Lagrangian" or "semigroup" identities) are optimized jointly, sometimes with conflict-aware updates such as PCGrad to reduce gradient interference (Woo et al., 8 Feb 2026, Zhong et al., 11 Mar 2026).
Auxiliary Discriminators/Distribution Alignment: In multimodal or distillation tasks, adversarial or TDA losses ensure endpoint distributions are plausible, mitigating mean-seeking bias under few-step regimes (He et al., 18 May 2026).

4. Sampling, Inference, and Implementation

Once the MeanFlow network or kernel is trained, sampling is universally non-iterative: starting from noise (or initial state), a single evaluation of the MeanFlow object suffices to traverse the interval $u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 5, e.g.:

$u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 6

or, for control systems,

$u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 7

eliminating the need for ODE solvers.

In discrete MeanFlow, sampling reduces to one categorical draw using the output transition kernel $u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 8 (Khan et al., 12 May 2026).

In some deployments, minor refinement steps (e.g., noise injection, fine-tuned decoder passes) may follow the main step to adjust for minor distributional mismatches or domain shifts, but do not fundamentally alter the one-step paradigm (Dai et al., 16 Jun 2026, Huang et al., 31 Jan 2026).

Implementation details emphasize normalized, stabilized attention modules, low-frequency Fourier time embeddings, careful handling of conditional information, and—in geometry—coordinate or tangent-space representations, with training batch sizes and network widths scaled per data modality and task (Woo et al., 8 Feb 2026, Hu et al., 17 Nov 2025, Geng et al., 19 May 2025).

5. Extensions and Unified Frameworks

MeanFlow parametrizations extend beyond first-order velocity fields:

Second-Order MeanFlows: Incorporate average acceleration and admit higher-order analogs of additive and differential identities, implemented by additional transformer heads and allowing even fewer sampling steps (Cao et al., 9 Aug 2025).
Riemannian and Structured Spaces: Unified framework supporting Euclidean, Riemannian, and product-structure (e.g., graph, molecule) spaces, with synchronized MeanFlow steps over mixed continuous–discrete domains (Woo et al., 8 Feb 2026, Zhong et al., 11 Mar 2026, Xu et al., 9 Apr 2026).
Generative Control and RL: One-step MeanFlow branches with stochastic perturbations (e.g., Gaussian noise for entropy), entropic mirror descent, and auxiliary advantage-weighted regression in RL, supporting exploration and multimodal action distributions in policy optimization (Wang et al., 20 May 2026, Kim et al., 22 May 2026, Fang et al., 22 Dec 2025).
Distillation and Latent Space: Parameterizations in semantically rich or compressed latent spaces (e.g., RAEs, SD-VAEs) are paired with specialized architectural and training strategies (e.g., mid-training with flow-matching teachers, trajectory-aware initialization) to address the instability and sample complexity of direct image-based models (Hu et al., 17 Nov 2025, Kim et al., 24 Nov 2025).

6. Empirical Validation and Impact

Empirical studies confirm that MeanFlow parametrizations—across generative models, control, RL, multimodal and structured domains—consistently yield:

Substantially reduced inference steps and wall-clock inference time (true one-step or few-step generation).
Model-consistent evolution, with exact dynamical/actuation structure respected at each step.
Robustness to discretization and improved extrapolation/generalization, especially under coarse time grids or limited supervision.
Statistically superior performance in standard benchmarks (ImageNet, MoleculeNet, MuJoCo, Meta-World, text-to-image, multimodal audio-visual tasks) compared to instantaneous velocity-based or multi-step baselines (Geng et al., 19 May 2025, Hu et al., 17 Nov 2025, Kim et al., 24 Nov 2025, Dong et al., 20 Mar 2026, Woo et al., 8 Feb 2026, Zhong et al., 11 Mar 2026, He et al., 18 May 2026, Wang et al., 20 May 2026, Kim et al., 22 May 2026, Fang et al., 22 Dec 2025).

MeanFlow models have demonstrated notable empirical gains: improved FID for one-step image generation (down to 2.03 on ImageNet 256×256 in latent space), state-of-the-art returns in RL with just a single network evaluation, and physically consistent, equivariant generation of joint discrete–continuous molecular graphs (Hu et al., 17 Nov 2025, Kim et al., 24 Nov 2025, Dong et al., 20 Mar 2026, Wang et al., 20 May 2026, Xu et al., 9 Apr 2026).

7. Trade-offs, Limitations, and Future Directions

While MeanFlow parametrizations represent a unifying framework connecting integration-free generative modeling and control, open challenges and nuanced trade-offs remain:

Stability vs. Expressivity: Full gradient propagation in higher-order MeanFlow, or large-interval losses, can induce instability; hybrid schedules and curriculum, as well as finite-difference approximations for Jacobian–vector products, are necessary but may slightly trade off ultimate expressiveness (You et al., 24 Aug 2025, Fang et al., 22 Dec 2025, Cao et al., 9 Aug 2025).
Multimodality and Mean-Seeking Bias: In extremely sparse or multimodal settings, MeanFlow may collapse to “mean-seeking” velocities; trajectory distribution alignment or adversarial objectives are required to restore sharpness (He et al., 18 May 2026).
Refinement and Domain Shift: When composing MeanFlow-generated latents with pretrained decoders, refinement strategies may be required to mitigate mismatch (Dai et al., 16 Jun 2026).
Empirical Risk Minimization: Generalization in low-data/OOD or high-dimensional regimes benefits from explicit loss weighting, dispersive/consistency regularizers, or equivariant backbones (Hu et al., 17 Nov 2025, You et al., 24 Aug 2025, Xu et al., 9 Apr 2026).
Theoretical Expressivity: Expressivity in second-order and modular frameworks has been rigorously established (e.g., $u(z_t, r, t) = \frac{1}{t - r} \int_r^t v(z_\tau, \tau) \, d\tau$ 9 computability for transformers), but implementation bottlenecks (attention scaling, normalization) must be rigorously controlled for provable guarantees (Cao et al., 9 Aug 2025).

Continued work is expected to elaborate the theoretical understanding of MeanFlow in stochastic, hybrid (e.g., Markov decision process), and highly-structured domains, to develop more stable and scalable solvers and loss schedulers, and to unify further the stochastic, discrete, and continuous aspects into single networks and training objectives.

Cited Works (by arXiv ID):

(Geng et al., 19 May 2025, Hu et al., 17 Nov 2025, Kim et al., 24 Nov 2025, Dong et al., 20 Mar 2026, Woo et al., 8 Feb 2026, Zhong et al., 11 Mar 2026, Cao et al., 9 Aug 2025, You et al., 24 Aug 2025, He et al., 18 May 2026, Wang et al., 20 May 2026, Kim et al., 22 May 2026, Fang et al., 22 Dec 2025, Dai et al., 16 Jun 2026, Huang et al., 31 Jan 2026, Khan et al., 12 May 2026, Xu et al., 9 Apr 2026, Yang et al., 8 Sep 2025, Kim et al., 8 Jun 2026)