Coarse-Graining Optimal Transport for Rollout-Free Post-Training of Flow Matching
Authors: Soon Hoe Lim, Aditi Gupta, Michael W. Mahoney, N. Benjamin Erichson
Organizations: Department of Mathematics, KTH Royal Institute of Technology · Nordita, KTH Royal Institute of Technology and Stockholm University · International Computer Science Institute · Department of Statistics, University of California at Berkeley · Lawrence Berkeley National Laboratory
Flow matching depends critically on the coupling used to pair source and target samples. Optimal transport (OT) couplings can reduce velocity conflict and straighten trajectories; but, in practice, challenges remain: the source comes from a continuous distribution, only a finite number of target samples data are accessible, and the neural network has finite capacity. For these reasons, making the coupling more OT-like may not consistently improve sampling quality, in particular in the low number-of-function-evaluation (NFE) regime. Motivated by this, we introduce ReCouple, a controllable framework that constructs coarse-grained OT couplings, to better understand rollout-free OT-based post-training. We also consider extensions of this basic framework: ReCouple+ extends this toward the continuous-source setting by increasing source resolution, converging to semi-discrete OT in the appropriate limit; and ReCouple++ further adds spatial regularization, motivated by the observation that straight trajectories need not correspond to spatially simple velocity fields. Empirical validation demonstrates the effectiveness of these methods, and it reveals an NFE-dependent interplay between coupling optimality, neural realizability, and generation quality.
Figures & tables
Figure 1: Generated samples on the 2D multimodal task across methods and NFEs. Post-training using the ReCouple family improves recovery of the target modes, particularly at low NFEs.
Table 1: Evaluation on the 2D multimodal task. We report the 2-Wasserstein distance ( W2 , lower is better) at different NFEs, together with the neural acceleration energy Eacc (lower is better).
Method
NFE 1
NFE 2
NFE 4
NFE 8
NFE 10
NFE 50
RF (backbone)
357.341
165.085
50.014
16.821
13.123
5.180
Teacher Reflow (MSE)
11.417
7.008
5.915
5.539
5.502
5.511
One-step distillation (MSE)
9.436
128.435
149.356
163.575
167.244
182.325
Fresh-independent FT
356.823
165.896
50.278
16.853
13.110
5.033
OT-CFM-FT ( Tong et al. (2023) )
232.657
90.181
28.572
13.338
11.269
5.137
AlignFlow ( Kong et al. (2026) )
164.919
58.269
23.050
12.861
11.142
5.175
Table 2: CIFAR-10 FID results with the Rectified Flow pretrained backbone. Lower is better. We fix R=32 and B=2048 for ReCouple, ReCouple-AS, ReCouple+, and ReCouple++. We use M=10 for ReCouple+ and ReCouple++. Results are averaged over three seeds.
Method
NFE 1
NFE 2
NFE 4
NFE 8
NFE 10
NFE 50
OT-CFM (backbone)
230.321
86.640
27.312
12.709
10.723
4.902
Teacher Reflow (MSE)
11.196
6.848
5.705
5.291
5.234
5.208
One-step distillation (MSE)
9.087
137.009
159.789
177.066
181.694
201.234
Fresh-independent FT
351.241
155.466
47.210
15.845
12.286
4.634
OT-CFM-FT ( Tong et al. (2023) )
228.462
87.273
27.764
12.771
10.717
4.745
AlignFlow ( Kong et al. (2026) )
163.365
57.590
22.516
12.260
10.568
4.788
Table 3: CIFAR-10 FID results with the OT-CFM pretrained backbone. Same setting as Table 2 .
Method
NFE 1
NFE 2
NFE 5
NFE 10
NFE 25
NFE 50
SiT-XL/2 FM
275.340
99.078
10.172
12.175
12.724
12.645
Minibatch OT ( Tong et al. (2023) )
253.866
90.371
7.845
9.036
9.437
9.360
SD-FM ( Mousavi-Hosseini et al., 2026 )
239.573
80.828
8.801
10.329
10.828
10.855
ReCouple
245.571
82.684
8.989
10.637
11.150
11.164
ReCouple-AS
246.547
88.086
7.275
7.051
7.432
7.539
ReCouple++ ( M=2 )
246.249
87.645
7.226
6.960
7.358
7.302
Table 4: ImageNet-256 FID results using the pretrained SiT-XL/2 backbone. All metrics are computed using 50K generated samples with Euler integration. Lower is better; the best result at each NFE is shown in bold. ReCouple++ uses pooling factor M=2 .
