acceptodds
Under review as a conference paper at ICLR 2027

Simultaneous Neural Optimal Transport

Abstract

Optimal Transport (OT) provides a principled framework for learning transformations between probability distributions from unpaired samples. In many applications, however, a single transformation must map several source distributions to a common target distribution. For example, image restoration may require handling different types of degradation without knowing the degradation of each input at inference time. Simply pooling the source distributions only encourages alignment with the target at the aggregate level and may leave individual sources misaligned. In this paper, we consider the simultaneous OT problem, which seeks a shared transport map that minimizes the average transport cost while aligning each source distribution with a prescribed target. We derive a max-min formulation of an unbalanced extension and propose SimNOT, a neural method based on this formulation. We illustrate its application to image restoration, where a single model handles multiple degradation types using a common collection of clean target images.

Primary area:
generative models
Keywords:
optimal transport, generative modeling

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

This paper proposes SimNOT, which uses a shared map to transport multiple source distributions to a common target distribution. The authors formulate a divergence-penalized unbalanced optimal transport objective, derive its semi-dual formulation, and establish approximation results for shared neural maps as well as deterministic map properties under specific conditions.

Critical strengths
  1. The paper clearly distinguishes mixture-level alignment from source-wise alignment and proposes a trainable unbalanced formulation based on the existing simultaneous optimal transport framework. The research problem and method motivation are clear.

  2. The method uses source distribution information during training but does not require source labels at inference time, making it suitable for applications with unknown degradation types.

Critical weaknesses
  1. The experiments are conducted only on CelebA images at 64×64 resolution with artificially constructed degradations. The lack of validation on higher-resolution images, real-world degradations, and other image domains limits the practical generalizability of the method.

  2. SimNOT uses an independent potential for each source distribution and requires source membership information during training. As the number of sources increases, the memory and computational costs of training will also increase. Therefore, although the method does not require source labels at inference time, it still relies on a strong source-grouping assumption during training.

Questions for authors
  1. If the capacity of the Pooled UOT potential is increased, or if the multiple SimNOT potentials partially share parameters, does the performance ranking remain unchanged? How much of the current gain comes from the additional capacity of the source-specific potentials?

  2. How sensitive is the method to the number of source distributions? How do performance and training cost change when the number of sources is reduced or increased from five?

  3. Under unseen degradation strengths, does the advantage of SimNOT come from better out-of-distribution generalization or from producing smoother outputs? Can content-preservation metrics be used to distinguish between these possibilities?

  4. Do the multiple potentials require separately tuned learning rates or regularization strengths during training? Is it fair to use the same optimization settings for all sources?

Reviewer_7K14OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a neural method for solving the simultaneous optimal transport problem proposed by Wang & Zhang et al. (2025). They derive a max-min formulation of an for an unbalanced simultaneous transport problem, where the total masses of the sources and target distribution are unequal. The proposed neural method is evaluated simultaneous transport from five Gaussian sources to a common Swiss roll target, and on image restoration (different degradation types to a target HR output).

Critical strengths
  • The authors present a very clear motivation for their solution for the simultaneous OT problem. The notation used is standardized, and sufficient background on optimal transport is provided to understand the paper.
  • Clear presentation of of the SimNOT algorithm. The presentation of theorems and empirical results are well-formulated and easy to follow.
  • The proposed method is evaluated across with synthetic toy experiment and more practical image-restoration task.
Critical weaknesses
  • Consider adding these citation to the Related Works L103-104: [1] and [2].
  • The evaluation would be stronger if it spanned additional practical datasets/settings. An example could be pulling from single-cell sequencing datasets, which often are analyzed across labs/protocols/machines. This setting may be unbalanced because cells may die or be created from one time point to the next.
  • Figure 1 is not referenced in the text. It would help strengthen the argument in Table 2 around Table 2, and how it's different from Table 1. There should be qualitative results that match the setting in Table 1.
  • It would be helpful to ablate certain factors of the input distributions, like the number of input distributions, or the initial overlap of these source distributions.

[1] Adrai, T., Ohayon, G., Elad, M., & Michaeli, T. (2023). Deep optimal transport: A practical algorithm for photo-realistic image restoration. Advances in Neural Information Processing Systems, 36, 61777-61791. [2] Tartavel, G., Peyré, G., & Gousseau, Y. (2016). Wasserstein loss for image synthesis and restoration. SIAM Journal on Imaging Sciences, 9(4), 1726-1755.

Questions for authors
  • Does learning a single transport map still make sense as the number of source distributions K increases?
  • How do we interpret Conditional UOT outperforming SimNOT on CelebA restoration? What is the intuition for the difference with the results in Table 2 on different scale transfers?
  • How does SimNOT compare against standard IR baselines? Is it possible to adapt OT barycenter-based solutions to this SOT setting? This may not be a fair comparison, but it would be interesting to see.
Reviewer_AN53OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SimNOT, a continuous neural solver for many-to-one Simultaneous Optimal Transport (SOT) from unpaired samples. The authors introduce a divergence-based unbalanced SOT formulation, derive an exact semi-dual objective (Theorem 1), and prove neural approximation consistency (Theorem 2) and deterministic Monge structure under quadratic-KL settings (Theorem 3). Using one shared generator and KK source-specific potentials during training, SimNOT performs blind inference and is validated on a 2D toy task and 64×6464\times 64 CelebA image restoration.

Critical strengths
  1. Proving zero relaxation gap (Lemma 1), the exact semi-dual objective (Theorem 1), and the deterministic Monge structure (Theorem 3) provides a solid mathematical foundation for continuous multi-source OT.
  2. Decoupling training potentials from the shared generator enables blind restoration, avoiding the catastrophic routing failure suffered by Conditional UOT under unseen downsampling factors.
Critical weaknesses
  1. Evaluating only on 64×6464\times 64 CelebA is overly simplistic for image restoration. Standard high-resolution benchmarks and comparisons against established unpaired restoration methods are missing.
  2. Despite the motivation that pooling misaligns individual sources, Pooled UOT outperforms SimNOT on in-distribution Bicubic and Bilinear downsampling (Table 3) and significantly beats SimNOT on strong OOD shifts like Bicubic ×5\times 5 (FID 40.31 vs. 151.24, Table 5) and Blur σ=1.75\sigma=1.75.
  3. Maintaining KK independent potential networks (e.g., Discriminator_large in CelebA) may scale the parameter count and training memory linearly with KK.
Questions for authors
  1. Theorem 3 justifies a deterministic Monge map Tθ(x)T_\theta(x), yet image restoration (e.g., severe deblurring/inpainting) is inherently ill-posed (one-to-many). Would incorporating a stochastic noise input Tθ(x,z)T_\theta(x, z) improve perceptual quality (FID/LPIPS)?
  2. Theoretically, if two sources share overlapping regions in X\mathcal{X} that require different transport directions, how does a single deterministic map Tθ(x)T_\theta(x) resolve this ambiguity without source labels?
  3. Experiments only test uniform weights (αk=1/K\alpha_k = 1/K). How does SimNOT behave when source datasets have highly imbalanced sample sizes or non-uniform αk\alpha_k?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.