acceptodds
Under review as a conference paper at ICLR 2027

Zero-Shot Blind Inpainting with Flow Matching

Abstract

The blind inpainting (BI) problem, a challenging variant of image inpainting, aims to reconstruct missing or corrupted regions when the corruption mask is unknown. Recent deep-learning approaches have predominantly adopted data-driven paradigms, whose generalization depends on the diversity and quality of paired training data. In this paper, we revisit zero-shot BI using pretrained generative priors. The only prior zero-shot method, RGI, explicitly optimizes a sparsity-regularized mask. We show that its solution is inherently sensitive to the unknown corruption budget and can become degenerate when the generative prior interpolates the corrupted observation. To address these limitations, we propose ZSBI-FM, a mask-free framework that models corruption as a structured sparse reconstruction residual. Our nested SCAD objective combines pixel-level sparsity with SLIC superpixel groups to handle both isolated and spatially coherent corruptions, while an unsupervised stopping criterion limits overfitting. Across diverse mask geometries and fillings, ZSBI-FM improves PSNR over the strongest competing method by – dB on FFHQ and – dB on ImageNet, and remains robust under additional measurement noise. On simulated BTAD, it also enables accurate defect localization without defective training images or pixel-level annotations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deep Generative Prior, Inverse Problem, Blind Inpainting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PqqvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the blind image inpainting problem, where the corruption locations are unknown. To address the reliance of supervised methods on paired training data and the joint optimization of corruption masks and generative latents in existing zero-shot approaches such as RGI, the authors propose modeling unknown corruptions as structured sparse reconstruction residuals. Based on this motivation, they develop ZSBI-FM, which integrates a pretrained flow-matching generative prior with SLIC-based superpixel grouping and a nested SCAD penalty at the pixel and group levels. An unsupervised early-stopping strategy is further adopted to mitigate corruption overfitting. Experiments on FFHQ and ImageNet demonstrate improved PSNR and SSIM across various corruption settings, along with promising robustness to additional measurement noise and applicability to industrial anomaly localization.

Critical strengths
  1. The paper models unknown corruptions through structured sparse residuals, avoiding explicit corruption-mask optimization and providing a meaningful alternative perspective for zero-shot blind image inpainting.
  2. The proposed framework integrates a pretrained generative prior, pixel-level robust penalties, superpixel-level structural constraints, and early stopping to jointly address image reconstruction, corruption suppression, and overfitting.
  3. ZSBI-FM consistently achieves superior PSNR and SSIM across diverse corruption settings on both FFHQ and ImageNet, demonstrating notable improvements in reconstruction fidelity over competitive baselines.
Critical weaknesses
  1. The additive sparse corruption model in Eq. (7) has already been considered in RGI (Sec. 3.1), making the conceptual novelty of this reformulation unclear. Moreover, if the generator can reproduce the corrupted observation, the proposed objective still admits a degenerate solution with zero residual. Thus, eliminating explicit mask variables does not fundamentally resolve corruption fitting or reconstruction ambiguity. The authors should clarify what advantages the mask-free formulation provides beyond alternative loss design.
  2. Proposition 2.1 analyzes a fixed latent variable under a hard mask budget, which does not establish that accurate corruption-budget specification is necessary for reliable reconstruction under RGI's actual penalized formulation. Furthermore, RGI already derives an equivalent saturating robust residual objective and provides conditional asymptotic recovery guarantees (Sec. 3.3 and Theorems 1 and 2). These results call into question the attribution of the identified limitations to explicit mask parameterization itself. The authors should reconcile their theoretical claims with these prior findings.
  3. In its linear regime, the nested SCAD objective primarily induces group-level rather than independent pixel-level sparsity. Moreover, saturation of the outer penalty eliminates the group's contribution to reconstruction gradients, potentially affecting clean pixels grouped with corrupted ones. The authors should better justify the claimed dual-level sparsity and robustness to imperfect SLIC grouping.
  4. Despite achieving higher PSNR and SSIM, ZSBI-FM consistently underperforms FLUX.2-dev in LPIPS across the aggregated corruption categories. A potential explanation is that SCAD saturation suppresses gradients from heavily corrupted regions, making reconstruction increasingly dependent on the generative prior while prioritizing consistency in intact regions? The authors should analyze this potential mechanism and provide additional visual comparisons and region-wise evaluations to clarify the sources of this trade-off.
Questions for authors

