Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
1. Paper overview
In one sentence: Learning from straight interpolations and then reflowing generated endpoint pairs makes image transport usable with very few neural evaluations, at the cost of additional training and possible full-solver quality loss. e01e02e03e04e09e10e16
| At a glance | What to know |
|---|---|
| Research problem | Source description Given samples from two distributions, learn a usable transport map without paired semantic correspondences. For generation the source can be Gaussian noise; for transfer it can be another data domain. The computational bottleneck is repeated evaluation of a large neural velocity model during ODE simulation. The paper asks whether training can shape trajectories so that very few evaluations suffice. e02e03e07 |
| Core mechanism | |
| A key reported result | CIFAR-10 one-step generation after distillation: 2-rectified + distillation: FID 4.85, IS 9.01, recall 0.50, NFE 1. FID ↓; inception score ↑; recall ↑; NFE ↓. Unconditional 32×32 generation, DDPM++ architecture, one model evaluation. The supplied text does not enumerate the evaluation sample count or FID reference split. Distilled 1-rectified: FID 6.18; distilled VP/sub-VP: 16.23/14.32. Distilled 3-rectified: FID 5.21, recall 0.51. Best FID and best recall occur in different rectified-flow rows. These comparisons do not equalize total training cost or establish superiority over every architecture in Table 1(b). e07e08e12e16 |
| Reading caution |
Core contributions
Figure 2. Reflow rebuilds the interpolation problem from the previous flow's endpoint pairing. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the panels from left to right. In (a), independently paired purple source points and red target points define straight interpolations, including conflicting directions through the middle. Panel (b) shows the flow obtained by fitting a single velocity at each state and time. The inset distinguishes the original endpoint X1 from the reassigned endpoint Z1. Now use the endpoint pairs of (b), rather than the original pairs, to construct the interpolations in (c). Rectifying those pairs produces (d). The panel formulas, Figure 2 caption and Algorithm 1 agree on this sequence: interpolate, fit and simulate, then interpolate the new coupling. e03e04e05e07
What it supports. The second regression problem is easier to follow with nearly constant velocity because its endpoint pairs already reflect the first flow's reassignment. This is the central mechanism behind reflow: learning a new coupling whose trajectory geometry supports coarse integration, rather than merely copying the original flow's endpoint function.
Where the evidence stops. This is a low-dimensional illustration. Non-crossing refers to trajectories meeting at the same state and time under a unique ODE solution. Marginal preservation and transport-cost guarantees apply to the ideal field; the picture does not guarantee exact neural fitting.
2. Motivation
2.1 The problem and the proposed response
Given samples from two distributions, learn a usable transport map without paired semantic correspondences. For generation the source can be Gaussian noise; for transfer it can be another data domain. The computational bottleneck is repeated evaluation of a large neural velocity model during ODE simulation. The paper asks whether training can shape trajectories so that very few evaluations suffice. e02e03e07
2.2 What this reading follows
An ODE generator can produce excellent images yet remain expensive because every sample requires many calls to its velocity network. Rectified flow attacks the geometry behind that cost: train on straight interpolations, simulate the resulting transport, and train again on its endpoint pairs. This illustrated reading follows the 7 September 2022 arXiv v1 preprint, rather than asserting equivalence to the catalog's ICLR 2023 publication. The most useful experimental distinction is between reflow, which changes the transport coupling, and distillation, which fits an existing endpoint map. The figures and tables show why fewer sampling steps can improve while fully integrated quality deteriorates. e01e02e03e04e09e10e16
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The foundational flow-matching classification is supported by the velocity-regression construction and its probability-flow connection. The WAM architecture, action-prediction paradigm and quadrant are not applicable: this paper defines transport and image/feature outputs, without joint future/action prediction, inverse dynamics or a control loop. A single U-Net velocity model does not establish a 'One Model' world-action architecture. e03e06e07e13e15
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Why straight training segments do not immediately give a one-step generator
During training, both endpoints are available, so constructing a straight segment is trivial. During generation, only its starting point is available. The network must therefore replace the unknown endpoint displacement with a prediction from the current state and time. Eq. (2) identifies the ideal prediction as a conditional mean, which reconciles conflicting directions where training segments overlap. A useful deduction follows from the paper's conditional-mean prescription at t=0: with initially independent endpoint pairs, adding the predicted displacement to the source state gives the target mean. That explains why the first flow can be poor under a single Euler step even though every training segment was straight. The ODE can still transport a rich distribution through later evaluations. Reflow is needed to improve the endpoint dependence and the usable trajectory geometry. e03e07e08
