acceptodds
Under review as a conference paper at ICLR 2027

LightSBB-M: Bridging Schrödinger and Bass for Generative Diffusion Modeling

Abstract

The Schrödinger Bridge and Bass (SBB) formulation, which jointly controls drift and volatility, is an established extension of the classical Schrödinger Bridge (SB). Building on this framework, we introduce LightSBB-M, an algorithm that computes the optimal SBB transport plan in only a few iterations. The method exploits a dual representation of the SBB objective to obtain analytic expressions for the optimal drift and volatility, and it incorporates a tunable parameter β > 0 that interpolates between pure drift (the Schrödinger Bridge) and pure volatility (Bass martingale transport). On a Gaussian case with a closed-form solution, we verify that the algorithm recovers the optimal joint plan and both optimal controls, not merely the terminal marginal. We show that LightSBB-M achieves the lowest 2- Wasserstein distance on synthetic datasets against state-of-the-art SB and diffusion baselines with an average of 19% improvement. We also illustrate the generative capability of the framework on an unpaired image-to-image translation task (adult → child faces in FFHQ). These findings demonstrate that LightSBB-M provides a scalable, high-fidelity SBB solver that outperforms existing SB and diffusion baselines across both synthetic and real-world generative tasks.

Primary area:
generative models
Keywords:
Machine Learning - Generative Modeling - Optimal transport

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

The paper proposes LightSBB-M, an alternating solver for the Schrödinger–Bass bridge problem, which controls both drift and volatility. It uses the SBB representation as a transformed constant-volatility Schrödinger bridge. An inner LightSB-M-style procedure learns a Gaussian-mixture potential, while a separate regression learns the inverse endpoint maps. Sampling uses the learned endpoint coupling and a deterministic transformation rather than numerical integration of the full stochastic-volatility SDE.

Experiments include a Gaussian problem with known optimal coupling and controls, three two-dimensional transport benchmarks, and adult-to-child translation in a pretrained ALAE latent space.

Critical strengths

The Gaussian validation is well motivated. Evaluating the joint plan, cross-covariance, drift, and volatility goes beyond terminal-marginal matching and addresses what an SBB solver is supposed to recover. The five-seed comparisons provide useful evidence that the added flexibility can improve control recovery.

The computational construction is promising. Solving a constant-volatility bridge in transformed coordinates avoids explicitly learning a large matrix-valued volatility field, and endpoint sampling avoids SDE discretization at inference. The paper also reports the additional training cost rather than presenting the method as uniformly cheaper than LightSB-M.

Critical weaknesses

The claim of computing the optimal SBB plan is not established for the implemented alternating procedure. Exact solution of each SB subproblem does not establish convergence or optimality of the changing outer problem. The authors acknowledge the missing convergence proof, but marginal Wasserstein plateaus are not an optimality certificate. Moreover, the theoretical invertibility conditions are not enforced by the learned potential and inverse network. The main image result uses beta*T=1, outside the strict condition used in the stated guarantees. Numerical inverse residuals, curvature diagnostics, SBB-system residuals, and primal–dual gaps would substantially strengthen the solver claim.

The mathematical justification contains issues that affect the admissibility and control interpretation. Proposition A.1 is stated for every continuous quadratic-growth potential, but this class does not ensure finite Moreau transforms: in one dimension, phi(y)=-beta*y^2 belongs to the stated class while T_beta^+phi=-infinity. The proposition requires a proper domain and additional finiteness conditions.

The displayed identity dX_t=D²Phi_t(Y_t)dY_t also omits time and Itô correction terms; Appendix C.3 gives a fuller expression that should be reconciled with the main text. Finally, the explanation for replacing the full SBB objective with a drift-only gap is not convincing. An endpoint-feasible SB with constant reference volatility is an admissible SBB candidate with zero volatility penalty, so its full cost cannot be below the true SBB optimum. The full objective and endpoint feasibility should be the primary optimality diagnostics.

Questions for authors

What numerical evidence establishes that the outer iteration approaches an SBB optimum rather than merely stabilizing its marginal error? Can you report SBB-system residuals or primal–dual gaps, together with sensitivity to initialization?

What is the correct admissible domain for Proposition A.1? Does the implemented algorithm perform an envelope projection, or otherwise ensure the required curvature and invertibility? Please report inverse-map residuals and minimum-Jacobian-eigenvalue statistics, particularly for the image experiment at beta*T=1.

Please reconcile the main-text dynamics with the Itô expressions in Appendix C.3. Which controls were used to obtain Table 1, and how are learned and exact states coupled when computing the drift recovery error?

