One-Step Flow Generation

• One-step flow generation is a generative modeling approach that compresses iterative diffusion processes into a single neural network evaluation.
• It leverages advances in flow matching and diffusion theory to bypass multi-step integration while preserving diversity and controllability.
• Empirical results in image synthesis, robotics, speech, and scientific simulations show significant speedups with competitive fidelity, despite trade-offs in extreme cases.

One-step flow generation is a class of generative modeling methodologies that compresses the entire iterative flow, diffusion, or transport process into a single neural network evaluation at inference. Building on the theoretical and empirical advances in flow matching and diffusion, one-step approaches seek to eliminate the computational bottleneck of multi-step numerical integration while maintaining the sample quality, diversity, and controllability characteristic of state-of-the-art generative models. This paradigm encompasses continuous and discrete data domains, providing principled frameworks, algorithmic techniques, and application-specific adaptations across vision, control, speech, and scientific data.

1. Mathematical Foundations and Key Objectives

Conventional flow-matching models [Lipman et al., 2023] define a time-dependent ODE or SDE that transports samples from a simple prior (usually Gaussian noise) to the data distribution through a velocity field $v_t(x)$: $$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$ Training is formulated as a regression to either (1) an instantaneous or (2) a marginal velocity target, with Monte Carlo samples drawn along an interpolation path between data and noise.

One-step flow generation bypasses iterative ODE integration by learning either:

• the mean/average velocity field over the entire flow trajectory, as in MeanFlow (Geng et al., 19 May 2025), or
• a direct mapping (generator) from noise to data, by learning the composition of the entire flow in a single function approximation (e.g., ODE-free approaches (Shou, 7 Apr 2026), probabilistic flow generator matching (Huang et al., 2024)), or
• a solution map of the velocity ODE (solution flow models (Luo et al., 17 Dec 2025)).

This leads to single-step inversion formulas, typically of the form: $x_0 = z_1 - u_\theta(z_1, 0, 1),$ where $z_1\sim p_1$ is noise, and $u_\theta$ is the learned global/average velocity field (Geng et al., 19 May 2025, Chen et al., 2 Mar 2026).

The theoretical underpinning is the MeanFlow identity: $$u(z_t, r, t) = v(z_t, t) - (t - r) d\,u(z_t, r, t),$$ enabling regression to targets involving both instantaneous velocities and their total derivatives.

2. Methodological Variants and Loss Formulations

Method	Loss Target	Model Purpose
MeanFlow	Average velocity $u(z_t,r,t)$	1-NFE synthesis
Flow Generator Match	Explicit-implicit FM surrogate	Probabilistic 1-step
Rectified MeanFlow	MeanFlow on straightened paths	Robust 1-step
Solution Flow/SoFlow	Solution map $f(x_t,t,s)$	ODE-free inverse
OT-MeanFlow	Mean velocity via OT coupling	High-D fidelity
SnapFlow	Self-distilled FM + consistency	Action generation
OFP (One-Step Flow Policy)	Interval-averaged velocity (control)	Robot policy

Loss functions are typically composed of regression (MSE or robust MSE) between the learned object and a function of the teacher's output, Jacobian-vector products (for consistency with flow derivatives), and, in distillation-based schemes, explicit consistency or shortcut losses (Luan et al., 7 Apr 2026, Huang et al., 2024).

OT-MeanFlow incorporates optimal-transport batch pairings to mitigate high-dimensional mode collapse and improve alignment of one-step displacements with sample geometry (Akbari et al., 26 Sep 2025, Shou, 7 Apr 2026).

3. Algorithms and Training Strategies

One-step models are trained either:

• from scratch, by directly fitting the average velocity or global mapping to ground-truth sampled flows (Geng et al., 19 May 2025, Luo et al., 17 Dec 2025), or
• via (self-)distillation, where an existing multi-step flow-matching model is used to define targets under shortcut/consistency or explicit endpoint mapping (Luan et al., 7 Apr 2026, Huang et al., 2024, Zhang et al., 28 Nov 2025).

Protocols include:

• Sampling coupled data-noise pairs, and randomly selected pairs of time indices $(r,t)$,
• Regression to average or shortcut velocities using the teacher's predictions at multiple time points,
• JVPs (Jacobian-vector products), required for MeanFlow/Rectified MeanFlow training (Geng et al., 19 May 2025, Zhang et al., 28 Nov 2025),
• ODE-free approaches learning direct maps with optimal transport pairings (Shou, 7 Apr 2026),
• Use of masking or conditional tokens for variable-length / structured data (e.g., video (Oladokun et al., 14 Mar 2026), discrete state kernels (Khan et al., 12 May 2026)).

A high-level pseudocode for MeanFlow training: $$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$9 SnapFlow and OFP employ online self-distillation and shortcut consistency with no external teacher (Luan et al., 7 Apr 2026, Li et al., 12 Mar 2026).

