acceptodds
Under review as a conference paper at ICLR 2027

FlowSGS: Improving Flow Priors for Inverse Imaging with Stochastic Interpolants

Abstract

Flow matching has emerged as the state-of-the-art generative model and has been used for plug-and-play (PnP) priors to solve inverse problems in computational imaging. However, existing flow-based inverse solvers make simplifying approximations in posterior sampling and/or assume linear forward models. In contrast, principled diffusion-based methods exist but are computationally expensive. To circumvent these problems, we introduce FlowSGS, a flow-based posterior sampling method using Split Gibbs Sampling (SGS) with annealed coupling, and make flow-specific interventions to the prior sampling and annealing steps, theoretically grounded in the Stochastic Interpolants (SI) framework. Using a novel timestep correction and accelerated warm-up with one-step flow models, our method is faster and more accurate than the state of the art across a range of imaging problems. We demonstrate the first use of flow-based samplers for a nonlinear inverse problem (Fourier phase retrieval) and apply one-step flow models to principled Bayesian inverse imaging.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Inverse problems, posterior sampling, flow matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_agjNOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose FlowSGS, a split Gibbs sampling (SGS) method for posterior sampling in imaging inverse problems with a flow-model prior. FlowSGS generalizes the PnP-DM method (developed for diffusion-model priors) to all stochastic interpolant (SI) models, including flow matching, which improves efficiency. The authors provide theoretical results about how the prior step in SGS is equivalent to solving the SI reverse-time SDE, how PnP-DM is a special case of FlowSGS, and convergence guarantees, although these theoretical results are largely direct extensions of results already established by PnP-DM. The authors validate on several linear inverse problems and one nonlinear inverse problem, showing improved image quality and computational efficiency over existing state-of-the-art SI-based methods for posterior sampling.

Critical strengths

The method itself is extremely sound and could be already practically useful. It is clear that an SGS method that supports any SI prior ranging from a flow-matching prior to a diffusion-model prior would be useful for a variety of inverse problems, and this, to the best of my knowledge, is the first paper to layout a framework for doing so and provide sound validation.

Critical weaknesses

There’s nothing surprising or novel about this paper. The method and theoretical results are obvious (at least in hindsight) extensions of PnP-DM.

Questions for authors

For the result about warmup invariance (Corollary 3), it is confusing what the tradeoff is between efficiency and image quality when using one-step flow models for acceleration. The manuscript states: “We show that this does not affect convergence: any nonpathlogical prior step sampler may be used during warm-up, including one-step flow-map models, and the convergence rate and approximation floor of the subsequent stationary stage are unaffected.” But there must be some tradeoff. Perhaps the answer is in the bound in Equation 9, but it would be nice to include some concrete statements about the tradeoff to aid users in deciding how many warmup prior steps to take.

This is more of a clarification question, but in Figure 4, it is confusing why Mode 2 would even show up in the posterior since it so clearly doesn’t make sense under the prior. Can you provide some intuitive explanation for why this mode appears?

3 comments

Thank you for the positive assessment and for describing the method as extremely sound and practically useful.

W: Extension of PnP-DM. We agree the SGS template comes from PnP-DM (Corollary 1), but the extension is not automatic. Please see G1 in the global response.

Q1: Warm-up tradeoff (Corollary 3). You are right that there is one, and it sits in the initialization term of Equation 9, C⋅KL(μw∥πρmin⁡)/(Ksmin⁡(ρmin⁡,δ)2)C\cdot\mathrm{KL}(\mu_w\Vert\pi_{\rho_{\min}})/(K_s\min(\rho_{\min},\delta)^2): a cheaper warm-up leaves a larger KL and needs more stationary iterations KsK_s for the same bound, while the rate and floor are unchanged. With defaults, MFM warm-up costs 0.38 dB at fixed KsK_s for a 9.6× speedup (Tables 6 and 10). New Figure 15 varies the number of warm-up iterations that use MFM on motion deblurring: as more iterations use MFM, the total NFE drops from 778 to 100, while LPIPS (0.164 to 0.151) stays within the error bars and PSNR within 0.15 dB. This is in accordance with Corollary 3, which states that the warm-up does not affect the final convergence. As already discussed after Corollary 3, the warm-up affects only the bound, leaving the convergence rate and error floor unaffected, and we now also point to Figure 15 there.