Figure 2: Qualitative comparison on ImageNet-256 using pretrained SiT-XL/2 at NFE 25. Each row shows samples generated from the same initial noise and class condition using the SiT baseline, Minibatch OT, ReCouple, ReCouple-AS, and ReCouple++. This matched comparison highlights the differences in visual fidelity and semantic consistency induced by each sampling strategy.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Sparse OT graph comparison at B=8 for ReCouple and RandomBlockSelect.
Figure 4: Gap to optimal coupling cost and OT-edge recall across R at B=8 . OT-edge recall is the fraction of edges in the global empirical OT matching that are contained in the admissible block graph.
Figure 5: Held-out gap to optimal coupling cost and OT-edge recall across B for various choices of R .
C8−C⋆↓
OT-edge recall ↑
R
ReCouple
RandomBlock&Select
ReCouple
RandomBlock&Select
1
1.991±0.772
1.840±0.292
0.277±0.051
0.058±0.024
8
0.797±0.319
1.418±0.114
0.355±0.076
0.064±0.023
16
0.674±0.236
1.351±0.094
0.377±0.066
0.062±0.022
32
0.525±0.141
1.283±0.094
0.417±0.058
0.063±0.022
Appendix
Table 5: Matched best-of- R candidate selection for the standard Gaussian source to four-mode Gaussian target experiment ( N=128 , B=8 ). ReCouple chooses the best of R random projection-induced block geometries, while RandomBlock&Select chooses the best of R independent geometry-blind balanced block geometries, using the same block-OT objective. Values are mean ± standard deviation over 50 independent candidate pools. The unrestricted global empirical OT cost is C⋆ . OT-edge recall is the fraction of edges of the global OT permutation retained by the selected block graph; the geometry-blind baseline is B/N=0.0625 .
Q=⌈BN⌉,Ib={(b−1)B+1,…,min{bB,N}},b=1,…,Q.
Appendix
Algorithm 3 ReCouple: general case
θ⋆=θr⋆,r⋆∈argmin1≤r≤RCN(πr).
Appendix
Algorithm 4 ReCouple+
Coupling
Quadratic cost ↓
Source/target representation
Prep. time (s) ↓
ReCouple+ ( B=512,M=2 )
0.7234
100 k source / 50 k target
159.55
Global SDOT
0.6692
continuous source / 50 k target
815.19
Appendix
Table 6: Coupling-construction diagnostics on the 2D task. Preparation times exclude neural post-training.
Method
Precision ↑
Recall ↑
Density ↑
Coverage ↑
sW2 ↓
Pretrained
0.9934
0.1138
0.9979
0.0423
0.4293
PersistentIndependent-FT
0.9824
0.2443
0.9843
0.0410
0.4163
AlignFlow
0.5872
0.1893
0.4590
0.0449
0.3654
OT-CFM-FT
0.5789
0.8830
0.3972
0.5273
0.0384
SDOT
0.8700
0.9559
0.7973
0.7264
0.0373
Noise-ReCouple
0.7038
0.9459
0.5610
0.6626
0.0292
Appendix
Table 7: Additional distributional metrics on the 2D task at NFE 1. Higher is better for precision, recall, density, and coverage; lower is better for sliced Wasserstein distance (sW2).
Method
Precision ↑
Recall ↑
Density ↑
Coverage ↑
sW2 ↓
Pretrained
0.5622
0.8803
0.3928
0.5298
0.0732
PersistentIndependent-FT
0.5492
0.8315
0.3796
0.5073
0.0748
AlignFlow
0.6503
0.9343
0.5044
0.6226
0.0609
OT-CFM-FT
0.9125
0.9781
0.8713
0.8437
0.0176
SDOT
0.9332
0.9948
0.8901
0.9137
0.0381
Noise-ReCouple
0.9131
0.9920
0.8650
0.9013
0.0171
Appendix
Table 8: Additional distributional metrics on the 2D task at NFE 8. Higher is better for precision, recall, density, and coverage; lower is better for sliced Wasserstein distance (sW2).
RF backbone
OT-CFM backbone
Method
Prep.
Post-train
Prep.