5.2 What is learned again during reflow
First simulate the learned ODE and save each starting state together with its generated endpoint. These pairs now contain the deterministic correspondence created by the flow. The next regression stage uses straight interpolations of those pairs, rather than returning to independent source and data samples. Algorithm 1 expresses this recursively, and the CIFAR-10 appendix implements each stage with four million generated pairs and 300,000 fine-tuning steps. Exact rectification preserves the relevant marginals and reduces convex transport costs, while the straightening theorem controls accumulated geometric error. Neural fitting and numerical simulation introduce approximations, so those ideal properties do not guarantee improved FID after every stage. The table makes this visible: additional reflow helps coarse sampling but worsens the fully integrated result. Distillation then fits the chosen endpoint map, with LPIPS used for the reported one-step image variant. e03e04e05e09e16
5.3 Changing the source distribution is only part of transfer
For image generation, Gaussian noise supplies the source distribution and images supply the target. Image translation replaces the noise with source-domain images, but the paper also changes what the regression penalizes. Eq. (20) uses a domain-classifier feature Jacobian to emphasize stylistic changes; the authors explicitly prioritize retaining source identity over exact target-distribution matching. The portrait-to-cat examples should therefore be read as demonstrations of that task-specific objective. Domain adaptation makes another change: it transports pretrained features and measures classification accuracy after mapping test-domain features toward training-domain features. Reader interpretation: these are distinct validation problems even though both use flows. Attractive translated portraits do not establish better recognition, and a recognition gain does not validate semantic fidelity in images. Neither experiment supplies evidence of action-conditioned dynamics, planning or physical execution. e07e13e14e15
5.4 Training and inference
During training
CIFAR-10 uses a DDPM++ U-Net at 32×32 resolution, Adam with learning rate 2×10⁻⁴, dropout 0.15 and EMA 0.999999. Each reflow stage generates four million endpoint pairs and fine-tunes for 300,000 steps. One-step distillation replaces L2 with LPIPS; its improvement therefore includes an objective change. e16
Translation uses AdamW, batch size 4, 1,000 epochs and EMA 0.9999, selecting learning rate by training loss. Domain adaptation uses AdamW, batch size 16, 50,000 iterations, learning rate 10⁻⁴ and OneCycle scheduling. e17e18
During inference
Start from a source sample and integrate the learned velocity over t∈[0,1]. Experiments use uniform-step Euler or adaptive RK45. A perfectly straight constant-speed flow reaches its endpoint in one Euler update; learned flows only approximate that ideal. e04e07
Generation returns an image; translation returns a transformed source image; adaptation classifies transported test features after 100 uniform steps. These are distribution-transfer operations, with no online environmental feedback or executed control actions. e07e14e15
5.5 Implementation flow
- Create regression examples
Draw endpoint pairs and interpolate their states. The target is the full displacement X1−X0, independent of interpolation time. Training uses these constructed examples directly; it does not differentiate through an ODE solve. e03e16
- Causalize the interpolation
The learned velocity approximates the conditional mean displacement at the current state and time. Unlike a sampled interpolation path, simulation needs no access to the target endpoint. Unique ODE trajectories cannot cross at the same state and time; Figure 2 depicts the resulting reassignment of endpoints. e03e05
- Reflow, then optionally distill
Simulate the current flow to collect paired starting and ending states, then fit the next flow to their linear interpolations. Reflow changes the pairing. Final distillation instead approximates the existing pairing with a direct map, avoiding a numerical trajectory at inference. e04e07e16
- Adapt the transport objective
For image translation, Eq. (20) weights the velocity residual through the Jacobian of a domain-classifier feature map h. This emphasizes style-sensitive changes and relaxes exact target-distribution matching. Domain adaptation instead learns a flow between final-hidden-layer feature distributions before classification. e13e15
6. Experiments & results
Rectified flow learns distribution transport by regressing straight-line endpoint displacements onto intermediate states. Reflow trains on the learned flow's endpoint pairs to make coarse numerical integration more accurate; final distillation further improves one-step generation. The CIFAR-10 results establish a speed–quality tradeoff, while translation and domain adaptation demonstrate broader uses of the transport construction. This report concerns the verified September 2022 v1 preprint.
