acceptodds
Under review as a conference paper at ICLR 2027

FlowCycle: Rethinking Cycle Consistency via Exact Bijective Latent Flows

Abstract

Bidirectional cross-modal translation often entails a compromise between inference efficiency and formal invertibility. Continuous diffusion-based approaches typically require computationally intensive ordinary differential equation (ODE) solvers and are susceptible to integration error accumulation, whereas CycleGAN-type architectures depend on approximate cycle-consistency objectives that provide only indirect and potentially decoupled constraints on reversibility. We introduce FlowCycle, an end-to-end framework that enables single-pass translation while ensuring exact reversibility in latent space. Specifically, we cast cross-domain transport as a discrete normalizing flow in a learned representation space, implemented via spatial affine coupling transformations. This design yields closed-form analytic inversion in one pass, obviates numerical solvers, and exposes an interpretable latent transition trajectory without requiring additional flow evaluations. Empirical results on synthetic digits (Multimodal MNIST), urban scenes (Cityscapes), and clinical multi-sequence brain MRI indicate state-of-the-art translation fidelity. Beyond quantitative performance, FlowCycle offers substantive practical advantages: in medical imaging, its traceable trajectories support transparent, clinician-auditable morphing for reliable synthesis of missing diagnostic sequences; in complex visual environments, its deterministic single-step inference facilitates real-time data simulation and robust cross-domain perception without computational bottlenecks. Our code are publicly available by https://anonymous.4open.science/r/FlowCycle-5634/

Primary area:
generative models
Keywords:
Cross-Modal Translation, Normalizing Flows, Invertibility, Latent Manifold Dynamics, Medical Image Computing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sGKsOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a single-pass bidirectional cross-modal image translation framework based on invertible latent flows. The framework employs modality-specific encoders and decoders connected by an invertible flow and introduces two additional regularization objectives: manifold consistency and trajectory regularization. By regularizing the intermediate representations of the invertible flow, the authors argue that the model can produce smooth and interpretable domain transition trajectories without iterative inference. Experimental results on synthetic MNIST, Cityscapes, and brain MRI demonstrate competitive translation fidelity and inference efficiency compared to existing adversarial, diffusion, and flow-matching approaches, while maintaining exact invertibility in the latent space.

Critical strengths
  • The proposed framework demonstrates competitive translation performance across three diverse datasets, including synthetic imaging, Cityscapes, and brain MRI. The deterministic inference and low computational cost also suggest practical advantages over iterative diffusion and flow-matching approaches.
Critical weaknesses
  1. The technical novelty of the proposed framework is limited. The core architecture, consisting of modality-specific encoders/decoders and an invertible latent flow, is already well established in prior work. Exact reversibility and single-pass inference are inherent properties of discrete normalizing flows rather than novel technical contributions. Therefore, the main methodological differences appear to be the proposed manifold-consistency and trajectory regularization losses.
  2. The paper lacks comparisons with closely related flow-based translation methods. While extensive comparisons are provided against diffusion and continuous flow-matching approaches, directly relevant normalizing-flow-based translation methods, such as AlignFlow [1], are missing. Given the substantial architectural similarities, these comparisons are essential to establish the actual methodological and empirical contributions of the proposed framework.
  3. The effectiveness of the proposed regularization losses is not convincingly demonstrated. In Table 2, removing TR actually improves 8 out of 12 reported SSIM/PSNR metrics. Furthermore, Appendix A indicates that both the manifold-consistency and trajectory-smoothness losses are disabled simultaneously in this ablation. Consequently, the individual contributions of these two regularization terms remain unclear. Considering that these losses constitute the main methodological additions, the lack of convincing empirical improvements significantly weakens the paper's technical contribution.
  4. The claimed interpretability and smoothness of the intrinsic trajectories are insufficiently validated. Intermediate states naturally exist in any multi-layer invertible flow, but their existence does not necessarily imply meaningful semantic interpolation. In Figure 4, particularly for Cityscapes, the decoded trajectories do not clearly exhibit smooth transitions between domains. Although the authors suggest a trade-off between translation fidelity and trajectory stability, no direct quantitative comparison between the regularized and unregularized trajectories is provided to support this claim.

[1] Grover et al., "AlignFlow: Cycle Consistent Learning from Multiple Domains via Normalizing Flows," AAAI, 2020.

Questions for authors
  1. Could the authors clarify the fundamental technical differences between FlowCycle and existing invertible flow-based translation frameworks, and provide comparisons with closely related methods such as AlignFlow?
  2. Could the authors provide independent ablations for the manifold-consistency and trajectory-smoothness losses? The current ablation disables both simultaneously, making it difficult to identify their respective contributions.
  3. Given that TR degrades most endpoint-fidelity metrics, could the authors provide quantitative evidence demonstrating the benefits of trajectory regularization, particularly regarding trajectory smoothness, semantic consistency, or topological stability?
  4. What evidence supports the interpretation of intermediate coupling states as meaningful domain transitions, rather than arbitrary intermediate representations of an invertible network?
6 comments

We thank the reviewer for the careful reading, and in particular for raising the comparison with flow-based translators, which led directly to the AlignFlow experiments now in Table 2 and Fig. 12. Table and section numbers refer to the revised PDF, where all changes are in red.

