acceptodds
Under review as a conference paper at ICLR 2027

Posterior Transport Matching for Inverse Problems

Abstract

Posterior sampling must preserve uncertainty while remaining consistent with the observation. We introduce Posterior Transport Matching (PTM), a training-free framework that pairs numerical transport maps with stochastic input updates, termed rematching. The key is to supply each map with an input distribution that it sends to the posterior. We characterize the minimum-energy isotropic diffusion realizing these compatible laws and identify its objective with path-space relative entropy. Exact probability-flow maps admit Gaussian rematching. For computation, entropy projection yields an objective for the Gaussian transition mean, combining an expected potential with a squared-distance penalty. Denoised observation energy approximates the potential; an analytic approximation requires only the observation operator. A single bound on output KL separates potential approximation from finite computation. Controlled experiments show how the appropriate transition depends on the map. Across eight FFHQ inverse problems, PTM achieves the highest reported PSNR across eight FFHQ inverse problems and favorable quality–cost tradeoffs, with further evaluation on ImageNet.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Inverse problems, posterior sampling, diffusion models, transport maps, training-free inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sGh8OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper considers a problem of sampling from a tilted distribution π0(x)∝p0(x)eE(x)\pi_0(x) \propto p_0(x)e^{E(x)}, where $ p_0$ can be sampled using a diffusion model.

The authors propose an approximate sampling scheme to sample from π0\pi_0 by alternating between two steps

Starting from x0∼N(0,τ0I)x_0\sim N(0,\tau_0 I) the algorithm alternates between two steps

  1. Given current xτkx_{\tau_k} map it to clean estimate zkz_{k} by applying map zk=Ψτk(xτk)z_k = \Psi_{\tau_k}(x_{\tau_k}), where Ψτk\Psi_{\tau_k} is the approximation of the flow map sending πτk\pi_{\tau_k} to π0\pi_0. Denote by ρk\rho_k the distribution of zkz_k.

  2. Noise back using kernel K to get xk+1=K(zk,⋅)x_{k+1} = K(z_k, \cdot) that approximates

K⋆=argminKEρkKL(K(z,⋅)∣∣N(z,τkI))K^\star = argmin_K \mathbb{E}_{\rho_k} KL(K(z,\cdot)||N(z, \tau_k I))

subject to Ψτk+1(xk+1)=π0\Psi_{\tau_{k+1}}(x_{k+1}) = \pi_0

In practice they use

  1. take as Ψτk(xτk)\Psi_{\tau_k}(x_{\tau_k}) the posterior mean approximation mτkm_{\tau_k}.
  2. Take as K=N(uk,τk+1I)K=N(u_k, \tau_{k+1} I), where uk=argminE(u)+∣∣u−zk∣∣2/2τk+1u_k = argmin E(u) + ||u-z_k||^2/2\tau_{k+1}

They show that the algorithm performs strongly on FFHQ and ImageNet.

Critical strengths

Algorithm presents stong empirical performance

Critical weaknesses

The paper is very hard to follow, and I am still not sure if I understood it correctly.

  1. If I understand correctly, the algorithm closely resembles the TMPD algorithm the authors cite; the posterior mean approximation formula is inherited from it, and the only difference is the further approximation of the kernel KK. However, I do not see the comparison between them in Tables 2 and 3.

  2. The choice of the kernel KK does look arbitrary; the approximation U∗U^* by Ut=E(Dt(x))U_t = E(D_t(x)) while standard breaks the premise that the kernel satisfies the condition on the algorithm.

Questions for authors

What is the purpose of the rematching step? When we have a posterior mean, with Tweedie we have access to the posterior score and can simulate the backward process (I believe this is what TMPD does). The ablation to see whether this helps is required.

Reviewer_B99COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Posterior Transport Matching, a training-free framework for solving inverse problems with pretrained diffusion denoisers. Its central idea is to jointly consider a numerical endpoint map and the input distribution required for that map to produce posterior samples. The authors characterize compatible input-law evolution through minimum-energy isotropic diffusion and relate the variational objective to path-space relative entropy.

For computation, PTM constructs local endpoint proposals using denoiser Jacobian products and observation-energy acceptance tests. It then optimizes the mean of a Gaussian rematching transition using either denoised observation energy or an observation-only approximation. Validation includes controlled Gaussian and mixture distributions and eight inverse problems on FFHQ and ImageNet. PTM reports the best PSNR across all eight FFHQ tasks and lower effective evaluation counts than DAPS

Critical strengths

A useful formulation of compatibility between numerical maps and stochastic transitions. The paper clearly illustrates why a conditional mean is not generally a transport map and explains how changing the input law can compensate for map distortion. The minimum-energy and entropy-projection formulations give this perspective a substantive mathematical foundation.

Strong reported reconstruction performance with detailed computational accounting. Table 2 shows substantial FFHQ improvements on several tasks; for example, motion-blur PSNR increases from 31.80 dB for DAPS to 33.78 dB for PTM. The evaluation counts include Jacobian products, inner solves, and particles, making the efficiency comparison more informative than forward-call counts alone. Direct FFHQ reruns of DAPS and SITCOM strengthen the empirical evidence.

Critical weaknesses

The auther is still comparing the DAPS which is silghtly old FAST-DIPS can be a good case for comparison.

