acceptodds
Under review as a conference paper at ICLR 2027

Coarse-Graining Optimal Transport for Rollout-Free Post-Training of Flow Matching

Abstract

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, the source comes from a continuous distribution, only a finite number of target samples data are accessible, and the neural network has finite capacity. Therefore, making the coupling more OT-like may not consistently improve sampling quality in the low number-of-function-evaluation (NFE) regime. Motivated by this, we introduce ReCouple, a framework that constructs coarse-grained OT couplings, to better understand rollout-free OT-based post-training. ReCouple+ extends this toward the continuous-source setting by increasing source resolution; and it converges to semi-discrete OT in the appropriate limit. ReCouple++ further adds spatial regularization, motivated by the observation that straight trajectories need not correspond to spatially simple velocity fields. Empirical validations demonstrate their effectiveness, while revealing an NFE-dependent interplay between coupling optimality, neural realizability, and generation quality.

Primary area:
generative models
Keywords:
Flow matching, optimal transport, coupling design, few-step generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_e34bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ReCouple, a blockwise OT coupling framework for rollout-free flow-matching post-training. ReCouple+ increases source resolution, and ReCouple++ adds annealed spatial regularization. Theoretical analysis and experiments show that coupling refinement affects generation differently across sampling budgets.

Critical strengths
  1. The framework separates transport geometry, source resolution, and spatial regularization, providing a useful way to study coupling design. The convergence analysis clearly distinguishes global SDOT from the fixed-block limit and makes the approximation's scope explicit.
  2. The CIFAR-10 experiments include two pretrained backbones and three training seeds. The ablations demonstrate a useful NFE-dependent trade-off: increasing source resolution can improve higher-NFE performance while worsening one-step generation. This supports the practical insight that endpoint transport cost alone is insufficient to rank trained generators.

Informativeness: The controlled coupling hierarchy provides useful insight into how transport resolution interacts with the sampling budget.

Overall assessment: The explicit coupling hierarchy, convergence characterization, and systematic NFE-dependent ablations provide useful methodological and empirical insight. These contributions support acceptance, while the remaining concerns limit the strength of the mechanistic explanation and component-level performance claims.

Critical weaknesses
  1. The scope of the mechanistic explanation should be clearer. Equation (7) assumes a smooth positive-density state distribution, whereas training uses fixed discrete pair banks. Its connection to these experiments remains heuristic. Measurements of finite-resolution ambiguity or learned spatial Jacobians would strengthen this interpretation.
  2. Additional controls would clarify the contribution of each component. Baselines with the same spatial penalty are missing. On ImageNet at NFE 10, FID improves from 10.637 for ReCouple to 7.051 for ReCouple-AS, while ReCouple++ reaches 6.960 (Table 4). The regularization benefit is encouraging, but the incremental benefit of pooling needs uncertainty estimates, and its advantage over regularized alternatives remains open.

Soundness: The main concerns involve mechanistic interpretation and attribution of gains. The coupling construction and the observed NFE-dependent trade-offs are reasonably supported.

Questions for authors
  1. Could the authors clarify the applicability of Equation (7) to discrete training banks and the evidence supporting the spatial-realizability explanation?
  2. Are results available for minibatch OT or SD-FM with the same regularization and comparable computational budgets?
  3. What is the run-to-run uncertainty of the ImageNet improvement over ReCouple-AS?
1 comment
AuthorsOpenReview ↗

Thank you for your review

Thank you for your positive feedback and constructive suggestions, which have helped us improve the manuscript. We have uploaded a revised version, with all changes highlighted in blue. We have also rerun the ImageNet-256 experiments with the public SiT-XL/2 flow matching backbone, a common Euler sampler and evaluator, 50K generated images per evaluation, and both CFG 1.5 and 4. The revised Sec. 4.3 (Tables 4–6) now contains full NFE sweeps and a complementary fixed-6K, three-training-seed NFE-10 comparison; App. E.4 (Tables 17–21) reports matched regularized controls, precision/recall, Jacobian diagnostics, and cost.