As noted in Section 4.3, these guarantees do not apply to nonlinear problems, which explains the MFM failure in FPR, where the non-convex likelihood pushes warm-up iterates off the data manifold and the one-step map fails.

Q2: Why Mode 2 in Figure 4? We agree that rotated images are not seen when we train the prior, but we would like to point out that the rotated mode exists for the following reason, in addition to the symmetry of the Fourier phase retrieval forward model. Rotated faces follow the same natural image statistics as upright faces, and have only a slightly lower prior log-likelihood. To verify this, we measure the prior log-likelihood of face images and their rotated counterparts in Figure 11 (3.3953.395 vs. 3.3843.384 nats per dimension, compared to −0.871-0.871 for pure noise). Although not seen during training, the rotated face gets a much larger prior log-likelihood than pure noise, which is not a natural image.

Thank you again for the supportive review.

Thank you to the authors for responding to the reviewer comments and concerns. They have adequately addressed my questions, including clarifying the differences from PnP-DM and the warmup tradeoff. I'll be happy to raise my score since I believe this is a practically useful paper.

Thank you for your positive feedback and helpful suggestions. The discussions of the warm-up tradeoff and the second mode in Fourier phase retrieval that you suggested add important context, and have significantly strengthened our manuscript.

Reviewer_M6RDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FlowSGS, a Bayesian posterior sampling framework for inverse imaging problems based on a pretrained flow-matching prior. It uses the Split Gibbs Sampling inspired by PnP-DM: the sampler alternates a likelihood step with a prior step and anneals the coupling parameter ρ. Using stochastic interpolants, the authors show that the prior step reduces to solving a velocity-based reverse SDE, and they extend PnP-DM’s convergence guarantee to flow-matching priors. Experiments on deblurring, super-resolution, CS-MRI, and Fourier phase retrieval report competitive or improved quality over diffusion- and flow-based baselines with fewer network evaluations.

Critical strengths
  1. The paper presents a novel time-step correction that significantly improves the quality of the reconstruction.

  2. Hybrid one-step acceleration using a Meta Flow Map for the expensive warm-up prior steps makes the sampling process very fast.

Critical weaknesses
  1. Insufficient motivation for posterior sampling in imaging inverse problems. The primary objective of this paper is to enable accurate posterior sampling for imaging inverse problems using a pretrained flow-matching prior. However, the experiments do not convincingly demonstrate why posterior sampling is necessary or advantageous for solving practical imaging problems. While the proposed method achieves improved PSNR for some reconstruction tasks and more accurately approximates the posterior in a toy Gaussian setting, these results do not establish the practical value of posterior sampling over conventional point-estimation methods. In particular, the paper does not demonstrate whether the proposed approach provides meaningful benefits in terms of uncertainty quantification, ambiguity resolution, or robustness to noise and model mismatch. The authors should provide experiments that explicitly demonstrate the practical benefits of posterior sampling and uncertainty quantification for imaging inverse problems, beyond improvements in reconstruction metrics alone.

  2. One of the paper's main claims is that the proposed method enables posterior sampling for nonlinear inverse problems, including Fourier phase retrieval. However, the evaluation relies on the FFHQ human-face dataset, which is not representative of typical scientific applications of phase retrieval, such as diffractive imaging. Although I understand that previous works have adopted similar benchmarks to leverage readily available pretrained generative priors, this experimental setup provides limited evidence of the method's applicability to realistic scientific imaging scenarios. I encourage the authors to evaluate their approach on a more representative dataset and discuss the challenges associated with applying pretrained generative priors in such settings. In particular, a significant concern is the potential for prior-induced hallucinations under domain shift: when the test distribution differs from the training distribution, the generative prior may produce visually plausible but physically incorrect structures that are unsupported by the measurements.