6.1 Read the original evidence
Table 1(a). The best distilled one-step FID and the best distilled recall belong to different rows. Original paper, p. 24 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the solver headings, because the three blocks answer different questions. The top block fixes inference to one evaluation; values outside parentheses are undistilled, and parenthesized values follow distillation. The middle block integrates each ODE with adaptive RK45. The bottom block reports much more expensive SDE simulations. Within each block, smaller FID and NFE are preferred, while larger inception score and recall are preferred. This crop preserves the shared DDPM++ architecture comparison; the paper's separate Table 1(b) uses other architectures. Read all metrics across one row before comparing models, so that a strong FID is not accidentally paired with another model's recall. e07e08e09e12e16
What it supports. Distilled 2-rectified flow achieves FID 4.85 and recall 0.50 with one evaluation. Distilled 3-rectified flow reaches recall 0.51 but has FID 5.21. Under RK45, the first flow instead leads the rectified variants with FID 2.58 and recall 0.57. The table therefore exposes different winners for different budgets and metrics.
Where the evidence stops. NFE denotes network calls, not measured latency. The caption equates NFE with steps, but multistage RK45 should not be interpreted as an equal number of Euler updates. No confidence intervals are reported, and the rows do not match total training cost.
Figure 14. One additional rectification makes a single translation update much more expressive in these examples. Original paper, p. 28 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as one fixed input, beginning with the painted portrait at the left. The next two columns show first- and second-flow outputs with 100 Euler steps; the final two use only one step. Compare within a row and at the same step count before comparing across budgets. The one-step first flow retains much of the human face, while the one-step second flow produces a more pronounced cat-like transformation. The 100-step outputs remain sharper in these displayed examples. Connect the visual to Eq. (20): translation emphasizes changes in domain-classifier features rather than demanding exact matching of the complete target image distribution. e13e14e17
What it supports. The examples show that reflow's computational benefit extends beyond Gaussian-to-image generation to image-to-image transport. A single second-flow update can produce recognizable target-style changes while retaining visible composition from the input. This is qualitative support for the intended translation behavior, not a measured rate of identity preservation.
Where the evidence stops. Only two selected inputs appear here. The random test split supplies held-out examples, but the source reports no aggregate translation score, uncertainty or controlled feature-loss ablation. Visual similarity cannot establish semantic identity preservation or generalization across all domains.
Table 2. The OfficeHome gain coexists with a slightly lower DomainNet mean than CORAL. Original paper, p. 29 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Here the output being evaluated is a class prediction after feature transport, rather than an image judged for appearance. Read across one dataset row at a time, with higher accuracy preferred. The method maps final-hidden-layer test features toward the training-domain feature distribution, then applies the classifier; inference uses 100 uniform flow steps. Compare the Ours column first with CORAL and then with the simpler ERM baseline. The OfficeHome and DomainNet rows must remain separate because they involve different datasets and class spaces. Preserve the displayed plus-or-minus terms, but do not assign them a confidence-interval or standard-deviation meaning that the source does not specify. e15e18
What it supports. OfficeHome reports 69.2±0.5 for rectified flow versus 68.7±0.3 for CORAL. DomainNet reports 41.4±0.1 versus CORAL's 41.5±0.2. Both rectified-flow means exceed ERM, but the table does not show an outright win over CORAL on both datasets. The accompanying broad performance claim needs this row-level qualification.
