Rectified Flow: Theory & Applications

• Rectified flow is a deterministic ODE-based generative framework that transports probability distributions along straight paths while preserving marginal properties.
• It employs regression on conditional mean displacements to achieve fast, high-fidelity sampling in image, speech, and time-series applications.
• Model variants like reflow reduce integration steps, offering state-of-the-art performance with significant speedups compared to diffusion models.

Rectified Flow

Rectified flow is a transport-based generative framework that parameterizes the evolution between two probability distributions as a deterministic ordinary differential equation (ODE) following straight trajectories in sample space. Originally formulated in the context of optimal transport and generative modeling, rectified flow has become foundational in a variety of machine learning applications, including fast image and speech generation, plug-and-play priors for 3D/2D synthesis, protein backbone design, efficient time-series forecasting, and brain anomaly correction.

1. Mathematical Formalism and Theoretical Properties

Let $\pi_0$ and $\pi_1$ be source and target distributions on $\mathbb{R}^d$. Rectified flow models a path between these via a time-indexed interpolation

$$X_t = (1-t) X_0 + t X_1, \qquad t \in [0, 1],\quad X_0\sim \pi_0,\, X_1\sim \pi_1$$

and seeks a velocity field $v_t(x)$ such that the deterministic ODE

$\frac{dZ_t}{dt} = v_t(Z_t),\qquad Z_0 = X_0$

transports $Z_0$ to $Z_1$, matching the marginal $\pi_1$ at $t=1$ (Liu et al., 2022, Bansal et al., 19 Oct 2024).

The "rectification" step computes, at each time $t$, the minimizer

$v_t(z) = \mathbb{E}[X_1 - X_0 \mid X_t = z],$

i.e., the conditional mean displacement given the interpolated point. This is equivalent to a least-squares regression and always preserves the marginal distributions at each $t$. The ODE paths $\{Z_t\}$ are strictly non-crossing if the coupling and the velocity field are well-specified, yielding a so-called "straight" flow between the distributions (Bansal et al., 19 Oct 2024).

Key structural results include:

• Marginal preservation: For all $t$, the law of $Z_t$ matches the law of $X_t$.
• Monotonic cost reduction: Each rectification step never increases convex transport costs, and recursive rectification iteratively straightens paths and reduces the expected cost (Liu, 2022).
• Existence and regularity: Under mild Carathéodory conditions or for strongly log-concave densities, solutions exist, are unique, and are regular, with improved convergence rates versus standard nonparametric regression (Mena et al., 5 Nov 2025).
• Wasserstein convergence: For straight flows and a discretization grid of size $N$, the error in Wasserstein-2 distance after Euler discretization is $O(1/N)$ (Bansal et al., 19 Oct 2024).
• Limiting case: In dimension 1, with standard Gaussian base distributions, rectified flow recovers the Monge map for quadratic cost optimally (Bansal et al., 19 Oct 2024).

2. Algorithmic Implementation and Reflow

Rectified flow is typically implemented as a neural ODE model. Training proceeds as follows:

• Draw $(X_0, X_1)$ from independent or model-generated couplings.
• Sample $t \sim \mathrm{Uniform}[0, 1]$, and interpolate $X_t = (1-t)X_0 + tX_1$.
• Minimize

  $$\mathcal{L}(\theta) = \mathbb{E}\left[\,\|X_1 - X_0 - v_\theta(X_t, t)\|^2\,\right]$$

  where $v_\theta$ is parameterized as a neural network.

• At inference, an initial sample from $\pi_0$ is evolved by time discretization (e.g., Euler integration with $N$ steps) via

  $x_{k+1} = x_k + \Delta t\, v_\theta(x_k, t_k)$

  where $\Delta t = 1/N$, and $t_k = k/N$ (Liu et al., 2022, Yang et al., 5 Jun 2024).

The "reflow" process refers to using generated pairings from an initial or previous rectified flow as new couplings, further straightening the paths and reducing the number of integration steps (NFEs) needed for high-fidelity samples. Reflow, often iterated once or twice, allows for deterministic sampling with as few as one or a handful of steps (Liu et al., 2022, Seong et al., 29 Oct 2025).

3. Rectified Flow in Generative Modeling

Rectified flow plays a prominent role in generative modeling, offering enhanced sample efficiency, inference speed, and deterministic ODE transport compared to diffusion models. Key applications and results include:

• Image synthesis: Large-scale rectified flow and progressive multi-resolution models achieve state-of-the-art FID and CLIP scores with $1$–$10$ steps, outperforming or matching diffusion models at orders-of-magnitude lower inference cost (Ma et al., 12 Mar 2025, Yang et al., 5 Jun 2024, Zhu et al., 1 Jun 2024).
• Plug-and-play priors: Rectified flow can serve as a plug-and-play loss, analogously to diffusion’s SDS/VSD, enabling text-to-3D synthesis and image inversion/editing with fewer network calls and superior convergence. The time-symmetry allows for exact latent inversion via ODE reversal (Yang et al., 5 Jun 2024).
• Speech synthesis: In text-to-speech, rectified flow achieves superior MOS and objective metrics over diffusion with $2$–$10$ steps, confirmed by ablations showing rectification enables higher fidelity in extremely few steps (Guo et al., 2023).
• Protein design: Application to SE(3) protein backbone flows requires domain-specific adaptions; rectified flow can cut NFEs $3$–$5\times$ while boosting designability and matched diversity at low steps, provided annealing and loss schedules are handled carefully (Chen et al., 13 Oct 2025).
• Brain anomaly correction and enhancement: In anomaly correction, rectified flows enable one-step unsupervised correction maps and efficient anomaly localization, substantially outperforming diffusion-based UAD baselines (Beizaee et al., 4 Aug 2025). In low-light RAW enhancement, physics-guided rectified flow enables per-pixel calibration and two-stage deterministic correction (Zeng, 10 Sep 2025).
• Time-series and V2I beam prediction: By modeling discrete index sequences as continuous ODE flows with rectified terminal constraints, rectified flow provides ultra-fast beam selection with smooth latent evolution and top-K accuracy exceeding RNN/LSTM baselines (Zheng et al., 25 Nov 2025).