Questions for authors

One question on the toy problem where the true posterior is available: I am curious to understand how the KL divergence of FlowSGS varies with the number of SDE discretization steps, the number of SGS iterations, and the final coupling ρ_min.

3 comments

Thank you for the thoughtful review and for highlighting the timestep correction and the speed of the hybrid acceleration.

W1: Why posterior sampling?

We certainly appreciate the desire for real-world wins through posterior sampling for inverse imaging. Respectfully, we see this work as the first step toward such applications, which we wholeheartedly intend to pursue as future work. As these advances will almost certainly require significant input from domain experts to even target the right kind of problem setup and posterior-related uncertainty, we believe that sharing a general method may still be of use to this rapidly developing problem space before we are able to devote considerable time to careful application.

We agree that additional uncertainty metrics should be included; we have added NLL and ECE with 64 posterior samples for linear inverse problems in Tables 8 and 9, as well as for Fourier phase retrieval in Table 7. Diffusion-based methods still outperform flow-based methods on these metrics, but FlowSGS is competitive among flow-based methods. Note that NLL and ECE assume a Gaussian distribution, which our posteriors are not, so these numbers cannot be considered fully representative. We refer to the toy problem with a ground truth posterior (Figures 2 and 13) for a better benchmark of posterior accuracy.

W2: Experiment setups. We appreciate this point and agree that our datasets and forward models are not realistic scientific settings. A major reason is that scientific imaging tasks usually have complicated data processing and noise statistics that vary across tasks, which makes it hard to separate the effects of the data from those of the method. This work focuses on the method itself, so we chose standard datasets and simplified forward models, following the majority of previous work. Our MRI experiments are conducted with real MRI images from the fastMRI dataset, which consists of real measured data.

We agree that hallucination under domain shift is an important problem for inverse imaging with generative priors. We have added this as a direction for future work in Section 5 (see blue text).

Q: KL vs. SDE steps, SGS iterations, ρmin⁡\rho_{\min}. Thank you for this question. We tested our method and PnP-DM with different SDE discretization steps and provide results in Figure 13. We observe that with N≥4N\geq4, our method maintains a low KL divergence while PnP-DM's KL divergence increases when N≤32N\leq32. We believe this is direct evidence that our method is able to produce a more accurate posterior than PnP-DM under similar resource budgets. Both methods converge within K=100K=100 SGS iterations (Figure 13b) and are stable for ρmin⁡≤0.02\rho_{\min}\leq0.02 (Figure 13c), where FlowSGS is closer to the analytic SGS floor than PnP-DM.

We hope these address your questions. Please let us know if we misunderstood any of your points.

Thank you very much for addressing my concerns. While some of them remain unaddressed, I am nevertheless happy to raise my score, as I believe this is a good paper.

Thank you for your positive feedback and your thoughtful suggestions for our work! Your suggestions on uncertainty metrics and the toy-problem KL study have really strengthened our manuscript. We believe the KL study on the toy problem directly connects our theory to the empirical behavior of FlowSGS across its hyperparameters.

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

This work extends the PnP-DM framework for solving inverse problems to be more general, supporting its use with flow models (instead of / as well as diffusion models). The proposed approach makes use of split Gibbs sampling (SGS) to split the prior steps from the likelihood steps. The prior steps incorporate the stochastic interpolants framework, while the likelihood can be a proximal step or use Langevin dynamics. The authors also introduce a timestep correction (to correct for the mismatch between the theoretically required starting time and the solver's discrete time grid) and one-step-flow acceleration during warm-up. The method is evaluated on a Gaussian posterior, several linear imaging problems as well as nonlinear Fourier phase retrieval. The paper reports strong reconstruction performance and computational savings from acceleration.

Critical strengths
  1. The empirical performance of the method is a strength. FlowSGS achieves good reconstruction results (either SOTA or close in each experiment) on linear and nonlinear inverse problems. The evaluation is broad and the Gaussian posterior experiment provides evidence that the method is useful in practice.

  2. Another strength is that the paper contains a useful generalisation of PnP-DM. The stochastic interpolants framework provides a principled extension of PnP-DM to flow-based priors. Although the underlying sampling approach is largely unchanged, this generalisation is useful and theoretically well-motivated.

