acceptodds
Under review as a conference paper at ICLR 2027

CyFM: Cylindrical Optimal Transport for Few-Step Complex-Valued Flow Matching

Abstract

Complex-valued MRI and audio signals are commonly modeled as two Cartesian channels, although their phase is circular. In Cartesian flow matching, straight paths between nearly opposite phases can pass close to the origin and induce arbitrarily large angular velocities. We formulate Cylindrical Flow Matching (CyFM) on amplitude and phase coordinates with the decoupled metric . Shortest-arc phase interpolation bounds the angular training target by , while exact minibatch Optimal Transport (OT) pairs noise and data jointly over whole fields. For equal-amplitude Cartesian endpoints with uniform phase differences, we derive an index-1 Pareto tail in peak angular velocity. Across exact bridges on synthetic copula fields, LibriSpeech spectrograms, and fastMRI knee wavefields, 42.6%-49.4% of target signal energy lies on independently coupled paths whose peak angular velocity exceeds ; no cylindrical target exceeds this bound. Under matched training budgets, CyFM with joint OT achieves lower pooled sliced than the best Cartesian baseline at every tested solver step on synthetic fields at three resolutions and on speech spectrograms. On knee MRI, its pooled error is 1.8x lower at one step; amplitude texture favors CyFM from four steps, and phase coherence from eight steps. Joint OT reduces synthetic few-step error by 3%-60%, whereas factorizing the coupling disrupts joint dependencies. At full MRI resolution, CyFM preserves a 2.1x single-step advantage, demonstrating that cylindrical geometry and joint transport provide an effective inductive bias for rapid, few-step generation of complex-valued physical fields.

Primary area:
generative models
Keywords:
flow matching, optimal transport, continuous normalizing flows, geometric deep learning, complex-valued representations, manifold learning, generative models

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_8FAbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies few-step generation of complex-valued fields. It models amplitude and phase using cylindrical geometry and combines this with optimal-transport coupling. The authors analyze angular-velocity behavior near zero amplitude under straight-line interpolation and compare paths and couplings on synthetic data, complex MRI fields, and audio representations. The task is distribution generation, rather than MRI reconstruction from measurement constraints.

Critical strengths
  1. The paper identifies a concrete, testable geometric problem. Under equal amplitudes and uniformly distributed phase differences, Cartesian straight-line paths produce heavy-tailed angular velocities, whereas shortest-arc paths avoid this particular conditional-path problem. Crossed ablations of geometry and coupling also help separate their effects. This provides an interpretable insight whose value goes beyond a score improvement.
  2. The failure cases expose an important limitation of the evaluation metrics. On high-resolution MRI, better pooled pixel-distribution scores at one sampling step do not translate into better spatial structure. The patch-reassembly experiment also illustrates dependencies that pooled distribution measures can miss. These results do not establish overall success, but they provide useful diagnostics and show that the authors report unfavorable findings.
Critical weaknesses
  1. The conditional-path properties are not yet sufficiently connected to the few-step advantage of the learned generator. Flow matching on manifolds and optimal-transport coupling have direct precedents. The added contribution must therefore come from what this specific geometry achieves for complex fields. Bounded angular velocity along a shortest-arc path does not imply bounded network predictions or smaller few-step integration error. The Cartesian network regresses endpoint differences, rather than the angular velocity induced by the straight path. The paper reports that network predictions can exceed the path bound and includes a tanh-constrained experiment, which are useful controls. However, changing geometry also changes the path, parameterization, and metric used to measure velocity error. The experiments do not yet directly connect these changes to errors in the learned field. The more-stable-regression claim on page 6 and parts of the mechanistic attribution on page 17 should therefore be limited to what is established.
  2. The main distribution metrics do not establish better generation of complete complex-valued fields. Sliced W2 over pooled pixels is insensitive to spatial rearrangement and cannot by itself characterize dependencies between locations. In the 320×320 MRI experiments, CyFM is worse than the reported Cartesian controls on all spatial metrics at every evaluated step count. This directly limits the main quality claim. The evidence supports better few-step fitting of certain marginal distributions, but not an overall improvement in complete MRI structure or audio-waveform quality. 3.In the k=4 column of Table 6 on page 17, lower Sliced W2 is better. CyFM+OT has 0.1121, while Cartesian+OT has 0.1166, but the latter is boldfaced. Please correct the best-value marker. This is a table-formatting error. It does not change the reported numbers or independently determine my recommendation.
Questions for authors