  • Equation (7) and empirical-bank applicability. Equation (7) is a fixed-time population small-noise expansion under a smooth positive-density state law, finite Fisher information, and bounded field derivatives (Sec. 3.3; App. D.6). It is not asserted for an atomic training bank: In a finite bank of affine paths, two paths with different velocities can coincide at at most one time, so conditional velocity ambiguity vanishes almost everywhere in time. This is precisely why we introduce finite-resolution risk and a separate learned-field penalty. As empirical evidence, our common-query Jacobian measurements show that AS reduces the estimated squared guided-field Jacobian norm for ReCouple at t=0.9t=0.9 from 15.459 to 3.729 (CFG 1.5) and 16.764 to 4.161 (CFG 4) (App. E.4, Table 20). CFG-1.5 FID nevertheless worsens at that checkpoint (Table 19). The diagnostic supports the intended control of learned-field spatial variation, without claiming that the population expansion applies directly to atomic banks or that lower Jacobian norm guarantees lower FID.

  • Regularized alternatives and compute. The matched OT+AS and SD-FM+AS controls and their FIDs are now provided in App. E.4, Table 17; the separate active-penalty results appear in Table 19. Regularized post-training requires extra network evaluations: the ImageNet timings are 3.780 versus 6.192 hours for minibatch OT without/with AS and 4.248 versus 5.489 hours for SD-FM (Table 21) on an NVIDIA GH200.

  • Uncertainty of the pooling increment. In the three-seed NFE-10 ImageNet comparison, for the paired contrast Δ=FID(ReCouple++)−FID(ReCouple-AS)\Delta=\mathrm{FID}(\mathrm{ReCouple{++}})-\mathrm{FID}(\mathrm{ReCouple\text{-}AS}), the mean differences are +0.036 at CFG 4 (with 95% confidence interval [−0.081, +0.153]) and +0.020 at CFG 1.5 ([−0.447, +0.487]). Neither interval establishes an incremental advantage from pooling at this checkpoint. The intervals reflect three paired optimization seeds and are not uncertainty estimates for the NFE sweeps (App. E.4, Table 17).

Reviewer_Yor5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces ReCouple, ReCouple+ and ReCouple++, a family of method to post-train flow matching (FM) models using optimal transport (OT) based coupling. The method is theoretically grounded and provides an optimality result of ReCouple+ as well as presents empirical validation on a synthetic dataset and image generation on CIFAR-10 and ImageNet datasets.

Critical strengths
  • The novel framework that provides controls over coupling geometry through block size, source resolution through pooling, and spatial regularization.
  • A theoretical result that establishes convergence to semi-discrete optimal transport in the appropriate limit.
  • Empirical validation that showcases the effectiveness of the method and achieves comparable and superior results in some cases.
Critical weaknesses

While the proposed methodology is theoretically interesting and sound, the empirical validation, performance attribution, hyperparameter selection procedures, and large-scale statistical reliability are insufficiently established.

  • The proposed methodology for a few steps refinement does not consistently transfer to improved practical few-step generation: For example, on CIFAR-10 ReCouple+ and ReCouple++ perform substantially worse than plain ReCouple at NFE 1-4. On ImageNet, the ReCouple-AS and ReCouple++ seem to be worse than the plain ReCouple at NFE 1-2. Some of these considerations are acknowledged in the manuscript, however it limit to the practical scope of the overall framework.
  • Statistical significance of the results: The evaluation metrics, only include mean estimates of the performance and standard deviations are not reported, which does not establish a reliable superiority of the different methods.
  • Hyperparameter selection strategies are underspecified: The manuscript does not explain the process of choosing the hyperparameters for ReCouple, it is unclear if they were tuned ona a separate validation set to prevent contamination onto the evaluation set.
  • Attribution of gains. In some cases the improvements seem to be coming from when adding the spatial regularization, however the manuscript does not consider the performance of the related methods (like minibatch OT and SD-FM) with the same regularization to establish what gains are coming from ReCouple itself.
Questions for authors
  • Can the authors clarify the hyperparameter selection strategy and if a separate validation set was used for tuning the hyperparameters?
  • Can the authors provide std along with the mean estimates of the metrics to establish the significance of the results?
  • Can the authors report how related methods would perform with the same spatial regularization?
3 comments
AuthorsOpenReview ↗

Thank you for the careful review

Thank you for your thoughtful feedback and constructive suggestions, which have helped us improve the manuscript. We have uploaded a revised version, with all changes highlighted in blue. We have rerun the ImageNet-256 experiments with the public SiT-XL/2 flow matching backbone, a common Euler sampler and evaluator, 50K generated images per evaluation, and both CFG 1.5 and 4. The revised Sec. 4.3 (Tables 4–6) now contains full NFE sweeps and a complementary fixed-6K, three-training-seed NFE-10 comparison; App. E.4 (Tables 17–21) reports matched regularized controls, precision/recall, Jacobian diagnostics, and cost.