Please see my main concerns in the Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to reviewer Pqqv - I

Response to W1: Novelty of the additive formulation and zero-residual degeneracy

  • Novelty of the model reformulation : We agree the sparse corruption model has been considered in the RGI paper; however, we also do not claim any novelty in this model reformulation in our paper. Our main novelty lies in our novel solution of the BI problem on top of this sparse corruption model, i.e., the group-sparsity SCAD loss formulation for BI in Eq. (11). In short, the forward model is old, but our optimization formulation in Eq. (11), and the whole algorithm pipeline (Algorithm 1 in Appendix E in the revised version) are novel.

  • Zero-residual degeneracy: We do not claim that the mask-free model in Eq. (7) and our associated group-sparsity SCAD loss formulation in Eq. (11) is free from corruption fitting. In fact, we have acknowledged such degeneracy of our formulation Eq. (11) in Sec. 3.3. Therefore, we integrate the ES-WMV early stopping module into our algorithm framework to mitigate this issue.

  • Advantages of the mask-free model: The mask-based model adopted in RGI suffers from the sensitivity of the specified sparsity level for the mask (as in Eq. (2)), as implied by Proposition 2.1. This carries on to the sensitivity of the regularization parameter λ\lambda if one were to adopt the RGI formulation Eq. (3) or similar. Wrong parameter specification can easily lead to over/under-estimation of the mask, which likely further impacts the image estimation. Our industrial application in Sec. 4.3, which concerns mask estimation, shows a clear performance advantage of our method compared to the RGI, even after we have tried to reasonably tune around their λ\lambda.

Response to W3: Nested SCAD and imperfect SLIC grouping

  • Pixel-level vs. group-level sparsity: Sorry for the confusion! We realize we did not completely sync up our introduction with the method part. We do not have “dual-level sparsity,” and the group-sparsity formulation in Sec. 3.2—our primary formulation—is a single-level sparsity, as the reviewer correctly pointed out, “In its linear regime, the nested SCAD objective primarily induces group-level rather than independent pixel-level sparsity.” We will revise our introduction accordingly to remove the confusion.

  • Robustness to imperfect SLIC grouping: We agree that saturation of the outer SCAD can eliminate the group’s contribution to the overall gradient. They likely happen for superpixels inside the mask region, as well as the mixing region—mixed with clean and corrupted pixels. To empirically study the effect of such saturation on performance, we test our method on three different levels of pixel counts on the same 100-image FFHQ subset used in Sec. 4.2, and summarize the results in the table below.

    SuperpixelsPSNR ↑SSIM ↑LPIPS ↓Mixed SP ↓NP-UE ↓
    51231.240.9330.0570.0730.036
    102431.230.9330.0570.0470.026
    204831.220.9330.0570.0320.020

    Here, Mixed SP is the proportion of superpixels over all superpixels, and NP-UE is a similar but finer metric that all considers the mixing ratio inside individual superpixels. Intuitively, these metrics decrease as we get finer segmentation with more superpixel budget, as shown in the table. Overall, our method is very robust to these variations in segmentation quality.

