Rectified Flows in Generative Modeling

• Rectified Flows are ODE-based generative models that create nearly straight transport paths between source and target distributions via explicit regression.
• They enable efficient generative modeling with one- or few-step Euler integration while ensuring theoretical guarantees on convergence and transport cost reduction.
• Variants such as Balanced Conic Rectified Flow, Cumulative Velocity Fields, and MixFlow extend the approach to applications like image synthesis, restoration, and missing-data imputation.

Rectified Flows constitute a class of ODE-based generative models that construct nearly straight transport paths between two probability distributions. This framework leverages explicit regression to match straight-line displacements, enabling efficient simulation of optimal transport-like trajectories for generative modeling, domain translation, and restoration. Rectified Flows offer significant improvements in computational efficiency, often permitting single- or few-step Euler sampling while maintaining high data fidelity and providing theoretical guarantees on convergence and regularity.

1. Mathematical Formulation and Fundamental Principles

Given two distributions, $\pi_0$ (source) and $\pi_1$ (target), a Rectified Flow defines a family of interpolants $X_t = (1-t)X_0 + tX_1$, where $(X_0, X_1)$ is a coupling of $\pi_0$ and $\pi_1$, and $t \in [0,1]$. The velocity field $v(x, t)$ governing the ODE

$\frac{dX_t}{dt} = v(X_t, t)$

is trained to approximate the conditional expectation

$v^*(x, t) = \mathbb{E}[X_1 - X_0 \mid X_t = x].$

The prototypical learning objective is a mean-squared error regression,

$\pi_1$0

with $\pi_1$1 parameterized by a neural network (Liu et al., 2022).

Marginal preservation ensures that if $\pi_1$2 and $\pi_1$3 is generated via the learned ODE, then $\pi_1$4 for all $\pi_1$5. This construction yields a deterministic flow map from $\pi_1$6 to $\pi_1$7, with all straightness and mass-conservation properties inherited from the interpolation.

Key theoretical properties include:

• Monotonic reduction in any convex transport cost along the learned coupling trajectory, guaranteeing descent in, e.g., Wasserstein-$\pi_1$8 distance (Liu et al., 2022, Liu, 2022).
• Existence and global uniqueness of the flow under mild regularity: for strongly log-concave or uniformly continuous densities, the ODE $\pi_1$9 admits a unique solution, and $X_t = (1-t)X_0 + tX_1$0 is a diffeomorphism pushing $X_t = (1-t)X_0 + tX_1$1 onto $X_t = (1-t)X_0 + tX_1$2 (Mena et al., 5 Nov 2025).
• In one dimension, the flow exactly recovers the Monge map for quadratic cost, i.e., the $X_t = (1-t)X_0 + tX_1$3-optimal coupling (Bansal et al., 2024).

2. Algorithmic Workflow and Practical Implementation

Rectified Flow implementation follows a supervised regression paradigm:

1. Sample Couplings: Draw $X_t = (1-t)X_0 + tX_1$4 (or a more informative coupling).
2. Linear Interpolants: For each $X_t = (1-t)X_0 + tX_1$5, form $X_t = (1-t)X_0 + tX_1$6.
3. Regression Target: The difference $X_t = (1-t)X_0 + tX_1$7 becomes the target for $X_t = (1-t)X_0 + tX_1$8.
4. Training: Minimize $X_t = (1-t)X_0 + tX_1$9 with respect to $(X_0, X_1)$0 by stochastic gradient descent.
5. Inference: At generation time, integrate $(X_0, X_1)$1, typically with a coarse discretization (single or few Euler steps), to obtain $(X_0, X_1)$2 as the generated or restored sample (Liu et al., 2022, Nakashima et al., 2024, Iizuka et al., 20 Mar 2026).

Reflow Iteration: The rectification process can be iterated—after an initial flow is learned, new couplings $(X_0, X_1)$3 are generated by integrating the flow, and retraining $(X_0, X_1)$4 on these straightened pairs further straightens the ODE trajectories. Convergence results indicate $(X_0, X_1)$5 straightness and transport-cost decrease after $(X_0, X_1)$6 reflow steps (Liu et al., 2022).

Divergence Suppression: To avoid coupling distortion from divergent vector fields, an offline correction step attenuates the local divergence during coupling generation. This yields empirically straighter paths and improved generative quality, with all suppression cost paid offline (Min et al., 18 May 2026).

3. Variants and Architectural Extensions

Balanced Conic Rectified Flow (BCRF) (Seong et al., 29 Oct 2025): Addresses drift and mode collapse inherent in naive (purely self-generated) reflow by mixing supervision from real images with conic-perturbed neighborhoods during rectification steps. Alternating between standard synthetic pair reflow and conic real-pair reflow, BCRF yields straighter paths, superior FID scores (from 12.21 to 5.98 in one-step on CIFAR-10), and drastically reduces the synthetic data requirement.

Cumulative Velocity Fields (Fan et al., 21 Apr 2026): Cumulative velocity $(X_0, X_1)$7 points directly from the current state to the clean (restored) target, improving energy efficiency by a factor of 3 under the Benamou–Brenier action and yielding smoother ODE trajectories, faster convergence, and more accurate few-step sampling.

MixFlow (Nayal et al., 10 Apr 2026): Introduces mixtures of unconditional and data-conditioned sources to “straighten” generative paths, reducing curvature and sampling steps required for high-fidelity generation. By linearly interpolating Gaussian sources conditioned on task-relevant signals ($(X_0, X_1)$8-FC), MixFlow achieves up to 22% lower trajectory curvature and 12% lower FID scores.

