acceptodds
Under review as a conference paper at ICLR 2027

Beyond Explicit Mirror Maps: Constrained Generative Modeling via Brenier Maps

Abstract

Mirror-lifting paradigms for constrained generation are concise and efficient, but they suffer from fundamental limitations. Analytical mirror functions are only val-id for specific convex constraint domains, lacking a unified construction method and failing to handle non-convex constraints. Neural approximate mirror functions broaden applicable scenarios yet brings uncontrollable geometric distortion and heavy training overhead. To address these issues, we propose a unified framework for constructing constrained generative models based on Brenier maps. Theoretically, we prove that mirror functions on convex domains constitute a special case of dual Brenier functions. Leveraging singular set theory, we break the convexity constraint of mirror maps and enable unified adaptation to arbitrary convex and non-convex domains. Algorithmically, we develop a universal multi-stage solving strategy for Brenier maps. Anchor sampling and geometric variational principles are adopted to capture precise geometric structures, and a controllable smoothing mechanism is introduced to preserve critical geometric properties and improve generalization. Experimental results show that on three classical convex constraint domains, our method achieves better fitting accuracy for mirror functions than neural approximation baselines. For the watermark image generation task and five other complex real-world constrained generation tasks involving both convex and non-convex scenarios, our method surpasses existing SOTA baselines in generation quality and constraint satisfaction with lower computational cost, substantially improving the accuracy and efficiency of constrained generation.

Primary area:
generative models
Keywords:
Mirror Maps, Brenier Maps, Constrained Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_c8RnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper generalized the mirror flow matching framework to general non-convex constraints. The idea is based on optimal transport. First learn a Brenier Map and then perform generative modeling. Since exact Brenier map could be difficult to compute, the author proposed to uniformly sample poins inside the constrain set, and then approximate the Brenier map.

Critical strengths

The proposed approach is interesting. It extended previous approaches beyond convex set.

Critical weaknesses

This paper lacks some theoretical justification.

  • Is it guaranteed that, the output samples will satisfy constraint, as in mirror map based approaches? Line 240 equation 5, it seems to me that the output of defined Brenier map, is a convex combination of sampled (uniform) points. And even if the constraint set CC is non-convex, the output of Brenier map is still a convex combination, which seems to me, would corresponds to a convex set instead of the original non-convex set CC. If I didn't mis-understand this part, how can the author guarantee feasibility? I also noticed in table 2, the reported CSR for proposed approach is not 100100%.

This is my main concern. If this can be addressed (including but not limited to, theoretically analyzing whether there is some constraint satisfaction ratio that can be guaranteed, and what would affect this constraint satisfaction, etc.) I will raise my score.

Questions for authors

line 132: I think there should be a more detailed introduction on what Σ\Sigma is. I see that there is an explanaiton around line 190, but I think for people not familiar with this "singular set theory", it could be confusing.

line 210: in general is it expensive to generate uniform samples?

As discussed in mirror flow matching (Guan et al., 2026), when data points get close to the boundary, the mirror map might send these samples to infinity, making the learning difficult or unstable. Is there a similar issue in the proposed approach? In particular, the mentioneed singular set, will it cause some issue like numerical instability, bad regularity, etc.?

Reviewer_R2gPOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework for constrained generative modeling based on Brenier maps. The main idea is to replace conventional mirror maps with smoothed Brenier maps, allowing the framework to handle both convex and non-convex constraints. The authors establish a theoretical connection between mirror maps and Brenier maps, and develop a numerical procedure based on anchor sampling, geometric variational optimization, and smoothing. The resulting maps are used to transform constrained data into an unconstrained space for training standard generative models, with generated samples subsequently mapped back to the constrained domain.

The method is evaluated on mirror map approximation, watermarked image generation, and several constrained generation tasks. The reported results show improvements in constraint satisfaction, generation quality, and computational efficiency compared with the considered baselines.

Overall, I have a positive impression of this paper. The motivation is clear, the proposed framework seems well developed, and the experimental results are encouraging. I also appreciate the effort to connect the theoretical analysis with practical constrained generation tasks. My main reservation is that this specific topic is outside my primary research area, and I am not sufficiently familiar with the optimal transport and mirror map literature to confidently assess the significance of the theoretical contributions relative to prior work. Therefore, I lean toward weak acceptance, but with relatively low confidence.

Critical strengths
  1. The paper is well organized and generally easy to follow, despite the mathematical complexity of the topic. I found the motivation for moving beyond conventional mirror maps reasonably clear, and the overall framework is explained well.

  2. The experimental evaluation is quite comprehensive. The authors consider different constrained domains and several generative model architectures, rather than focusing on a single setting. The results are generally consistent and suggest that the proposed approach is effective across different tasks.

Critical weaknesses
  1. The method involves several preprocessing steps, including anchor sampling and numerical optimization. Although the paper reports computational advantages over NAMM, I would have liked a clearer discussion of the scalability of these steps, especially in higher-dimensional settings.

  2. Some of the theoretical arguments are quite technical. While the overall motivation is clear, I found it somewhat difficult to understand how the theoretical guarantees translate into practical behavior when using a finite number of anchor points and a nonzero smoothing coefficient.

Questions for authors
  1. How does the computational cost of constructing the Brenier map scale with the data dimension and the number of anchor points? I am particularly interested in whether the proposed procedure remains practical for substantially higher-dimensional generation tasks.

  2. The paper discusses the effect of the smoothing coefficient on geometric distortion. In practice, how sensitive is the final generation performance to this parameter? Is there a general strategy for choosing it across different constraint domains?

  3. The experiments cover several types of constraints, including non-convex ones. Are there particular geometric properties of the constrained domain that make the proposed Brenier map difficult to construct or numerically unstable?

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

This paper presents to use a finite-anchor and semi-discrete optimal-transport map to contruct a constrained generative model. This is to lift the training data onto a transformed space, where a conventional generator is trained there and mapped back to original data space.

Authors validate their approach on diffusion, flow matching, VAEs, as different types of generative models, and on watermarking task and other 5 additional contraint tasks. The experiments show superior results on most evaluated metrics.

Critical strengths

General design. The proposed method can be applied to vairous generative model families, such as, diffusion, flow matching, VAEs, etc. This means it has practical flexibility and is a genral design.

Good experimental results. The proposed approach is validated on several application tasks, which shows superior results on most of the metrics, suggesting that the design has good practical use.

Critical weaknesses

Reduction in overall training cost is not as effective as the paper claims. Tab. 3 and 4 show the counting task with flow matching is about more than 11 hours for the proposed method, while the unconstrained basline is less than 6 hours. This means despite having a faster generator training, the whole system still requires more time in this case.

Reported numbers does not match prior paper. For NAMM, in tab. 4 the constraint distance is 5.2213, which is not found in the paper, and has an obvious gap to the reported numbers in the original paper.

Questions for authors
  1. Could you elaborate more on the effectiveness on time cost calculation, to see if what I observed is indeed an weakness case?
  2. Where can I find the reported number in original paper, or is it something you implemented?

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your thoughtful reviews. The authors have provided detailed responses to the concerns raised.

I would appreciate it if you could revisit the author responses and clarify whether your main concerns have been adequately addressed, particularly which issues remain critical to your overall assessment.

Given the differences in the initial recommendations, I also encourage you to consider and discuss one another's key arguments to help clarify the remaining disagreements and develop a more consistent understanding of the paper's merits and limitations.

Please feel free to update your ratings if your assessments have changed following the discussion.

Thank you for your continued contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.