Response to W4: LPIPS Gap and the Explanation

  • Performance of FLUX.2-dev is for reference only. ZSBI-FM employs the lightweight, FLUX.2-klein backbone, and our results show that ZSBI-FM+FLUX.2-klein consistently outperforms FLUX.2-klein in all metrics we use. FLUX.2-dev is a substantially more powerful backbone that FLUX.2-klein is distilled from. So, it does not make sense to compare ZSBI-FM+FLUX.2-klein and FLUX.2-dev; we include the performance of FLUX.2-dev for reference only.
  • SCAD saturation might impact the performance of heavily corrupted region. We appreciate the reviewer sharing this insightful thought! During the rebuttal, we have computed the superpixel-level SCAD gradient and analyzed their average magtidudes over the mask region, off-mask region, and mixing region. We observe consistently that the mask region has substantially smaller gradients relative to the off-mark region, with those of the mixing region lying in between, confirming the intuition—we include the results as Fig. 11 of Appendix D.8.
AuthorsOpenReview ↗

Response to W2

We thank the reviewer for this insightful comment!

The necessity of corruption-budget specification for RGI's actual penalized formulation.

  • Yes, Theorems 1 & 2 of RGI have already established that corruption-budget specification is not needed for asymptotic optimality of z\mathbf z and M\mathbf M, under the idealized assumption that GG learns the true image manifold perfectly, i.e., y=G(z)+s\mathbf y = G(\mathbf z) + \mathbf s for sparse s\mathbf s.

  • However, in practice, GG never learns the manifold perfectly, given the small yet finite FID’s typically reported about generative models in the literature and the explicit coverage test in, e.g., Table 2 of the reference below.

Wan, Y., Devera, R., Zhang, W., & Sun, J. (2026). Saving foundation flow-matching priors for inverse problems. In Proceedings of the 43rd International Conference on Machine Learning.

This implies that a **realistic** model is $\mathbf y \approx G(\mathbf z) + \mathbf s$, where $\approx$     allows approximation errors that might be dense. With this practical modification, both Theorems 1 & 2 of RGI collapse; in particular, one needs to set an error-dependent lower bound for $\lambda$ for exact recovery to be possible—this can be figured out from their typical $\ell_1$ shrinkage-style proofs. In other words, in practice, one still needs to find an appropriate tradeoff $\lambda$—not too large nor too small, partially echoing the difficulty of setting the sparsity level $n_0$ in Eq (2), as it is unknown in practice. 

The equivalent saturating robust residual objective in Theorems 1 & 2 of RGI

In our paper, we use Proposition 2.1, which does not depend on the idealized nor realistic data assumptions above to hold, to discuss RGI’s limitation in its sensitivity to the sparsity parameter kk and motivate our mask-free formulation. RGI’s Theorems 1 & 2 prove the asymptomatic equivalence of their formulations to the ℓ0\ell_0-based mask-free formulation: min⁡z∈Sd  ∥y−G(z)∥0\min_{\mathbf z \in S^d} \; \| \mathbf y - G(\mathbf z)\|_0, which we also set to solve directly. In summary, our Proposition 2.1 and RGI’s Theorems 1&2 motivate the same mask-free formulation from different angles, but RGI and we approximate the ℓ0\ell_0 loss with different functions—more importantly, we also propose to exploit the group-sparsity structure for practical corruptions.

We will incorporate these discussions into the revised version.

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

This paper investigates the problem of zero-shot blind inpainting (BI) under conditions where the mask is unknown. The authors note that existing zero-shot methods, such as RGI, are sensitive to the unknown corruption budget when the mask is explicitly optimized, often leading to performance degradation. To address this, the authors propose ZSBI-FM—a framework that eliminates the need for explicit mask optimization by leveraging a flow-matching prior (FMPlug). This approach models unknown corruption as a structured sparse reconstruction residual. It constructs a nested SCAD (Smoothly Clipped Absolute Deviation) optimization objective that incorporates both pixel-level sparse constraints and SLIC superpixel-based group sparse constraints. Furthermore, an unsupervised early-stopping strategy is employed to mitigate overfitting to measurement noise. Experiments on synthetic degradations using the FFHQ and ImageNet datasets demonstrate that this method achieves strong PSNR and SSIM results; the study also explores its application to industrial anomaly localization.