Critical weaknesses
  1. Novelty qualification: The paper makes several claims that appear insufficiently qualified, particularly regarding novelty, the absence of approximations, and posterior sampling for nonlinear inverse problems. These claims should be more carefully situated within the existing literature and the limitations of the proposed approach.

  2. On a related note, the method is largely an extension of PnP-DM to flow-based priors, retaining the same underlying split Gibbs sampling framework. The additional contributions, particularly timestep correction, appear to be relatively straightforward numerical modifications. It is unclear whether these constitute a sufficiently substantial methodological advance.

Questions for authors
  1. [Clarification] Please reconsider the claim that this is the first flow-based posterior sampler demonstrated on nonlinear inverse problems, particularly in light of Parikh et al. (2026), D-Flow SGLD. Earlier works have also addressed nonlinear inverse problems using flow priors, including Ben-Hamu et al. (2024), D-Flow, and Webber et al. (2026), MPC-Flow (although these are not posterior samplers). The final sentence of the abstract is also potentially misleading, as it could be interpreted as claiming both that the entire inverse problem is solved with a single NFE and that this is the first application of flow models to nonlinear inverse problems. Please clarify the scope of these claims.

  2. [Clarification] The introduction states that the method avoids approximations used by existing flow-based samplers. However, the implementation involves finite-step SDE integration, approximate Langevin sampling, and a nonzero coupling parameter ρmin⁡\rho_{\min}, meaning the stationary SGS distribution is not exactly p(x∣y)p(x\mid y). Furthermore, Theorem 1 establishes exact SGS conditional sampling for the KL-optimal diffusion coefficient, whereas the experiments use a regularised coefficient. Does this modification preserve the required conditional distribution, or does it introduce an additional approximation? Also, could the authors clarify which approximations are avoided and how the remaining approximations affect posterior accuracy?

  3. [Future work] Figure 5(c) partially separates the improvements due to flow matching and timestep correction. Could the authors also evaluate PnP-DM with an equivalent timestep correction, to better isolate the benefits of flow-based priors from the benefits of the numerical modifications?

  4. [Clarification/Future work] A key claim is that this method is a good posterior sampler (as opposed to just solving an inverse problem effectively). For nonlinear phase retrieval, the reconstruction quality is reported using the best of eight samples selected against ground truth. Could the authors additionally report performance without oracle-based selection, and provide quantitative evidence of posterior accuracy beyond identifying distinct modes?

5 comments

Thank you for the detailed review and for recognizing the empirical strength and the generalization of PnP-DM.

W1 / Q1: Novelty claims and nonlinear problems. We thank the reviewer for these references and will clarify the scope of our claims by including them, with the following context. D-Flow SGLD (Parikh et al., 2026) is a posterior sampler, but its only nonlinear task is a 2D toy problem with a scalar measurement, which could not feasibly scale to an image; its imaging tasks are linear. D-Flow (Ben-Hamu et al., 2024) and MPC-Flow (Webber et al., 2026) are single-reconstruction methods rather than posterior samplers. D-Flow is tested with a nonlinear latent autoencoder rather than an actual nonlinear inverse problem, and MPC-Flow's only nonlinear task is a learned-kernel deblurring. In light of D-Flow SGLD, we withdraw the claim of "first flow-based posterior sampler on a nonlinear problem". We now discuss D-Flow SGLD, D-Flow and MPC-Flow in Section 1 and have revised the last sentence in the abstract (see updated text in blue). We could not find available code for D-Flow SGLD and D-Flow.

We particularly appreciate the pointer to nonlinear deblurring. We are currently working to add it as a test and to add MPC-Flow, which has official code, to our comparisons, and hope to provide results during the rebuttal period.