1.Which controlled results directly support easier few-step integration of the learned field, beyond improved conditional target angular velocities? 2.Given the deterioration in spatial metrics on 320×320 MRI, what is the main generation-quality claim that the authors consider established? 3.Do the few-step and speedup claims refer to integration steps, network evaluations, or measured runtime? How do differences in evaluation cost affect the comparison?

2 comments

We thank the reviewer for the precise, constructive review and for recognizing both the testable geometric problem and the insights from our crossed ablations and failure analyses. All revisions below are incorporated in the updated manuscript.

  • W1: Our novelty is what the cylindrical geometry gives complex-valued fields, while the manifold and OT machinery are prior work that we cite (Chen & Lipman, 2024, and Bose et al., 2024). We contribute (i) a proof that the cylindrical angular target is bounded by π\pi for any coupling, against an index-1 Pareto tail for Cartesian bridges, with 42.6–49.4% of signal energy above π\pi on real data (Table 1), (ii) the finding that whole-field OT must stay joint, because factorising it over amplitude, phase, or patches changes the target distribution (Section 5.4), and (iii) an exact one-step result (Proposition, Appendix B): one solver step keeps the mean data amplitude on the cylinder but shrinks it in the plane, which gives a 1.7–2.4× lower one-step amplitude error in all four blocks. At 320×320320\times320, where OT is inert, the one-step advantage holds even without coupling (pooled 0.058 vs. 0.135), so it comes from the geometry. All of this runs on a standard real-valued U-Net, because the geometry lives in the transport.

  • W1, Q1: We narrowed three claims in the revised manuscript where conditional-path bounds went beyond what was established for the learned field:

    • Page 6: removed "This bounded target provides an inherently more stable regression objective for neural networks," since the bound holds for conditional targets rather than network predictions (Appendix G).
    • Page 18 (item 1, submitted Page 17): replaced "Bounding the angular target by π\pi therefore still buys a single-step advantage at 102,400 dimensions" with "The cylindrical geometry therefore retains a single-step advantage on the pooled metric at 102,400 dimensions," because the gain sits in amplitude rather than the angular bound (Appendix B).
    • Pages 18–19 (item 3, submitted Page 17): replaced "the failure is a property of the coordinate representation, not of the pairing" with "[the OT collapse] does not explain the failure, and neither does the coupling," since no experiment isolates the representation as the cause.

    To answer Q1 directly, Appendix B proves an exact link to the learned field: for any input parameterisation, one Euler step returns the conditional mean of the endpoint, which attenuates amplitude in the plane under phase uncertainty but preserves it on the cylinder. Measured in data space, one-step amplitude error is lower for CyFM across all four blocks with seeds separated.

  • W2: Pooled sliced W2W_2 cannot characterize spatial dependencies by itself, as the patch-reassembly experiment shows (the metric reads 0.000 while cross-seam correlation drops from 0.979 to 0.044, Section 5.4). That is why our knee tables pair it with amplitude lag-one, circular phase lag-one, and radial-spectrum gaps, and why our revised claim keeps marginal and spatial results separate.

