acceptodds
Under review as a conference paper at ICLR 2027

SPIR: Symmetric-Probe Image Restoration with Flow-Matching Priors

Abstract

Flow-matching priors provide clean-image estimates from degraded inputs, but a limited evaluation budget leaves a choice between a single central prediction and aggregation over a noisy neighborhood. We introduce Symmetric-Probe Image Restoration (SPIR), which averages predictions from opposite Gaussian probes and applies a data-consistency correction. The symmetric construction excludes odd perturbation responses while retaining nonlinear neighborhood information. Reusing the probes keeps the sampled prior estimator fixed within each time step. Our analysis identifies conditions under which symmetric probes improve estimation and characterizes the stability of data correction. Experiments on four datasets and five restoration tasks demonstrate competitive fidelity and perceptual quality, with task-dependent trade-offs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Restoration, Inverse Problems, Flow Matching, Symmetric Probes, Data Consistency

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_wvhmOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper proposes Symmetric-Probe Image Restoration (SPIR), an unsupervised solver for linear inverse image restoration. It uses symmetric Gaussian probes with a pretrained flow-matching prior, averaging predictions to cancel first-order odd responses while preserving nonlinear effects. Combined with data-consistency corrections and early stopping, this approach improves reconstruction efficiency.
  • The paper theoretically analyzes the MSE conditions and local error orders of symmetric sampling. Experiments across multiple benchmarks and inverse tasks, compared with multiple diffusion- and flow-based solvers, demonstrate competitive reconstruction quality (PSNR/SSIM), inference speed, and memory efficiency.
Critical strengths
  1. Sound Theoretical Foundation: The paper applies classical antithetic variates to flow-matching endpoint prediction, deriving an explicit variance condition for symmetric sampling to outperform independent sampling. It also rigorously establishes the non-expansiveness of the data-correction operator and a residual bound.
  2. Practical Efficiency: The method requires neither backpropagation nor complex ODE solvers, offering low memory usage and fast inference. Compared with diffusion- and flow-based methods requiring higher-order derivatives or multi-step sampling, it is more deployment-friendly.
Critical weaknesses
  1. Limited Novelty and Theoretical Concerns: The method mainly combines antithetic sampling with existing flow-matching frameworks. Frozen probes provide negligible empirical gains, while their interaction with updated states raises concerns about the theoretical validity of Proposition 1.

  2. Perceptual Quality Limitations: Despite improving PSNR in some cases, symmetric averaging can degrade LPIPS and oversmooth high-frequency details, with SPIR underperforming strong baselines on perceptual metrics for certain tasks.

Questions for authors
  • Frozen Probes: Given the lack of clear performance gains and potential violations of the independence assumption, what justifies using frozen probes rather than resampling?

  • Perceptual Quality: Does symmetric averaging lead to oversmoothing and loss of fine textures despite improving pixel-level fidelity? Could this limit its effectiveness in perceptually demanding tasks such as super-resolution and texture restoration?

  • Generalization: Does the method remain effective on diverse, high-resolution natural images and more complex or ill-conditioned degradation operators, where the local smoothness assumption may be less reliable?

1 comment

I thank the authors for their detailed response and additional experiments, which have satisfactorily addressed most of my concerns. Taking into account the rebuttal and the discussion with other reviewers, I believe my initial assessment remains appropriate. Therefore, I will maintain my original score.

Reviewer_Mac5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SPIR allows high fidelity, training free restoration by averaging flow-model predictions from opposite Gaussian perturbations and then enforcing measurement consistency. The paper analyzes when symmetric sampling reduces estimation variance, and how we can take advantage of it. Across five restoration tasks and four datasets, SPIR achieves strong fidelity and competitive perceptual quality.

Critical strengths
  • The method is simple and practical: it averages predictions from opposite noise inputs and then corrects the result to match the measurement.
  • The analysis for the stability of the measurement correction step is very clear and gives you the intuition why symmetric probes should reduce prediction variance.
  • Generally strong PSNR and SSIM results and a useful comparison against independent probes. across 5 inversion tasks.
Critical weaknesses
  • The variance proof assumes the current image estimate is independent of the sampled probes. In the actual algorithm, probes are reused and therefore influence later estimates, so the proof does not fully justify the complete iterative method (especially since probe reuse shows no consistent empirical benefit).
  • The computational comparison is incomplete because matching methods by NFE accounts only for flow-model evaluations and ignores SPIR’s additional conjugate-gradient corrections, meaning that equal NFE does not necessarily imply equal runtime or total computational cost.
Questions for authors
  • How does the variance-reduction analysis apply to the full SPIR algorithm, where reused probes influence later iterates and are therefore no longer independent of the current estimate?

  • Could the authors report wall-clock runtime or total computation, including the conjugate-gradient corrections, rather than comparing methods only by NFE?

  • Could the authors evaluate independent versus symmetric probes across more datasets and restoration tasks to show that the benefit is not specific to the reported inpainting setting?