Imputation and Inpainting (Yu et al., 16 May 2025): Rectified Flows can be extended to missing-data imputation by learning flows that minimize the mutual information between imputed data and missingness masks, effectively sampling from the conditional law and outperforming prior GAN- and round-robin-based imputers in MMD, FID, and PSNR.

Discriminative–Generative Bridging (Fan et al., 21 Apr 2026): Hybrid frameworks such as IR-Flow unify discriminative (single-step) and generative (progressive) restoration by multi-level flows and multi-step consistency terms, directly interpolating between degraded and clean images. Empirically, this enables SOTA deraining and denoising with only 1–4 function evaluations.

4. Theoretical Guarantees and Statistical Properties

Optimal Transport Approximation and Uniqueness:

Rectified Flow recovers the Monge optimal map in one dimension and converges to the unique OT plan under commutativity and regularity in higher dimensions. However, when the learned velocity field is simply the conditional mean, the method produces a multi-objective convex descent for generic costs, not necessarily the $(X_0, X_1)$9-optimal plan for a specific $\pi_0$0 (Liu et al., 2022, Liu, 2022, Bansal et al., 2024, Hertrich et al., 26 May 2025). Imposing a gradient constraint can yield the optimal transport under stringent conditions, but disconnectedness and nonrectifiability cause breakdowns (Hertrich et al., 26 May 2025).

Sample Complexity:

Rectified Flows trained by squared loss along linear paths achieve order-optimal sample complexity $\pi_0$1—matching the minimax rate for mean regression and improving over the $\pi_0$2 rate of unconstrained flow matching (Sahoo et al., 28 Jan 2026).

Statistical Estimation Theory:

Nonparametric regression or density-based plug-in estimators for the velocity achieve statistical rates that can surpass classical nonparametric transport rates, especially under strong log-concavity or Hölder regularity. In dimension $\pi_0$3, kernel regression-based $\pi_0$4 for the rectified map attains MSE $\pi_0$5 (Mena et al., 5 Nov 2025).

Failure Modes:

Deterministic training without stochastic perturbation of interpolants can induce memorization of specific training pairings, even in the presence of intersecting straight lines in data space, leading to poor generalization. Noise injection during training breaks such deterministic bijection and restores fidelity to the true optimal transport (Reu et al., 20 Oct 2025).

5. Applications Across Modalities

Rectified Flow has demonstrated practical effectiveness in numerous domains:

• Image Generation: High-resolution synthesis with single-step or few-step ODE inference; outperforms diffusion and GANs in data efficiency and sample quality (Liu et al., 2022, Liu et al., 2024).
• Image-to-Image Translation and Restoration: Plug-in reformulations (I2I-RFR) enable existing discriminative backbones to gain continuous-time refinement with minimal architectural changes, improving both perceptual quality and LPIPS, SSIM, and PSNR metrics (Iizuka et al., 20 Mar 2026).
• Audio, Protein, and Scientific Simulation: Audio generation (e.g., FlashAudio) leverages rectified flows with strategic time-reweighting and pairing; multiscale fluid flow modeling achieves accurate posterior and spectra preservation using only 4–8 ODE steps (Liu et al., 2024, Armegioiu et al., 3 Jun 2025). In protein backbone design, rectified flows accelerate low-NFE inference by tuning coupling and annealing strategies (Chen et al., 13 Oct 2025).
• Medical Imaging and Anomaly MAP Correction: One-step rectified flows yield fast correction maps for anomaly detection and localization in brain MRIs, with sharper segmentation and higher test efficiency than diffusion-based UAD baselines (Beizaee et al., 4 Aug 2025).
• Sensor Data Generation: Efficient, Transformer-based rectified flows for unconditional LiDAR scan generation match quality benchmarks with two-step inference and tensorized representations (Nakashima et al., 2024).
• Missing Data Imputation: Rectified Flows minimizing mutual information produce sharper and less biased inpainted images and features than GANs and diffusion inpainting (Yu et al., 16 May 2025).

6. Empirical Performance and Limitations

Empirical highlights across domains include:

• One- to few-step inference typically reaches or matches state-of-the-art distortion and perceptual metrics, with significant speedup: e.g., 10–22x in scientific simulation (Armegioiu et al., 3 Jun 2025), 400x in audio (Liu et al., 2024), and low-NFE SOTA in image restoration (Fan et al., 21 Apr 2026).
• Methods such as Balanced Conic Rectified Flow reduce synthetic data requirements by an order of magnitude and improve stability.
• Robustness to out-of-distribution (OOD) inputs is enhanced via linear, parametric interpolation and energy-efficient (cumulative) velocity learning.
• Practical limitations include noise schedule tuning, sensitivity to coupling choice (especially in protein design (Chen et al., 13 Oct 2025)), and the possibility of degenerate convergence under inadequate noise injection or disconnected support in the data (Reu et al., 20 Oct 2025, Hertrich et al., 26 May 2025).
• Future research directions center on optimizing coupling selection and divergence correction, exploring non-Gaussian source laws, improving theoretical convergence rates for high-dimensional data, and integrating rectified flows with diffusion or score-based frameworks for hybrid generative modeling (Min et al., 18 May 2026, Nayal et al., 10 Apr 2026).

7. Theoretical and Practical Context

Rectified Flows stand at the intersection of optimal transport, continuous-time generative modeling, and ODE-based sample generation. They distinctively differ from traditional diffusion by constructing deterministic trajectories with explicit straightness guarantees and by requiring only standard regression rather than function-space optimization over velocity fields or potentials (Liu et al., 2022, Mena et al., 5 Nov 2025). While not a panacea for all transport problems (due to the non-optimality in some settings), the framework is compelling for its simplicity, analytic tractability, statistical rate guarantees, and computational efficiency—justifying its use across a wide spectrum of modern ML workflows.