  • W2, Q2: The established claim is better few-step fitting of pixel-marginal distributions at matched network evaluations, carried by amplitude: at every k≤8k \le 8 on synthetic fields and speech, and at k=1k=1 on knee MRI at both resolutions (Appendix K). On knee 64×6464\times64, CyFM also leads from k=4k=4 in 7 of 9 comparisons across Table 2's three spatial measures. As stated in the conclusion, we claim no overall improvement in complete 320×320320\times320 structure or in audio quality. At 320×320320\times320, an endpoint phase target (new Appendix J, see our reply to Reviewer 8t4Q on 320×320320\times320) shrinks the spatial gaps several-fold, leaving CyFM no worse than any Cartesian arm on any spatial measure at k=8k=8.

  • W3: Thank you for catching this. We have corrected the bold marker in Table 7 (submitted Table 6).

  • Q3: Both "few-step" and the "25×25\times fewer solver steps" in Section 5.2 refer to solver steps kk. Because every flow arm uses the same Heun solver, kk steps cost 2k−12k - 1 network evaluations of the same U-Net in every arm, and one evaluation takes the same time in both geometries: averaged over kk, 0.515 vs. 0.513 ms on knee 64×6464\times64, with a cylinder/plane ratio of 1.00 also at 320×320320\times320 and with the larger backbone (one RTX 4070 Ti SUPER, measured for this reply). Step-count reductions therefore carry over directly to wall-clock time. All three units on knee 64×6464\times64:

Solver steps kk14100
Network evaluations (both)17199
ms per field, CyFM / Cartesian0.52 / 0.523.61 / 3.58102.0 / 101.7

We hope these revisions and clarifications address the reviewer's questions.

Thank you for authors' response and clarifications which address several of my concerns. I would keep its original rate.

Reviewer_8t4QOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, the authors focus on flow matching for complex-valued fields such as MRI wavefields and STFT spectrograms. Instead of standard Cartesian (Re, Im) interpolation, the authors propose Cylindrical Flow Matching (CyFM). Each pixel is represented in amplitude-phase coordinates on [0, inf) x S1

Critical strengths
  • The authors identify a limitation of Cartesian interpolation for complex-valued signals and motivate the use of a cylindrical manifold. Their results offer a useful theoretical perspective on the interaction between signal geometry and Flow Matching dynamics.
  • Sec 3 and Appx A show that the method substitutes a metric by deleting the A^2 factor, rather than inheriting a geometry.
  • Appreciate authors reporting the 320 x 320 reversal, the unexplained k = 1 to k = 2 rise (Appx G), the two-epoch speech probe (Appx F), and the post-selection caveat in Tab. 5
  • The authors evaluate their approach on synthetic complex-valued fields, audio STFT representations, and fastMRI knee data (several domains)
Critical weaknesses
  • L320-321 The authors mention dataset in generic terms "synthetic copula fields, LibriSpeech speech STFT, and fastMRI knee". No citation provided, so difficult to judge the results.
  • Can authors disucss if their method can scale beyond 64 x 64 ?
  • Table 2 doesn't provide standard deviations, hence its difficult to understand if results are statistically significant. Did the authors perform t-test? This is because Cartesian outperforms/is comparable in several cases too
  • Pooled sliced W2 compares pixelwise marginal distributions and ignores spatial arrangement. Appx G states that at k = 1 “a single Euler step transports amplitude almost completely while phase barely evolves”. The generated phase at k = 1 is therefore close to the uniform prior.
  • L26 says "amplitude texture favors CyFM from four steps, and phase coherence from eight steps." but in Table 2, we see Cartesian better at k=4 in some cases
  • Can the authors include generated MRI visualizations, reconstructed audio examples? This is for qualitative assessment
  • For quantitative assessment, authors need to report domain-specific metrics like psnr / ssim for MRI, etc. Currently they only report quantitative metric as pooled sliced Wasserstein-2 distance
  • Why does CyFM underperform on full-resolution MRI?
Questions for authors

please see weakness

1 comment

We thank the reviewer for the careful reading, including the appendices. Every revision below is in the updated manuscript.

  • Citations: We added a citation for each dataset in the main text: LibriSpeech (Panayotov et al., 2015), fastMRI (Zbontar et al., 2018, and Knoll et al., 2020), and ESPIRiT coil combination (Uecker et al., 2014). The synthetic copula fields are our own, and Appendix C gives their generator and all splits.

