acceptodds
Under review as a conference paper at ICLR 2027

Better Trajectories, Not More: Paired Posterior Correction for Flow-Matching Inverse Problems

Abstract

We propose *PPC-Flow* (Paired Posterior Correction), a training-free flow-matching solver for linear image inverse problems. The strongest flow-matching solvers re-inject noise and average independent trajectories, which multiplies their cost and gives up the speed advantage of flow priors. We trace their error to two sources, a bias that builds up late in the trajectory and the variance of the re-injected noise. The bias arises because the proximal weight follows a fixed schedule that vanishes near the terminal time, which leaves the pull of the network toward the prior uncorrected in the final steps. Averaging independent trajectories reduces only the variance, and only at the Monte Carlo rate. *PPC-Flow* counters the bias with a posterior velocity correction, which scales up the proximal correction, and suppresses the variance by antithetic pairing of the re-injection noise, both without additional network evaluations. On the CelebA and AFHQ-Cat benchmarks, *PPC-Flow* achieves state-of-the-art PSNR and SSIM among flow-matching solvers while using only 40% of the NFE budget. Code and video are available at https://anonymous.4open.science/r/ppc-flow-C40F.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
flow matching, image inverse problems, training-free inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xMu8OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a new method to solve imaging inverse problems using a pretrained flow matching model (unspecific to the considered inverse problem). The method is build on FLOWER, which itself was built on PNP-FLOW. Both FLOWER and PNP-FLOW iterations consist of 1) a data consistency step; 2) a renoising step unsing random noise; 3) a flow-based denoising step. Usually in the renoising step 2, one draws independent noise samples and then averages over them.

Contributions:

  • The main modification with respect to FLOWER that is proposed in this paper is to use antithetic noise (ie ε\varepsilon and −ε-\varepsilon, where ε∼N(0,I)\varepsilon \sim \mathcal{N}(0, I)). This helps to improve the image restoration, as the experiments show.
  • The second modification with respect to FLOWER is the ``Posterior velocity correction'' but I am not even sure it is really different than what is done in FLOWER or if it's a different way to write it.

Overall assesment The paper is well written and easy to follow, and the numerical results are reasonably competitive. However, I find the contribution limited (little methodological novelty, not much theoretical understanding of why such a method would solve a MAP problem / do posterior sampling). There have been so many papers over the last two years trying to slightly improve the state of the art for inverse problems. While this is certainly a valuable objective, I believe the inverse problem community would benefit more from studies that try to understand what these solvers actually do and what guarantees they provide.

Critical strengths
  • Using antithetic noise improves the restoration performance at a fixed NFE compared to FLOWER, which is one of the state-of-the-art methods
  • The use of antithetic noise is justified
  • The paper is clear
Critical weaknesses
  • Lack of novelty compared to FLOWER
  • Visually, the results are not so impressive: the method seems to suffer from the same kind of oversmoothing than it's predecessors
  • Comparison to recent diffusion baselines are missing (see below) and it would be good to evaluate on another dataset than Celeba and AFHQ-Cat that are very centered on a face.
Questions for authors
  1. How does the velocity correction step you propose lead to a different ``denoising'' update than the one of FLOWER? Could you also please provide numerical results of your method but without the antithetic noise (say just one random sample) and compare it to FLOWER-1 (with refinement) so that we see the effect of the velocity correction.

  2. Since the antithetic noise reduces the variance, can you observe a reduced variability between runs compared to FLOWER? Is this desirable?

  3. Do you have an insight on why the reconstructions are oversmoothed visually?

  4. Since it has clearly been established that flow matching and diffusion both build a path of denoiser (Dσ)(D_\sigma) for noise levels σ∈(0,+∞)\sigma \in (0, +\infty), it would be necessary to also compare to the SOTA methods that were proposed in the diffusion framework (using the same networks as for flow matching). Could you compare your method to DAPS [1]?

  5. Why do you use flows trained with OT couplings? What does it bring to your method?

[1] Zhang, B., Chu, W., Berner, J., Meng, C., Anandkumar, A., & Song, Y. Improving diffusion inverse problem solving with decoupled noise annealing. CVPR 2025

14 comments
AuthorsOpenReview ↗

Response to Reviewer xMu8 (1/4)

Thank you for the careful review and the direct questions. We agree that the community needs to understand what these solvers actually compute, and we tried to make that the center of our answers. New material in the revised PDF is shown in blue.