Critical strengths
  • Strong Zero-Shot Robustness and Application Potential: Requiring no paired training data, ZSBI-FM achieves favorable PSNR and SSIM scores across various synthetic degradation and noise scenarios, while also demonstrating potential for applications in industrial anomaly localization.

  • Theoretical Analysis and Innovation in Mask-Free Modeling: The authors analyze the sensitivity of explicit mask optimization (as seen in RGI) to the corruption budget, as well as the limitations of L1L_1 relaxation. By employing a structured residual sparsity constraint—thereby avoiding the explicit estimation of an unknown mask—the approach offers a novel perspective on zero-shot blind image restoration.

Critical weaknesses
  • Although the paper provides numerous quantitative metrics to validate the effectiveness of the proposed method, it lacks sufficient visual comparisons for the primary blind inpainting task. This makes it difficult to intuitively assess the structural fidelity, texture quality, and presence of artifacts in the reconstructed results, and also limits further evaluation of the method's generalization capability.

  • As a zero-shot optimization method based on pre-trained generative priors, ZSBI-FM requires iterative optimization for each test sample. Compared to data-driven models that can complete inference in a single forward pass, this may result in higher inference latency. This computational cost—and its impact on efficiency in practical applications—warrants further discussion.

Questions for authors
  • Current experiments are primarily based on synthetic degradations, and industrial anomaly localization is also conducted using simulated BTAD data. Could the authors further discuss the applicability of this method to scenarios involving real-world image degradations, as well as the potential limitations that might arise when actual degradations do not conform to the assumption of structured sparse residuals?

  • How sensitive is ZSBI-FM to the SCAD hyperparameters, such as λ\lambda and aa, across corruption regions with different textures and scales? Does the method require manual parameter tuning for different corruption types?

6 comments
AuthorsOpenReview ↗

Response to Reviewer jCMW - I

Response to W1: Qualitative reconstruction comparisons

We appreciate the suggestion! During the rebuttal, we added representative visual comparisons across different corruption types to the main text as well as Appendix D.

Overall, we find the comparison of the visual results is consistent with the comparison of the quantitative metrics. Also, compared to other methods, our method tends to produce better visual reconstruction inside the mask region. We will add more detailed analysis of the visual results in our revision.

Response to W2: Computational cost

We agree that zero-shot methods, which our ZSBI-FM belongs to, tend to be slower than supervised-learning-based methods due to the need for iterative optimization. But supervised-learning-based methods can easily suffer from robustness issues, as our cross-corruption-type evaluation clearly shows.

So, how to choose between the two families of methods ultimately depends on the quality-and-speed tradeoff the end user is willing to strike for their particular applications, and there is no universal winner here.

Response to Q1: Real World Applications

Yes, our current evaluation focuses on synthetic data, because the availability of the clean image allows us to calculate reference-based quantitative metrics. We appreciate the reviewer’s suggestion to stretch the evaluation into some real-world applications.

During the rebuttal, we evaluate ZSBI-FM on naturally degraded historical photographs and paintings and include the visual results in the newly added Fig. 10 in Appendix D.7. Unlike the synthetic corruption patterns considered in Sec. 4.1, these images exhibit naturally occurring scratches, cracks, and surface damages.

Our method largely mitigates these artifacts while almost preserving the underlying structures and appearance, providing encouraging qualitative evidence of its applicability beyond synthetic corruptions.

Group sparsity of corruption is a critical assumption for our method to work. Absent additional assumptions, our method might not work well for dense or global degradations, such as widespread color fading, illumination changes, or severe blur, which are intrinsically highly ill-posed. We will clarify this in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer jCMW - II

Response to Q2: Parameter Sensitivity

We use a set of global hyperparameters for all corruption types in our experiments. This might be suboptimal, and corruption-specific tuning likely can improve the performance further. But, even with the suboptimal global hyperparameter choice, our method already almost uniformly beats competing methods.

