acceptodds
Under review as a conference paper at ICLR 2027

Branched Optimal Transport Amortization

Abstract

Methods of Branched Optimal Transport (BOT) mimic the economy and efficiency of natural tree-like structures, such as those found in rivers and biological systems. These methods are widely applicable for designing efficient networks in society, from river basins and blood vessels to mail and gas distribution systems. However, they remain understudied in the context of designing deep generative models, particularly at a large scale. Standard continuous-time generative models, such as the flow matching approach, fail to capture the inherent hierarchical and branching patterns present in real-world data. Current models provide no mechanism for flows to merge or share pathways to minimize total transport cost. Inspired by the "economy of scale" principle in BOT, we introduce a novel, scalable branched flow-matching algorithm designed to solve the branched optimal transport problem in high dimensions. Our method adapts the Benamou-Brenier continuous-time optimal transport formulation to learn branched generative flows. These flows allow probability mass to aggregate along common pathways before branching out to diverse targets. Parametrized by neural networks, our method effectively learns complex branched generative processes. We demonstrate its effectiveness on challenging high-dimensional tasks in biology and image generation.

Primary area:
generative models
Keywords:
branched optimal transport, flow matching, deep generative models, continuous-time optimal transport, hierarchical generative modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NTmbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Branched Optimal Transport Amortization (BOTA), a method for learning flow-based generative models whose trajectories exhibit hierarchical, tree-like branching structures. Inspired by the economy-of-scale principle of Branched Optimal Transport (BOT), the method encourages particles to share transport pathways before branching toward their respective destinations. Analogous to minibatch optimal transport, BOTA follows a two-stage procedure: (1) optimizing discrete particle trajectories using a BOT-inspired transport cost and (2) amortizing the optimized trajectories into a neural velocity field through flow matching.

A central challenge is that theoretical BOT requires atomic probability measures, whereas continuous normalizing flows operate on continuous distributions. To address this incompatibility, the authors introduce a differentiable soft-atomic relaxation that approximates the mass-dependent BOT cost using soft assignments of particles to proxy centers. The resulting discrete optimization combines this transport cost with a terminal matching objective. The proposed method is evaluated on Gaussian mixtures, simulated biological differentiation data (TedSim), and image datasets (MNIST and FFHQ). The experiments demonstrate branching-like trajectory structures and reduced soft-atomic transport costs compared with standard flow matching, while maintaining comparable distribution matching performance.

Critical strengths

Novelty. The paper presents a methodological contribution by introducing a differentiable, soft-atomic approximation of Branched Optimal Transport (BOT) and amortizing the resulting optimized particle trajectories into a neural velocity field. While the mass-dependent transport objective originates from classical BOT formulations, the general paradigm of incorporating optimal transport solutions into flow matching has been explored in prior work [1]. Furthermore, branching generative flows have already been investigated in Branched Schrödinger Bridge Matching [2], which also employs trajectory optimization followed by flow matching-based amortization. Thus, the primary novelty lies in the specific mass-dependent soft-atomic objective and its discrete numerical optimization, rather than the broader ideas of branching trajectories or trajectory optimization followed by amortization.

[1] Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport, TMLR, 2024.

[2] Branched Schrödinger Bridge Matching, ICLR, 2026.

Critical weaknesses

Insufficient theoretical justification. The proposed soft-atomic objective approximates BOT through particle clustering, but continuous normalizing flows cannot represent exact merging and splitting of trajectories. The theoretical connection between the proposed relaxation and true BOT remains insufficiently established, with no guarantees of optimality or convergence.

Insufficient experimental validation and baseline comparisons. The experiments primarily demonstrate visually plausible branching trajectories and reductions in the proposed soft-atomic cost, without establishing whether the learned trajectories recover meaningful hierarchical structures. Moreover, important baselines such as OT-CFM, ReFlow, and existing branching generative models are missing or insufficiently compared, making it difficult to assess the advantages of the proposed approach.