W1. Lack of novelty compared to FLOWER

Flower is an excellent solver and a natural starting point for us, and PPC-Flow keeps its proximal step. Our contribution is a deeper look at what happens along the trajectory, which gives a definite answer to a question Flower leaves open. In its Appendix 8.5.4, Flower reports that averaging more trajectories saturates after about five runs. Our analysis explains why, and shows how the NFE budget should be spent instead. The analysis applies to any flow solver that alternates a network prediction, a proximal data step, and noise re-injection, such as PnP-Flow and Flower. It reveals three behaviors:

  • A spring-like pull-back. The proximal step moves the predicted endpoint toward the measurement, but the next network call pulls it back toward the prior, like a stretched spring recoiling. Figs. 2a and 5 show this effect. Since the proximal weight fades as t→1t \to 1, the pull-back wins in the late steps. This bias is shared by all runs, so averaging cannot remove it. In fact, about 80% of the single-trajectory error is shared across runs (Table 16).
  • Short memory. Per-step contractions erase early corrections and the initial latent, as Lemma 2 shows. Correcting only the early steps changes almost nothing, and pairing the initial latents has no effect (Table 12). So what matters at the output is the late steps and the re-injected noise.
  • A near-linear noise response. Even after 100 network calls, more than 90% of the variance caused by re-injected noise is odd in the noise (Table 15).

Each mechanism targets one of these behaviors. Velocity correction over-relaxes the late data step to undo the pull-back, and a prior-free cap keeps every residual mode contracting (Proposition 1). On SR×2 it cuts the shared error by 15%, which no rescaling of the proximal weight can do, as we show in Post 2/4. Antithetic pairing flips the sign of the re-injected noise, where the variance actually comes from, and cancels its linear part with a single pair. After one pair, the variance is below 2% of the MSE (Table 16).

In short, the analysis turns "average more trajectories" into "make each trajectory better". With only two trajectories, PPC-Flow beats Flower at the same budget, at five times the budget, and even at its averaging ceiling, on all five CelebA tasks. Here is Table R1 from Appendix D.1, with PSNR in dB:

TaskFlower, K=2 (2T)Flower, K=10 (10T)Flower ceiling (∞\infty)PPC-Flow, K=2 (2T)Gain vs. K=2 / ceiling
SR×232.79433.18833.27934.061+1.27 / +0.78
Random inpainting33.66634.15534.27334.906+1.24 / +0.63
Denoising32.80433.26433.38533.482+0.68 / +0.10
Box inpainting31.41931.83931.98132.084+0.67 / +0.10
Deblurring35.39435.76135.85335.889+0.50 / +0.04

We estimate the ceiling by running Flower with five antithetic pairs, which is already very close to the limit of infinitely many runs.

AuthorsOpenReview ↗

Response to Reviewer xMu8 (2/4)

Summary. […] I am not even sure it is really different than what is done in FLOWER or if it's a different way to write it.

Q1. How does the velocity correction step you propose lead to a different "denoising" update than the one of FLOWER? Could you also please provide numerical results of your method but without the antithetic noise (say just one random sample) and compare it to FLOWER-1 (with refinement) so that we see the effect of the velocity correction.

No, velocity correction is not Flower written in another way. It gives a different update. In the "denoising" step, velocity correction keeps the network prediction mθ,tm_{\theta,t} and adds a shift along the measurement residual (Eq. 11). This shift is computed from the measurement, not the network, so it needs no extra network call. Combined with the proximal step, this scales Flower's update by ωt=1+(1−t)αt/rt≥1\omega_t = 1 + (1-t)\alpha_t/r_t \ge 1. Proposition 1 gives the resulting step:

x⋆=mθ,t+ωt(zt⋆−mθ,t).\mathbf{x}^\star = m_{\theta,t} + \omega_t(z_t^\star - m_{\theta,t}).

Flower is the case ωt=1\omega_t = 1. On each observed mode it moves only a fraction wj(t)<1w_j(t) < 1 of the way toward y\mathbf{y}. It never crosses the measurement, so the next network call pulls the iterate back toward the prior. PPC-Flow uses ωt>1\omega_t > 1 under ωtwmax⁡(t)<2\omega_t w_{\max}(t) < 2, so every residual mode still contracts. For example, on SR×2 at t=0.9t = 0.9, Flower moves 83% of the way to y\mathbf{y}, while PPC-Flow moves 183% and the next pull-back lands near y\mathbf{y}. Fig. 5 confirms this on SR×2: at t=0.9t = 0.9, velocity correction lowers the endpoint error on the observed pixels by a quarter. Here ωt\omega_t is not a free knob. It grows only in the late steps, where the pull-back dominates, and a prior-free cap keeps every residual mode contracting for any network (Proposition 1).