  • Inconsistent transfer to few-step generation. We agree that increasing coupling resolution does not yield a uniform practical improvement. CIFAR-10 exhibits a low-NFE versus high-NFE crossover as M increases (App. E.3, Tables 14–16), while ImageNet shows guidance and checkpoint dependence (Sec. 4.3). Our contribution is a controllable rollout-free OT approximation, its limiting connection to SDOT, and evidence that endpoint transport cost alone does not predict generation quality for a learned field and finite sampling budget. The results support studying these trade-offs rather than a universally better few-step generator.

  • Statistical reliability. We now report mean ± sample SD over three training runs for the 2D task (Table 1 and App. E.1), over three post-training seeds for CIFAR-10 (Tables 2–3 and App. E.2, Tables 12–13), and over three optimization seeds for every main ImageNet post-trained method in the fixed-6K NFE-10 study (Table 6; App. E.4, Table 17). The unchanged ImageNet backbone instead varies sampling seeds. Every reported FID for ImageNet uses 50K generated images; precision/recall and matched one-seed controls are reported separately (App. E.4, Tables 17–18).

  • Hyperparameter selection. We provide details on the block size, projection count, pooling and regularization parameters and sensitivity studies (App. E.1–E.4). The reference settings were chosen heuristically to facilitate controlled ablations of the block size BB, projection count RR, pooling factor MM, and spatial regularization strength. We did not use a separate held-out real-image set for hyperparameter tuning. The ablations in App. E.3 characterize the sensitivity to these choices and their NFE-dependent trade-offs. The ImageNet sweeps select an EMA checkpoint by NFE-5 benchmark FID, separately at each CFG. By contrast, the three-seed study uses a fixed 6K checkpoint with no FID-driven selection. These protocols are explicitly separated. For ImageNet, B=N=4096B=N=4096 within each class, so projection search is redundant; this setting tests pooling and regularization at full-block resolution rather than a multi-projection benefit (App. E.4).

  • Attribution to spatial regularization. The revision contains matched minibatch OT+AS and SD-FM+AS controls under the same ImageNet penalty recipe (App. E.4, Table 17). At CFG 4, matched NFE-10 FID changes from 10.026 to 9.859 for minibatch OT and 9.548 to 9.535 for SD-FM; at CFG 1.5, both worsen. These comparisons are descriptive and show that regularization is not exclusive to ReCouple. A complementary fixed-coupling CIFAR-10 ablation isolates the effect of regularization more directly: holding the pooled plan at M=5M=5, increasing its calibrated penalty ratio from 0 to 1 improves NFE-10 FID from 11.04 ± 0.11 to 10.33 ± 0.15, and NFE-50 FID from 5.76 ± 0.04 to 5.00 ± 0.13 (App. E.3, Tables 14–16). Thus spatial control can improve generation without changing the coupling, although its FID benefit is not universal on ImageNet.

Reviewer_Yor5OpenReview ↗

Reviewer reply

Thank you for your response and for the additional experiments.

  1. Few-step generation. Thank you for the clarification, however I think the inconsistent improvements across NFEs outlined in the initial review remains a limitation of the practical scope of the framework.

  2. Statistical significance. I acknowledge the additional ImageNet experiments. However, it is unclear why the full NFE sweeps and regularized competing-method controls remain single-seed, so the reliability of these comparisons remains insufficiently established. Reporting the available standard deviations with the means in the main part of the manuscript would improve the clarity and presentation.

  3. Hyperparameter selection. I appreciate the clarification of the selection process and the hyperparameter sensitivity studies. However, it remains unclear how the hyperparameters were chosen for the final evaluation. To what extend the choses were informed by the evaluation metrics, it seems like in some cases the checkpoint selection relyed on the evaluation set, does this lead to inflated performance metrics?

  4. Attribution of gains. The matched minibatch OT+AS and SD-FM+AS experiments address my request.

With the above considerations, I will maintain my score.

Thank you for the follow-up! We address the remaining concerns below.