  • 320×320: CyFM scales beyond 64×6464\times64. Appendix E trains and evaluates on the full 320×320320\times320 matrix (102,400 complex dimensions), where CyFM keeps its single-step advantage (0.056 vs. 0.120, seeds separated). Minibatch OT's cost reduction falls to 0.5% at this size (Section 5.3), a general limit of OT-coupled flows (Weed & Bach, 2019). Part of the spatial gap came from the angular-velocity target of the main tables: its least-squares optimum is a conditional mean, so where training pairs turn by nearly +π+\pi and −π-\pi, the turns cancel and those pixels move slowly. The new Appendix J retrains every arm with an endpoint phase target, (cos⁡θ1,sin⁡θ1)(\cos\theta_1, \sin\theta_1), whose optimum (the circular mean) does not cancel, and the Cartesian arms are retrained the same way as a control. CyFM + OT then leads on pooled sliced W2W_2 at every k≤8k \le 8 with seeds separated from all four Cartesian arms, its spatial gaps shrink several-fold, and at k=8k=8 it is no worse than any Cartesian arm on any spatial measure. Appendix E names a candidate cause for the remaining gap: unmeasured background (16.7% of coefficients are exactly zero).

320×320k=1k=2k=4k=8k=100
Pooled sliced W₂0.0707* / 0.12030.0738* / 0.12230.0620* / 0.11660.0438* / 0.08720.0538 / 0.0629
Amplitude texture0.1815 / 0.0582*0.0283 / 0.0051*0.0027 / 0.00480.0011* / 0.00790.0006 / 0.0014
Phase coherence0.0438 / 0.0089*0.0718 / 0.0297*0.0689 / 0.0521*0.0711 / 0.07330.0771 / 0.0671*
Radial spectrum1.0331 / 0.8160*0.4441 / 0.1649*0.1735 / 0.1372*0.1564 / 0.16620.0990 / 0.0706*

CyFM + OT (endpoint) / best of four Cartesian arms, mean of 5 seeds, lower is better, * marks seeds separated (p=0.008p = 0.008).

  • Table 2: We added both standard deviations and significance tests. Table 2, Table 7 (320×320320\times320), and the values behind Figure 2 now report the five-seed mean ± standard deviation. The new Appendix K gives Welch's tt-test and the exact Mann–Whitney test of CyFM + OT against both Cartesian arms for every block, metric, and kk (for example, knee 64×6464\times64 at k=1k=1: 0.0593±0.00330.0593 \pm 0.0033 vs. 0.1062±0.00100.1062 \pm 0.0010, Welch p=1.4×10−6p = 1.4\times 10^{-6}). The asterisks in Table 2 mark cells where all five seeds of the best arm lie below every seed of each arm of the other geometry, which for five seeds against five is an exact Mann–Whitney p=0.008p = 0.008. Where Cartesian is better or comparable, the paper makes no CyFM claim, and Appendix K lists those cases.

  • Metrics: Alongside pooled sliced W2W_2, Table 2 reports amplitude lag-one, circular phase lag-one, and the radial spectrum against a held-out patient cohort, with CyFM leading on knee 64×6464\times64 in 7 of 9 comparisons for k≥4k \ge 4. PSNR and SSIM score paired reconstructions against ground truth and therefore do not apply to unconditional samples, which have no ground truth. Conditional reconstruction is left to future work.

  • Phase at k=1: The generated phase stays close to the prior at k=1k=1, as our Proposition (Appendix B) predicts: one Euler step returns the conditional mean, in which opposing phase targets cancel. The k=1k=1 gain comes from amplitude: amplitude-only W2W_2 favors CyFM across all four blocks with seeds separated (knee 64×6464\times64: 0.086 vs. 0.147), while phase alone shows no significant advantage on speech or knee. With the endpoint phase target (see the 320×320320\times320 point above), the phase does move at k=1k=1, cutting its spatial-coherence gap from 0.45 to 0.16 on knee 64×6464\times64 and from 0.74 to 0.04 at 320×320320\times320.

