Rectified-Flow: Deterministic Generative Modeling

• Rectified-flow is a deterministic generative modeling framework that constructs an ODE-based velocity field to transport samples along nearly straight, efficient paths.
• It achieves fast, high-fidelity sample generation, anomaly correction, and optimal transport through one-step inference and reflow strategies.
• Leveraging high-capacity architectures and iterative trajectory straightening, the method guarantees marginal preservation and convergence of transport costs.

Rectified-Flow Framework

Rectified-flow is a deterministic generative modeling paradigm that explicitly constructs an ordinary differential equation (ODE) to transport samples between two distributions—typically a tractable base (e.g., Gaussian noise) and an empirical target—using nearly straight, efficiently computable paths. The framework was introduced to address the computational inefficiency and stochasticity inherent to diffusion-based generative models, providing a scalable, theoretically principled alternative for both generative modeling and structured transport tasks. Rectified-flow underpins algorithms for fast sample generation, image-to-image translation, anomaly correction, and optimal transport, and has been extended to infinite-dimensional settings, multi-modal transports, and domain-conditional tasks.

1. Mathematical Foundations and Core ODE Formulation

Rectified-flow seeks to learn a velocity field $v(x, t)$ such that the ODE

$$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$

transports samples from a source distribution $\pi_0$ (e.g., Gaussian) to a target distribution $\pi_1$ (e.g., data) along straight (minimal energy) trajectories. The construction proceeds by defining a reference linear interpolant

$X_t = (1-t) X_0 + t X_1$

for paired samples $(X_0, X_1) \sim \pi_0 \times \pi_1$. The ground-truth velocity at each interpolated location is $X_1 - X_0$, but since $(X_1, X_0)$ are not recoverable at arbitrary $X_t$, the optimal velocity field is shown to be $v^*(x, t) = \mathbb{E}[X_1 - X_0 \mid X_t = x]$. This is obtained as the solution to the regression problem

$$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$0

The solution to this ODE, under mild conditions of uniqueness, guarantees that at each $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$1, the marginal law of $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$2 matches that of the reference interpolant $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$3 (Liu et al., 2022).

Extensions generalize this construction to functional Hilbert spaces, where the continuity equation and superposition principle still ensure marginal preservation and cost monotonicity under mild infinite-dimensional restrictions (Zhang et al., 12 Sep 2025).

2. Straightness, Reflow, and One-Step Inference

A central property of rectified-flow is the straightening of transport paths. The expected squared deviation of the ODE's velocity from the reference direction quantifies curvature: $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$4 Successive "reflow" steps—where a new velocity network is trained on ODE-generated pairs from the previous flow—progressively straighten $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$5, reducing $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$6 at an $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$7 rate over $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$8 iterations until Euler integration with a single step becomes near-exact (Liu et al., 2022). The limiting case enables direct, one-step mapping $$\frac{dZ_t}{dt} = v(Z_t, t), \qquad Z_0 \sim \pi_0, \quad t \in [0,1]$$9 without recursive ODE solvers, greatly enhancing computational efficiency in high-resolution image generation, domain transfer, or anomaly localization (Beizaee et al., 4 Aug 2025).

Conditional and one-step versions can be distilled via specialized loss functions, e.g., empirical $\pi_0$0 matching or MeanFlow average-velocity regression (Zhang et al., 28 Nov 2025), or by explicitly minimizing the error over a single ODE step.

3. Implementation and Architectural Considerations

Rectified-flow architectures typically employ high-capacity U-Nets or transformer-based modules as velocity field approximators. Time $\pi_0$1 is encoded using sinusoidal or learned embeddings, injected at each scale (Liu et al., 2022, Beizaee et al., 4 Aug 2025). For conditional tasks (e.g., image restoration or translation), conditioning variables are concatenated to the input or injected via feature modulation.