W1 / Q1: Limited novelty; differences from invertible flow-based translators.

Response. We thank the reviewer for the opportunity to clarify this. Exact invertibility is shared with normalizing flows, but FlowCycle is fundamentally different from AlignFlow and other flow-based translators in both the problem it solves and its mathematical structure:

  • Problem setting: paired vs. unpaired. FlowCycle is designed for paired cross-modal translation, where corresponding images of the same subject or scene are available (MRI/PET, T1/FA, Photo/Labels), and uses this pairing to learn a specific correspondence (L_cross). AlignFlow addresses unpaired domain alignment, where no correspondence exists and only the two marginal distributions are matched.
  • Mathematical structure: a bijection between modality latents vs. a shared prior. AlignFlow maps each domain by its own flow to one shared Gaussian latent in pixel space and translates by composing one flow with the inverse of the other through that prior. FlowCycle instead learns modality-specific latents with encoders and decoders and places a single bijection f between them, with no shared prior. This structure is what lets the round trip split exactly into an architectural part, guaranteed by f, and a trained part (Eq. 4); the losses follow from that split (new Prop. 1, App. E). Another reviewer also found this decomposition useful for seeing which part of the round trip is guaranteed. No such decomposition exists for AlignFlow.
  • What the structure yields: one shared, progressive path. Because f is a flow between the two latents that starts as the identity, its own states form one path that both directions share exactly and that progresses toward the target. In AlignFlow, every translation passes through the Gaussian prior, so its intermediate states are noise-like (Fig. 12).

Against a capacity-matched non-invertible bridge trained with the same losses, only FlowCycle keeps the latent round trip and the forward/return path exact, at no cost in fidelity (App. D, Table 4, Fig. 14). Table 1 summarizes these differences across all methods. [Revised: Sec. 1; Table 1; Sec. 3.3; App. D, E.]

W2: Comparison with AlignFlow.

Response. Thank you for this suggestion. Although AlignFlow targets a different, unpaired setting (W1), it is the closest flow-based translator, and the comparison makes the fundamental differences concrete. It is now in Table 2 under the same splits and protocol and in the property comparison of Table 1, and its native states are shown in new Fig. 12. For a comparison within the paired setting, Table 2 also includes the supervised pix2pix and per-direction U-Net baselines. It trails FlowCycle in all six directions (e.g. SSIM 0.17 vs. 0.71 on Photo→Labels, 0.49 vs. 0.75 on T1→FA), and its states pass through noise-like patterns of its Gaussian latent, whereas the states of FlowCycle progress toward the target. [Revised: Table 1; Table 2; App. C, Fig. 12.]

W3 / Q2 / Q3: Effect of the regularization losses.

Response. We thank the reviewer for asking what each regularization term contributes; the revision now makes this explicit. TR is designed for latent consistency and for the trajectory, not for endpoint fidelity, and it delivers both with endpoint fidelity essentially unchanged (SSIM differs by about 0.02 at most; Ours remains best or second-best in SSIM in all six directions of Table 2). The roles of its two terms follow from the round-trip decomposition (new Prop. 1, App. E), which separates them:

  • L_lat is exactly the condition-(ii) term of the bound. Its effect is measured directly: the test residual of condition (ii) drops from ≈1 without TR to 0.10 / 0.57 / 0.35 on MNIST / Cityscapes / ADNI.
  • L_traj acts only on the path between fixed endpoints and leaves the bound unchanged. Among the many flows with the same endpoints, it selects the one whose successive states change least, which is what makes the raw states progress toward the target (Figs. 9, 14).

Because the two terms act on different quantities, one on the latent residual and one on the path, their contributions are separable by construction, and each is measured on the quantity it acts on. [Revised: Table 2; Sec. 3.3; App. E, Prop. 1; App. C, D.]

W4 / Q4: Are the intermediate states meaningful transitions?

Response. We agree that the existence of states alone does not imply semantic interpolation, and have accordingly revised "interpretable" to "inspectable". What distinguishes the states of FlowCycle is measured rather than assumed: they are the operator's own states, both directions share them exactly (Table 4, Fig. 14), and they progress toward the target in the raw decoded images without any post-processing (Figs. 9, 14), unlike the oscillating states of RevGAN and the noise-like states of AlignFlow and diffusion (Figs. 10–12). For Cityscapes, Fig. 14 shows the raw Labels→Photo states, which change progressively toward the target scene. We will further validate their clinical meaning through biomarker annotation by medical experts along the trajectory and a test against real disease progression on longitudinal data. [Revised: Sec. 3.4; App. C, D; Sec. 5.]

We hope the added comparison and analysis clarify both the novelty and the role of the losses. If any concern remains, we would be glad to address it during the discussion phase, and we would be grateful if the reviewer would take the revision into account in the final assessment.

I thank the authors for their detailed responses and additional experiments. I consider W2 and W4 sufficiently addressed. While I remain unconvinced that the intermediate states represent semantically meaningful transitions, I acknowledge their visual coherence compared to the baselines and appreciate the clarification from "interpretable" to "inspectable."

Regarding W1 and W3, my main concern remains. FlowCycle shares the general encoder–invertible core–decoder structure with prior work such as RevGAN, which also supports paired translation. This makes empirical validation of the proposed regularization terms particularly important for establishing the methodological contribution.