Also,The practical approximation is insufficiently connected to the posterior-preservation theory. The ideal matching potential in Equation (13) depends on both the incoming distribution and the next numerical map. The implementation replaces it with E∘DtE\circ D_t, or further approximates it by the observation-only objective. Equation (16) retains the resulting potential error through βz\beta_z, but the paper does not establish that this error is small for the implemented image sampler or verify the required curvature condition. The controlled experiments themselves show that correction can worsen accuracy for correctly conditioned maps or at larger variances. Moreover, image benchmarks measure returned reconstructions, sometimes after particle selection and averaging, rather than posterior calibration. The manuscript acknowledges this distinction, but it remains a material limitation of the central practical claim. Cohort-level ablations isolating rematching from the endpoint proposal and output-selection rules would substantially strengthen the evidence.

The evaluation protocol limits confidence in generalization and some comparisons. Appendix F states that FFHQ solver configurations were screened on the first 16 images, which remain included in the reported 100-image cohort. This overlap creates potential selection bias, especially for smaller improvements. In addition, several comparisons use published aggregates rather than paired reruns; FFHQ SITCOM uses independent observation realizations, and ImageNet phase retrieval uses reference-based rotation alignment whose equivalence to baseline evaluation is unclear. These qualifications are disclosed, but a held-out evaluation with matched protocols is needed to establish the robustness of the reported advantages.

Questions for authors

ractical posterior accuracy: Can you provide evidence that E∘DtE\circ D_t approximates the map- and incoming-law-dependent matching potential for the actual numerical maps? In a tractable setting using the implemented proposal and acceptance rule, how do posterior moments, mode weights, and distributional error compare across Gaussian, analytic, and denoiser-based rematching?

Attribution of reconstruction gains: With the endpoint map, schedule, particle count, and output rule fixed, what are the cohort-level results for Gaussian rematching versus analytic and sampled rematching? How much of the improvement remains without particle averaging or selection?

Reviewer_C3tbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a method for solving inverse problems using pretrained diffusion models without retraining. The method combines a numerical transport map with a stochastic update called rematching. The transport map turns noisy inputs into clean estimates, while rematching prepares the inputs for the next noise level. Instead of computing the exact transport map, the authors use a cheaper approximation. Rematching is designed so that this approximate map can still produce samples from the target posterior distribution. The authors develop a mathematical framework to describe how this works.

For practical use, they simplify rematching to a Gaussian step and optimize its center. They also propose cheaper ways to compute the updates using a denoiser or the observation model. The numerical map uses a conditional-mean estimate and an energy-based check to improve consistency with the observations. The authors test PTM on simple two-dimensional distributions with known posteriors and eight image reconstruction tasks using FFHQ and ImageNet. PTM achieves the highest PSNR on all eight FFHQ tasks.

Critical strengths
  • The paper introduces an interesting framework that connects numerical transport maps with stochastic rematching. A key idea is to treat the input distribution as part of the sampler to preserve the target posterior.
  • The theory is carefully developed. The authors provide results on minimum-energy rematching, establish a connection to path-space KL divergence, and derive an error bound for the practical approximation.
  • PTM achieves the best performance on several tasks. It also uses fewer denoiser and observation evaluations than DAPS across all eight FFHQ tasks.
  • The method handles both linear and nonlinear inverse problems without retraining. It uses automatic differentiation of the observation model rather than requiring problem-specific formulas.
  • The paper provides detailed experimental settings and computational costs, which helps with reproducibility and makes comparisons easier to assess.
Critical weaknesses
  • Gap between theory and experiments. The controlled experiments use exact probability-flow map constructions, while the image experiments use approximate denoiser-based maps. The error bound in Equation (16) is not evaluated, and its usefulness at image scale remains unclear.

  • Limited ablation studies. The gains are not clearly separated from the conditional-mean step, energy-based acceptance test, or particle selection and averaging. HDR uses Gaussian rematching without the optimized center correction. Figure 5 also shows only a small difference between the two variants on two selected examples. Comparisons with closely related methods such as MMPS and TMPD would help isolate the contribution of rematching.

  • Limited posterior validation. The image experiments focus on reconstruction metrics rather than posterior accuracy. Higher PSNR may result from averaging or overly concentrated outputs rather than better posterior sampling. This makes it difficult to connect the reported gains to the paper's focus on posterior preservation.

  • Mixed ImageNet results. PTM leads PSNR on five of eight tasks, while SITCOM achieves higher SSIM on five. DAPS also performs better on FID for several tasks.

  • Experimental limitations. Some FFHQ settings were screened on the evaluation cohort, and configurations vary considerably across tasks. Uncertainty estimates are not reported for the main image comparisons. FID estimates based on only 100 images may be unreliable.

Questions for authors
  1. Theory and implementation: Could the authors test the complete PTM algorithm on a simple distribution with a known posterior? This would help show whether rematching works with the approximate numerical map used in the image experiments.

  2. Ablation studies: Could the authors separate the contributions of the conditional-mean step, energy-based acceptance test, and rematching? How much improvement comes from optimized rematching compared with plain Gaussian rematching?

  3. Theoretical guarantees: Can the authors evaluate the error bound in Equation (16), even on simple examples? How informative is the bound at image dimensions, and how do its assumptions apply to the practical algorithm?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.