On 1. The absence of uniform gains across NFEs is a practical limitation that our study explicitly shows. Our controlled experiments show that improving transport fidelity alone does not ensure better generation. They also demonstrate benefits: in the 3-seed CFG-4 study (see below), all four ReCouple variants achieve lower mean FID than the backbone and minibatch OT at every evaluated NFE. Plain ReCouple has the lowest mean at NFE 5 and 10, while ReCouple++ leads at NFE 2, 25 and 50. These results are meaningful, support the framework’s value, and characterize how coupling choice and spatial regularization interact with the sampling budget.

On 2. The remaining 3-seed evaluations were still running when we submitted our previous response. They are now complete: Table A below reports the full CFG-4 sweep, including ReCouple+. The previously reported three-seed CFG-1.5 NFE-10 results are unchanged. The regularized controls and the CFG-1.5 sweep remain single-seed comparisons but we are happy to provide the full 3-seed results (means and standard deviations) in the final version.

Table A. ImageNet-256 CFG-4 sweep (runs are over 3 random seeds).

MethodNFE 1NFE 2NFE 5NFE 10NFE 25NFE 50
SiT backbone275.411 ± 0.22098.925 ± 0.09410.163 ± 0.04612.215 ± 0.01712.711 ± 0.06212.614 ± 0.123
Minibatch OT250.963 ± 0.49487.640 ± 0.3718.793 ± 0.04110.079 ± 0.10410.539 ± 0.09810.448 ± 0.081
SD-FM233.104 ± 0.95181.019 ± 1.2488.255 ± 0.0739.445 ± 0.10210.006 ± 0.17810.033 ± 0.226
ReCouple235.772 ± 1.68481.828 ± 1.1618.162 ± 0.0869.410 ± 0.02910.006 ± 0.03110.007 ± 0.068
ReCouple-AS234.378 ± 0.45980.810 ± 1.3518.322 ± 0.0359.468 ± 0.0309.988 ± 0.0359.932 ± 0.028
ReCouple+235.784 ± 0.64782.166 ± 0.3148.237 ± 0.0989.479 ± 0.09010.066 ± 0.11210.038 ± 0.147
ReCouple++234.425 ± 0.45380.035 ± 0.3218.297 ± 0.1009.504 ± 0.0779.952 ± 0.0629.909 ± 0.106

*Recall that SD-FM provides a semi-discrete OT baseline approached by ReCouple+ in the limit studied in the paper.

On 3. Reference hyperparameters were chosen heuristically and fixed before the additional-seed runs. For ImageNet, these include B=N=4096 within each sampled class pool, M=2 for pooled methods, perturbation scale 0.05 times latent RMS, and initial penalty-to-CFM ratio 0.25. They were held constant across seeds, NFEs and guidance scales. No separate held-out real-image tuning set was used.

The earlier single-seed sweeps selected checkpoints by NFE-5 FID using 10K generated images and the benchmark reference statistics, which can produce optimistic selected estimates. Table A instead reports means and sample SDs over three runs, with each run evaluated at its fixed 6k-update EMA checkpoint, without FID-driven checkpoint selection.

The checkpoint distinction explains changes from the earlier sweep: ReCouple++’s CFG-4 NFE-50 FID was 9.362 at its selected 4K checkpoint, versus 9.909 ± 0.106 at fixed 6K over three seeds. We will clearly separate these protocols and incorporate the completed tables and corresponding qualifications into the final version.

We hope that the above addresses your remaining concerns and would appreciate your reconsideration of the overall assessment.

Reviewer_NM2fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReCouple, a rollout-free post-training framework that constructs coarse-grained OT couplings for flow matching. The authors further introduce ReCouple+ to increase source resolution and ReCouple++ to add spatial Jacobian regularization. The methods are evaluated on 2D multimodal generation, CIFAR-10, and ImageNet-256.

Critical strengths
  1. Interesting motivation and analysis. The paper studies how coupling optimality, neural realizability, and generation quality interact. The analysis of SDOT shows that straight trajectories can still require spatially sharp velocity fields, which explains why a more OT-like coupling does not always help few-step generation.

  2. Broad experiments and useful ablations. The methods are tested from 2D data to ImageNet-256 with SiT-XL/2. The ablations on block size, pooling factor, and spatial regularization clearly show the NFE-dependent behavior of different couplings.