To further test the sensitivity of the cross-type performance to the global hyperparameter, we fix a=3.7a=3.7, as recommended by Fan and Li (2001), and sweep both the inner λpixel\lambda_{\mathrm{pixel}}, outer λgroup\lambda_{\mathrm{group}} in the SCAD loss, as well as the SLIC superpixel count (KK), using the same 100-image FFHQ subset as in our robustness evaluation (Sec. 4.2). We get the following results:

  • Superpixel grouping: Varying (KK) from 512 to 2048 results in a maximum PSNR difference of only 0.028 dB, suggesting that our method is relatively insensitive to SLIC granularity within the evaluated range.
  • Pixel-level SCAD: Varying λpixel\lambda_{\mathrm{pixel}} from 0.2 to 0.6 produces a maximum PSNR difference of only 0.077 dB, indicating stable reconstruction performance across these settings.
  • Group-level SCAD: Varying λgroup\lambda_{\mathrm{group}} from 0.1 to 0.4 produces a maximum PSNR difference of 0.238 dB. Although the group-level parameter has a somewhat larger influence, reconstruction performance remains relatively stable across the evaluated range.

The SSIM and LPIPS results show similarly small variations. Overall, these experiments suggest that the overall performance is reasonably robust to the choice of the global hyperparameters. The detailed results are reported in Appendix D.4, Table 21.

Superpixel count sweep

SuperpixelsPixel λGroup λPSNR ↑SSIM ↑LPIPS ↓
5120.40.131.1410.9336120.05697
10240.40.131.1270.9334720.05725
20480.40.131.1130.9334740.05719

Pixel-level λ sweep

SuperpixelsPixel λGroup λPSNR ↑SSIM ↑LPIPS ↓
5120.20.131.1620.9336370.05745
5120.30.131.1370.9336860.05730
5120.40.131.1410.9336120.05697
5120.50.131.1600.9333820.05726
5120.60.131.0850.9333070.05736

Group-level λ sweep

SuperpixelsPixel λGroup λPSNR ↑SSIM ↑LPIPS ↓
5120.40.131.1410.9336120.05697
5120.40.231.0390.9333870.05755
5120.40.330.9030.9328230.05766
5120.40.431.0540.9332130.05731

Thank you for the detailed response and the additional experimental analysis. I agree with the authors that the relationship between quality and speed involves a trade-off, with no absolute winner. However, I suggest including an analysis of this aspect in future versions—even if it only covers basic inference overhead—as it would make the paper more comprehensive. Furthermore, the additional visual comparisons undoubtedly enhance the credibility of the proposed method and significantly improve the overall quality of the paper.

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

This paper proposes a blind image inpainting framework based on a pre-trained flow matching model. Specifically, the authors model the corruption of the degraded image as structured sparse reconstruction instead of explicitly estimating the degradation mask. The proposed algorithm utilizes a latent optimization strategy with an unsupervised stopping criterion to mitigate the overfitting issue. Experiments on multiple datasets demonstrate the competitive performance of the proposed algorithm compared to the exsiting baselines.

Critical strengths
  1. The motivation for discarding an explicit degradation mask is reasonable and could be more practical in real-world applications.
  2. The reconstruction results, especially on the FFHQ dataset, demonstrated superior performance compared to the competing baselines.
Critical weaknesses
  1. The proposed method is largely built on FMPlug, with a nested SCAD loss to address the blind inpainting settings. Given that the novelty is limited, there is no direct comparison with FMPlug. Thus, it is unclear whether the performance comes from the proposed loss or directly from FMPlug itself.
  2. Another main issue is that no ablation study is presented in the paper. This paper proposed pixel-level SCAD, SLIC-based group-level SCAD, and another early stopping strategy. However, it is unclear how much contribution can be attributed to each component, making it hard to justify the overall contribution of the paper.
  3. The authors should justify more why making explicit mask optimization into residual optimization is better.
Questions for authors

Please refer to the weakness above. Besides, there should be a figure or algorithm describing the whole framework.