Where the evidence stops. The exact domain partitions, repetition count and meaning of ± are not stated. Because novel-domain features participate in transport learning, the setup should not be silently treated as a protocol with no target-domain access.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| CIFAR-10 one-step generation after distillation Unconditional 32×32 generation, DDPM++ architecture, one model evaluation. The supplied text does not enumerate the evaluation sample count or FID reference split. | 2-rectified + distillation: FID 4.85, IS 9.01, recall 0.50, NFE 1. FID ↓; inception score ↑; recall ↑; NFE ↓ | Distilled 1-rectified: FID 6.18; distilled VP/sub-VP: 16.23/14.32. Distilled 3-rectified: FID 5.21, recall 0.51. Best FID and best recall occur in different rectified-flow rows. These comparisons do not equalize total training cost or establish superiority over every architecture in Table 1(b). e07e08e12e16 |
| CIFAR-10 one-step reflow without distillation Table 1(a), DDPM++ with Euler N=1, before the parenthesized distillation results. | 1/2/3-rectified: FID 378/12.21/8.15; recall 0.0/0.34/0.41. FID ↓; recall ↑ | The second flow follows one additional reflow stage; the third follows another. Reflow strongly improves coarse integration, but undistilled one-step performance remains below the distilled endpoints. e08e10 |
| CIFAR-10 generation with adaptive RK45 Unconditional DDPM++ ODEs, adaptive full simulation in Table 1(a). | 1-rectified: FID 2.58, recall 0.57, NFE 127. FID ↓; recall ↑; NFE ↓ | 2/3-rectified: FID 3.36/3.96 and NFE 110/104. VP/sub-VP ODE: FID 3.93/3.16. VP SDE: FID 2.55 at NFE 2000. Extra rectification reduces neural evaluations but worsens fully integrated sample quality. NFE counts calls, not measured wall-clock speed; no uncertainty is reported. e09e10 |
| High-resolution unconditional image generation 1-rectified flow; 256×256 examples from LSUN Church, LSUN Bedroom, CelebA-HQ and AFHQ Cat in Figure 11. | The figure displays recognizable scenes and faces across all four datasets. Qualitative generated samples | No quantitative high-resolution baseline comparison is provided in this figure. The samples demonstrate application beyond CIFAR-10, but do not establish aggregate fidelity, coverage or a high-resolution state-of-the-art result. e19 |
| Unpaired portrait-to-cat image translation Figure 14; random 80%/20% training/test split within the source datasets, 512×512 images, test inputs, Euler N=1 or N=100. | The displayed second-flow one-step outputs show stronger cat-like changes than first-flow one-step outputs. Qualitative target-style transformation | Both flows also produce sharper displayed transformations with 100 steps. These selected examples support the intended transfer behavior; the source supplies no aggregate translation score or measured identity-retention rate. e13e14 |
| Domain adaptation on OfficeHome and DomainNet Transport novel-domain test features toward training-domain features, then classify; 100 uniform flow steps. Exact domain partitions are not stated. | OfficeHome 69.2±0.5; DomainNet 41.4±0.1. Classification accuracy ↑, as reported in Table 2 | CORAL: 68.7±0.3 and 41.5±0.2; ERM: 66.5±0.3 and 40.9±0.1, respectively. The method leads the OfficeHome row but falls slightly below CORAL on DomainNet. The ± convention and repetition count are unspecified; statistical superiority is not established. e15 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 8. Reflow changes the sampling-budget curve, while distillation adds another intervention. Original paper, p. 24 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. In panel (a), read the two plots together: FID measures one aspect of sample quality and recall tracks coverage. The horizontal sampling-step axes are logarithmic, as is the FID axis. Green, orange and purple identify the first three rectified flows; stars show their separately distilled one-step versions. In panel (b), the horizontal axis changes to training iterations in units of ten thousand. Circles, squares and crosses indicate one, two and three Euler steps, respectively. Red dashed markers show reflow boundaries. Keep these two axes conceptually separate: moving right in (a) spends more computation per sample, while moving right in (b) spends more training computation. e08e09e10e16
What it supports. Reflow markedly improves the low-step regime, but the first flow catches up and becomes preferable with sufficiently fine integration. Continued fitting within each reflow stage also changes the results. The figure supports a practical sampling tradeoff; it does not by itself isolate the effect of changing endpoint pairings from additional training.
Where the evidence stops. The training trajectories are not a matched-compute control. Distilled stars also include the appendix's one-step LPIPS objective change. A mechanism claim requires comparisons that hold model capacity, training budget and distillation loss fixed.
Figure 9. Pixel trajectories become more linear, while the aggregate straightness measure remains nonzero. Original paper, p. 24 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left plot as a diagnostic of trajectory geometry, using Eq. (3): smaller values mean less discrepancy between instantaneous velocity and endpoint displacement. The initial green point, orange second-flow curve and purple third-flow curve correspond to successive stages; the red dashed lines mark reflow. The right panels show individual sampled pixel values as interpolation time t runs from zero to one. A straight line in a pixel trace indicates constant velocity for that coordinate. Compare the bending in the first-flow panel with the more linear traces in the second. These are projections of image trajectories, not separate examples of whole-image quality. e04e05e11
What it supports. The diagnostic supports the proposed geometric effect: reflow produces straighter sampled trajectories and reduces the aggregate discrepancy relative to the starting model. Yet the plotted discrepancy is nonzero, and the third-stage curve fluctuates during training. This fits an approximate learned mechanism, without implying exact one-step transport.