4. Empirical Results and Domain-Specific Applications

Published empirical results demonstrate that one-step flow generation matches or exceeds the performance of multi-step flow-matching models in numerous domains:

• Image Synthesis (ImageNet, CIFAR-10): FID scores of 3.43 (MeanFlow-XL/2) (Geng et al., 19 May 2025), 2.87 (Rectified MeanFlow, 64$\times$64) (Zhang et al., 28 Nov 2025), and 3.08 for Flow Generator Matching on CIFAR-10 (Huang et al., 2024), substantially closing the gap with strong 50–100-step flows.
• Vision-Language-Action (VLA) and Robotics: SnapFlow achieves 98.75% average closed-loop success, 9.6$$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$0 denoising speedup over baseline 10-step policies (Luan et al., 7 Apr 2026); One-Step Flow Policy (OFP) provides $$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$1 acceleration and surpasses 100-step diffusion policies in average task success (Li et al., 12 Mar 2026); Mean-Flow based One-Step VLA shows 8.7$$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$2 (SmolVLA) to 83.9$$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$3 (Diffusion Policy) speedups on real robot benchmarks (Chen et al., 2 Mar 2026).
• Speech and Audio: MeanFlow-TSE outperforms multi-step TSE models in speaker extraction, with an SI-SDR of 18.80 dB (clean), $$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$4 faster than diffusion models (Shimizu et al., 21 Dec 2025); DSFlow achieves high naturalness MOS at one step in TTS, with a reduced parameter footprint (Lin et al., 3 Feb 2026).
• Scientific Domains: Cardiac Mesh Flow enables anatomically-coherent heart mesh synthesis over cardiac cycles from a single pass (Ma et al., 3 May 2026); EchoLVFM demonstrates one-step echocardiogram video generation with explicit EF control at $$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$5 speedup (Oladokun et al., 14 Mar 2026); one-step physical field generators achieve $$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$6–$$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$7 acceleration over FEM in path-dependent simulations (Zhou et al., 22 Jun 2026).
• Discrete Generative Modeling: Discrete MeanFlow develops exact one-step finite-state generators parameterized by transition kernels satisfying discrete MeanFlow identities (Khan et al., 12 May 2026).

5. Trade-offs, Limitations, and Theoretical Guarantees

The principal benefit of one-step flow generation is the drastic reduction in sampling latency (up to 100$$\frac{d}{dt}x(t) = v_t(x(t)),\quad x(1)\sim q_1,\quad x(0)\sim q_0.$$8 speedup), enabling real-time or edge deployment in computationally-constrained environments.

Key trade-offs and limitations include:

• Slight fidelity loss in highly nonlinear or multimodal settings when compressing too aggressively to one step (notable in tasks with large transport curvature, stacking in robotics (Chen et al., 2 Mar 2026), or video reconstruction sharpness (Oladokun et al., 14 Mar 2026)).
• Robustness to rare modes: vanilla one-step models risk mode collapse; SubFlow introduces sub-mode conditioning to restore full mode coverage at negligible FID cost (Lin et al., 14 Apr 2026).
• Computational burden in training: certain algorithms (e.g., OT-MeanFlow) require batchwise OT solvers (cubic in batch size), but acceleration strategies are available (Akbari et al., 26 Sep 2025).

Theoretical results establish gradient equivalence between MeanFlow/FGM losses and the original flow-matching divergence (Huang et al., 2024), guarantee recovery of target distributions at zero loss, and demonstrate non-asymptotic convergence rates for simulation-free generators (Ding et al., 2024). Discrete MeanFlow kernels are shown to recover CTMC transition laws to high precision (Khan et al., 12 May 2026).

6. Architectural and Implementation Principles

One-step flow generation methods leverage modular neural architectures:

• Transformer backbones dominate in vision and vision-language settings (Geng et al., 19 May 2025, Luan et al., 7 Apr 2026, Chen et al., 2 Mar 2026, Li et al., 12 Mar 2026).
• U-Net and DiT backbones are common for image and audio, adapted with time interval or flow tokens (Luo et al., 17 Dec 2025, Akbari et al., 26 Sep 2025, Lin et al., 3 Feb 2026).
• Conditional embeddings (class, task, phenotypic variables, clinical markers) enable controllable synthesis (Ma et al., 3 May 2026, Shimizu et al., 21 Dec 2025).
• Step-aware tokens and efficient parameterizations support compact, high-speed models suitable for resource-constrained inference (Lin et al., 3 Feb 2026, Zhu et al., 2024).
• Architectures are frequently plug-compatible with classifier-free guidance and self-distillation mechanics (Huang et al., 2024, Luan et al., 7 Apr 2026).

Notably, methods such as SnapFlow, DSFlow, and OFP achieve their improvements without requiring architectural changes to baseline models, instead relying on training loss reshaping or plug-in regularizers for shortcut or consistency (Luan et al., 7 Apr 2026, Lin et al., 3 Feb 2026, Li et al., 12 Mar 2026).

7. Extensions, Limitations, and Future Research Directions

Current research extends one-step flow generation to numerous modalities (images, 3D point clouds (Akbari et al., 26 Sep 2025), meshes (Ma et al., 3 May 2026), video (Oladokun et al., 14 Mar 2026), audio (Shimizu et al., 21 Dec 2025), speech (Lin et al., 3 Feb 2026)), and structural forms (conditional, variable-length, discrete finite-state (Khan et al., 12 May 2026)). Emerging focus areas include:

• Adaptive multi-step refinement: recovering some local curvature with 2–4 steps for hard cases without reverting to high-NFE sampling (Zhang et al., 28 Nov 2025).
• Robustness: sub-mode conditioning (SubFlow (Lin et al., 14 Apr 2026)) and geometry-informed couplings (OT-MeanFlow (Akbari et al., 26 Sep 2025, Zhou et al., 22 Jun 2026)) address failure modes in diversity and path curvature.
• The integration of adversarial and perceptual losses to sharpen outputs in cross-space or latent-to-pixel models (Wang et al., 18 Jun 2026).
• Scaling to larger, more realistic domains (e.g., full-resolution, multi-modal text-to-image pipelines (Huang et al., 2024)).
• Addressing theoretical limits: further analysis of error decompositions, convergence rates, and uniqueness under various architectural and coupling constraints (Ding et al., 2024).

Recent works continue to advance the practical and theoretical limits of flow-based one-step generation, establishing it as a cornerstone paradigm for high-efficiency, high-fidelity generative modeling across disciplines.