Training pipelines consist of:

• Paired sample generation (e.g., real or synthetically corrupted pairs for anomaly detection (Beizaee et al., 4 Aug 2025))
• Batched sampling of $\pi_0$2
• Computation of interpolated states and velocity targets
• Optimization via mean squared error or task-augmented losses, e.g., bond-length constraints in molecular design (Zhang et al., 2024), or physics-guided per-pixel noise regularization (Zeng, 10 Sep 2025)

Recursive reflow, distillation to compact models (Zhu et al., 2024), and data-efficient hedging strategies (e.g., balanced conic reflow incorporating real and generated pairs (Seong et al., 29 Oct 2025)) are employed to trade off straightness, data support, and compute cost. Extensions employ latent variable modeling for multimodal flows (Guo et al., 13 Feb 2025) or velocity-field averaging using rectified trajectories (Zhang et al., 28 Nov 2025).

4. Practical Applications

Rectified-flow has been deployed across a spectrum of domains:

• High-efficiency generative modeling: One-step or few-step synthesis of high-fidelity images, outperforming diffusion models in FID, recall, and wall-clock efficiency (e.g., ImageNet, FFHQ, LSUN) (Liu et al., 2022, Zhu et al., 2024, Zhang et al., 28 Nov 2025, Seong et al., 29 Oct 2025).
• Anomaly correction and segmentation: The REFLECT framework directly maps abnormal brain MRI images toward the healthy manifold in a single latent-space flow, achieving SOTA unsupervised anomaly localization with gains up to 6.4% Dice over diffusion-based approaches (Beizaee et al., 4 Aug 2025).
• Image-to-image translation & restoration: Reformulations like I2I-RFR enable ODE-based refinement with minimal architectural changes, improving perceptual detail and preserving simplicity (Iizuka et al., 20 Mar 2026); IR-Flow achieves SOTA distortion–perception trade-offs in few steps (Fan et al., 21 Apr 2026).
• Physics-informed and scientific ML: Conditioning on instrument noise models for RAW image enhancement (Zeng, 10 Sep 2025) and for high-resolution, multiscale PDE modeling (Armegioiu et al., 3 Jun 2025), rectified-flow allows effective posterities with dramatic inference speedups and improved statistical fidelity over diffusion baselines.
• Domain-adaptive and multimodal modeling: Used for domain-invariant transfer (Office-Home, DomainNet) (Liu et al., 2022) and unified vision-language understanding/generation (JanusFlow) (Ma et al., 2024).
• Structure-based drug design: FlowSBDD leverages plug-and-play loss terms for ligand generation, attaining SOTA binding and diversity with fast deterministic sampling (Zhang et al., 2024).
• Optimal Transport: The “rectified” and “c-rectified” flow approaches provide an interior, monotonic reduction of convex costs, converging to OT solutions under marginal constraints (Liu, 2022).

5. Algorithmic Innovations and Comparisons

The practical strength of rectified-flow lies in its algorithmic flexibility:

• Performance: Rectified-flow consistently reaches or surpasses the performance of diffusion models in terms of FID, IS, and inference efficiency. For example, one-step variants achieve FID=2.87 on ImageNet 64×64, surpassing two-stage distillation baselines (Zhang et al., 28 Nov 2025). Small models (≤15.7M params) achieve leading FID on CIFAR-10 with efficient resource usage (Zhu et al., 2024).
• Trajectory accuracy and drift: Balanced conic reflow and noise optimization (VRFNO) mitigate generated-data drift and reduce curvature, enhancing one-step accuracy and data fidelity (Seong et al., 29 Oct 2025, Dai et al., 14 Jul 2025).
• Conditioning and guidance: Variants like FlowChef generalize classifier-free, linear inverse, and masked editing guidance to rectified flow, leveraging the deterministic ODE dynamics for memory- and compute-efficient control without extensive backpropagation (Patel et al., 2024).
• One-step distillation: Direct modeling of mean velocity (MeanFlow) and reflow-meanflow hybrids eliminate repeated network evaluations in both training and inference (Zhang et al., 28 Nov 2025). Annealing reflow and flow-guided distillation further facilitate capacity-matched compression (Zhu et al., 2024).

