On the Convergence and Straightness of Rectified Flow

Alessandro Rinaldo; Purnamrita Sarkar; Saptarshi Roy; Vansh Bansal

arxiv: 2410.14949 · v7 · pith:M5T3Y5W3new · submitted 2024-10-19 · 💻 cs.LG · stat.ML

On the Convergence and Straightness of Rectified Flow

Vansh Bansal , Saptarshi Roy , Purnamrita Sarkar , Alessandro Rinaldo This is my paper

Pith reviewed 2026-05-23 18:20 UTC · model grok-4.3

classification 💻 cs.LG stat.ML

keywords rectified flowflow matchingwasserstein convergencediscretization errorstraight trajectoriesgenerative modelspiecewise straightnessone-step sampling

0 comments

The pith

Rectified flows achieve exact one-step sampling when rectification produces straight trajectories, as a new bound ties all discretization error to the piecewise straightness parameter γ_{2,T}.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces the piecewise straightness parameter γ_{2,T} that quantifies how much a flow's trajectories bend away from straight lines. It proves the first Wasserstein convergence bound on discretization error for any flow model in terms of this parameter, showing that curvature directly controls sampling accuracy. Under sufficient conditions that depend on problem geometry, a single rectification step produces a perfectly straight or Monge-optimal coupling. When those conditions hold, the next rectified flow has γ_{2,T} exactly equal to zero, removing discretization error entirely and enabling flawless single-step generation.

Core claim

We establish the first Wasserstein convergence bound that explicitly links the discretization error of any general flow-model to γ_{2,T}, proving that minimizing curvature is the key to achieving high-fidelity, one-step sampling. We identify sufficient conditions under which a single rectification step yields a perfectly straight or Monge optimal coupling, making the subsequent flow perfectly straight with γ_{2,T}=0 and eliminating discretization error.

What carries the argument

The piecewise straightness parameter γ_{2,T}, which measures deviation from straight trajectories across a flow and directly bounds its Wasserstein discretization error.

If this is right

When γ_{2,T} reaches zero the discretization error bound collapses to zero, permitting exact one-step sampling.
The convergence bound holds for any general flow model, not only rectified flows.
A single rectification can produce a Monge-optimal coupling under the identified conditions.
Straightening trajectories through rectification is the mechanism that reduces the number of required sampling steps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Training losses could be augmented to penalize γ_{2,T} explicitly rather than relying solely on flow matching.
The same straightness analysis could diagnose sampling cost in score-based diffusion models by defining an analogous curvature measure.
In practice the geometric conditions may fail on complex data, suggesting a need for adaptive numbers of rectification steps.
The bound offers a diagnostic tool to predict how many steps a trained flow model will need before training finishes.

Load-bearing premise

The sufficient conditions for a single rectification to produce a perfectly straight coupling must hold for the given data and noise distributions.

What would settle it

On a distribution pair satisfying the stated geometric conditions, such as linear transport between two Gaussians, check whether the two-rectified flow produces exactly zero one-step sampling error in Wasserstein distance.

read the original abstract

Flow Matching has become a cornerstone of modern generative models like Stable Diffusion 3, largely due to the efficiency of its Rectified Flow (RF) variant. The success of RF hinges on iteratively learning straight trajectories, pushing generation towards fewer sampling steps. However, the theoretical link between path geometry and sampling efficiency has been underexplored. This paper fills this gap by introducing a novel \textit{Piecewise Straightness} parameter, $\gamma_{2,T}$. We establish the first Wasserstein convergence bound that explicitly links the discretization error of \textit{any} general flow-model to $\gamma_{2,T}$, proving that minimizing curvature is the key to achieving high-fidelity, one-step sampling. Building on this theory, we establish the first theoretical framework to analyze the straightness of RF. We begin by offering intuitive geometric arguments for simple cases before identifying sufficient conditions under which a single rectification step (1-RF) yields a perfectly straight or even a Monge optimal coupling. While whether these sufficient conditions are met depends on the problem geometry, they enable the first concrete proofs in this area. Critically, fulfilling these conditions makes the subsequent flow (2-RF) perfectly straight ($\gamma_{2,T}=0$). This eliminates the discretization error in our bound and makes flawless, single-step sampling possible.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a new straightness parameter γ_{2,T} and an explicit Wasserstein bound tying it to discretization error, but the route to zero-error one-step sampling rests on geometry-dependent conditions whose reach is unclear.

read the letter

The new parameter and the bound that links curvature to sampling error are the actual additions here. They formalize something people have been saying informally about rectified flows, and the derivation looks like a reasonable way to quantify how much straightness buys you in terms of fewer steps. The geometric arguments for the simple cases are straightforward and help motivate the sufficient conditions for a single rectification step to produce a straight or Monge coupling. That part is useful for anyone trying to understand why RF works better than plain flow matching in practice. The bound itself applies to general flow models, which is broader than prior work on this specific family. The main limitation is that the conditions needed for γ_{2,T} to hit zero after one rectification are stated to depend on problem geometry, and nothing in the abstract or claims shows they hold for the kinds of measures that appear in image or latent-space models. If those conditions fail, the bound remains valid but the advertised flawless one-step regime does not follow. The paper does not appear to contain checks or counter-examples on real data distributions, so the practical scope stays conditional. Readers working on theoretical analysis of flow-based generative models will find the framework worth reading; people mainly interested in immediate training tricks will get less out of it. The work is coherent on its own terms and engages the literature directly enough that it deserves a serious referee to check the proofs and see whether the conditions can be made more explicit or verified.

