One-Step Generative Transport

• One-Step Generative Transport is a family of methods that directly maps a source distribution to a target distribution in a single transport step, bypassing iterative procedures.
• It leverages techniques such as optimal transport, local conditioning, and finite-interval learning to maintain multimodality and calibration while significantly reducing latency.
• Empirical applications range from image synthesis and motion prediction to Bayesian inverse problems, demonstrating enhanced efficiency and precision.

Searching arXiv for the cited one-step generative transport papers to ground the article in current metadata. arXiv search query: (Bella et al., 28 Apr 2026) OR (Shou, 7 Apr 2026) OR (Han et al., 12 May 2026) OR (Moreno-Muñoz et al., 21 May 2026) OR (Luo et al., 17 Dec 2025) OR (Akbari et al., 26 Sep 2025) OR (Geng et al., 19 May 2025) OR (Huang et al., 19 Jun 2026) OR (Zhou et al., 22 Jun 2026) OR (Li et al., 11 Mar 2026) OR (Shi et al., 21 Jun 2026) OR (Li et al., 12 Mar 2026) OR (Cheng et al., 16 Mar 2026) OR (Lin et al., 20 May 2026) OR (Luo et al., 8 Jun 2026) OR (Ding et al., 2024) One-step generative transport denotes a family of generative methods that replace the tens to hundreds of denoising or ODE evaluations used by diffusion, score, and standard flow-matching models with a single forward pass, a single large transport step, or a directly learned endpoint map. Across the recent literature, the common objective is to transport a simple, approximate, or condition-specific source distribution to a target law without iterative sampling, while retaining multimodality, calibration, and bounded latency. The field now spans direct map learning, average-velocity and solution-map models, local-transport conditioning, constrained posterior transport, and Wasserstein- or OT-guided formulations (Geng et al., 19 May 2025, Shou, 7 Apr 2026, Bella et al., 28 Apr 2026).

1. Conceptual foundations

A recurring diagnosis in this literature is that one-step generation is not intrinsically inaccurate; rather, it fails when the model is asked to solve a global transport problem in one shot. “FlowS: One-Step Motion Prediction via Local Transport Conditioning” argues that “single-step integration is accurate when the underlying transport problem is local,” and attributes one-step failure to the combination of long displacement and multimodal mode discovery from a scene-agnostic base such as $Z\sim\mathcal N(0,I)$ (Bella et al., 28 Apr 2026). This viewpoint shifts the question from whether one step is too coarse to whether the transport geometry has been conditioned into a regime where one large step is faithful.

A complementary diagnosis appears in “ODE-free Neural Flow Matching for One-Step Generative Modeling.” There, the central obstruction is not path length but inconsistent coupling: if source samples and target samples are paired independently, the squared-loss minimizer becomes a regression-to-the-mean solution, and the learned endpoint collapses to $\mathbb E_{p_1}[\mathbf x_1]$ (Shou, 7 Apr 2026). This makes one-step transport a coupling-sensitive problem. The literature therefore converges on two broad principles: one-step transport becomes plausible when the source is already close to the target manifold, and when the transport supervision preserves coherent source–target pairings.

These principles explain why many recent one-step models depart from the classical “Gaussian noise to data” template. Some begin from learned anchors near plausible futures, some begin from the observed corrupted input itself, some restrict transport to a measurement-consistent affine subspace, and some replace white noise with a prior-aligned Gaussian reference. The shared implication is that source design is part of the transport problem, not merely an implementation detail (Li et al., 11 Mar 2026, Shi et al., 21 Jun 2026, Cheng et al., 16 Mar 2026).

2. Mathematical formulations

One major formulation learns the endpoint map directly. In OT-NFM, the flow map is parameterized as

$$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$

so one-step generation is simply $F_\theta(1,\mathbf x_0)$ for $\mathbf x_0\sim p_0$ (Shou, 7 Apr 2026). “Characteristic Learning for Provable One Step Generation” instead learns a characteristic map $\hat g_{0,T}$ derived from the probability transport ODE, so inference becomes $Z_T=\hat g_{0,T}(Z_0)$ in one evaluation (Ding et al., 2024). “Solution Flow Models for One-Step Generative Modeling” pushes the same idea further by learning the bi-time ODE solution map

$f_\theta(x_t,t,s),$

with one-step sampling given by $f_\theta(x_1,1,0)$ (Luo et al., 17 Dec 2025).