6. Extensions and Theoretical Properties

Rectified-flow theory guarantees marginal preservation along ODE trajectories, strict decrease of all convex transport costs, and flexibility to optimize for specific $\pi_0$3-costs (in c-rectified flow) (Liu, 2022, Liu et al., 2022). The approach extends to infinite-dimensional Hilbert spaces with functional superposition principles, removing strong measure-theoretic constraints required by prior functional models (Zhang et al., 12 Sep 2025).

Algorithmic variants have introduced variational multi-modal velocity fields (Guo et al., 13 Feb 2025), velocity discrepancy modeling for inversion-free path-aware editing (Beaudouin et al., 1 Sep 2025), and exact higher-order ODE solvers to improve inversion/editing accuracy with pre-trained transformers (Wang et al., 2024). The "anchored classifier guidance" method addresses the limitations of vanilla classifier approaches in rectified flow by guaranteeing stable convergence with off-the-shelf discriminators and fixed-point solvers (Sun et al., 2024).

7. Limitations and Future Directions

Key limitations include the potential for residual path curvature in high-complexity domains, which may require multiple reflow passes or post-hoc correction (Liu et al., 2022, Zhang et al., 28 Nov 2025), drift when using purely synthetic (generated) couplings (Seong et al., 29 Oct 2025), and increased inference latency from anchoring or gradient-based guided solvers (Sun et al., 2024). Extensions to multimodal or highly non-linear transport tasks often necessitate multi-modal modeling or advanced mean/variance prediction.

Open directions include scalable one-step modeling for ultra-high-res and multimodal domains, deeper integration of functional extensions for spatio-temporal or PDE-valued data (Zhang et al., 12 Sep 2025), discrimination-aware or target-specific flows, and further theoretical characterizations of the straightening and monotonicity properties in non-Euclidean or nonconvex domains.

---

References:

• "Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow" (Liu et al., 2022)
• "REFLECT: Rectified Flows for Efficient Brain Anomaly Correction Transport" (Beizaee et al., 4 Aug 2025)
• "Balanced conic rectified flow" (Seong et al., 29 Oct 2025)
• "Flow Straighter and Faster: Efficient One-Step Generative Modeling via MeanFlow on Rectified Trajectories" (Zhang et al., 28 Nov 2025)
• "SlimFlow: Training Smaller One-Step Diffusion Models with Rectified Flow" (Zhu et al., 2024)
• "Variational Rectified Flow Matching" (Guo et al., 13 Feb 2025)
• "Rectified Flow: A Marginal Preserving Approach to Optimal Transport" (Liu, 2022)
• "Flow Straight and Fast in Hilbert Space: Functional Rectified Flow" (Zhang et al., 12 Sep 2025)
• "Physics-Guided Rectified Flow for Low-light RAW Image Enhancement" (Zeng, 10 Sep 2025)
• "Rectified Flow for Vision-Aided mmWave V2I Beam Prediction" (Zheng et al., 25 Nov 2025)
• "IR-Flow: Bridging Discriminative and Generative Image Restoration via Rectified Flow" (Fan et al., 21 Apr 2026)
• "JanusFlow: Harmonizing Autoregression and Rectified Flow for Unified Multimodal Understanding and Generation" (Ma et al., 2024)
• "Delta Rectified Flow Sampling for Text-to-Image Editing" (Beaudouin et al., 1 Sep 2025)
• "Improving Image-to-Image Translation via a Rectified Flow Reformulation" (Iizuka et al., 20 Mar 2026)
• "Taming Rectified Flow for Inversion and Editing" (Wang et al., 2024)
• "Steering Rectified Flow Models in the Vector Field for Controlled Image Generation" (Patel et al., 2024)
• "Straighten Viscous Rectified Flow via Noise Optimization" (Dai et al., 14 Jul 2025)
• "RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance" (Sun et al., 2024)
• "Rectified Flow For Structure Based Drug Design" (Zhang et al., 2024)
• "Rectified Flows for Fast Multiscale Fluid Flow Modeling" (Armegioiu et al., 3 Jun 2025)