acceptodds
Under review as a conference paper at ICLR 2027

What Do Flow-Based Inverse Solvers Approximate? A Posterior-Transport View

Abstract

A growing family of training-free solvers-FlowDPS, FLOWER, PnP-Flow and their diffusion ancestors (DPS, DAPS)-repurpose a pretrained flow-matching prior to solve imaging inverse problems by adding a measurement-guidance term to the deterministic probability-flow ODE. Despite strong empirical results, what these per-step corrections actually approximate—and how far the resulting samples are from the true posterior —has not been characterized. We give a posterior-transport account of flow-based inverse problem solving. Our starting point is a simple but consequential fact: for a deterministic flow prior, Bayesian conditioning is realized entirely by a reweighting of the source distribution, not by a drift correction; pushing the reweighted source through the unmodified velocity field yields exact posterior samples. From this we define a canonical minimum-energy corrector that exactly transports the unconditional source to the posterior. Existing guidance methods can then be placed within the same correction framework as local surrogate correctors-differing in their endpoint estimate and mobility—and their discrepancy from the canonical corrector controls the terminal posterior bias in Wasserstein distance. A controlled D study with a closed-form posterior confirms the theory decisively: source reweighting matches the two-sample Monte-Carlo reference on every metric, whereas trajectory guidance incurs substantially larger error-– that reference on sliced- and – on energy distance and MMD—and collapses posterior modes across all tested guidance strengths. Guided by the analysis we propose a cheap, principled velocity-correction solver that is competitive across two in-domain priors (AFHQ, CelebA) and two out-of-distribution settings while, unlike point-estimate source-space optimizers, producing diverse conditional samples whose sample-based uncertainty correlates with reconstruction error (moderately so once the inpainting mask is controlled for).

Primary area:
generative models
Keywords:
flow matching, posterior sampling, inverse problems, optimal transport, uncertainty quantification, image restoration

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_m1hgOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Summary

  • This paper shows that deterministic flow based inverse solvers can be made possible entirely by reweighting the starting point without changing drifting term.
  • This paper provide a posterior W2 distance bound by velocity field approximation error.
Critical strengths

Strength

  • The claim that reweighting the source can achieve posterior sampling is intuitative. Every posterior sample correspond to one source noise in deterministic flow. And it is indeed possible to achieve exact posterior sampling by resampling the source noise.
  • Theorem 1 provide interesting relationship between the posterior W2 error and velocity field approximation error.
  • The toy example looks very nice and clearly show the advantage of the proposed approach.
Critical weaknesses

Weakness

  • The main empirical validation on image restoration is not quite sufficient. The authors adopt simple AFHQ-Cat dataset and unconditional diffusion model, with PSNR + LPIPS metric, 3 operators and a handful baseline. This is clearly simpler than the FLOWER baseline (CelebA + AFHQ cat, 5 operators, 3 metrics and 7 baselines). Visual comparsion of the restored images between baseline and proposed approach is encouraged.
  • The empirical advantage of the proposed approach over FLOWER is unclear. From Table 2, it is clear that the proposed approach achieves a perception-distortion trade-off (See [CVPR 18 The Perception-Distortion Tradeoff]) in favor of FLOWER. However, this different trade-off does not sustain an advantage.
  • It might be a little bit misleading to say "Bayesian conditioning is realized entirely by a reweighting of the source distribution, not by a drift correction". Drifting correction like DPS can also achieve exact bayesian sampling given correct drifting term, at least for Gaussian source Flow matching optimal transport (FM-OT, see [ICLR 23 Flow Matching for Generative Modeling]). Reweighing the source noise is one way to achieve posterior sampling but dirfting is also a valid way (See Eq 11 of [ICLR 22 diffusion posterior sampling for general noisy inverse problems]).
  • Some previous works think DPS is maximizing posterior, and thus lead to smaller sample diversity. It might be benefical to discuss this point of view [ICLR 25 Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior] [ICML 26 Local MAP Sampling for Diffusion Models] [ICLR 24 From posterior sampling to meaningful diversity in image restoration]
  • AI USE STATEMENT should be in main text not Appendix?
Questions for authors

Questions

  • See weakness.