Missing ablations and unclear practical benefits. Despite the newly added analysis of the branching parameter α\alpha, the paper lacks systematic ablations on key hyperparameters such as KK, NN, σ\sigma, and λ\lambda. Furthermore, although BOTA achieves lower branched transport costs, it does not demonstrate meaningful improvements in generation quality or sampling efficiency over standard Flow Matching, despite requiring approximately twice the training time.

Questions for authors
  • Theoretical justification: Given that continuous normalizing flows cannot represent exact merging and splitting of trajectories, in what sense does the proposed soft-atomic relaxation approximate true Branched Optimal Transport? Can the authors provide stronger theoretical justification or clarify the limitations of this approximation?
  • Validation of hierarchical structure: The experiments mainly demonstrate visually plausible branching trajectories and reductions in the proposed soft-atomic cost. Can the authors quantitatively evaluate whether the learned trajectories recover meaningful hierarchical structures, for example, using the ground-truth lineage information available in TedSim?
  • Baseline comparisons and practical benefits: Can the authors provide comparisons against relevant OT-based methods, such as OT-CFM and ReFlow, as well as stronger comparisons with existing branching generative models? Beyond reducing the proposed transport cost, what practical advantages does BOTA offer in terms of generation quality, computational efficiency, or downstream applications?
  • Hyperparameter sensitivity: Can the authors provide systematic ablations on the number of proxy centers KK, particle count NN, bandwidth σ\sigma, and terminal matching weight λ\lambda? How sensitive are the learned branching structures and distribution matching performance to these choices?
2 comments

Dear Reviewer, thank you for your feedback.

Answers to Weaknesses

  1. Connection to branched optimal transport.

    BOTA's contribution is the soft-atomic, mass-dependent objective (Section 4), which encourages shared routes. Within a Lipschitz ODE flow, branching takes the form of distinct, nearby paths traveling together before diverging.

    Lemma 1 (p. 5, lines 239–254; Appendix A, pp. 11–12) establishes the hard-assignment limit as σ→0\sigma\to0 for fixed centers and unique nearest-center assignments. For the mathematical objective without numerical stabilizers, the limiting cost equals the atomic BOT cost when each occupied cell represents one atom with a common velocity. Stage 1 optimizes paths for fixed source–target assignments with an endpoint-matching penalty (Section 5.1, p. 6, lines 290–298).

  2. Evaluation and baseline comparisons.

    Table 1 (p. 7) compares BOTA with BSBM, CNF, CFM, and FM. BOTA achieves approximately 2.3×2.3\times lower soft-atomic cost than FM and CFM, with comparable endpoint matching (Section 6.2, p. 8, lines 388–395; Figure 8b, p. 14).

    We have now added minibatch OT-CFM and ReFlow comparisons in a supplementary TedSim time-8-to-time-12 experiment. All methods are evaluated with the same BOT cost (α=0.5\alpha=0.5, shared bandwidth and number of integration steps). Values are mean ±\pm sample SD over five training seeds on the same simulation.

    MethodBOT cost ↓
    FM115.78 ± 5.74
    OT-CFM92.36 ± 2.96
    ReFlow121.60 ± 8.17
    BOTA80.29 ± 10.60

    BOTA achieves the lowest mean cost, 13.1% below OT-CFM.

  3. Practical benefits and parameter analysis.

    On FFHQ, BOTA's FID is comparable to FM: 10.46±1.8010.46\pm1.80 versus 10.61±1.3910.61\pm1.39 (Appendix Table 2, p. 12). Appendix Table 3 (p. 15) reports 425 s for offline optimization + 360 s for amortization, versus 360 s for FM. The reported sampling protocol uses the same network size and integration procedure. These submitted results demonstrate control over transport geometry alongside comparable endpoint and sample quality.

    The α\alpha ablation shows how the branching exponent controls shared-trunk geometry (Figure 6, p. 13). Question 4 details the parameter settings and their roles.

Answers to Questions

  1. In what sense does the relaxation approximate BOT?

    Please see our response to Weakness 1 above.

  2. Does BOTA recover the TedSim hierarchy?

    Please see our response to Weakness 2 above.

  3. What comparisons and practical benefits are established?

    Please see our responses to Weaknesses 2 and 3 above.