Please clarify the full-objective comparison. If a negative gap occurs, what marginal infeasibility, truncation, or estimation error explains it? Can the complete SBB objective and marginal residuals be reported together for all Gaussian comparisons?

Can image identity preservation be evaluated with an independent face-recognition representation rather than cosine similarity of ALAE latent codes? Please also report uncertainty for FID and clarify the contributions of evaluation seeds versus variation across individual faces.

How is the terminal approximation implemented? Are samples drawn at T or at T-delta, and how sensitive are endpoint accuracy and objective estimates to delta?

Reviewer_mck2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LightSBB-M, a practical solver for the established Schrödinger–Bass bridge problem with jointly controlled drift and volatility. It alternates Gaussian-mixture Schrödinger-bridge fitting in transformed coordinates with inverse-map learning, enabling direct conditional endpoint sampling. Validation comprises a Gaussian problem with analytic plan/control references, three two-dimensional transport benchmarks, and adult-to-child face translation in a pretrained 512-dimensional ALAE latent space.

Critical strengths
  1. The computational construction makes joint drift/volatility modeling accessible through an efficient SB model. The transformed-coordinate formulation and Algorithms 1–2 combine a tractable bridge potential with a learned or approximate inverse, avoiding a separate diffusion-matrix network and SDE stepping for endpoint generation (§§3–4; Appendix B.1). This is a useful practical extension of LightSB-M if the learned maps satisfy the intended SBB structure.

  2. The Gaussian evaluation examines quantities that endpoint matching alone cannot distinguish. At β = 10, the methods share mean terminal SW₂ = 0.038, while LightSBB-M improves reported normalized squared drift error from 0.021 to 0.010 and volatility error from 0.021 to 0.0014 (Table 1; Appendix C.3). Comparing the joint plan and both controls against an analytic reference is directly relevant to the paper's central SBB claim, despite the residual errors discussed below.

  3. The reported empirical gains extend across distinct endpoint-generation settings. Table 2 has the lowest mean W₂ on all three synthetic tasks, with a reproducible approximately 18.8% average decrease relative to each task's strongest reported baseline; Table 3 increases outputs estimated at age 18 or younger to 76.0%, versus 19.3% for LightSB-M. These results support practical promise, while the image result trades lower latent similarity for stronger age transformation and the uncertainty of its small FID gain is unreported.

Critical weaknesses
  1. The Gaussian-mixture approximation is a major, untested limitation. The method uses a Gaussian-mixture potential to compute both the drift and the endpoint sampler (Section 4.1). It uses 50 components for 2D problems but only 10 for the 512-dimensional face latent space (Appendices C.5–C.6). There is no study of how accuracy or training cost changes with the number of components, so it is unclear whether this approximation can handle more complex distributions. Fast sampling is less compelling if the underlying bridge is not learned accurately.

  2. The experiments do not show why SBB is worth the extra complexity. Beyond the 2D Gaussian test in Section 5.1, the paper mainly evaluates the generated endpoint distribution (Table 2) or image quality and age translation (Table 3), rather than showing that the learned drift and volatility improve the transport itself. Training costs about 4–9 times as much as LightSB-M (Table 5), but there is no compute-matched comparison or ablation showing that the gains come from learning volatility rather than extra training. DSBM variants appear in the 2D table using previously reported numbers, but not in the image study. BMS and ASBS are not compared; Since they use different target information, they should at least be discussed rather than treated as directly interchangeable baselines.

  3. Inconsistencies in the mathematical statements Equation (15) is inconsistent with Equation (4) in a simple case: for $ T =\varepsilon=1$ and one Gaussian component with $ r_1 =0$ and Σ1=1\Sigma_1=1, Equation (4) gives a Brownian transition with zero drift, while Equation (15) gives $ s_ \theta(0,y)=-2y$ (Section 4.1). Proposition A.1 is also too broad as stated: ϕ(y)=−βy2\phi(y)=-\beta y^2 satisfies its stated growth condition but gives $ T_ \beta^+[\phi]= -\infty$. Moreover, the proof of the volatility bound mixes β\beta and β−δ\beta-\delta; under the original-β\beta interpretation, $ T =\varepsilon=1$, β=2\beta=2, δ=1/2\delta=1/2, and ϕ(y)=0.45y2\phi(y)=0.45y^2 give σ0=5.5>β/δ=4\sigma_0=5.5>\beta/\delta=4 (Appendix A.3). These issues need correction or more precise assumptions.

Questions for authors
  1. Gaussian-mixture approximation. How do accuracy, training time, and memory change as the number of Gaussian components increases, particularly in the 512-dimensional experiment? Can you compare with a more flexible potential model to show that the mixture approximation is not the main bottleneck?