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

The paper proposes SPIR, a training-free flow-matching solver for linear inverse problems. At each time step, the flow's destination prediction is averaged over a Gaussian probe and its negative (antithetic pairs), and then corrected with a data-proximal step solved by CG. The authors show that pairing cancels the odd part of the predictor's response and give a condition under which it estimates the Gaussian-averaged destination better than independent probes. Experiments on four datasets and five tasks show competitive or better fidelity than the baselines.

Critical strengths
  1. The method is simple, needs no backpropagation through the network, and has low memory and latency (Appendix G).
  2. The experiments are broad (four datasets, five tasks, eleven baselines), with standard deviations and bootstrap intervals.
Critical weaknesses
  1. Limited novelty and unclear motivation. The destination estimator (Eq. 6), the proximal step (Eq. 13) and the schedule (Eq. 9) come from FLOWER (Pourya et al., 2026), and PnP-Flow already averages predictions over random perturbations (a better discussion about FLOWER and how the method is similar should be provided). Thus, the new element is antithetic pairing, a classical variance-reduction technique. The problem is that the paper shows that pairing estimates Qt(x)Q_t(x) better (Fig. 2a), but never explains why a better estimate of Qt(x)Q_t(x) should give better restoration. Its own results suggest the benefits of the object are weak: Independent has lower LPIPS than Symmetric (Fig. 2c), and the perceptual effect of pairing changes sign across tasks (Table 3). Related work that averages a denoiser over noisy copies of the iterate is also missing, e.g., SNORE [1], score-distillation methods such as RED-diff [2] and, for flow priors, FLAIR [4].

  2. The gains are hard to attribute to pairing. Pairing adds about 0.3–0.4 dB PSNR in Table 3, yet SPIR beats FLOWER by 2.6 dB (FFHQ SR) and 1.9 dB (AFHQ SR). The remaining gap likely comes from other aspect of the solver, namely the schedule, the inner updates, or per-task tuning of N (Tables 4–7), while FLOWER uses 100 steps everywhere. The NFE-matched ablation covers only one task, and its Center baseline spends the extra evaluations on more inner updates.

Questions for authors
  1. Which pretrained flow model does SPIR use for each dataset? Is it the same network as the other flow baselines? Diffusion and flow baselines use different priors, which confounds the comparison.
  2. How were SPIR's hyperparameters (N, M, K, tolerance, grid) selected? Were the baselines tuned with similar effort?
  3. How does SPIR with K=1 and 2N time steps (same NFE as K=2) compare? How does SPIR with K=1 compare to FLOWER under the same schedule?
  4. FLOWER and PnP-Flow outputs are averaged over five samples, which helps PSNR and hurts LPIPS. Can you report single-sample results? Otherwise, this may explain part of SPIR's LPIPS advantage (e.g., 0.071 vs. 0.131 on AFHQ denoising).
  5. Why are OT-ODE and PnP-Diff N/A for deblurring, given that both can handle them?
  6. The theoretical contribution is limited. Proposition 1 is the standard antithetic-variance identity. Proposition 2 only bounds the distance to the center prediction (not to Q_t(x) or the clean image), and the bound scales with d, so it is vacuous at image scale. It also suggests the symmetric estimate is close to the center prediction, which seems at odds with the motivation of capturing nonlinear information. What does Proposition 2 add? Can you say anything about the inner loop C∘D̄ (convergence or fixed points), given that Sec. 4 criticizes PnP-Flow and FLOWER on this point?
  7. Please move a few visual results from Appendix A into the main text given that the paper is an image restoration one

References

[1] M. Renaud, J. Prost, A. Leclaire, N. Papadakis. Plug-and-Play image restoration with Stochastic deNOising REgularization. ICML 2024.

[2] M. Mardani, J. Song, J. Kautz, A. Vahdat. A Variational Perspective on Solving Inverse Problems with Diffusion Models. ICLR 2024.

[3] J. Erbach, D. Narnhofer, A. Dombos, B. Schiele, J. E. Lenssen, K. Schindler. Solving Inverse Problems with FLAIR. NeurIPS 2025.

1 comment
AuthorsOpenReview ↗

Response Part (2/3)

Q5: Why are OT-ODE and PnP-Diff N/A for deblurring, given that both can handle them?

A5: Thank you for raising this question. Both OT-ODE and PnP-Diff support deblurring. The earlier missing entries resulted from a local numerical implementation issue. We have corrected it and added the missing deblurring results to Table 1 of the revised paper. The results are also reported below (PSNR / SSIM / LPIPS).

MethodAFHQ-CatCelebAFFHQImagenette2
OT-ODE27.38 / 0.720 / 0.12232.82 / 0.918 / 0.03127.32 / 0.775 / 0.10823.38 / 0.544 / 0.199
PnP-Diff28.59 / 0.761 / 0.11733.99 / 0.925 / 0.02428.95 / 0.818 / 0.15130.02 / 0.871 / 0.035

Q7: Please move a few visual results from Appendix A into the main text given that the paper is an image restoration one

A7: Thank you for this suggestion. We have moved the AFHQ-Cat and CelebA qualitative results from Appendix A to Figure 2(a–b) on page 8 of the main text, with a reference in Section 5.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.