I appreciate the additional analysis and clarifications, including that TR is not intended to improve endpoint fidelity. However, my central concern regarding W3 remains unresolved, as the requested independent ablations have not been provided. Specifically:

  1. Latent consistency: The reported reduction in latent residuals is encouraging, but it remains unclear whether this translates into improved image-space cycle consistency without a direct with/without TR comparison.

  2. Trajectory quality: The intermediate states appear visually coherent, but the effect of TR on trajectory smoothness or progression has not been isolated through a with/without TR comparison.

  3. Independent contributions: Although the two losses target different quantities, they update shared parameters and may interact during optimization. Their mathematical distinction does not replace independent ablations of LlatL_{\mathrm{lat}} and LtrajL_{\mathrm{traj}}.

Since these were central concerns in my original review and remain insufficiently addressed, I will maintain my original rating for now.

Reviewer_c8FoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper inserts a discrete invertible flow between the encoders and decoders of two domains, yielding a bidirectional latent mapping with an analytic inverse. The method combines paired reconstruction, latent consistency, image cycle, and trajectory constraints, and displays the cross-domain transition process by decoding intermediate states.

Critical strengths
  1. The paper distinguishes the exact invertibility of the latent core from the image-cycle error after encoding and decoding, making the mathematical properties of the bidirectional mapping clearly.
Critical weaknesses
  1. The added losses are a combination of existing methods. Paired reconstruction, latent consistency, and cycle constraints are all common designs. This paper should include a theoretical analysis of why these components are used.
  2. Visualizing intermediate states does not equal to a new interpretability mechanism. Layer-by-layer decoding only shows how the representation changes and does not guarantee that these changes correspond to interpretable semantic processes. It cannot by itself support an interpretability contribution.
Questions for authors

Please refer to Weakness.

3 comments

We thank the reviewer for the constructive and focused comments. Both points concern how the paper justifies its design and its claims, and they helped us make the design rationale and the claims more explicit. Table and section numbers refer to the revised PDF, where all changes are in red.

W1: The losses combine existing designs; a theoretical analysis of why they are used is needed.

Response. We appreciate this request and are glad to make the rationale explicit: the losses were designed around the structure of the round trip, and the revision now derives this formally (new Prop. 1, App. E). Each loss targets one term of the round-trip error, and the decomposition follows from the exact invertibility of f. With condition (i) exact by construction, Prop. 1 bounds the image round-trip error by

‖T_B→A(T_A→B(x_A)) − x_A‖ ≤ ‖D_A(E_A(x_A)) − x_A‖ + K · ‖Π_B(β̂) − β̂‖,

where the first term on the right is L_rec, the second is L_lat, and K is the Lipschitz constant of D_A ∘ f⁻¹.

  • L_rec controls the reconstruction term; L_lat and L_cyc control the condition-(ii) term, in latent and in image space.
  • L_cross, the only paired term, fixes where f maps, i.e. the endpoint fidelity that the round trip alone cannot determine.
  • L_traj acts on the path and leaves the bound unchanged: among the flows with the same endpoints, it selects one whose successive states change least.

What distinguishes this from the common designs is what the losses no longer need to do: in a CycleGAN, the cycle loss must also learn the inverse, whereas here the inverse is exact, so the consistency terms act only on the encoder–decoder residual. App. D confirms this empirically: with the same losses, a non-invertible bridge loses the exact round trip by orders of magnitude. [Revised: Sec. 3.3; App. E, Prop. 1; App. D, Table 4.]

W2: Visualizing intermediate states is not a new interpretability mechanism.

Response. We thank the reviewer for this point. We agree on the wording: decoding states alone does not by itself establish that they correspond to semantic processes, so we now use "inspectable" instead of "interpretable" and "mechanistic". The contribution, however, is not a visualization added to a model, but a trajectory with three properties that no other method combines, each of which we measure:

  • Intrinsic: the states are those of the translation operator itself, at no extra cost, rather than a path reconstructed after the fact.
  • Consistent: both directions share the same path exactly; a non-invertible bridge departs from it by orders of magnitude (App. D, Table 4, Fig. 14).
  • Progressive: the states move toward the target already in the raw decoded images, without any post-processing (Fig. 9), while the states of RevGAN oscillate and those of AlignFlow and diffusion pass through noise (Figs. 10–12).

These properties make the path a sound basis for semantic analysis, which we will pursue directly: medical experts will annotate biomarkers along the trajectory, and longitudinal data will test whether it follows real disease progression. [Revised: Abstract; Sec. 1; Sec. 3.4; App. C, D; Sec. 5.]

Thank you again for the comments; they made the paper's argument more rigorous and its claims more precise. We would be happy to expand on Prop. 1 or any other point during the discussion.

Thank you for the response and the additional explanation. I also have different views on some points. For W2, changing "interpretable" to "inspectable" is reasonable. However, I think the progressive behavior is partly caused by the loss, so it cannot be described as an intrinsic property. I think a distinction should be made between those resulting from training and the model itself.

We sincerely thank the reviewer for the follow-up, which goes to the core of our method. The reviewer is right: the final trajectory is the joint result of the model and the loss. We clarify the part of each below, and the revision now states the distinction explicitly (Sec. 3.4, App. D, in red).