Critical weaknesses
  1. Limited improvements in the ultra-low-NFE regime. No ReCouple variant produces usable samples at NFE 1–2. On CIFAR-10, all variants have FIDs above 100 at NFE 1 (Table 2). On ImageNet-256, ReCouple++ is worse than SD-FM at NFE 1–2 (Table 4). Moreover, no single variant works well across NFEs. ReCouple is better at low NFEs but nearly doubles the backbone FID at NFE 50 on CIFAR-10, while ReCouple++ shows the opposite trend.

  2. High computational cost relative to the improvements. On CIFAR-10, ReCouple++ (477.49 +171.85 min) requires about 6.6× the total preparation and post-training time of OT-CFM-FT (0 + 97.73 min) for a single run (Table 9), but improves FID at NFE 50 only from 5.14 to 4.78 (Table 2). The paper does not report the cost on ImageNet-256.

  3. Non-standard evaluation on ImageNet-256. All ImageNet results use a classifier-free guidance scale of 4.0. Under this setting, the SiT-XL/2 baseline has an FID of 12.6 at NFE 50, far from the FID of about 2 reported under standard settings [1], and its FID even gets worse as NFE increases. This suggests that the high guidance scale reduces sample diversity. Since ReCouple mainly improves recall (Table 16), its FID gains may come from offsetting over-guidance rather than better generation. Results under the standard guidance scale are needed.

[1]. SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers, ECCV 2024

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Thank you for your review

Thank you for your feedback and constructive suggestions, which have helped us improve the manuscript. We have uploaded a revised version, with all changes highlighted in blue. We have rerun the ImageNet-256 experiments with the public SiT-XL/2 flow matching backbone, a common Euler sampler and evaluator, 50K generated images per evaluation, and both CFG 1.5 and 4. The revised Sec. 4.3 (Tables 4–6) now contains full NFE sweeps and a complementary fixed-6K, three-training-seed NFE-10 comparison; App. E.4 (Tables 17–21) reports matched regularized controls, precision/recall, Jacobian diagnostics, and cost.

  • NFE 1–2 quality and cross-budget trade-offs. We retain NFE 1–2 results as diagnostic comparisons. The high absolute FIDs at NFE 1–2 are unsurprising for these undistilled flow-based models sampled with one or two coarse Euler steps: their training loss does not directly optimize a one-step flow map. This is a limitation of the tested setup (which we intentionally study and expose here) rather than of flow-based generation generally; the teacher-based reflow and one-step-distillation controls achieve much better one-step quality, with distillation explicitly targeting that budget (Tables 2–3).
    Even in this regime, coupling choice matters. In the rerun ImageNet CFG-4 experiments, ReCouple++ has the lowest reported FID at every NFE from 2 through 50, including 78.840 versus 99.025 for the backbone at NFE 2 and 9.362 versus 12.578 at NFE 50. At CFG 1.5, ReCouple-AS leads at NFE 1 and ReCouple++ at NFE 2, although absolute one- and two-step quality remains poor (Sec. 4.3, Tables 4–5). On CIFAR-10, ReCouple++ attains the lowest mean NFE-50 FID among the evaluated rollout-free post-training methods under both backbones, while plain ReCouple is better at NFE 1–2 (Tables 2–3). These crossovers are central evidence for our thesis: transport resolution must be considered jointly with neural realizability and numerical resolution, not optimized in isolation.

  • Computational cost. We agree that the most heavily pooled configuration is expensive. On CIFAR-10/RF, ReCouple++ requires 477.49 + 171.85 = 649.34 minutes of preparation and mean post-training versus 97.73 minutes for OT-CFM-FT (6.64×), for NFE-50 FID 4.781 versus 5.137 on an NVIDIA GH200 (App. E.2, Table 11). However, ordinary ReCouple's coupling preparation takes 30.53 minutes versus 542.46 minutes for global empirical OT in the same implementation, a 17.8× reduction; the hierarchy explicitly exposes the cost of increasing source resolution. We now also report ImageNet post-training times (3.623 hours for ReCouple; 5.641 hours for ReCouple++), excluding one-time preparation and evaluation (App. E.4, Table 21).