A second formulation learns a finite-interval or average velocity rather than an infinitesimal field. “Mean Flows for One-step Generative Modeling” defines

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

so the exact finite-interval update is $\mathbb E_{p_1}[\mathbf x_1]$0, and one-step generation is $\mathbb E_{p_1}[\mathbf x_1]$1 (Geng et al., 19 May 2025). Variants of this pattern appear in AlphaFlowTSE for mixture-to-target speech transport, in OFP for action generation, and in NullFlow for inverse problems on measurement-consistent subspaces (Li et al., 11 Mar 2026, Li et al., 12 Mar 2026, Shi et al., 21 Jun 2026).

A third formulation keeps the flow interpretation but makes one-step sampling feasible by changing the transport geometry. FlowS replaces the base $\mathbb E_{p_1}[\mathbf x_1]$2 with learned scene-conditioned anchors $\mathbb E_{p_1}[\mathbf x_1]$3 and then predicts a semigroup-consistent displacement

$\mathbb E_{p_1}[\mathbf x_1]$4

so the single step acts as local correction rather than global relocation (Bella et al., 28 Apr 2026). NullFlow similarly confines all trajectories to

$\mathbb E_{p_1}[\mathbf x_1]$5

with source samples $\mathbb E_{p_1}[\mathbf x_1]$6; because the velocity lies in $\mathbb E_{p_1}[\mathbf x_1]$7, the flow remains measurement-consistent without a separate data-fidelity step (Shi et al., 21 Jun 2026).

These formulations differ in parameterization, but all replace numerical integration of a local rule by a learned finite-time transport object.

3. Main design patterns

The recent literature repeatedly returns to a small number of design patterns.

Design pattern	Representative papers	Core mechanism
Consistent source–target pairing	(Shou, 7 Apr 2026, Akbari et al., 26 Sep 2025)	Replace independent couplings with OT-based pairings
Finite-interval transport learning	(Geng et al., 19 May 2025, Luo et al., 17 Dec 2025)	Learn average velocity or solution map directly
Local transport conditioning	(Bella et al., 28 Apr 2026, Sun et al., 13 May 2026)	Start near plausible targets so one step is local refinement
Structural subspace restriction	(Shi et al., 21 Jun 2026, Cheng et al., 16 Mar 2026)	Choose source/reference on the correct feasible geometry
Self-consistency across scales or intervals	(Bella et al., 28 Apr 2026, Li et al., 11 Mar 2026, Li et al., 12 Mar 2026)	Enforce large-step coherence with smaller-step or teacher targets

The first pattern is optimal transport or coherent coupling. OT-NFM proves that direct map regression under the product coupling leads to mean collapse, and fixes it with optimal transport pairings, minibatch OT, or LOOM (Shou, 7 Apr 2026). OT-MF makes the same point inside the Mean Flow framework: by replacing independent pairing with OT-based coupling, one-step generators preserve the fidelity and diversity of the original multi-step flow more faithfully (Akbari et al., 26 Sep 2025).

The second pattern is learning the right transport quantity. MeanFlow argues that one-step generation should target the finite-interval quantity actually needed at inference rather than the instantaneous tangent field (Geng et al., 19 May 2025). SoFlow learns the ODE solution function itself and avoids JVP-based consistency constraints (Luo et al., 17 Dec 2025). OFP transfers this principle to robot action generation by learning interval-averaged velocity over an OT action path, then enforcing self-consistency and self-guided sharpening (Li et al., 12 Mar 2026).

The third pattern is conditioning transport to be local or straight. FlowS formalizes local transport conditioning by requiring learned anchors $\mathbb E_{p_1}[\mathbf x_1]$8 to lie much closer to the ground-truth future than a Gaussian sample does, so one-step Euler error is directly suppressed (Bella et al., 28 Apr 2026). DirectTryOn makes a parallel argument for virtual try-on: because conditional entropy is low once person and garment are fixed, the ideal transport should be close to straight, and one-step sampling and multi-step sampling are equivalent when the condition uniquely determines the target (Sun et al., 13 May 2026).

The fourth pattern is source design as preconditioning. In function-space Bayesian inversion, the reference cannot be naive white noise in the infinite-dimensional limit; “Preconditioned One-Step Generative Modeling for Bayesian Inverse Problems in Function Spaces” replaces it with a prior-aligned anisotropic Gaussian reference and proves Lipschitz regularity of the resulting one-step transport (Cheng et al., 16 Mar 2026). For path-dependent physical fields, a geometry-based non-Gaussian empirical source reduces crossings among conditional transport paths and makes one-step Euler sampling viable without distillation (Zhou et al., 22 Jun 2026). NullFlow makes the same move algebraically by randomizing only in the null space of the forward operator (Shi et al., 21 Jun 2026).

A broader theoretical unification appears in “Self-Consistent Generative Paths via Admissible Random Variational Transport,” which treats one-step, shortcut, and MeanFlow models as “large-step degeneracies of path self-consistency” and evaluates them by a random fixed-point path residual rather than endpoint matching alone (Luo et al., 8 Jun 2026).