Q2: Which approximations are avoided? We agree the wording was too broad and will clarify it. Like every MCMC sampler, FlowSGS has numerical approximations: finite SDE and Langevin steps, and a nonzero ρmin⁡\rho_{\min}. What it avoids are likelihood approximations: Tweedie-based guidance at intermediate xt\mathbf{x}_t (DPS, FlowDPS), measurement-space projection (PnP-Flow) and trajectory steering (FlowChef), which Figure 2 shows visibly distort the posterior. For a linear-Gaussian likelihood the SGS target at coupling ρ\rho is the exact posterior with noise covariance σ2I+ρ2AA⊤\sigma^2\mathbf{I}+\rho^2\mathbf{A}\mathbf{A}^\top, so the ρmin⁡\rho_{\min} bias has a closed form and vanishes as ρ→0\rho\to0. For the SDE discretization, we show its impact on the KL divergence in the toy problem in new Figure 13a. We observe that our method maintains a low KL divergence for N≥4N\geq4 SDE steps. On the regularized coefficient: you are right that Theorem 1 needs the exact wtKLw^{\mathrm{KL}}_t, and any other wtw_t preserves the marginals but not the conditional, so it is an additional approximation. It is small because the irregularity is at t→1t\to1 while the stationary prior step covers only t≲0.05t\lesssim0.05, and the unregularized coefficient gives a similar result (29.21 vs. 29.23 dB, Table 11). We also report the KL divergence for every diffusion coefficient (including the unregularized wtKLw^{\mathrm{KL}}_t) in new Figure 14. With the regularization, the KL divergence of the KL-optimal coefficient is in fact better (2.81×1042.81\times10^4 vs. 2.90×1042.90\times10^4).

W2 / Q3: Timestep correction. We hugely appreciate the suggestion to also test PnP-DM with our timestep correction (see updated Figure 5c). We show that PnP-DM receives a major boost in accuracy when combined with our simple yet effective timestep correction, particularly at low discretization steps (LPIPS from 0.895 to 0.224 at 8 steps). In fact, PnP-DM benefits more from the timestep correction than our method does. This further shows that our novel improvement is very effective for SGS samplers with low step counts in their prior-step SDEs.

Q4: Oracle selection; posterior accuracy. We followed the best-of-8 protocol of DPS and PnP-DM, but agree it conflates sampling with selection. Our new Table 7 reports metrics over the mean image of all 8 posterior samples, as for all other tasks. For posterior accuracy beyond mode identification, we refer to the toy-problem KL study in Figures 2 and 13, and have added uncertainty metrics (NLL, ECE) in Tables 7, 8 and 9.

We hope these changes address your concerns and would be grateful if you would reconsider your score.

Thank you for your clear response to my questions. I appreciate the detailed and concise reply.

I am convinced that your revised manuscript now adequately qualifies the contribution within the literature (W1). I am also more convinced of the value of the timestep correction contribution, in light of the ablation with PnP-DM that you have carried out (W2).

As your rebuttal has clarified my questions and addressed the weaknesses, I'd be happy to raise my score.

AuthorsOpenReview ↗

Follow-up to Q1: MPC-Flow comparison

Thank you for your positive feedback!

We have added MPC-Flow (Webber et al., 2026) using its official code as a baseline on all tasks. The results are in Tables 1 and 5–9 of the revised draft (labeled in blue), and the MPC-Flow and FlowSGS rows of Table 1 are reproduced below.

FlowSGS outperforms MPC-Flow in PSNR, SSIM and LPIPS on every task (the only exception is FlowSGS (MFM) on phase retrieval, where the one-step warm-up fails, as discussed in Section 4.3). FlowSGS (Linear) is ahead by 1.7–6.1 dB PSNR on the linear problems, and on Fourier phase retrieval by 11.3 dB with the oracle selection and 2.5 dB on the mean of 8 samples (Table 7). We note that MPC-Flow is the fastest method among all (Table 6). We also report its uncertainty metrics (NLL, ECE) in Tables 7, 8 and 9. On compressed sensing MRI (Cartesian), MPC-Flow has the lowest data-fit residual (47.14), but this is below that of the ground truth (52.41, Table 9), which suggests it overfits the measurement noise.