Here is the single-trajectory comparison you asked for, without pairing and at the same TT NFEs. Both columns keep Flower's refinement step, so the gain shows the effect of velocity correction alone. The PSNR values in dB come from Table 14:

TaskFlower-1, with refinement (TT)Ours, K=1K=1, no pairing (TT)Gain
SR×232.34633.081+0.735
Random inpainting33.12333.629+0.506

The gain stays positive on the other three tasks as well (Appendix D.5). On SR×2, one corrected trajectory is already on par with Flower-5, which averages five. Interestingly, most of the MSE drop lies on unobserved pixels, 84% on SR×2 and all of it on random inpainting (Table 4). The correction anchors the observed pixels, and the network then does a better job on the missing ones.

Could a larger λt\lambda_t do the same? No. Since wj(t)=λtσj2/(λtσj2+σn2)<1w_j(t) = \lambda_t \sigma_j^2 / (\lambda_t \sigma_j^2 + \sigma_n^2) < 1 for any λt>0\lambda_t > 0, rescaling λt\lambda_t never crosses the measurement. We checked this by scaling λt\lambda_t with constant factors from ×0.5\times 0.5 to ×16\times 16, with velocity correction off and K=2K=2. Table R5 in Appendix D.5 gives the PSNR gain in dB:

TaskBest λt\lambda_t rescalingPPC-Flow
SR×2+0.38+0.81
Random inpainting+0.01+0.67
Box inpainting0+0.15
Denoising0+0.12
Deblurring0+0.05

The best factor stays well below velocity correction on every task, so no setting of Flower's λt\lambda_t reproduces it. Our related work section already draws this distinction: Flower's refinement fades near the terminal time, exactly where it matters. To make this distinction clearer, we revised Section 3.2 to state it explicitly right before Proposition 1, and Appendix D.5 now gives the full λt\lambda_t-rescaling control and the single-trajectory results.

AuthorsOpenReview ↗

Response to Reviewer xMu8 (3/4)

Overall. […] not much theoretical understanding of why such a method would solve a MAP problem / do posterior sampling […]

Q3. Do you have an insight on why the reconstructions are oversmoothed visually?

Q2. […] can you observe a reduced variability between runs compared to FLOWER? Is this desirable?

These three questions share one answer, so we group them. With γ=0\gamma = 0, the network predicts the conditional mean mθ,tm_{\theta,t}, and the proximal step takes the mean of a local Gaussian posterior, as Theorem 1 shows. These annealed solvers are therefore best seen as approximate minimum-MSE (posterior-mean) estimators, not global MAP solvers. With γ=1\gamma = 1, they draw local conjugate posterior samples instead (Proposition 4). We now state this in Section 2.2 and Appendix D.2. Beyond this single-step view, the paper gives three trajectory-level results. Lemma 2 shows that early corrections fade through per-step contractions. Proposition 1 shows that the capped update contracts every observed residual mode for any prior. And Proposition 2 gives an exact variance identity that predicts the pairing gain with no free parameter.

Smoothness. This is a deliberate choice, not a side effect. PPC-Flow sits on the fidelity side of the perception–distortion trade-off (Blau & Michaeli, CVPR 2018). It targets the minimum-MSE estimate, so it does not invent details that the measurement cannot determine. Many inverse problems, such as MRI and microscopy, need exactly this, because hallucinated details are not acceptable there (Bhadra et al., IEEE TMI 2021, and Belthangady & Royer, Nature Methods 2019). That said, within this regime PPC-Flow still keeps more high-frequency detail than Flower-5. On all five tasks it retains more high-frequency energy and has a lower high-frequency error (Appendix D.2). When sharp textures matter more than fidelity, the γ=1\gamma = 1 mode of Appendix A.5 can be used, and velocity correction helps there as well.

Variability. Yes, it drops a lot. With fewer network calls than Flower-5, PPC-Flow has less than half of its run-to-run variance, and less than a tenth of that of Flower-1 (Tables 15 and 16). For a point estimate, this is desirable: the variance is part of the MSE, so lowering it raises PSNR and makes repeated runs agree. No uncertainty information is lost either. Pairing keeps each trajectory's marginal unchanged (Section 3.3), and with γ=0\gamma = 0 the spread across runs is not a calibrated posterior uncertainty in the first place. Flower makes the same observation in its Appendix 8.5.3: γ=0\gamma = 0 "led Flower to generate samples concentrated in higher-probability regions, while failing to capture the tails of the posterior". For uncertainty quantification, the γ=1\gamma = 1 mode is the right tool.

Reviewer_y8LoOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The authors propose PPC-Flow, a training-free flow-matching solver for image inverse problems. It corrects bias caused by insufficient data-consistency updates late in the trajectory through posterior velocity correction, and reduces trajectory variance by applying antithetic pairing to the re-injected noise.