Response.

  • Guaranteed by the invertibility. Exact invertibility at the representation level fixes the path and the transition: f is a bijection between the two modality representations, so both directions traverse one path. This holds without the trajectory regularization (TR): latent round-trip error ≤ 3×10⁻⁴, path difference ≤ 2×10⁻⁵ (Table 4), and the decoded states already progress toward the target (detour on MNIST / Cityscapes / ADNI: 1.30 / 1.60 / 1.64, against 2.85 / 10.58 / 5.12 for RevGAN). A loss cannot create this: with the same losses, a non-invertible bridge keeps a round-trip error of 0.6–0.9 (Table 4, Fig. 14).
  • Contributed by the loss. The encoders and decoders are not exactly invertible, so information is lost when a state is decoded and re-encoded (condition (ii), Prop. 1). TR is added to keep this loss as small as possible: on ADNI it lowers the condition-(ii) residual from 1.01 to 0.35, raises the image-level cycle from 28.1 to 34.4 dB, and lowers the change between successive decoded states from 0.024 to 0.008 (Table 9).

We hope this resolves the concern and would be grateful if the reviewer would kindly consider raising the final rating. [Revised: Sec. 3.4; App. D, Tables 4, 8, 9, Figs. 14–16.]

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

The paper proposes FlowCycle for paired, bidirectional cross-modal image translation. It connects the latent spaces of two modality-specific autoencoders with an affine-coupling normalizing flow, which gives exact latent invertibility, single-pass inference, and intermediate states that are presented as an interpretable translation trajectory. On synthetic MNIST, Cityscapes and ADNI T1↔FA, the method reports the best SSIM/PSNR in most directions against GAN, diffusion and flow-matching baselines, with inference speed comparable to CycleGAN.

Critical strengths
  1. A coherent framework for invertible image translation. The paper focuses on invertible image translation rather than on a new generative paradigm, and the framework is mathematically consistent at the level of the latent operator. The decomposition in Eq. 4, which separates the exact latent inversion (i) from the manifold-consistency residual (ii), is a useful way to show which part of the round trip is guaranteed and which part still depends on training.

  2. Competitive results with single-pass inference. The method achieves competitive SSIM/PSNR on three different domains, with inference speed comparable to CycleGAN (Fig. 6c). This makes it a reasonable option for settings that need fast, deterministic bidirectional translation.

Critical weaknesses

Major 1. The practical value of exact invertibility is not isolated.

The paper evaluates cross-modal translation. This is distribution-to-distribution mapping overall, but at the level of individual samples it leans towards one-to-one mapping. The comparison therefore should not lack representative supervised-learning baselines. Ordinary supervised models can learn bidirectional mappings, although they do not guarantee exact mathematical invertibility.

Exact invertibility is a key novelty of this paper. The authors should therefore:

  • explain what value it brings;
  • compare how invertible and non-invertible methods differ on this kind of task.

If exact invertibility serves only cross-modal translation, it is hard to see it as a contribution of the paper.

The evidence in the paper makes this concern sharper:

  • The guarantee covers only the latent operator. Eq. 4 makes clear that f−1(f(z))=zf^{-1}(f(z))=z is only one of the two conditions. The image-level round trip also depends on the re-encoding projection of the target modality. Exact latent invertibility therefore does not, by itself, make the image translation exactly invertible. The 10−510^{-5} latent error in Fig. 5 holds regardless of what ff learns.

  • Invertibility alone does not seem to improve fidelity. RevGAN, whose latent core is exactly invertible, performs almost identically to the non-invertible CycleGAN in Table 2 (SSIM 0.8266 vs. 0.8280 on MNIST; 0.5693 vs. 0.5566 on Cityscapes photo→labels). FlowCycle's gains may therefore come from its training recipe (autoencoders + paired L1 + adversarial losses) rather than from invertibility.

  • Key comparisons are missing:

    • pix2pix (the authors use the pix2pix version of Cityscapes);
    • a supervised U-Net trained for each direction;
    • most importantly, the same encoders, decoders, data and budget with the flow replaced by a non-invertible latent network of equal capacity (plus a separate reverse network).

    Such an ablation would show whether exact invertibility translates into better endpoint quality, smaller image-cycle error, better preservation of source details, or more robustness to noisy pairs.

  • Exactly invertible latent translation already exists: RevGAN (van der Ouderaa & Worrall, 2019), AlignFlow (Grover et al., 2020), and Rombach, Esser & Ommer (NeurIPS 2020, Network-to-Network Translation with Conditional Invertible Neural Networks).

  • The adversarial baselines may be handicapped. Appendix A reports λid=5\lambda_{id}=5 for CycleGAN/RevGAN. An identity loss assumes similar appearance across domains and may hurt these baselines on labels↔photo and T1↔FA. It is also unclear whether they received paired supervision.

Major 2. Intermediate network states are not necessarily a unique or interpretable transport trajectory.

The authors repeatedly emphasize that FlowCycle yields an interpretable, intrinsic transformation trajectory without extra flow computation. What is puzzling is why the intermediate outputs of different layers of a neural network should be interpretable as a true transport trajectory from the source distribution to the target distribution. For most invertible mappings, even when the final mapping is the same, the intermediate path is not uniquely determined.