4. Domain-specific instantiations

In autonomous motion prediction, one-step transport is driven by latency constraints. FlowS uses a learned prior that emits $\mathbb E_{p_1}[\mathbf x_1]$9 calibrated anchor trajectories per agent and a step-consistent displacement field. On the Waymo Open Motion Dataset, the ensemble model achieves Soft mAP $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$0 and mAP $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$1 at about 75 FPS, while the single model reaches Soft mAP $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$2 and mAP $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$3 with about 13.25 ms/scene on a single A100 (Bella et al., 28 Apr 2026).

In virtual try-on, one-step transport exploits strong conditional structure. DirectTryOn argues that VTON outputs are highly constrained by person and garment inputs, so the transport path can be straightened by pure conditional transport, garment preservation loss, self-consistency, and then one-step distillation. Under the unpaired setting, it reports VITON-HD FID $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$4, KID $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$5, and DressCode FID $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$6, KID $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$7, with $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$8 s end-to-end inference on a PPU-810E (Sun et al., 13 May 2026).

In image compression and reconstruction, one-step transport often appears as a local latent correction. FlowCodec decodes a bitrate-constrained latent $$\mathbf x_t = F_\theta(t,\mathbf x_0), \qquad F_\theta(0,\mathbf x_0)=\mathbf x_0,$$9, then applies one near-terminal update

$F_\theta(1,\mathbf x_0)$0

using a pretrained MMDiT prior, with trainable parameters below $F_\theta(1,\mathbf x_0)$1 of the generative backbone (Huang et al., 19 Jun 2026). NullFlow instead turns reconstruction into posterior transport restricted to $F_\theta(1,\mathbf x_0)$2, so the learned flow never leaves the measurement-consistent subspace and needs no separate data-fidelity correction (Shi et al., 21 Jun 2026).

In speech and robotics, the source is often the observed signal or a temporally correlated warm start rather than Gaussian noise. AlphaFlowTSE transports directly from the observed mixture spectrum $F_\theta(1,\mathbf x_0)$3 to the target speech $F_\theta(1,\mathbf x_0)$4, eliminating the need for mixture-ratio prediction and yielding one-step inference $F_\theta(1,\mathbf x_0)$5 (Li et al., 11 Mar 2026). OFP uses interval-averaged action transport, self-consistency, self-guidance, and warm start from the suffix of the previous action chunk to minimize transport distance in control space (Li et al., 12 Mar 2026).

In scientific computing, several papers adapt one-step transport to structured physical laws. One-Step Flow Matching for path-dependent stress fields uses a geometry-informed empirical source and a latent spatiotemporal transformer to generate full $F_\theta(1,\mathbf x_0)$6-frame $F_\theta(1,\mathbf x_0)$7 stress-field sequences in one ODE step (Zhou et al., 22 Jun 2026). “Two-Step Diffusion” uses a different compromise: Stage I is a Meanflow-style one-step global transport, and Stage II is a near-identity corrector trained with a mini-batch $F_\theta(1,\mathbf x_0)$8 objective once the geometry has become local enough for stable OT (Shen et al., 27 Jan 2026). This suggests that even when strict one-shot transport is insufficient, a global-then-local decomposition can retain one-step efficiency at the dominant stage.

5. Empirical landscape

On ImageNet $F_\theta(1,\mathbf x_0)$9, the strongest reported one-step number in the provided literature is W-Flow’s FID $\mathbf x_0\sim p_0$0, with throughput on one H100 GPU of 77.88 images/sec for W-Flow-XL/2 and an abstract-level claim of approximately $\mathbf x_0\sim p_0$1 faster sampling than multi-step diffusion models with similar FID scores (Han et al., 12 May 2026). MeanFlow reports FID $\mathbf x_0\sim p_0$2 at 1-NFE for ImageNet $\mathbf x_0\sim p_0$3 trained from scratch, while SoFlow improves the matched DiT comparison further, reaching FID $\mathbf x_0\sim p_0$4 for SoFlow-XL/2 at 1-NFE and $\mathbf x_0\sim p_0$5 at 2-NFE (Geng et al., 19 May 2025, Luo et al., 17 Dec 2025).

On synthetic transport and small-scale image generation, OT-NFM shows that one-step map learning can be competitive when coupling is coherent: on Gaussian $\mathbf x_0\sim p_0$6 2-Moons it reports $\mathbf x_0\sim p_0$7 at 1 NFE, outperforming 100-NFE OT-CFM on that task, and it qualitatively avoids the mean-collapse seen without OT pairings (Shou, 7 Apr 2026). OT-MF similarly improves one-step Mean Flows across 2D distributions, MNIST, image-to-image translation, and point clouds by replacing independent couplings with OT-based ones (Akbari et al., 26 Sep 2025).