Post-train
Teacher Reflow (MSE)
445.53†
105.89±0.61
284.98†
108.22±3.15
One-step distillation (MSE)
445.53†
107.41±3.63
284.98†
105.41±2.73
Fresh-independent FT
0.00
96.64±3.30
0.00
96.10±3.27
OT-CFM-FT ( Tong et al. (2023) )
0.00
97.73±2.83
0.00
97.88±1.86
AlignFlow ( Kong et al. (2026) )
30.41
118.53±6.19
30.41
115.91±4.02
Appendix
Table 9: Wall-clock cost comparison for the CIFAR-10 experiments. Times are reported in minutes. Preparation is the one-time method-specific cost for constructing the required coupling, pair bank, or teacher targets and is shared across compatible runs. Post-training time is reported as mean ± sample standard deviation over three training seeds. FID evaluation, backbone pretraining, and scheduler waiting are excluded. † Teacher preparation uses the recorded historical allocation wall time with four concurrent GPUs and is therefore not directly hardware-normalized against the other preparation times.
Method
NFE 1
NFE 2
NFE 4
NFE 8
NFE 10
NFE 50
RF (backbone)
0.166
0.291
0.174
0.115
0.094
0.011
Teacher Reflow (MSE)
0.010
0.021
0.026
0.022
0.023
0.021
One-step distillation (MSE)
0.006
0.246
0.359
0.362
0.360
0.393
Fresh-independent FT
2.154
1.828
0.247
0.136
0.091
0.040
OT-CFM-FT ( Tong et al. (2023) )
0.619
0.370
0.053
0.064
0.038
0.030
AlignFlow ( Kong et al. (2026) )
0.832
0.194
0.101
0.066
0.061
0.041
Appendix
Table 10: Sample standard deviation of CIFAR-10 FID with the Rectified Flow pretrained backbone, corresponding to Table 2 . Standard deviations are computed over three seeds. For the RF backbone row, the standard deviation reflects sampling-seed variability; for the fine-tuned methods, it reflects variability across training seeds.
Method
NFE 1
NFE 2
NFE 4
NFE 8
NFE 10
NFE 50
OT-CFM (backbone)
0.141
0.035
0.161
0.040
0.023
0.010
Teacher Reflow (MSE)
0.149
0.102
0.101
0.114
0.115
0.113
One-step distillation (MSE)
0.012
0.133
0.130
0.160
0.158
0.233
Fresh-independent FT
1.864
1.150
0.309
0.124
0.090
0.010
OT-CFM-FT ( Tong et al. (2023) )
0.596
0.287
0.142
0.110
0.121
0.059
AlignFlow ( Kong et al. (2026) )
1.078
0.462
0.172
0.075
0.054
0.052
Appendix
Table 11: Sample standard deviation of CIFAR-10 FID with the OT-CFM pretrained backbone, corresponding to Table 3 . Standard deviations are computed over three seeds. For the OT-CFM backbone row, the standard deviation reflects sampling-seed variability; for the fine-tuned methods, it reflects variability across training seeds.
Method
Ablation
B
M
R
η
NFE 1
NFE 2
NFE 4
ReCouple
(B,R)
512
1
4
0
137.42±1.04
48.89±0.36
23.48±0.13
512
1
16
0
137.42±1.04
48.89±0.36
23.48±0.13
512
1
32
0
137.42±1.04
48.89±0.36
23.48±0.13
1024
1
4
0
130.07±0.26
47.02±0.05
22.99±0.12
1024
1
16
0
130.07±0.26
47.02±0.05
22.99±0.12
1024
1
32
0
130.07±0.26
47.02±0.05
22.99±0.12
Appendix
Table 12: CIFAR-10 ablations with the RF backbone at low NFEs. For ReCouple, we jointly vary the block size B and number of projection proposals R . For ReCouple+ and ReCouple++, we vary the pooling factor M and spatial-regularization strength η , respectively, while fixing the remaining hyperparameters to their default values. Each cell reports FID as mean ± sample standard deviation over three seeds (lower is better).