  • Guidance and diversity. The reruns directly address the concern about CFG 4. In the three-seed NFE-10 ImageNet comparison, ReCouple improves FID over the backbone at CFG 4 (9.410 ± 0.029 versus 12.215 ± 0.017), but not at CFG 1.5 (7.941 ± 0.078 versus 7.535 ± 0.088); the unchanged backbone’s NFE-50 FID falls from 12.578 at CFG 4 to 2.477 at CFG 1.5, where it outperforms all considered post-trained methods. The precision/recall analysis is informative and gives more insight: at NFE 10, ReCouple raises recall from 0.275 to 0.348 at CFG 4 and from 0.525 to 0.569 at CFG 1.5, with lower precision at both scales (App. E.4, Table 18). Thus, the coupling changes the precision–recall trade-off under both guidance settings, although improved FID is guidance-dependent. We explicitly distinguish these results from published SiT evaluations with different numerical solvers and sampling budgets (Sec. 4.3; App. E.4).

Overall, our main contribution is a rollout-free, controllable transport framework—blockwise coupling resolution (BB), source pooling (MM), and spatial-field regularization—together with theory and experiments that identify when improving the coupling helps a finite neural flow at a given numerical budget. We test this mechanism directly rather than claim that one variant dominates at every NFE, guidance scale, or task.

Thank you for your detailed response and additional experimental results, which resolve part of my concerns. However, the limited improvements at ultra-low NFEs, substantial computational overhead, and the lack of consistent gains under CFG 1.5 remain concerning. So I will maintain my rating.

Thank you for the follow-up. We would like to clarify a few points.

While absolute sample quality at NFE 1–2 remains limited for the Euler-sampled flow models considered here, plain ReCouple actually leads to substantial relative improvements in this regime on CIFAR-10, compared with the OT-CFM-FT and SD-FM baselines (Tables 2–3):

BackboneNFEOT-CFM-FTSD-FMReCoupleReCouple++
RF1232.657159.871123.159187.080
RF290.18158.98245.72267.905
OT-CFM1228.462157.251121.547184.021
OT-CFM287.27357.95845.01865.635

FID (lower is better); three-seed means. Standard deviations are reported in App. E.2, Tables 12–13.

Across these comparisons, ReCouple lowers FID by approximately 22–23% relative to SD-FM and 47–49% relative to OT-CFM-FT. ReCouple++ instead performs better at higher NFEs on CIFAR-10, reaching NFE-50 FIDs of 4.781 (RF) and 4.342 (OT-CFM). This crossover is an important finding: coupling refinements that improve higher-step generation need not improve one- or two-step sampling. On ImageNet-256, ReCouple++ achieves the lowest reported NFE-2 FID at both CFG 4.0 and 1.5, while ReCouple-AS achieves the lowest NFE-1 FID at CFG 1.5 (Tables 4–5). These comparisons further illustrate that the preferred coupling configuration depends on the sampling budget and guidance scale.

The computational overhead also differs substantially across configurations. For CIFAR-10/RF, ordinary ReCouple requires 30.53 minutes of one-time coupling preparation and 105.56 minutes of post-training (136.09 minutes for one run, versus 97.73 minutes for OT-CFM-FT). The much larger ReCouple++ cost arises from source pooling and spatial regularization (Table 11). We intentionally expose and report this cost–quality trade-off rather than claiming an equal-compute advantage.

Regarding CFG 1.5, we agree that the absence of consistent FID gains is a limitation, but it also reveals an informative trade-off. In the ImageNet-256 comparison at NFE 10, the pretrained backbone has lower FID than ReCouple (7.535 versus 7.941), even though ReCouple improves recall from 0.525 to 0.569, accompanied by a reduction in precision from 0.739 to 0.711 (Tables 6 and 18). Thus, changes in coupling can improve distributional coverage without improving FID, and their effectiveness depends on the guidance setting. We view this guidance-dependent behavior as an important empirical finding that helps delineate when improved transport coupling translates into better generation.

More broadly, our main contribution is a principled, rollout-free framework for constructing and studying transport couplings, with explicit controls over coupling resolution, source pooling, and spatial regularity, supported by theory and controlled experiments. Beyond the demonstrated empirical improvements, the framework helps explain why better transport approximation does not necessarily yield improved few-step generation. The observed NFE-dependent trade-offs and guidance sensitivity are therefore central findings of our work.

We hope that the above addresses your remaining concerns and would appreciate your reconsideration of the overall assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.