On reconstruction tasks, NullFlow reaches LPIPS $\mathbf x_0\sim p_0$8 with NFE $\mathbf x_0\sim p_0$9 on FFHQ inpainting, compared with LPIPS $\hat g_{0,T}$0 for PnP-Flow at 500 NFE and LPIPS $\hat g_{0,T}$1 for DiffPIR at 200 NFE, while still producing posterior samples from a single network evaluation (Shi et al., 21 Jun 2026).

On robot manipulation, OFP demonstrates that one-step transport can outperform multi-step action generation. Across 56 simulated manipulation tasks, one-step OFP reaches $\hat g_{0,T}$2 average success, versus $\hat g_{0,T}$3 for DP3 at 100 steps and $\hat g_{0,T}$4 for a 100-step flow policy, while reducing action chunk generation time to $\hat g_{0,T}$5 ms from $\hat g_{0,T}$6 ms and $\hat g_{0,T}$7 ms, respectively (Li et al., 12 Mar 2026).

On Bayesian inverse problems in function spaces, the operator-learning one-step sampler with a prior-aligned reference produces a $\hat g_{0,T}$8 posterior sample in $\hat g_{0,T}$9 s/sample and matches posterior mean and variance summaries within roughly $Z_T=\hat g_{0,T}(Z_0)$0 relative $Z_T=\hat g_{0,T}(Z_0)$1 error across Darcy, Advection, Reaction–Diffusion, and Navier–Stokes examples (Cheng et al., 16 Mar 2026).

These results do not show a single dominant mechanism; rather, they show that one-step transport is empirically strongest when the task structure makes transport short, straight, conditionally low-entropy, or geometrically preconditioned.

6. Limitations, controversies, and open directions

A first limitation is that not all one-step claims are equally strict. “Generative Modeling by Value-Driven Transport” explicitly says that exact one-step generation is possible “in principle” because optimal OT paths are straight, but also states that the level of precision needed for such one-step generation is hard to achieve in practice; its strongest empirical result is efficient few-step transport, especially reducing $Z_T=\hat g_{0,T}(Z_0)$2 training steps to about $Z_T=\hat g_{0,T}(Z_0)$3 sampling steps with little loss (Moreno-Muñoz et al., 21 May 2026). This distinguishes true one-NFE endpoint-map methods from methods whose theory admits one-step sampling but whose evidence mainly supports few-step use.

A second limitation is that source and coupling choice are often indispensable rather than optional. OT-NFM proves that independent pairing induces mean collapse in direct map learning (Shou, 7 Apr 2026). The function-space Bayesian paper shows that white-noise references become incompatible with the infinite-dimensional limit and that prior-aligned anisotropic references are needed for mesh-stable transport (Cheng et al., 16 Mar 2026). This suggests that one-step transport is unusually sensitive to the geometry of its reference law.

A third limitation is that computational savings at inference may shift difficulty into training. Exact or approximate OT couplings can be expensive (Shou, 7 Apr 2026, Akbari et al., 26 Sep 2025). Two-Step Diffusion makes this explicit by postponing explicit $Z_T=\hat g_{0,T}(Z_0)$4 optimization until after a Meanflow initializer has simplified the geometry (Shen et al., 27 Jan 2026). The broader implication is that one-step transport often relies on a separation between an easy-to-sample learned map and a harder offline procedure that shapes its geometry.

A fourth issue is diagnostics. Endpoint quality can hide path inconsistency, teacher dependence, or compression artifacts. The SCGP framework formalizes this by bounding one-step generation quality with a residual term $Z_T=\hat g_{0,T}(Z_0)$5, teacher quality, distillation error, and proxy error, thereby reframing one-step evaluation as terminal path self-consistency testing rather than endpoint matching alone (Luo et al., 8 Jun 2026).

Finally, recent theory suggests a more favorable mathematical picture than early pessimism implied. For PDE-induced target measures, “On the Regularity and Generalization of One-Step Wasserstein-guided Generative Models for PDE-Induced Measures” proves doubling conditions and Hölder continuity of the optimal transport map from a uniform source to the target, and derives excess-risk and robustness estimates for one-step Wasserstein-guided learners such as DeepParticle (Lin et al., 20 May 2026). A plausible implication is that one-step generative transport is most principled when the population transport map itself is regular, the source is geometrically compatible, and the learned correction is tested against an admissible refinement.

In that sense, the contemporary literature no longer treats one-step generative transport as merely an acceleration trick. It is increasingly framed as a problem of choosing the correct transport object, the correct source law, and the correct notion of self-consistency so that a single finite-time map can stand in for an otherwise long generative path.