4 comments

Response to W1 & W3: Novelty and Comparison with FMPlug.

We appreciate the reviewer's concern about novelty. We would like to take the chance to clarify the novelty as follows.

  • FMPlug is a generic method for using pretrained foundation flow-matching models for solving inverse problems. In this sense, it can be applied to BI, which is also an inverse problem.

  • To apply FMPlug to any inverse problems, one needs to make a few critical problem-specific choices, including the forward model, the data-fidelity loss, and any extra regularization term if necessary.

  • Our careful choice of the forward model. The typical forward model used for BI is the mask-based forward model: y=(1−M)⊙x+M⊙c\mathbf{y} = (\mathbf{1} - \mathbf{M}) \odot \mathbf{x} + \mathbf{M} \odot \mathbf{c}, which is used in RGI; Yan, 2013 for example. If one applies FMPlug to this mask-based forward model, they will arrive at formulations very similar to RGI. We identified a couple of critical issues with the RGI formulation in Proposition 2.1 and its subsequent discussions. These limitations of RGI and the mask-based forward model have motivated us to focus on the mask-free forward model, i.e., the sparse corruption forward model instead: y=x+n=x+M⊙(c−x),wherex∈M\mathbf{y} = \mathbf{x} + \mathbf{n} = \mathbf{x} + \mathbf{M} \odot (\mathbf{c} - \mathbf{x}), \quad \text{where} \quad \mathbf{x} \in \mathcal{M}. To the best of our knowledge, we are the first to develop a BI algorithm based on this mask-free forward model.

  • Our careful integration of different ingredients toward a practical BI algorithm. Having decided to focus on the mask-free forward model, we (1) develop the superpixel-based double-SCAD loss to deal with practical corruptions that are often group-sparse; (2) incorporate an early-stopping strategy to counter the overfitting issue—without this, global optimization of our BI formulation would lead to bad reconstruction; RGI has entirely overlooked this issue.

To sum up, the performance gain comes from our judicious choice of forward model and critical algorithm ingredients, on top of FMPlug.

Our table below directly compares applying FMPlug to the mask-based forward model (RGI-FM) and our final algorithm (Group-Sparsity SCAD), controlling the FMPlug component. It is clear that our method also benefits substantially from other algorithm components.

CasePSNR ↑SSIM ↑LPIPS ↓
RGI-FM29.37 ± 2.690.919 ± 0.0170.070 ± 0.018
Group-Sparsity SCAD31.24 ± 2.800.933 ± 0.0170.057 ± 0.020

Response to W2: Ablation Study

Thank you for this suggestion! During the rebuttal, we have studied several variants of our algorithm to isolate the contributions of individual components:

  • RGI-FM is the baseline, with FMPlug applied directly to the mask-based forward model; Pixel-Level SCAD and Group-Sparsity SCAD to explore the benefit of group-sparsity loss (vs pixel-sparsity loss); For these methods, we run the algorithms until final convergence, so we expect overfitting to the corruption. Finally, the “+ES” versions also include the early stopping strategy. From the table below, it is clear that our sparse corruption, whether pixel-level or at the group level, and the early stopping strategy, all contribute to the final success. A caveat to interpreting the effect of early stropping: it does not improve the performance at the final convergence; instead, it picks up near-peak performance before the flow-matching model starts to overfit the corruption after a reasonable reconstruction of the clean image. | Method | PSNR (dB) ↑ | SSIM ↑ | LPIPS ↓ | DISTS ↓ | CLIP-IQA ↑ | MUSIQ ↑ | |:---|---:|---:|---:|---:|---:|---:| | RGI-FM | 22.79 | 0.890 | 0.152 | 0.672 | 0.139 | 69.26 | | Pixel-Level SCAD | 24.96 | 0.909 | 0.128 | 0.118 | 0.645 | 69.24 | | Pixel-Level SCAD + ES | 30.90 | 0.933 | 0.058 | 0.047 | 0.704 | 65.63 | | Group-Sparsity SCAD | 28.52 | 0.925 | 0.099 | 0.083 | 0.672 | 68.76| | Group-Sparsity SCAD + ES | 31.14 | 0.934 | 0.057 | 0.045 | 0.711 | 66.23 |