  4. How sensitive is BOTA to KK, NN, σ\sigma, and λ\lambda?

    Appendix A.4 (p. 15, lines 758–773) reports K=50K=50, σ=0.6\sigma=0.6, a stage-1 particle count of N=10,000N=10{,}000, and λ=20\lambda=20, with centers following the particle cloud. KK controls spatial resolution, NN affects mass estimates, and λ\lambda balances endpoint matching against path cost. Large σ\sigma blurs assignments; very small σ\sigma makes them nearly one-hot, weakening positional gradients away from nearest-center ties (Section 8, p. 9, lines 450–452).

    Figure 6 (p. 13) provides a qualitative ablation over α∈{0.1,0.5,0.9}\alpha\in\{0.1,0.5,0.9\}: smaller values favor shared trunks.

Thank you again for your thoughtful review. We hope these responses clarify the contribution and would appreciate your reconsideration of the score.

Thank you for the detailed response and additional experiments. I appreciate the new OT-CFM and ReFlow comparisons, which strengthen the empirical evidence that BOTA reduces the proposed soft-atomic transport cost while maintaining comparable endpoint matching. I also acknowledge the clarification that BOTA primarily aims to control transport geometry rather than improve generation quality or sampling efficiency.

However, two central concerns remain:

1. Insufficient theoretical justification. Lemma 1 (Section 4, p. 5) establishes convergence of soft assignments to hard nearest-center assignments as the bandwidth approaches zero. However, this does not establish that minimizing the proposed objective approximates solutions to the original BOT problem. In particular, convergence or optimality guarantees for the resulting transport trajectories remain unclear.

2. Lack of validation of meaningful hierarchical structure. In response to Question 2, the authors refer to their additional baseline comparisons, which demonstrate lower soft-atomic transport costs but do not establish whether the learned trajectories recover TedSim's ground-truth lineage relationships. Importantly, the authors explicitly acknowledge in Section 8 (p. 9, lines 454–456) that direct evaluation against TedSim's ancestor–descendant coupling is left to future work. This missing validation is particularly concerning given the paper's central motivation of learning meaningful hierarchical generative processes.

Overall, while the rebuttal strengthens the empirical evidence for the proposed objective, lower transport cost alone does not establish theoretical fidelity to BOT or recovery of meaningful data hierarchies. As these central concerns remain unresolved, I maintain my current rating of weak rejection.

Reviewer_NkFQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

There is a strong body of literature demonstrating that the joint distribution between data and noise used in generative models based on continuous-time normalizing flows (CNFs) matters for downstream performance: the standard independent coupling x0,x1∼N(0,I)pdatax_0,x_1\sim N(0,I)p_\text{data} induces a distribution over paths that are less "straight" then the OT coupling obtained by optimally matching x0,x1x_0,x_1, typically done within each mini-batch (see Tong et al 2023).

The paper introduces the use of branched optimal transport (BOT) couplings in CNF models, and proposes a scalable method to find such couplings on high-dimensional data. The motivation for why such couplings may help is sound: they are inductive biases that encourage distributions over generative paths to group into semantically meaningful "clusters". The BOT cost differs from the standard Benamou-Brenier formulation in that it introduces a cost in trajectories spreading out through the sample space: raising the density by a value 0<α<10<\alpha<1 makes it so that merging mass is cheaper than letting it spread out. What made it "click" for me was the simple observation that ∣m1+m2∣α≤∣m1∣α+∣m2∣α|m_1 + m_2|^\alpha \leq |m_1|^\alpha + |m_2|^\alpha for 0<α<10<\alpha<1 and positive mass -- I wish the authors spent a few lines delivering that intuition. Anyways, the challenge with BOT couplings is that computing them is non-trivial.

BOT is effectively weighting the path cost by the density of samples at each time tt: this is clearly challenging, since flow matching does not directly provide a simulation-free estimate of density. Therefore, the method relies on simulation-based continuous-time normalizing flows (CNFs). The velocity is defined as a "mixture", and weighted by the probability of each "attractor branch" via an exponential kernel. Lemma 1 shows that as the σ\sigma in this kernel goes to zero, one recovers (an approximation of) the BOT cost in (12).