  • The intermediate states are not identifiable. For any invertible gg, f=(ϕ2∘g−1)∘(g∘ϕ1)f=(\phi_2\circ g^{-1})\circ(g\circ\phi_1) leaves the end-to-end map unchanged while changing the intermediate states arbitrarily. By contrast, the path of flow matching is pinned down by a prescribed probability path, and displacement interpolation in optimal transport is pinned down by a minimal-energy principle.The paper does constrain the path through Ltraj\mathcal{L}_{traj} (adversarial + smoothness on states decoded by DmixD_{mix}). This can encourage a smooth, visually plausible path, but it cannot establish uniqueness or physical/semantic meaning. A smooth visual progression alone is therefore not evidence of mechanistic interpretability.

  • Two controlled cases would test the claim:

    • Case A: both modalities are identical and the pairing is the identity. Ideally ff collapses to the identity with static intermediate states. If instead the blocks change the representation and then undo it, the trajectory does not carry the claimed meaning.
    • Case B: the marginals are identical but the pairing differs. For example, a symmetric distribution remains unchanged under rotation. Here different trajectories can satisfy the same endpoint constraint, so what additional assumption determines the "true" trajectory?
  • The visual evidence may partly come from the decoding and visualization pipeline:

    • Inconsistent description of DmixD_{mix}. Sec. 3.3 says it is trained adversarially on intermediate states via Ltraj\mathcal{L}_{traj}. Appendix B says it is "supervised only on source/target images, not intermediate states, and used only for visualization".
    • The change seems to occur when switching decoders. In Fig. 4(a), Steps 1–4 (decoded by DmixD_{mix}) remain green, PET-like images; only "Translated" (decoded by DBD_B from the same final code) becomes white and MRI-like. In Fig. 4(b), Steps 1–3 are nearly identical to the input photo.
    • Hand-crafted blending. For Fig. 8, Appendix B describes a luminance blend with m=ℓ/4m=\ell/4, a distance-ranked smoothstep reveal mask and Gaussian feathering.
    • Ground-truth orientation. Appendix A states that "diffusion orientation comes from acquisition and is not predicted", which suggests that the colour in the DEC-FA images comes from ground-truth acquisitions.

Major 3. The deterministic-bijection assumption and the scope of applicability are not discussed.

Most of the related work and baselines follow common generative-model paradigms, which mainly model a conditional distribution in which one input corresponds to multiple outputs. Five of the seven baselines (DDPM, CFM, RectFlow, MeanFlow, DiT) are stochastic conditional generators. FlowCycle, by contrast, is a point-to-point mapping at the operator level, which is precisely what guarantees its invertibility. At the probabilistic level it aims at distribution-to-distribution transport. This cannot address the one-to-many generation problem.

More precisely:

  • What the model can and cannot represent. For a fixed xAx_A, the model defines pθ(xB∣xA)=δ(xB−TA→B(xA))p_\theta(x_B\mid x_A)=\delta\big(x_B-T_{A\to B}(x_A)\big). A deterministic map can in principle still match the marginals, T#PA=PBT_\#P_A=P_B. However, this does not imply learning the correct conditional distribution PB(⋅∣xA)P_B(\cdot\mid x_A), let alone covering multiple plausible outputs for the same input.
  • Real tasks are often one-to-many. Labels→photo is one-to-many, and FlowCycle's outputs in this direction are visibly blurry (Fig. 3). T1 structure does not uniquely determine FA microstructure either.
  • Bijectivity forces information to be preserved. Every detail of the source must survive in β^=f(α)\hat\beta=f(\alpha), which is why condition (ii) cannot hold exactly in general.

I do not see this as a defect that the paper must solve. It is a modeling assumption that should be stated and justified as a scope of applicability. The comparison and the claims should also be framed accordingly: PSNR/SSIM structurally favour a deterministic, mean-seeking predictor over stochastic generators.

Minor. Evaluation scope and presentation.

  • Metrics. PSNR/SSIM only measure per-pair fidelity to the paired ground truth. They show that T(xA)≈xBT(x_A)\approx x_B, but not that T#PA≈PBT_\#P_A\approx P_B. The current experiments, which are based on paired image translation, demonstrate the effectiveness of deterministic translation. However, they are not sufficient to examine target-distribution coverage, generation diversity and conditional uncertainty.
    • Marginal distribution alignment should be evaluated with FID/KID or modality-appropriate feature distances, plus task metrics such as mIoU / FCN-score for photo→labels.
    • Robustness to imperfect pairing (registration errors, misaligned or shuffled pairs) is a separate question from one-to-many coverage. It would show whether weakly paired samples act merely as noise, or whether some pairing relations are ignored by the model.
  • Baseline setup. DDPM and DiT appear to generate from noise conditioned on the source, so their chaotic intermediate states (Fig. 4) are expected. Source-to-target bridge models, e.g., I2SB (Liu et al., ICML 2023), DDBM (Zhou et al., ICLR 2024) or BBDM (Li et al., CVPR 2023), would be fairer trajectory baselines. Some baseline numbers are unusually low (e.g., DDPM's SSIM of 0.19/0.21 on MNIST/ADNI), and the sampling settings are inconsistent (DiT uses 1,000 ancestral steps, DDPM 50 DDIM steps).
  • Variant selection in Table 2. The qualitative results use the TR variant ("Ours"), which is below "Ours w/o TR" in 4 of 6 directions. The claim that PSNR is best in 5 of 6 directions relies on the w/o TR variant.
  • Presentation.
    • Variant naming is inconsistent: TR / λreg\lambda_{reg} in the text vs. MS / λs\lambda_s in the Table 2 caption.
    • The asterisks in Fig. 4 are not explained.
    • There is a broken citation, "DiT (?)", in Appendix A.
    • Rombach et al. (2022, LDM) is cited under the invertible-network paradigm.
    • Several phrases overstate the evidence, e.g., "clinician-auditable" and "mechanistic interpretability".
Questions for authors