  • Parameter sensitivity: We also conduct sensitivity analyses for the SLIC superpixel count (K), pixel-level λ\lambda, and group-level λ\lambda. The maximum PSNR variations are 0.028 dB, 0.077 dB, and 0.238 dB, respectively, across the tested ranges (see Table 6). These results indicate relatively stable reconstruction performance under moderate parameter variations.

    Superpixel count sweep

    SuperpixelsPixel λ\lambdaGroup λ\lambdaPSNR ↑SSIM ↑LPIPS ↓
    5120.40.131.1410.9336120.05697
    10240.40.131.1270.9334720.05725
    20480.40.131.1130.9334740.05719

    Pixel-level λ\lambda sweep

    SuperpixelsPixel λ\lambdaGroup λ\lambdaPSNR ↑SSIM ↑LPIPS ↓
    5120.20.131.1620.9336370.05745
    5120.30.131.1370.9336860.05730
    5120.40.131.1410.9336120.05697
    5120.50.131.1600.9333820.05726
    5120.60.131.0850.9333070.05736

    Group-level λ\lambda sweep

    SuperpixelsPixel λ\lambdaGroup λ\lambdaPSNR ↑SSIM ↑LPIPS ↓
    5120.40.131.1410.9336120.05697
    5120.40.231.0390.9333870.05755
    5120.40.330.9030.9328230.05766
    5120.40.431.0540.933213*0.05731

Response to Q1: Algorithm description

We appreciate the suggestion! We have added Algorithm 1 in Appendix E of the revised manuscript to provide a clear description of the proposed ZSBI-FM method.

Thanks for the additional experiments and clarification. The author only partially addressed my concerns. My main concern still lies with the limited novelty of the proposed framework based on FM-Plug. The authors justify that they propose a task-specific forward process modeling using masking as a residual. However, I do not get a very direct intuition from this. I still tend to keep my rating.

We appreciate the reviewer's follow-up. We would like to emphasize that FMPlug is a general inverse-problem-solving framework, rather than an existing solution to blind inpainting. Unlike conventional inpainting, blind inpainting involves unknown corruption locations, which cannot be handled by directly applying the standard known-mask data-fidelity formulation.

A natural extension of FMPlug to blind inpainting is to replace the GAN inversion component in RGI with FMPlug while keeping its explicit-mask formulation. This application work yields our RGI-FM baseline. However, as discussed in Proposition 2.1 and the subsequent analysis, the explicit-mask formulation presents important challenges. Our approach instead addresses the unknown-corruption problem through a residual-based objective with nested pixel-level and group-sparsity SCAD penalties, avoiding explicit mask optimization. We further incorporate early stopping to mitigate corruption fitting during optimization.

Therefore, our methodological contribution goes beyond simply applying FMPlug to blind inpainting: we develop a blind-inpainting-specific optimization formulation and algorithmic components for handling unknown, spatially structured corruptions. Our comparison with RGI-FM, using the same pretrained flow-matching model, demonstrates the effectiveness of these methodological choices.

Comments on the paper

1 comment
Area_Chair_PNAgOpenReview ↗

Please review the authors’ responses and provide feedback

Dear Reviewers,

The authors have submitted their responses. Please review them carefully, assess whether they address your concerns, and leave a follow-up comment for the authors.

Thank you for your time and thoughtful feedback.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.