4. Theoretical Considerations and Limitations

Relation to Optimal Transport (OT)

Rectified flow has deep ties to optimal transport: at each $t$, it solves a convex quadratic cost problem under a fixed marginal constraint. For some special cases (e.g., independent Gaussians with commutative covariances), a single rectification coincides with the $2$-Wasserstein OT map (Hertrich et al., 26 May 2025, Bansal et al., 19 Oct 2024). However, in general, rectified flow only approximates the OT solution, and convergence to the true Monge map fails unless strong support, regularity, and rectifiability conditions are satisfied. Counterexamples show that zero flow-matching loss under a gradient constraint does not guarantee optimality; non-optimal fixed points exist with disconnected supports or singular flows (Hertrich et al., 26 May 2025).

Statistical Properties

Rectified flow maps can be consistently estimated via regression or density-ratio estimation. For smooth, log-concave densities, rates approach those of one-dimensional regression rather than $d$-dimensional ones due to the structure of ODE integration. Local uniform rates and central limit theorems for the estimated flow map are available under both compact and noncompact support (Mena et al., 5 Nov 2025).

Failure Modes

Deterministic (noise-free) training can induce "memorization" of arbitrary training pairings when interpolant lines intersect, confirmed by both Gaussian and real-image (CelebA) experiments. The gradient variance is minimized for such memorizing vector fields, while noise injection breaks this effect and promotes generalization to entropic OT (Reu et al., 20 Oct 2025). Noise in interpolant sampling is thus essential for preventing ill-defined solutions.

Boundary Condition Violation

Standard neural parameterizations may violate theoretical boundary conditions $v(x,0)=C-x$ and $v(x,1)=x$, leading to error amplification and bias in stochastic sampling. Enforcing these boundaries via minimal code changes (mask-based or subtraction-based parameterizations) yields 8–9% FID improvements on ImageNet and stabilizes both ODE and SDE sampling (Hu et al., 18 Jun 2025).

5. Model Variants and Extensions

Progressive and Multi-Resolution Flows

Partitioning both time and image resolution into coarse-to-fine stages combined with spatial cascading of transformer modules reduces computation in early stages and preserves capacity for fine details, enabling 40% speedup over vanilla rectified-flow transformer baselines in high-resolution synthesis (Ma et al., 12 Mar 2025).

Balanced and Hybrid Reflow

Empirical evidence demonstrates that standard k-reflow may drift from the target data manifold and saturate on generated samples. Alternating real data pairs and synthetic pairs, leveraging conic Slerp perturbations for local supervision, maintains alignment with the real data distribution and achieves superior FID/IS with substantially fewer generative pairs (Seong et al., 29 Oct 2025).

Conditioned and Domain-Agnostic Flows

Rectified flows are easily conditioned on external information for domain adaptation and upstream tasks, e.g., vision-aided beam prediction in V2I communication, physics-based noise models in RAW enhancement, or attention-based representations in disentangled semantic editing (Zheng et al., 25 Nov 2025, Dalva et al., 12 Dec 2024, Zeng, 10 Sep 2025).

6. Practical Impact and Empirical Highlights

• One-step or few-step sampling: Rectified flow models routinely provide high-fidelity samples with as few as 1–10 steps (vs. 50–1000 in diffusion), enabling $10\times$–$20\times$ speedups in inference across domains (Liu et al., 2022, Armegioiu et al., 3 Jun 2025, Ma et al., 12 Mar 2025, Yang et al., 5 Jun 2024).
• Sample quality: On CIFAR-10 and ImageNet, rectified flow and related derivatives achieve FID and IS competitive with (or better than) best diffusion or consistency models matched at similar or even lower function evaluations (Yang et al., 5 Jun 2024, Ma et al., 12 Mar 2025, Seong et al., 29 Oct 2025).
• User preference and editing: In image editing, attention-based rectified flow transformers achieve high text-image alignment and user satisfaction scores, with disentangled style and content control (Dalva et al., 12 Dec 2024).
• Domain coverage: Demonstrated success ranges from high-dimensional generative modeling (images, speech), low-dimensional OT/Gaussian transport, to domain-adaptive learning and structured temporal/sequential forecasting.

7. Comparison with Related Frameworks

Rectified flow forms part of the broader family of continuous normalizing flows and flow-matching methods. Its distinguishing feature is the regression-based "straight-line" transport objective and deterministic ODE sampling, in contrast to stochastic SDE-based diffusion models. It admits reinterpretation as a special flow-matching form, and "rectified diffusion" extends its principle beyond straight paths, showing that first-order ODE accuracy, not straightness per se, is the essential property for efficient generation (Wang et al., 9 Oct 2024). Despite its efficiency and tractability, standard rectified flow is not in general a black-box solver for OT maps; high-dimensional generative modeling remains its primary domain of empirical success.

---

The rectified flow framework thus occupies a central position at the intersection of generative modeling, optimal transport, and efficient neural ODEs, offering a theoretically grounded and computationally efficient alternative to score-based and normalizing flow methods. Its algorithmic simplicity—regression against path velocities, deterministic ODE integration, and marginal preservation—renders it widely applicable for both research and production tasks requiring fast, high-fidelity generative modeling and structured distributional transport.