Central question. The paper achieves a strict latent bijection at the architectural level and shows competitive cross-modal translation performance. However, it has not yet demonstrated the unique value of exact invertibility compared with ordinary supervised bidirectional translation. Nor has it established a necessary link between the network's discrete internal states and a semantically meaningful cross-modal transport trajectory. Could the authors clarify whether these two properties are mathematical features of the architecture, or capabilities that independent experiments show to be practically meaningful for cross-modal translation?

A. Questions whose answers could change my assessment

  1. Value of invertibility (Major 1).
  • Can you provide results, even partial, for pix2pix / a supervised U-Net, and for an ablation that replaces ff with a non-invertible latent network of equal capacity while keeping everything else fixed?
  • In what concrete respect does exact invertibility help: endpoint quality, image-cycle error, preservation of source details, or robustness to noisy pairs?
  1. Difference from prior invertible translation. What does FlowCycle add beyond RevGAN, AlignFlow and network-to-network cINN translation? What is the basis for "Path ✗" for RevGAN in Table 1, given that its invertible core also consists of sequential blocks?

  2. Visualization protocol (Major 2).

  • Which description of DmixD_{mix} (Sec. 3.3 or Appendix B) produced the figures?
  • Are the brain trajectories in Fig. 1 and Fig. 4(c) produced with the m=ℓ/4m=\ell/4 blending described in Appendix B?

If the progressive morphing is largely produced by the visualization pipeline, I would lower my rating.

  1. DEC-FA orientation. Is the direction-encoded colour of the T1→FA results in Fig. 1, Fig. 3(c) and Fig. 4(c) taken from ground-truth acquisitions? Are the SSIM/PSNR values computed on scalar FA or on RGB DEC-FA images?

  2. Identifiability of the trajectory.

  • What does the trajectory look like in the identity case (Case A)?
  • Are trajectories consistent across random seeds?
  • What assumption, if any, makes the trajectory unique in Case B?
  1. Scope (Major 3).
  • Why did you choose a deterministic bijection, and for which tasks do you consider this assumption appropriate?
  • Please report the test-set residual of condition (ii), ∥ΠB(β^)−β^∥/∥β^∥\|\Pi_B(\hat\beta)-\hat\beta\|/\|\hat\beta\|.
  • Appendix B shows that the relative distance from the final transported code to the paired target code only decreases from 1.29 to 0.83/0.78. Does this mean the decoder, rather than ff, does most of the work?
  1. Final variant. Which variant in Table 2 is the proposed method? Please re-mark best/second-best results accordingly.

  2. Baseline setup.

  • Were CycleGAN/RevGAN given paired supervision, and why was λid=5\lambda_{id}=5 used?
  • What were the training budgets and tuning procedures for the diffusion/flow-matching baselines?

B. Clarification requests

  1. Is the "±" in Table 2 the standard deviation over test samples or over training runs? Are there results over multiple random seeds?
  2. The Cityscapes training split has 2,677 pairs, versus the standard 2,975. Why?
  3. Downstream classification (Fig. 6a,b). No baseline translator is compared under the same protocol. Moreover, the translator was trained on 12,000 pairs, and the synthetic PET domain contains class-conditional hotspot enhancement. Could the gain reflect the extra data seen by the translator?

C. Suggestions for future work (not decisive for my rating)

  1. Evaluate marginal distribution alignment (FID/KID) and robustness to weak or noisy pairing, as two separate experiments.
  2. Extend to unpaired training, higher resolutions and 3D volumes, and to stochastic outputs for one-to-many mappings.
  3. Support clinical claims such as "clinician-auditable" with a reader study.
4 comments

We sincerely thank the reviewer for an exceptionally thorough review. The two central questions, what exact invertibility buys in practice and whether the trajectory progression depends on the visualization, go to the heart of the paper, and answering them strengthened the revision: the capacity-matched non-invertible bridge (Major 1(b)) and the raw-state figures (Figs. 9, 14) were added in direct response, and they now make the paper's central claims directly measurable. Below, "Ours" is the paper's main model; table and section numbers refer to the revised PDF, where all changes are in red.

Major 1: the practical value of exact invertibility

Summary. We thank the reviewer for raising this central question; the revision now answers it explicitly and measures it directly. One-way fidelity is a natural outcome of the recipe, not what FlowCycle sets out to improve. What exact invertibility provides is consistency: one operator for both directions, a lossless latent round trip, and one path that both directions traverse. Standard benchmarks do not measure this, so, following the reviewer's suggestion, we test it directly against a capacity-matched non-invertible bridge (see (b)). [Revised: Abstract; Sec. 1; Sec. 5.]