Method
Ablation
B
M
R
η
NFE 8
NFE 10
NFE 50
ReCouple
(B,R)
512
1
4
0
16.99±0.17
15.94±0.20
11.49±0.16
512
1
16
0
16.99±0.17
15.94±0.20
11.49±0.16
512
1
32
0
16.99±0.17
15.94±0.20
11.49±0.16
1024
1
4
0
16.53±0.08
15.45±0.08
10.95±0.11
1024
1
16
0
16.53±0.08
15.45±0.08
10.95±0.11
1024
1
32
0
16.53±0.08
15.45±0.08
10.95±0.11
Appendix
Table 13: CIFAR-10 ablations with the RF backbone at higher NFEs. Experimental settings are identical to Table 12 . Each cell reports FID as mean ± sample standard deviation over three seeds (lower is better).
Method
Setting
Rel. cost reduction (%) ↑
Norm. empirical-OT gap ↓
ReCouple
B=512
3.62
0.256
ReCouple
B=1024
3.84
0.211
ReCouple
B=2048
4.04
0.169
ReCouple+
M=2
3.91
–
ReCouple+
M=5
3.96
–
ReCouple+
M=10
3.97
–
Appendix
Table 14: Coupling diagnostics for the CIFAR-10 RF ablations. Relative cost reduction is 100(1−C/Cind) , where C is the average sum of squared pixelwise endpoint differences and Cind is the cost of independent pairing using the same source pool and empirical target marginal for each M . For M=1 , we additionally report the normalized endpoint-cost gap to global empirical OT, (C−COT)/(Cind−COT) . For each B , R∈{4,16,32} yields the same assignment; all ReCouple++ settings share the M=5 ReCouple+ plan.
Flow matching models learn to transport samples from a simple prior distribution to a complex data distribution. When prior-data pairs are coupled via optimal transport (OT), the learned trajectories are straight and non-crossing, enabling fast, even single-step, generation. However, computing the OT coupling in high dimensions is intractable, and existing methods attempt to solve the OT problem, at the cost of persistent bias or significant overhead. Rather than solving for the OT coupling, we reformulate the problem. Once the prior is treated as a design choice rather than a fixed input, the OT coupling between prior and data is no longer unique. Many priors admit an OT-optimal identity coupling to the data, leaving us free to choose one that is also tractable to sample. We identify low-frequency projection of natural images as such a choice. The identity coupling between data and its low-frequency representation is empirically OT-optimal, the prior is structured enough to be sampled by a lightweight model at inference, and the remaining flow-matching task reduces to synthesizing high-frequency detail. Interpolating the prior with Gaussian noise further improves generation quality while preserving the OT coupling. The approach requires no modifications to the flow model itself, and integrates naturally with latent-space models, classifier-free guidance, and one-step generation frameworks. Across all benchmarks, our method reduces trajectory curvature by more than 2× compared to existing flow matching methods, yielding better generation quality in the few-step regime.
In generative models, Optimal Transport (OT) is used to improve Flow Matching (FM) by reducing noise-data coupling cost. However, different noise-to-output assignments can yield nearly equal costs, raising a key question. Is cost alone sufficient to guide coupling design? We address this question by separating transport cost from routing, i.e., the destination reached by each noise sample. We show numerically how FM and OT can differ in routing while remaining close in cost. We examine its consequences in learned neural FM. Using the exact FM routing as an oracle, we further construct a routing-aware training coupling and find that it yields a directionally consistent improvement in generation over a cost-matched, cost-only counterpart. Our findings highlight what cost minimization can overlook and motivate using both cost and routing to evaluate the design of OT-based FM couplings. Code will be released upon acceptance.
Eungyeol Han, Jong-Seok Lee
School of Integrated Technology, Yonsei University
A central challenge in scaling flow models is maintaining stability in training and integration. In flow matching, this stability depends critically on how source samples are paired with data. Independent random pairing introduces substantial stochasticity, motivating minibatch optimal transport (OT) to construct online source-target couplings that reduce this randomness (Tong et al., 2024). However, as applications extend to larger and higher-dimensional datasets, small minibatches can limit these benefits, while scaling the coupling to substantially larger minibatches at every iteration is computationally prohibitive. Moreover, how OT coupling affects the full stochastic gradient and integration remains theoretically unresolved. In this work, we develop a scalable online entropic OT approximation that enables substantially larger source-target couplings while preserving the benefits of OT. We close the theoretical gap for both unregularized and entropic OT by proving strict reductions in full stochastic-gradient variance and integration error relative to independent pairing, and further characterize how these improvements strengthen as the online coupling minibatch grows. Experiments on single-cell gene interpolation-flow and image-flow training validate these predictions, showing that larger online couplings consistently improve training stability and final generation performance.