Where the evidence stops. Selected pixel traces do not characterize every image coordinate or sample. Straightness is not FID, recall or a test of semantic correctness, and the theoretical best-iterate bound does not promise monotonic improvement at every training checkpoint.
7. Analysis & limitations
7.1 What the evidence leaves open
The exact-flow guarantees require rectifiability and ideal velocity estimation. Straightness alone is not multidimensional optimality. The authors caution against excessive reflow because estimation errors accumulate; full-solver CIFAR-10 scores exhibit that tradeoff. e04e05e09
Translation uses a random 80%/20% split and qualitative examples, including one-step results. Neither identity preservation nor target fidelity receives an aggregate numerical evaluation or a controlled Eq. (20) ablation. e13e14
Figure 8 intertwines additional training and reflow; LPIPS further changes the distillation objective. The reported comparisons do not isolate straightening under matched total training budgets. e10e16
7.2 Questions for discussion
- Under equal training budgets, how much one-step improvement remains attributable to reflow?
- Does feature-weighted translation preserve identity better than plain velocity regression on a fixed test split?
8. Reproducibility audit
8.1 Requirements and known gaps
A CIFAR-10 reproduction needs the DDPM++ configuration, endpoint-pair generation, checkpoint initialization and solver settings. The appendix specifies several optimizer settings, but omits hardware, software versions, initial training duration, evaluation sample count and exact tolerances; some settings are deferred to prior work. e07e16
Translation additionally requires the domain classifier and its selected representation; the source does not explicitly state its freezing schedule. Proposed checks should separate reflow from extra fitting and compare feature-weighted versus plain regression on identical held-out inputs. e13e14e17
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Separate reflow, extra training and the distillation loss
Reader-proposed check; not executed. Starting from one fixed DDPM++ CIFAR-10 checkpoint, compare one reflow stage with an equal-update continuation on the original independent-pair objective. Follow the appendix's four-million-pair and 300,000-update reflow recipe, record the additional pair-generation cost, and keep optimizer and model settings equal. Evaluate fixed source seeds at one, two and three Euler steps and with the same RK45 tolerances; measure FID, recall, endpoint deviation from the finely solved flow and S(Z). Separately distill both checkpoints with L2 and LPIPS under equal budgets. The mechanism is supported if reflow lowers geometric and coarse-solver errors beyond continued training before LPIPS is introduced; gains appearing only after LPIPS would weaken that attribution. e04e07e08e09e10e11e16
Check 2: Test whether feature weighting preserves identity during translation
Reader-proposed check; not executed. Fix the random 80%/20% MetFace-to-AFHQ-cat split and the same DDPM++ training budget. Compare plain Eq. (1) regression with Eq. (20), using the same domain-classifier representation and documenting an explicitly fixed feature-network policy as a reproduction choice. Evaluate first and second flows at one and 100 Euler steps on identical held-out portraits. Predefine blinded human ratings of target-style recognizability and source composition/identity retention, alongside numerical endpoint deviation from the 100-step output. Feature weighting earns support only if it improves the retention–style tradeoff across the held-out set, rather than merely producing a few favorable examples. Report failures and rater disagreement. e03e13e14e17
8.3 Reading coverage
Visual audit: The title/author/version page, all numbered Figures 1–21, both quantitative tables and the listed supporting method, theory and experiment pages were visually inspected. Appendix pages 37–38 were inspected for training and reproduction details. All six final original crops were separately viewed and checked for complete labels, legends and table headings. Figure 2's pairing sequence was cross-checked against Eqs. (1)–(2) and Algorithm 1; no claim-relevant disagreement was found. Figure 9 was checked against Eq. (3), and the translation figure against Eq. (20) and its test protocol. The FID/recall juxtaposition in the introduction is resolved using Table 1(a) and p. 25; Table 2's DomainNet ordering is preserved. The Table 1 caption's broad step/NFE equivalence is qualified in the table caution. All supplied text, including pages outside this visual list, was read in ten individual chunks. Separate supplements and the ICLR publication edition remain unverified; code was not inspected and no reproduction was run.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 18, 22, 23, 24, 25, 26, 27, 28, 29, 37, 38, 39, 40, 41. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract and Section 1 Introduction (pp. 1–4)
- Sections 2.1–2.3.1: rectification, reflow, distillation and nonlinear extensions (pp. 4–12)
- Sections 3.1–3.5: marginal preservation, transport costs, straightening, optimal couplings and probability-flow ODEs (pp. 12–18)
- Section 4 Related Works and Discussion (pp. 18–22)
- Sections 5.1–5.4: toy examples, image generation, translation and domain adaptation (pp. 22–29)
- References (pp. 30–36)
- Appendix A, Algorithms 2–4 and all additional figure captions (pp. 37–41)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Recovered-edition limitation: arXiv v1 preprint from 2022. Catalog cites ICLR 2023; the recovered preprint is not asserted to be identical to the blocked ICLR publication PDF. Title and all three authors match.