(a) Supervised baselines (pix2pix, a U-Net per direction).

Response. Thank you for suggesting these; they add a useful supervised reference point. Both are added under the same splits and protocol (SSIM):

MRI→PETPET→MRIPhoto→LabelLabel→PhotoT1→FAFA→T1
pix2pix0.8330.9310.6530.4070.7390.839
U-Net (L₁, per direction)0.8700.9500.7330.5140.7520.850
Ours0.8680.9480.7070.4260.7520.822

Ours exceeds pix2pix in 5/6 directions and is on par with the per-direction U-Net on MNIST and T1→FA; the U-Net is ahead on Cityscapes and FA→T1. The U-Net serves as an upper reference for one-way fidelity: it is two separately trained models, one per direction, with no inverse and no shared path. FlowCycle stays close to it with a single invertible operator that in addition guarantees the consistency shown in (b), which no per-direction model can provide. [Revised: Sec. 4.2; Table 2; App. D, Table 6.]

(b) Same encoders/decoders/budget with a non-invertible bridge.

Response. This suggestion turned out to be the cleanest way to isolate what invertibility contributes, and we are grateful for it. We replaced f by two independent latent networks g_AB, g_BA of equal total size (1.83M vs. 1.84M), with everything else fixed.

DataBridgeA→B / B→A SSIMlatent round-trip err.forward/return path diff.
MNISTOurs / NonInv0.868 / 0.948 vs 0.868 / 0.9472×10⁻⁴ vs 0.811×10⁻⁶ vs 0.011
CityscapesOurs / NonInv0.707 / 0.426 vs 0.683 / 0.4033×10⁻⁴ vs 0.905×10⁻⁶ vs 0.031
ADNIOurs / NonInv0.752 / 0.822 vs 0.754 / 0.8231×10⁻⁴ vs 0.606×10⁻⁶ vs 0.018

Ours matches or exceeds NonInv at the endpoints (e.g. +0.02 SSIM in both Cityscapes directions) and, on top of that, keeps the latent round trip and the forward/return path exact. NonInv loses both by orders of magnitude, although it is trained with the same cycle losses and TR, so training alone does not recover this consistency. This also explains the reviewer's observation that RevGAN performs on par with CycleGAN: the benefit of invertibility does not show in endpoint metrics, but in consistency, which those metrics do not measure. Since the intermediate states have no ground truth, this consistency, together with their progression toward the target, is what we evaluate. New Fig. 14 shows it directly: the forward and return paths of Ours coincide, those of NonInv do not. [Revised: Sec. 4.4 "Consistency from invertibility"; App. D, Table 4, Fig. 14.]

(c) The guarantee is latent-only; the 10⁻⁷ in Fig. 5 holds whatever f learns.

Response. We agree with both observations, which are an accurate reading of the method. Exactness lives in the latent operator by design, because the latent is where the two modalities are linked; the image level additionally passes through the decoders. And the 10⁻⁷ is indeed architectural rather than learned. The key point, which the revision now states explicitly, is that this is exactly what a non-invertible bridge cannot obtain from the same training (round-trip error 0.6–0.9, see (b)): the consistency is guaranteed by construction rather than approximated by training. As requested, the test residual of condition (ii), i.e. the trained part of the round trip, is 0.10 / 0.57 / 0.35 on MNIST / Cityscapes / ADNI. [Revised: App. D, Table 5.]

Major 1 (continued)

(d) Prior invertible translators (RevGAN, AlignFlow, cINN).

Response. We thank the reviewer for the opportunity to sharpen the positioning. Exact invertibility is shared with these methods; what distinguishes FlowCycle is what invertibility carries: intermediate states that both directions share, that agree exactly between the two directions, and that progress visibly toward the target. This makes the trajectory itself usable, e.g. for following biomarkers along the path. None of the prior methods offers this combination:

  • RevGAN is exactly invertible, but its states oscillate (per-state SSIM 0.39 → 0.30 → 0.71 → 0.22 → 0.67 on ADNI; Figs. 10–11).
  • AlignFlow is exactly invertible, but its states pass through noise-like patterns of its Gaussian latent (SSIM 0.01–0.07 mid-path; new Fig. 12), and it trails in fidelity (e.g. 0.17 vs. 0.71 on Photo→Label, Table 2).
  • cINN (Rombach et al., 2020) samples target codes; it is not a deterministic bijection.

We add AlignFlow to Table 2 and turn Table 1 into a property comparison of all methods. [Revised: Sec. 1; Table 1, Sec. 3.4; Table 2; App. C, Figs. 10–12.]

(e) λ_id = 5; paired supervision?

Response. Thank you for catching this; the training regime of the baselines is now stated explicitly. Following the original papers, CycleGAN and RevGAN are trained unpaired. To confirm that the identity loss does not handicap them, we retrained both with λ_id = 0; each changes little from its own λ_id = 5 run (SSIM):

MRI→PETPET→MRIPhoto→LabelLabel→PhotoT1→FAFA→T1
CycleGAN, λ_id = 5 / 00.828 / 0.8280.943 / 0.9460.557 / 0.5740.305 / 0.3510.680 / 0.6790.813 / 0.799
RevGAN, λ_id = 5 / 00.827 / 0.8270.940 / 0.9440.569 / 0.5370.305 / 0.3340.661 / 0.6640.798 / 0.805

[Revised: App. A; App. D, Table 7.]