The method in 5.1 proposes to learn (approximate-)BOT couplings via a simulation-based method, which computes optimal assignments and routing on each minibatch by directly differentiating through the simulation of each trajectory. This is clearly not scalable to high dimensional spaces. Most CNFs use the adjoint method, but this possibility is not discussed in the paper. This "phase" does not immediately yield a generative model, it only learns the optimal continuous-time trajectories. The actual generative model is learned via "branched flow matching", which requires re-using the intermediate states from the learned BOT trajectories.

The experiments are quite unconvincing: GMM shows the emergence of interesting structures in the path, but it's unclear how this should help in actual generative modeling applications. Results on the Tedsim dataset show marginally better performance of BOT, but the experimental protocol (e.g. how K and alpha were chosen) is unclear and introduces additional hyperparameters. Results on image data are qualitative at best, and the only quantitative ones (on MNIST) do not show significant improvements.

References:

  • Tong et al 2023; Improving and generalizing flow-based generative models with minibatch optimal transport.
Critical strengths
  • Introducing BOT to generative modeling is a novel and interesting contribution.
Critical weaknesses

Methodology:

  • The proposed method does not recover the BOT coupling, and the authors do not make it clear enough. In section 5.1 the paper explicitly says that source and targets are coupled independently, and only the intermediate states are optimized! This is very different from an OT coupling, which induces a coupling at the endpoints as well.
  • It is unclear to me how scalable the approach in Section 5.1 really is, since it involves optimizing entire sampled trajectories.

Experiments:

  • No substantial evidence that the claims in the introduction ("BOT perspective is powerful for generative modeling because it encodes shared prefixes of computation") are supported in practice.
  • Limited experimental results, poor documentation of how each method was tuned.

Writing:

  • important notation is missing: what is ItI_t in (7)? This is crucial to understanding the method. In a broader sense, I don't see the point of introducing the atomic version (7) before the "continuous" version in (9), only to then resort back to a discrete approximation in (12).
  • in section 5.1, the paper justifies using an MSE loss by saying that it is preferable to a heavier probabilistic loss. But this is effectively just a Gaussian NLL loss on the terminal points with a fixed standard deviation, which is a probabilistic loss.
  • the syntax is off in a few sentences throughout the paper. For example: "Branched Benamou–Brenier formulation studied by Brasco et al. (2011), which is also equivalent to the formulation of Xia (2003). " (??)

Minor: the paper only cites two out of the three concurrent works that independently introduced flow matching. It's missing "Building Normalizing Flows with Stochastic Interpolants" from Albergo and Vanden-Eijnden.

Questions for authors

Suggestion for future work: it is unclear to me why BOT couplings would improve performance in generative models, compared to flow matching with OT coupling. However, I see a strong argument to adapt this framework to trajectory inference for applications where the trajectory between source and target samples have a physical meaning -- essentially extending experiments on biological data and making them more rigorous, and abandoning dubious claims for general purpose applications. The BranchSBM paper would be a good starting point for experiments of this kind.

2 comments

Dear Reviewer, thank you for your feedback.

Answers to Weaknesses

  1. Endpoint coupling.

    Stage 1 performs BOT-inspired path optimization for fixed, independently sampled source–target pairs. It optimizes their velocities and uses a soft terminal penalty for approximate endpoint matching. We will make this scope explicit in Section 5.1.

    Nearby particles contribute to shared proxy masses, so their trajectories interact through the soft-atomic cost. Its concave mass weighting, (m1+m2)α<m1α+m2α(m_1+m_2)^\alpha<m_1^\alpha+m_2^\alpha for positive masses and 0<α<10<\alpha<1, favors transporting mass together along shared routes.