Critical strengths

1、The authors argue that averaging multiple trajectories can reduce noise variance but cannot eliminate systematic bias shared across trajectories. The key is therefore to improve the trajectories themselves rather than simply increase their number. 2、The paper provides a detailed experimental analysis, including PSNR, SSIM, and LPIPS metrics, as well as component ablations, error decomposition, comparisons across computational budgets, and bootstrap confidence intervals. 3、The authors provide separate theoretical analyses of posterior velocity correction and antithetic noise pairing.

Critical weaknesses

1、Velocity correction does not yield substantial gains across all tasks. 2、Lacks of extensive qualitative visualizations.

Questions for authors

I am not an expert in this area, but I find the paper’s idea of iteratively correcting flow-matching trajectories to improve reconstruction promising.

1 comment
AuthorsOpenReview ↗

Response to Reviewer y8Lo

Thank you for the encouraging review and for your kind words about our ablations and theory. We were especially glad that you find "the paper’s idea of iteratively correcting flow-matching trajectories to improve reconstruction promising." New material in the revised PDF is shown in blue. Below we address both of your suggestions.

W1. Velocity correction does not yield substantial gains across all tasks.

PPC-Flow gives a clear gain on every task. On all ten task–dataset settings, PPC-Flow with two trajectories beats Flower-5, which averages five, by up to about 1 dB in PSNR (Table 9). Velocity correction costs no extra network evaluation and improves nine of the ten settings (Table 11).

How much velocity correction adds depends on the task, as expected from its design. It strengthens data consistency on the observed pixels, which helps the network infer the missing pixels in later steps. It therefore helps most on super-resolution and random inpainting, where most of its error reduction lands on unobserved pixels (Table 4). On fully observed tasks such as denoising and deblurring, pushing harder toward y\mathbf{y} also amplifies the measurement noise, so validation selects a mild strength. On these tasks and on box inpainting, the two mechanisms share the work. On CelebA, antithetic pairing alone already beats Flower-5 on these three tasks, and on AFHQ-Cat velocity correction adds a further 0.35 dB on denoising and 0.14 dB on deblurring.

W2. Lacks of extensive qualitative visualizations.

Thanks for the suggestion. We added more qualitative comparisons in Appendix D.9.

Thank you again for the encouraging review.

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

For training-free posterior sampling in linear inverse problems, PPC-Flow decomposes per-image MSE into a late-step bias and a trajectory variance, addressing them with a contractive posterior velocity correction and antithetic trajectory pairing. At 2T NFE it outperforms Flower5-OT at 5T NFE across all ten CelebA and AFHQ tasks.

Critical strengths

1.The budget-allocation question is clear and is tested directly. 2.The two main derivations are correct under their assumptions and are used in the method.

Critical weaknesses
  1. I am concerned that the bridge-variance envelope is not directly stress-tested. Both λt and the correction envelope rely on the Gaussian assumption, yet the ablations never perturb σr²(t), replace it with a constant, or use an empirical estimate, leaving the sensitivity of the gain to this scaling unclear.

  2. Results beyond K=2 remain limited. While Figure 6(a) compares antithetic pairing at K up to 10, it lacks exact numerical tables and full PPC-Flow with velocity correction is not systematically evaluated at K>2 across tasks.

Questions for authors
  1. I am quite curious about how sensitive are the results to the σr²(t) envelope? A control scaling it (e.g., ×0.5, ×2) or replacing it with a constant would clarify the role of the bridge model.
2 comments
AuthorsOpenReview ↗

Response to Reviewer AeAT (1/2)

Thank you for the positive assessment of our budget-allocation setup and derivations. Both experiments you suggested, on the σr2(t)\sigma_r^2(t) envelope and on K>2K > 2, turned out to be very informative, and we added them to Appendices D.7 and D.8. New material in the revised PDF is shown in blue.

W1 / Q1. I am quite curious about how sensitive are the results to the σr²(t) envelope? A control scaling it (e.g., ×0.5, ×2) or replacing it with a constant would clarify the role of the bridge model.

We ran all the controls you suggested on CelebA. In both λt\lambda_t and αt\alpha_t, we scaled rt=σr2(t)r_t = \sigma_r^2(t) by ×0.5\times 0.5 and ×2\times 2, or replaced it with constants or with the empirical endpoint error r^(t)\hat r(t). Two things came out clearly. First, in the over-relaxation factor ωt\omega_t, rtr_t cancels out below the stability cap, so the over-relaxation factor itself does not depend on its scale. Second, in the proximal step λt=rt\lambda_t = r_t, the bridge model acts as a smooth time-decay schedule, not as a calibrated variance.