  2. Need for SBB and experimental comparisons. What is gained by learning volatility, beyond better endpoint metrics? Please compare against LightSB-M with matched training compute, and report at least one non-Gaussian diagnostic of the learned transport or controls. Why are DSBM methods not included in the image experiment, and how do you position the method relative to BMS and ASBS given their different target-information settings? For the image results, please also report variability in FID and age-translation success.

  3. (Related to weakness 3) Are these calculations correct under your definitions? If not, please point out where my interpretation is wrong. Otherwise, could you correct Equation (15) and Proposition A.1, and clarify whether the implementation uses Equation (15) as written?

Reviewer_D9CGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents LightSBB-M, a novel idea for solving the Schrödinger–Bass Bridge problem, which involves optimizing both drift and volatility in stochastic transport. The method transforms the problem into a simpler Schrödinger Bridge problem with constant volatility. It combines Gaussian-mixture-based bridge matching with a neural network to learn the inverse transport map. This allows the model to generate samples without directly simulating stochastic differential equations.

The authors evaluate their method using a Gaussian problem with a known solution to check how accurately it recover the optimal transport plan, drift, and volatility. They have tested the method on three synthetic datasets and an adult-to-child face translation task. The results show that LightSBB-M performs better than some existing Schrödinger Bridge and flow-based methods.

Critical strengths
  1. Evaluation of the optimal transport plan and controls

One strength of this paper is that the authors evaluate not only the final results. They have also compared the learned transport plan, drift, and volatility with a known Gaussian solution. The results in Section 5.1 and Table 1 show that LightSBB-M can learn these properties accurately. This provides good support for the proposed methodology.

  1. Efficient method with good results

Another strength is that the proposed method avoids directly learning the full volatility matrix, making the approach easier to implement.

Critical weaknesses

I have identified some weaknesses of this paper.

The paper explains the mathematical background of the proposed method but it does not show that the training process is not converge always.For the main image translation experiment uses (\beta=1) and (T=1). But the theory requires (\beta T>1).The authors discuss this issue in the appendix, but the theoretical result does not fully support this setting.

This method has primarily been tested on three simple two-dimensional datasets, a Gaussian problem, and a task involving the transformation of an adult's face into that of a child. While the results obtained are promising, further research is required to demonstrate that this method also performs effectively across various high-dimensional tasks

Questions for authors

1.The main FFHQ experiment uses β T to one but Proposition A.1 requires β T greater than one. Why does the method still work when β T equal, to one? Provide results that show the learned transformation remains stable and invertible.

2.Figure 5 shows that the Wasserstein distance stabilises after three to five iterations. Does the stability of Wasserstein distance also mean that the SBB objective, transport plan and optimal controls have converged? Provide evidence that supports this conclusion.

3.The authors use an age estimation model and latent‑space similarity to evaluate the generated faces. Will the results remain similar if a different age estimation model or an independent face recognition model were used? Were the baseline methods compared using computational resources?

Comments on the paper

1 comment

We thank the reviewer for reading our rebuttal and for the follow-up. Please let us add the following clarifications.

Gaussian-mixture parameterization. LightSBB-M uses the same parameterization as LightSB (Korotin et al., ICLR 2024) and LightSB-M (Gushchin et al., ICML 2024). Korotin et al. (2024, Thm. 3.4) prove that Gaussian-mixture potentials are universal approximators of Schrödinger bridges, and that the training objective equals, up to a constant, the KL divergence to the true bridge. Since LightSBB-M solves a Schrödinger bridge in transformed coordinates, both properties carry over to that step.

We chose this form because the SBB relies on the derivatives of hh more than the SB does. Beyond the drift, the volatility σ=ε(I+εβ−1D2log⁡h)\sigma=\sqrt\varepsilon(I+\varepsilon\beta^{-1}D^2\log h) requires the full Hessian, and the Bass map XT=YT+εβ−1∇log⁡hT(YT)X_T=Y_T+\varepsilon\beta^{-1}\nabla\log h_T(Y_T) requires the gradient. With a mixture, all of them are exact and available in closed form, at a cost linear in dd and JJ thanks to diagonal covariances. A neural potential could replace the mixture, since the construction only uses hh, ∇log⁡h\nabla\log h and D2log⁡hD^2\log h, but each Hessian would then cost dd backward passes per sample. In other words, it is this mixture representation that makes the method light.

Evaluation setting. The 512-dimensional ALAE latent translation is the high-dimensional benchmark used by LightSB and LightSB-M. We evaluate on the same task, with additional baselines from four method families (Appendix C).

We hope these points clarify that the mixture is a deliberate solver choice, while the contribution of the paper is the Schrödinger–Bass bridge, which learns the volatility jointly with the drift.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.