Motion deblurringGaussian deblurringSuper-resolution (4×)
MethodPSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
MPC-Flow25.580.7420.22527.940.8010.17125.810.7530.204
FlowSGS (VP)28.930.8400.14729.580.8440.13928.500.8240.157
FlowSGS (GVP)29.010.8410.14929.410.8430.14428.590.8290.161
FlowSGS (Linear)29.230.8430.14829.610.8450.14429.090.8400.159
FlowSGS (MFM)28.850.8320.13929.450.8390.13129.060.8410.142
CS-MRI (Cartesian, 8×)CS-MRI (radial, 8×)Fourier phase retrieval
MethodPSNR ↑SSIM ↑Data-fitPSNR ↑SSIM ↑Data-fitPSNR ↑SSIM ↑LPIPS ↓
MPC-Flow26.310.67947.1432.070.81249.1225.100.6620.259
FlowSGS (VP)31.770.81951.73234.710.86449.75834.480.9250.082
FlowSGS (GVP)31.990.82653.42534.900.86951.15731.590.8990.128
FlowSGS (Linear)32.370.84251.93034.820.86950.93836.350.9390.059
FlowSGS (MFM)32.570.84451.92434.810.86950.91614.880.4740.584

Comments on the paper

3 comments
AuthorsOpenReview ↗

Global response

We thank the reviewers for their careful reading. They found FlowSGS extremely sound and practically useful (agjN), a useful, theoretically well-motivated generalization of PnP-DM with SOTA or near-SOTA results (KSAG), and a source of a novel timestep correction and a very fast hybrid acceleration (M6RD).

We have edited the draft with revised text in blue, and have added figures and tables for new experiments (caption highlighted in blue). The shared concerns are the scope of our claims, what is new beyond PnP-DM, and evidence for posterior accuracy.

G1. Beyond PnP-DM. FlowSGS inherits PnP-DM's SGS template (Corollary 1), but the extension is not automatic. What is new is flow-specific: (1) a velocity-only prior step for any SI schedule, via schedule-specific time inversion and rescaling (Theorem 1); omitting either costs 5.5–22 dB (Table 10, Figure 5b). (2) The timestep correction. The few-step sampling ability of flow priors cannot be unlocked with PnP-DM's prior-step discretization; removing the correction costs 9.8 dB (Table 10). PnP-DM also benefits from our correction, even more than FlowSGS (updated Figure 5c). (3) Warm-up invariance (Corollary 3), which licenses one-step flow maps during annealing for a ~10× speedup. We note that together these novel designs put our method at a favorable position in the time–performance tradeoff space (see Figure 1b).

G2. New experiments. We thank the reviewers for the suggested experiments. The revised draft now includes:

  • PnP-DM with timestep correction (Figure 5c)
  • Toy-problem KL vs. SDE steps, SGS iterations and ρmin⁡\rho_{\min} with the analytic SGS floor (Figure 13)
  • Toy-problem KL for each diffusion coefficient (Figure 14)
  • Effect of the number of MFM warm-up iterations (Figure 15)
  • Prior log-likelihood of upright vs. rotated faces (Figure 11)
  • FPR metrics without oracle selection (Table 7)
  • Uncertainty metrics (NLL, ECE) with 64 posterior samples on linear inverse problems (Tables 8 and 9)

We are also working to add MPC-Flow as a baseline and will try to post the results during the rebuttal period.

AuthorsOpenReview ↗

Update: MPC-Flow baseline

We would like to thank all reviewers for their careful read-through of our rebuttals and their positive feedback.

As mentioned in our responses, we have now added MPC-Flow (Webber et al., 2026) as a baseline on all tasks (Tables 1 and 5–9, new rows in blue) using its official implementation. FlowSGS (Linear) outperforms MPC-Flow in PSNR, SSIM, and LPIPS on every task, including nonlinear Fourier phase retrieval. See our reply to Reviewer KSAG for detailed results.

AuthorsOpenReview ↗

Final response

Dear Area Chair and reviewers,

As the rebuttal period is approaching its end, we would like to thank the reviewers for their helpful comments and positive feedback. We sincerely appreciate their time and the improvements to the manuscript that they have assisted us in making. We also thank the Area Chair for their time in handling our submission.

Best regards,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.