  2. Computational cost and scalability.

    We optimize V∈RT×N×DV\in\mathbb{R}^{T\times N\times D} through Xk+1=Xk+ΔtVkX_{k+1}=X_k+\Delta tV_k, without neural-network, flow-density, or divergence evaluations. Differentiation uses discrete updates; no continuous adjoint is used. Each solver iteration costs O(TNKD)O(TNKD): linear in NN for fixed T,K,DT,K,D, matching FM's batch-size scaling with a fixed architecture (Appendix A.4, p. 15, lines 779–795). Memory is O(TND)O(TND) for velocities and up to O(TNK)O(TNK) for particle–center weights.

    Amortization has the same per-update computational complexity as FM, and sampling has the same complexity under the reported architecture and integration settings. Appendix Table 3 (p. 15) reports 425 s for optimization + 360 s for amortization on TedSim, versus 360 s for FM. The additional trajectory optimization is performed once offline.

  3. Practical benefits.

    BOTA achieves lower transport cost at similar endpoint agreement on TedSim (Section 6.2) and comparable FID on FFHQ (Table 2). We will clarify that shared prefixes refer to trajectory geometry; inference retains per-particle network evaluations. Please see our response to Question 1 below for the supplementary biological experiments.

  4. Parameter settings and experimental documentation.

    Appendix A.4, “Proxy atoms, KK and σ\sigma” (p. 15, lines 758–769) reports K=50K=50 and σ=0.6\sigma=0.6, with centers following the particle cloud. KK controls spatial resolution and can change the cost scale. As σ\sigma grows, assignments approach uniformity; very small σ\sigma makes them nearly one-hot, weakening positional gradients away from nearest-center ties (Section 8, p. 9, lines 450–452).

    Appendix A.4, “Solver and amortization” (p. 15, lines 771–773) reports a stage-1 batch of N=10,000N=10{,}000, λ=20\lambda=20, and T=50T=50 on TedSim (T=10T=10 for Gaussians). TedSim uses α=0.5\alpha=0.5 in Figures 3 (p. 8) and 8c (p. 14). Figure 6 (p. 13) gives a qualitative ablation over α∈{0.1,0.5,0.9}\alpha\in\{0.1,0.5,0.9\}: smaller values favor shared trunks. We will organize these settings by experiment and distinguish fixed parameters from ablations.

  5. Notation and presentation.

    We will address these writing issues in the revision: define the missing notation, clarify the relationship between the atomic, continuous, and discrete formulations, correct the MSE/Gaussian NLL explanation and specify the loss-averaging convention, and revise unclear sentences, including the Brasco et al. sentence. We will also add the missing Albergo and Vanden-Eijnden citation.

Answers to Questions

  1. Trajectory inference and the BranchSBM suggestion.

    Trajectory inference is a natural application of the transport objective, and we appreciate the suggested direction, including BranchSBM.

    We have conducted additional exploratory experiments on three biological datasets using a local-mass variant of BOTA. These yield 23.5% lower mean transport cost on mouse, 31.4% on Veres, and 3.8% on Trametinib relative to FM (α=0.5\alpha=0.5). The table reports this cost as mean ±\pm sample SD over five seeds; lower is better. FM has lower endpoint errors in these runs.

    DatasetMethodLocal-mass cost ↓
    MouseFM3.627 ± 0.012
    MouseBOTA2.775 ± 0.188
    VeresFM197.975 ± 1.701
    VeresBOTA135.837 ± 5.591
    TrametinibFM55.061 ± 0.921
    TrametinibBOTA52.991 ± 2.503

    Evaluation settings. Costs are evaluated in 2, 30, and 50 dimensions for mouse, Veres, and Trametinib, respectively, with a common bandwidth derived from training data per dataset.

We appreciate your careful reading and constructive feedback. We hope the responses above clarify our contribution and would be grateful for your reconsideration of the score.

I thank the authors for their detailed response and additional experiments. However, my main concern remains unresolved: it is unclear what practical benefits the BOT perspective provides in the context of generative modeling. The additional biological experiments are interesting, but I remain unconvinced of the motivation for the proposed approach in generative modeling and therefore maintain my original score.

Reviewer_KEi6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a two-stage framework for learning generative flows inspired by BOT. Unlike standard optimal transport, which penalizes transport linearly in the transported mass, BOT introduces a concave mass-dependent cost that encourages particles to share transport pathways.