Referee Report

2 major / 0 minor

Summary. The paper introduces a Piecewise Straightness parameter γ_{2,T} and establishes the first Wasserstein convergence bound explicitly linking the discretization error of any general flow-model to γ_{2,T}, proving that minimizing curvature is key to high-fidelity one-step sampling. It provides a theoretical framework for RF straightness, with geometric arguments and sufficient conditions (dependent on problem geometry) under which 1-RF yields a perfectly straight or Monge optimal coupling, making 2-RF perfectly straight (γ_{2,T}=0) and eliminating discretization error.

Significance. If the bound derivation is correct and the sufficient conditions hold for typical data distributions, the result would supply the first explicit theoretical connection between path geometry (via γ_{2,T}) and sampling efficiency in flow models, offering a principled explanation for the empirical success of rectified flow in few-step generation. The parameter γ_{2,T} itself is a useful quantifiable measure of relevant geometry, but the geometry-dependent conditions limit the scope of the one-step regime claim.

major comments (2)

[Abstract] Abstract, second paragraph: the claim that 1-RF produces a perfectly straight (or Monge) coupling under sufficient conditions that 'depend on the problem geometry' is load-bearing for the assertion that 2-RF has γ_{2,T}=0 and enables flawless single-step sampling; no verification or counter-example analysis is supplied for high-dimensional image or latent distributions, leaving the advertised regime conditional rather than general.
[Abstract] Abstract, first paragraph: the Wasserstein bound is asserted to apply to 'any general flow-model' and to prove that minimizing curvature is the key to one-step sampling, but the derivation of the bound from γ_{2,T} and the proof steps cannot be checked; this is the central claim and requires explicit verification that the bound does not reduce to a self-referential or fitted quantity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on the manuscript. We respond point by point to the major comments below.

read point-by-point responses

Referee: [Abstract] Abstract, second paragraph: the claim that 1-RF produces a perfectly straight (or Monge) coupling under sufficient conditions that 'depend on the problem geometry' is load-bearing for the assertion that 2-RF has γ_{2,T}=0 and enables flawless single-step sampling; no verification or counter-example analysis is supplied for high-dimensional image or latent distributions, leaving the advertised regime conditional rather than general.

Authors: The abstract already states that the sufficient conditions depend on the problem geometry, so the one-step regime is presented as conditional rather than universal. The contribution consists of geometric arguments for simple cases together with the identification of these sufficient conditions that enable the first concrete proofs of perfect straightness for 2-RF. No empirical verification on high-dimensional image or latent distributions is provided because the work is primarily theoretical; the framework itself is intended to support such analysis when the geometry of a given data distribution satisfies the conditions. revision: no
Referee: [Abstract] Abstract, first paragraph: the Wasserstein bound is asserted to apply to 'any general flow-model' and to prove that minimizing curvature is the key to one-step sampling, but the derivation of the bound from γ_{2,T} and the proof steps cannot be checked; this is the central claim and requires explicit verification that the bound does not reduce to a self-referential or fitted quantity.

Authors: The bound is obtained by expressing the discretization error of an arbitrary flow as a function of the integrated deviation from straightness measured by γ_{2,T}. Because γ_{2,T} is defined directly from the path geometry (independent of the sampling procedure), the resulting inequality is not self-referential; it recovers the standard Wasserstein error for piecewise-linear paths when γ_{2,T}=0 and shows that any positive curvature inflates the error. The derivation therefore supplies an explicit, non-circular link between path curvature and sampling error that holds for general flow models. revision: no

Circularity Check

0 steps flagged

No circularity; new parameter and bound derived independently

full rationale

The paper introduces the novel parameter γ_{2,T} to measure piecewise straightness and derives a Wasserstein convergence bound that links discretization error to this quantity for general flow models. It then states sufficient (geometry-dependent) conditions under which rectification yields γ_{2,T}=0. The provided text shows no reduction of the bound or central claim to a self-definition, fitted input renamed as prediction, or load-bearing self-citation chain. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the newly introduced measure γ_{2,T} and standard assumptions from optimal transport theory; no free parameters fitted to data are mentioned in the abstract.

axioms (2)

standard math Properties of Wasserstein distance and optimal transport couplings
Invoked to establish the convergence bound linking discretization error to straightness.
domain assumption Existence and properties of flow models and rectification procedures
Assumed for the geometric arguments and analysis of RF straightness.

invented entities (1)

Piecewise Straightness parameter γ_{2,T} no independent evidence
purpose: Quantifies the curvature of trajectories in a piecewise manner to bound discretization error
Newly defined in the paper to connect path geometry to sampling efficiency; no independent evidence provided.

pith-pipeline@v0.9.0 · 5772 in / 1510 out tokens · 41469 ms · 2026-05-23T18:20:25.663336+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

We establish the first Wasserstein convergence bound that explicitly links the discretization error of any general flow-model to γ_{2,T}, proving that minimizing curvature is the key to achieving high-fidelity, one-step sampling.
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

under some regularity conditions ... 1-RF yields a straight coupling ... 2-RF produces a straight flow ... 1-RF yields the Monge map

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Tessellations of Semi-Discrete Flow Matching
cs.LG 2026-05 unverdicted novelty 7.0

Semi-discrete Flow Matching produces terminal assignment regions that are topologically simple (open, simply connected, homeomorphic to the ball under assumption) yet geometrically distinct from optimal transport Lagu...