- 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
- 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.
- 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.
- 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
- 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.
- 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.
- 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.
- 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
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 and , under the idealized assumption that learns the true image manifold perfectly, i.e., for sparse .
-
However, in practice, 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 and motivate our mask-free formulation. RGI’s Theorems 1 & 2 prove the asymptomatic equivalence of their formulations to the -based mask-free formulation: , 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 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.
Response to reviewer Pqqv - I
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 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.
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.
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.