Critical strengths
  1. The proposed two-stage approach to learn structured transport trajectories is very direct and makes sense intuitively. Since standard flow matching does not explicitly encourage the sharing of transport paths among samples with related destinations, so the author proposed to incorporate a carefully-designed transport objective into generative modeling. This is appealing from a computational perspective. Separating discrete trajectory optimization allows the potentially expensive routing optimization to be performed offline, while retaining a conventional continuous-time generative model for inference.
  2. The synthetic Gaussian experiments illustrate a visible change in trajectory structure when the branching parameter varies. The results of TedSim experiments are also encouraging: BOTA reports lower transport cost under its proposed surrogate.
Critical weaknesses

BOTA learns meaningful branching trajectories, but the experiments do not directly verify this. In particular, TedSim experiment provides ground-truth lineage information, yet the evaluation mainly relies on the proposed soft-atomic cost and endpoint distribution metrics. A lower cost under the method's own objective does not necessarily mean that the learned trajectories recover the true branching structure.

Questions for authors
  1. Since TedSim provides ground-truth lineage information, could the authors directly evaluate how well BOTA recovers the actual branching relationships compared with the baselines?
  2. How sensitive are the learned branching structures and reported cost improvements to the choice of proxy atoms, particularly K and \sigma?
2 comments

Dear Reviewer, thank you for your feedback.

Answers to Weaknesses

  1. Evaluation of branching trajectories.

    The TedSim dataset provides expression snapshots and ancestry labels, but no ground-truth continuous cell trajectories. The labels cannot specify paths between observations.

    To assess the learned paths, we use a criterion motivated by branched optimal transport. To our knowledge, we are the first to introduce this differentiable soft-atomic relaxation of the branched transport cost for amortized generative flows (Section 4). We also use it as an evaluation metric because it encodes the preference for transporting mass efficiently along shared routes. Endpoint metrics assess whether the target distribution is matched, but do not characterize the paths taken. Scoring BOTA, FM and CFM with the same cost and settings therefore compares their performance under a common transport criterion.

    Under this criterion, BOTA outperforms FM and CFM on TedSim: 174.81±5.02174.81\pm5.02 versus 405.08±1.00405.08\pm1.00 and 408.88±1.00408.88\pm1.00, respectively (Section 6.2; Figure 8b). This demonstrates BOTA's quantitative advantage under the BOT-motivated objective, alongside the qualitative shared-trunk geometry in Figure 8.

Answers to Questions

  1. Does BOTA recover the TedSim lineage?

    Please see our response to Weakness 1 above.

  2. How sensitive is BOTA to KK and σ\sigma?

    We discuss their qualitative effects in Appendix A.4, “Proxy atoms, KK and σ\sigma” (p. 15, lines 758–769), and summarize the bandwidth failure modes in Section 8 (p. 9, lines 450–452). The reported settings are K=50K=50 and σ=0.6\sigma=0.6, with centers following the particle cloud.

    As σ\sigma grows relative to particle–center distances, assignments approach 1/K1/K and lose spatial discrimination. As σ→0\sigma\to0, they become one-hot and their positional gradients vanish away from nearest-center ties. KK affects spatial resolution and can change the cost scale.

    Separately, Appendix Figure 6 (p. 13) reports an ablation over the branching exponent α∈{0.1,0.5,0.9}\alpha\in\{0.1,0.5,0.9\}. It illustrates how α\alpha controls aggregation: smaller values favor shared trunks, while larger values produce less aggregated paths.

Thank you again for your careful review. We hope these clarifications address your concerns and would appreciate your reconsideration of the score.

Thank you for the clarifications. I understand that TedSim does not provide ground-truth continuous trajectories, and I appreciate the additional explanation of the proxy-atom parameters. However, my concern remains that the evaluation does not independently validate the learned branching structure. Because of unavailability of the exact trajectories, the ancestry labels could still support a lineage-level structural evaluation. Moreover, the discussion of K and σ provides useful intuition but does not establish empirical sensitivity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.