  • Abstract: "From four steps" marks where CyFM starts to lead on amplitude texture (seeds separated at k=4k=4), not a lead on every measure. The Cartesian wins at k=4k=4 are on phase coherence (claimed from eight steps) and the pooled metric (claimed at one step), and the revised abstract states the count instead: 7 of 9 spatial comparisons from k=4k=4.

  • Qualitative samples: For MRI, the new Appendix I shows unconditional knee samples from both geometries using a larger backbone (15.4M parameters). For audio, we add none: the generated items are 0.26-s STFT segments, too short for a listening test.

We hope these revisions address the reviewer's concerns.

Reviewer_gMDpOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CyFM combines amplitude–phase flow matching under a cylindrical metric with whole-field minibatch optimal transport. It evaluates few-step generation on synthetic fields, speech spectrograms, and MRI.

Critical strengths

The shortest-arc construction provides a clear bound on angular targets. Five-seed experiments separate geometry and coupling effects, and the paper explicitly reports limitations at full MRI resolution.

Critical weaknesses
  1. The headline metric does not establish field quality. Pooled sliced W2W_2 ignores spatial arrangement. At 320×320320\times320, CyFM’s single-step marginal advantage coexists with worse performance on every reported spatial criterion at every tested step count (Appendix Table 6). Thus, the claimed acceleration does not establish more faithful MRI generation.
  2. The theoretical connection to numerical difficulty is incomplete. Divergent angular velocities describe Cartesian bridges after a coordinate transformation; the actual Cartesian regression target remains bounded and constant along each bridge. Bounding cylindrical angular targets alone does not establish easier learning or lower integration error (Sections 3–5).
Questions for authors

Do the few-step gains persist under whole-field distribution metrics and assessments of anatomical or reconstructed-audio quality? Can the authors directly relate angular-target behavior to measured integration errors or properties of the learned vector fields? Can the equality condition in Section 5.4 be corrected and its implications for the coupling argument clarified?

1 comment

We thank the reviewer for the thoughtful assessment and for highlighting both the clean separation of geometry and coupling effects and our explicit reporting of full-resolution MRI limitations. All revisions below are incorporated in the updated manuscript.

  • W1, Q1: Beside pooled sliced W2W_2, every MRI table already reports three spatial measures: the gaps to the data in amplitude lag-one autocorrelation, circular phase lag-one coherence, and the radial power spectrum. On knee 64×6464\times64, CyFM leads on these from k=4k=4 in 7 of 9 comparisons. The new Appendix I also shows unconditional knee samples from both geometries.

  • W1 (320×320): We retrained every arm with an endpoint phase target (new Appendix J). The main tables regress the angular velocity, whose least-squares optimum is a conditional mean: where training pairs turn by nearly +π+\pi and −π-\pi, the turns cancel and those pixels move slowly. The endpoint target instead predicts (cos⁡θ1,sin⁡θ1)(\cos\theta_1, \sin\theta_1), whose optimum points to the circular mean and does not cancel. Everything else is unchanged, and the Cartesian arms are retrained with an endpoint target as a control. CyFM + OT then leads on pooled sliced W2W_2 at every k≤8k \le 8, with seeds separated from all four Cartesian arms, and its spatial gaps shrink several-fold (the phase-coherence gap at k=1k=1 falls from 0.74 to 0.04). At k=8k=8, it is no worse than any Cartesian arm on any spatial measure and leads on amplitude texture. Section 5.2 now states the acceleration strictly as fewer solver steps to a fixed pixel-marginal error. The full comparison at 320×320320\times320:

320×320k=1k=2k=4k=8k=100
Pooled sliced W₂0.0707* / 0.12030.0738* / 0.12230.0620* / 0.11660.0438* / 0.08720.0538 / 0.0629
Amplitude texture0.1815 / 0.0582*0.0283 / 0.0051*0.0027 / 0.00480.0011* / 0.00790.0006 / 0.0014
Phase coherence0.0438 / 0.0089*0.0718 / 0.0297*0.0689 / 0.0521*0.0711 / 0.07330.0771 / 0.0671*
Radial spectrum1.0331 / 0.8160*0.4441 / 0.1649*0.1735 / 0.1372*0.1564 / 0.16620.0990 / 0.0706*

Each cell reports CyFM + OT (endpoint) / the best of the four Cartesian arms (velocity and endpoint, with and without OT), averaged over five seeds. Lower is better. Bold marks the lower value, and * indicates complete seed separation (p=0.008p = 0.008).

  • W2, Q2: We thank the reviewer for this question. An exact result now makes the link (Proposition, Appendix B): one Euler step returns the conditional mean of the endpoint. In the plane, averaging over uncertain endpoint phases shrinks the amplitude, whereas on the cylinder the mean amplitude is preserved. The measured one-step amplitude W2W_2 confirms this in all four blocks with seeds separated (knee 64×6464\times64: 0.086 vs. 0.147). Section 5.2 now states this link, and for space the proof and the per-block split are in Appendix B.

  • Q3: Equality holds if and only if some coupling has both coordinate projections optimal. Product measures suffice, though they are not necessary. Our coupling argument does not depend on this condition, since with our product prior the factorised plan transports onto the product of the amplitude and phase marginals, which differs from the data law whenever amplitude and phase are dependent (Table 3).

We hope these revisions address the reviewer's concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions for the Area Chair

Dear Area Chair,

We thank you and the reviewers for the careful assessment of our submission. We have replied to every point in the three reviews, and all revisions are in the updated manuscript. The main changes are:

  • Full-resolution MRI (320×320) (all reviewers). We added Appendix J, which retrains every arm with an endpoint phase target. We also retrained the Cartesian arms the same way as a control. With this target, CyFM + OT leads on pooled sliced W₂ at every solver step, and its seeds are separated from all four Cartesian arms. Its spatial gaps shrink several-fold. At k = 8, it is no worse than any Cartesian arm on any spatial measure.
  • Link between geometry and the learned field (gMDp, 8FAb). A new exact result in Appendix B shows that one Euler step returns the conditional mean of the endpoint. Under phase uncertainty, this mean keeps the amplitude on the cylinder but shrinks it in the Cartesian plane. The measured one-step amplitude error is 1.7–2.4× lower for CyFM in all four data blocks, with seeds separated.
  • Statistical significance (8t4Q). All main tables now report the five-seed mean ± standard deviation. The new Appendix K gives Welch's t-tests and exact Mann–Whitney tests for every block, metric and step count.
  • Evaluation and presentation (8t4Q, 8FAb). We added citations for all datasets and unconditional knee MRI samples (Appendix I). We also clarified that the speedup refers to solver steps, which cost the same number of network evaluations and the same wall-clock time in both geometries.
  • Narrowed claims (gMDp, 8FAb). We removed or rewrote three statements that went beyond the evidence. Our established claim is now stated as better few-step fitting of pixel-marginal distributions, kept separate from the spatial results. We also clarified the equality condition in Section 5.4 and fixed the bold marker in Table 7.

Reviewer 8FAb acknowledged that our response addresses their concerns. Reviewers gMDp and 8t4Q have not yet responded to our rebuttal.

We also note that Reviewer 8t4Q's rating of 1 (clear rejection) is hard to reconcile with their own soundness score of 2 ("minor or fixable concerns") and with weaknesses that are mostly requests for citations, statistics and additional metrics, some of which (PSNR/SSIM) require paired ground truth and do not apply to unconditional generation. We responded to each of these points but have not received a reply.

We would be grateful if you could look at these revisions and discuss them with the reviewers, especially the new results in Appendices B, J and K.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.