Major 2: are the intermediate states a meaningful trajectory?

(a) The intermediate states are not unique.

Response. We thank the reviewer for this fundamental question. Indeed, many paths connect the same endpoints, and none has ground truth. This is exactly why FlowCycle does not leave the path unspecified but selects it by an explicit minimal-change principle: the flow starts as the identity, and the trajectory term penalizes the change between successive decoded states. The selected path is shared exactly by both directions (Major 1(b)), and it progresses toward the target more steadily than the states of the other methods (Figs. 10–12), which oscillate (RevGAN), pass through noise (AlignFlow, DDPM, DiT) or follow a separate path per direction (CFM). Building on this, medical experts will annotate biomarkers along the trajectories to track how clinically relevant content evolves along the shared path, and longitudinal data will test the path against real disease progression. [Revised: App. C, Figs. 10–12; Sec. 5.]

(b) Is the progression produced by the decoding and visualization pipeline?

Response. We thank the reviewer for checking this carefully, as it lets us state the point unambiguously: the progression does not depend on the visualization. It is produced by the model and is already present in the raw decoder outputs, before any display step is applied.

  • Raw states show the progression. New Fig. 9 shows ADNI states decoded by the native decoder D_B as raw scalar FA, with no colouring, blending or display decoder; new Fig. 14 shows the raw states of all three datasets, one decoder per row. In both, the states progress toward the target.
  • The progression is a training objective. The trajectory term is applied to the states as read out by D_B, penalizing the L₁ change between successive states (Sec. 3.3, Eq. (8), Alg. 1). The progression is therefore learned by the model, not created when the states are displayed.
  • The display steps come after inference and change no result. In Figs. 1, 4, 8, 10, 11, the states are rendered by the display decoder G of App. B, fitted after training on endpoints only, and the brain panels (Figs. 1, 4(c), 8) are drawn in DEC colour with the fixed blending schedule of App. B (m = ℓ/4), which is agnostic to the prediction. These steps only present FA the way it is read in clinical practice. Neither enters training or any reported metric, which use the scalar outputs of the native decoders.

[Revised: Sec. 3.3, Eq. (8), Alg. 1; App. C, Fig. 9; App. D, Fig. 14.]

(c) DEC-FA orientation; are SSIM/PSNR computed on scalar FA?

Response. Thank you for asking; this is now stated explicitly. The model predicts scalar FA; the colour comes from the subject's measured principal diffusion direction and is used for display only. SSIM and PSNR are computed on the raw scalar FA images. [Revised: App. C, Fig. 9.]

Major 3: the deterministic bijection and its scope

(a) Why a deterministic bijection, and for which tasks?

Response. We thank the reviewer for this question, which helped us state the scope of the method more sharply. Once trained, FlowCycle generates the paired target modality from either modality and can cycle back to the source, with no information lost in the latent space. This is the setting it targets: tasks in which the content that matters must survive translation. In medical imaging this content is the biomarkers; in general vision it is the semantics, which is why we evaluate Cityscapes Photo↔Labels, where many photometric appearances share one semantic layout. By design, FlowCycle is not an any-to-any generator: it preserves what the target representation encodes and does not invent what it does not. The tree shadows in the teaser photo, for example, carry no semantic label and are therefore not kept. The revision states this scope explicitly. [Revised: Sec. 5.]

(b) Condition (ii) cannot hold exactly in general.

Response. Agreed, and this is precisely why the method separates the round trip into an exact part (condition (i), by construction) and a trained part (condition (ii)). Table 5 now reports the test residual of the trained part, and further limiting the information loss of the encoders and decoders is a direction we will pursue. [Revised: App. D, Table 5; Sec. 5.]

(c) PSNR/SSIM favour deterministic, mean-seeking predictors.

Response. We agree that these metrics are incomplete, and the point is well taken for the stochastic baselines in particular. Against the deterministic translators, to which this bias does not apply (pix2pix, U-Net, CycleGAN, RevGAN, AlignFlow), FlowCycle remains best or competitive. Standard metrics also capture only the endpoints, not the trajectory quality and cross-direction consistency where FlowCycle differs most; as future work we will develop qualitative and quantitative metrics for both and compare all methods on them. [Revised: Table 2; App. D, Table 6; Sec. 5.]

Minor

We thank the reviewer for the many detailed minor comments, which we have addressed. Some overlap with the Major points above (evaluation scope and metrics: Major 3; identity loss and pairing of the baselines: Major 1(e)). In addition, we revised the variant selection in Table 2, the presentation (naming, Fig. 4, citations and wording), and the requested clarifications (Table 2 statistics, Cityscapes split, training budgets). [Revised: Table 2; Sec. 3.3, 3.4, 4.1, 4.2; Fig. 4; App. A, C.]

We are grateful for the time and care the reviewer invested; the revision is clearer and better supported for it. We hope the new experiments, in particular the non-invertible bridge and the raw trajectory figures, resolve the main concerns. If so, we would appreciate it if the reviewer would consider revisiting the assessment, and we would welcome any further questions during the discussion period.

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' efforts to address my concerns. I will maintain my rating of 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.