- The extraction limitation was addressed by inspecting all numbered figures, both tables and all pages supporting retained method, training and numerical claims. All ten supplied text chunks were read individually.
- Code and external referenced works were not inspected; no experiments were reproduced. The ICLR publication PDF was not supplied or compared.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, equal-contribution footnote and arXiv version stamp
Title and authors Xingchao Liu, Chengyue Gong and Qiang Liu match the catalog. All three list University of Texas at Austin; Liu and Gong contributed equally. The inspected version is arXiv:2209.03003v1, 7 September 2022.
Go to primary source ↓e02PDF p. 2, Introduction, 'The Transport Mapping Problem' and inference-cost discussion
The task is to transport empirical source and target distributions. Repeated neural drift evaluations make continuous-time generation expensive.
Go to primary source ↓e03PDF pp. 4–6, Figure 2, Section 2.1, Eq. (1), Algorithm 1 and Eq. (2)
Regress endpoint displacement on interpolated states and uniform time. The population minimizer is the conditional mean displacement. Figure 2 illustrates rewiring independent endpoint pairings into non-crossing ODE trajectories, then rectifying the resulting coupling again.
Go to primary source ↓e04PDF pp. 7–8, Section 2.2, Eq. (3), 'Reflow, straightening, fast simulation' and 'Distillation'
Straightness means constant velocity along each trajectory. S(Z) integrates squared discrepancy between endpoint displacement and instantaneous velocity. A perfectly straight flow admits an exact single Euler update. Reflow changes the coupling; distillation fits an existing endpoint map. Excessive reflow can accumulate estimation error.
Go to primary source ↓e05PDF pp. 13–16, Definitions 3.2/3.4 and Theorems 3.3, 3.5–3.10
For rectifiable processes with the required existence and uniqueness, exact rectification preserves marginal laws and does not increase convex transport costs. Theorem 3.7 bounds accumulated straightness and intersection errors, yielding a best-iterate O(1/K) statement. Straightness need not imply optimal transport in multiple dimensions.
Go to primary source ↓e06PDF pp. 9–10, Section 2.3, Eq. (6) and Section 2.3.1; p. 18, Proposition 3.11
General time-differentiable interpolation permits velocity regression with marginal preservation, but lacks the canonical linear construction's general cost-reduction and straightening guarantees. Probability-flow ODEs fit the extended framework, with the initial-distribution approximation discussed on p. 10.
Go to primary source ↓e07PDF pp. 22–23, Section 5 'Algorithm', Section 5.1 and Section 5.2 'Experiment settings'
Experiments begin with independent pairs, generate endpoint couplings for reflow, and use constant-step Euler or adaptive SciPy RK45. Image generation uses standard Gaussian input and DDPM++ U-Net drift; Table 1(a) compares the shared architecture. RK45 tolerances are referred to prior work rather than enumerated.
Go to primary source ↓e08PDF p. 24, Table 1(a), 'ODE One-Step Generation' rows, all columns; p. 25, first paragraph
At NFE 1, distilled 2-rectified flow reports IS 9.01, FID 4.85, recall 0.50; distilled 3-rectified flow reports 8.79, 5.21, 0.51. Distilled 1-rectified, VP and sub-VP FIDs are 6.18, 16.23 and 14.32. Undistilled 1/2/3-rectified FIDs are 378/12.21/8.15 with recalls 0.0/0.34/0.41. Parentheses denote distillation.