Justification

  • For now I recommend to reject this paper. I do like the theoretical insights but the current empirical validation does not support the proposed approach derived from analysis.
Reviewer_D8ryOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the use of pretrained flow matching models for posterior sampling in the context of solving inverse problems. They first theoretically relate posterior sampling to the reweighting of the initial distribution (which is a direct consequence of the deterministic nature of flow matching models); although this result is interesting, they do not propose any mechanism for such measurement-aware sampling of the source distribution. Instead, they come back to the velocity correction perspective and define an optimal condition for the correction velocity and interpret what current solvers approximate in relation to their optimal velocity correction. Eventually, they propose a new correction method whose main advantage seems to be its efficiency.

Critical strengths

The topic is certainly interesting, and the mathematical results, although not surprising, are interesting as well!

Critical weaknesses

For me, the whole message of the paper is a bit self-contradictory. The authors first emphasize source reweighting but then also propose a velocity correction pipeline, which does not outperform current baselines in terms of reconstruction quality.
The novelty of the theoretical results and the proposed method is limited in my view.

Questions for authors
  1. I would be curious to see a similar compute budget (in terms of seconds) for each compared method, and what the ranking of the methods will be. i.e., reducing the number of steps of other methods to reach the evaluation time of the proposed method.
Reviewer_T7AnOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a posterior-transport view for flow-based inverse problem solvers. They first theoretically characterize exact posterior transport from a Bayesian point of view – for a deterministic flow-based prior, instead of injecting measurement guidance interleaved with the reconstruction trajectory, one can incorporate the conditioning information directly into the seed distribution through posterior reweighting, while leaving the original flow dynamics unchanged. Later, they also used a controlled 2D experiment to validate the effect of source reweighting. Building on this perspective, the authors reinterpret existing interleaving methods as surrogate mechanisms of performing the posterior corrections through the gradient updates. Motivated by this, they proposed their interleaving method with normalized measurement-gradient correction along the trajectory and compared that with other methods based on AFHQ and CelebA tasks. Overall, this is a well-motivated work with a solid theoretical foundation and a thoughtful discussion of posterior conditioning in flow-based inverse problems.

Critical strengths

This is a strong and conceptually interesting work with a solid mathematical foundation. The authors first introduce a posterior-transport perspective for solving inverse problems and derive a posterior-reweighting formulation with rigorous theoretical analysis. They further validate this formulation using a controlled 2D toy problem where the posterior distribution can be directly evaluated. Building on this perspective, the paper interprets existing interleaving-based methods as approximate mechanisms for performing posterior correction through step-wise gradient updates. This provides a new and theoretically grounded way to understand and unify an important family of flow-based inverse problem solvers.

Critical weaknesses

There is a gap between the theoretical analysis and the practical image-restoration method. For real-world image restoration, directly performing posterior correction in the seed space is difficult due to the high-dimensional sampling problem. Therefore, the authors instead design a practical surrogate that performs step-wise measurement-gradient corrections along the flow trajectory. This formulation appears closely related to FlowChef (https://arxiv.org/pdf/2412.00100), which similarly estimates the endpoint and applies a gradient-based correction at each flow step, with the main additional ingredients here being the per-image gradient normalization and mobility scaling. The author should include this method as a baseline in the paper and discuss the distinction and novelty relative to FlowChef more clearly. The sampling strategy used for image-level uncertainty is also questionable. In Table 4, the authors generate only 8 reconstructions for each observation and use their mean and pixelwise standard deviation as the reconstruction and uncertainty estimate. First, 8 samples may be insufficient to reliably characterize the underlying posterior distribution. Second, for ill-posed inverse problems, there may exist multiple well-separated regions of the solution space that are all consistent with the same observation and carry substantial posterior probability. In such cases, a pixelwise variance can reflect overall dispersion but does not capture the underlying multimodal structure or the relative likelihood of different plausible solutions.

Questions for authors

Aside from the interleaving family, there is also a line of plug-in based methods https://openreview.net/forum?id=ZnmofqLWMQ (Diffusion Inversion) and https://arxiv.org/abs/2405.16749 (DMPlug). They don’t involve any intermediate gradient update step, instead they use the stacked diffusion model as a fixed function and optimize on the seed space. These methods therefore seem conceptually closer to your source-space characterization than methods that modify the diffusion trajectory at every sampling step. Could the authors clarify how such seed-space optimization methods relate to the proposed source-reweighting view? In particular, although optimizing a seed does not in itself constitute posterior sampling, does operating directly in the source space avoid some of the posterior distortion attributed to trajectory-level guidance? More broadly, could source-space optimization or sampling methods provide a more scalable approximation to the theoretically correct source-reweighted posterior than the proposed trajectory correction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.