On the first point, rtr_t cancels in the ratio αt/rt\alpha_t / r_t of Eqs. 16–17:

αt=min⁡(a,aˉt)rt,ωt=1+(1−t)min⁡(a,aˉt).\alpha_t = \min(a,\bar a_t) r_t, \qquad \omega_t = 1 + (1-t)\min(a,\bar a_t).

Below the cap, the extra correction therefore tapers as 1−t1-t. The only place rtr_t enters is the cap aˉt\bar a_t of Proposition 1, which is an exact algebraic contraction bound and does not need the Gaussian assumption. Scaling αt\alpha_t below the cap is the same as scaling the strength aa, and the gain changes smoothly. On CelebA SR×2, a=8,12,24a = 8, 12, 24 gives gains of 0.74,0.81,0.860.74, 0.81, 0.86 dB. We also clarified this point in Appendices A.3 and B.10.

On the second point, rtr_t matters through the proximal weight λt=rt\lambda_t = r_t, which PPC-Flow inherits from Flower. Table R7 in Appendix D.7 gives the PSNR change in dB relative to λt=rt\lambda_t = r_t, with velocity correction off:

λt\lambda_tSR×2RandomBoxDenoisingDeblurring
0.5rt0.5 r_t−0.661−0.040−0.050−1.220−0.547
2rt2 r_t+0.329+0.011−0.299−0.260−0.249
Constant 0.5−0.581−0.001−2.671−13.317−24.662
Constant 0.01−0.620−0.016−2.510−7.203−4.303
Measured r^(t)\hat r(t)−4.925−0.275−1.180−9.856−3.602

Here, 0.50.5 is the time average of rtr_t, and r^(t)\hat r(t) is the endpoint error measured on validation images, which is much smaller than rtr_t. Scaling rtr_t by ×0.5\times 0.5 or ×2\times 2 has a modest effect on four of five tasks. On box inpainting, denoising, and deblurring, replacing its time decay with a constant hurts much more, and the smaller r^(t)\hat r(t) degrades all five tasks. So rtr_t works as a smooth step-size schedule, and Appendix A.1 now says this for natural images.

We also measured the velocity-correction gain under each λt\lambda_t schedule, with aa kept at its default value (Table R8 in Appendix D.7). The gain stays positive on SR×2, random and box inpainting under every schedule. On all five tasks, it also grows as λt\lambda_t weakens from 2rt2 r_t to rtr_t, 0.5rt0.5 r_t, and r^(t)\hat r(t). This fits its role of compensating for a weak proximal step. With a doubled λt\lambda_t, the proximal step already does more of the work, so on denoising and deblurring a smaller aa would be selected.

AuthorsOpenReview ↗

Response to Reviewer AeAT (2/2)

W2. Results beyond K=2 remain limited. […] full PPC-Flow with velocity correction is not systematically evaluated at K>2 across tasks.

Thanks for this suggestion. We extended the evaluation to K=2,4,10K = 2, 4, 10 on all five CelebA tasks, for Flower, antithetic pairing alone, and full PPC-Flow. The paired methods use K/2K/2 antithetic pairs. Here are the PSNR results in dB, from Table R9 in Appendix D.8:

TaskFlower (K=2/4/10K = 2 / 4 / 10) ↑Pairing only, a=0a=0 (K=2/4/10K = 2 / 4 / 10) ↑PPC-Flow (K=2/4/10K = 2 / 4 / 10) ↑
SR×232.794 / 33.030 / 33.18833.249 / 33.268 / 33.27934.061 / 34.085 / 34.092
Random inpainting33.666 / 33.970 / 34.15534.238 / 34.261 / 34.27334.906 / 34.929 / 34.948
Box inpainting31.419 / 31.717 / 31.83931.933 / 31.952 / 31.98132.084 / 32.111 / 32.139
Denoising32.804 / 33.083 / 33.26433.363 / 33.376 / 33.38533.482 / 33.497 / 33.507
Deblurring35.394 / 35.615 / 35.76135.837 / 35.846 / 35.85335.889 / 35.899 / 35.906

The numbers support the K=2K=2 choice. After one antithetic pair, the remaining variance is below 2% of the MSE (Table 16). So PPC-Flow reaches its plateau at K=2K=2 and does not need more trajectories. Flower keeps improving with KK because its trajectories still carry variance. Yet even at K=10K=10, with five times the NFEs, it stays below PPC-Flow at K=2K=2 on all five tasks. The gain of velocity correction over pairing alone is almost the same at every KK.

Thank you again for both suggestions. They made the paper more complete.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.