Go to primary source ↓e09PDF p. 24, Table 1(a), 'ODE Full Simulation' and 'SDE Full Simulation' blocks; Table 1 caption
RK45 1/2/3-rectified flows report NFE 127/110/104, FID 2.58/3.36/3.96 and recall 0.57/0.54/0.53. VP and sub-VP ODE FIDs are 3.93 and 3.16 with NFE 140 and 146. VP SDE reports FID 2.55, recall 0.58 and NFE 2000. NFE is defined as neural-network calls.
Go to primary source ↓e10PDF p. 24, Figure 8(a–b), caption and 'Results on few and single step generation'
Figure 8 separates Euler-step sweeps from reflow training trajectories and uses stars for distilled models. The text states that reflow improves small-step performance but worsens large-step results because velocity-estimation errors accumulate.
Go to primary source ↓e11PDF p. 24, Figure 9 and caption; p. 7, Eq. (3); p. 25, 'Reflow straightens the flow'
The CIFAR-10 diagnostic plots straightness against training iterations and sample pixel values against time. The second flow has visibly more linear pixel trajectories; the third-stage straightness curve is not monotonic at every checkpoint. This is a geometric diagnostic, not an image-quality metric.
Go to primary source ↓e12PDF p. 4, contribution-result paragraph; pp. 24–25, Table 1(a) and one-step-results discussion
The introduction juxtaposes FID 4.85 and recall 0.51. The detailed table and p. 25 assign those best values to different distilled models: 2-rectified flow has recall 0.50, while 3-rectified flow has recall 0.51 and FID 5.21.
Go to primary source ↓e13PDF pp. 25 and 27, Section 5.3, objective motivation and Eq. (20)
Translation seeks stylistic change while retaining source identity, explicitly relaxing exact matching of the target image distribution. Eq. (20) applies the feature Jacobian to the velocity residual; h is a domain-classifier representation fine-tuned from an ImageNet-pretrained model.
Go to primary source ↓e14PDF pp. 27–28, Section 5.3 experiment settings, Figure 14 and results
AFHQ, MetFace and CelebA-HQ are split randomly 80%/20%; images are resized to 512×512 and examples start from test data. Figure 14 compares two inputs under 1- and 2-rectified flows with N=100 and N=1. Translation evaluation is qualitative.
Go to primary source ↓e15PDF pp. 28–29, Section 5.4 and Table 2, OfficeHome and DomainNet rows
Test-domain final-hidden-layer features are transported toward training-domain features, then classified. Inference uses 100 uniform steps and the text names DDPM++. OfficeHome: ours 69.2±0.5, CORAL 68.7±0.3, ERM 66.5±0.3. DomainNet: ours 41.4±0.1, CORAL 41.5±0.2, ERM 40.9±0.1. The uncertainty convention, repetitions and exact domain partitions are not specified here.
Go to primary source ↓e16PDF p. 37, Appendix A, Algorithms 2–4 and 'Experiment Configuration on CIFAR10'
CIFAR-10 uses 32×32 images, DDPM++, Adam learning rate 2e−4, dropout 0.15 and EMA 0.999999. Each reflow dataset contains four million endpoint pairs; the model is fine-tuned for 300,000 steps. Few-step distillation samples a discrete time grid; one-step distillation substitutes LPIPS for L2.
Go to primary source ↓e17PDF p. 38, Appendix A, 'Experiment Configuration on Image-to-Image Translation'
Translation uses DDPM++, AdamW with betas (0.9,0.999), weight decay 0.1, dropout 0.1, batch size 4, 1,000 epochs and EMA 0.9999. Learning rate is selected by lowest training loss from five listed candidates. The appendix repeats the random 80%/20% split.
Go to primary source ↓e18PDF p. 38, Appendix A, 'Experiment Configuration on Domain Adaptation'
Domain adaptation uses AdamW, batch size 16, 50,000 iterations, learning rate 1e−4, weight decay 0.1 and OneCycle scheduling.
Go to primary source ↓e19PDF p. 26, Figures 10–11; pp. 39–41, Figures 17–21 and captions
Additional inspected visuals show image generation at 256×256, latent interpolation, endpoint extrapolation, reconstruction with different solvers, and image-translation trajectories. They provide qualitative examples rather than aggregate quantitative tests.
Go to primary source ↓8.5 Primary sources
Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow ↗
PDF · 17,533 extracted words
Source fingerprint
d81790cb6988ef480002adcc84000d4f44855b11fb868001fd53aad133bc3a74