acceptodds
Under review as a conference paper at ICLR 2027

Fine-Grained Image Tampering Localization via Semantic Proposals and Adaptive Residual Refinement

Abstract

Proactive image protection enables subsequent tampering to induce observable responses in pretrained vision models, yet these responses are often too coarse and spatially inconsistent to support fine-grained localization. We investigate this observability-to-localization gap in a proactive protection setting, where a protected reference and its manipulated counterpart are available. We propose SPARR (Semantic Proposal and Adaptive Residual Refinement), a response-interpretation framework that converts manipulation sensitive responses into accurate tampering masks while keeping the upstream protection mechanism unchanged. SPARR first exploits paired response variations under shared spatial prompts to derive robust semantic proposals, and then uses paired RGB residuals to refine response-induced regions toward the actual manipulated structures. To account for sample-dependent reliability, SPARR further introduces an adaptive routing strategy that evaluates whether additional refinement is supported by ground-truth-free evidence and selectively accepts or rejects it. This design enables SPARR to improve localization without relying on a fixed refinement behavior across different manipulation patterns. We evaluate SPARR under controlled same-protection comparisons, five conventional image tampering benchmarks, and four generative-AI editing settings. The results show that reliable downstream response interpretation can substantially improve fine-grained localization without modifying the upstream protection mechanism, while maintaining strong performance across diverse editing settings. The code will be made publicly available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Tampering Localization, Multi-Prompt Response, Fine-Grained Localization

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_XzwsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the gap between manipulation observability and fine-grained localization in proactive image protection. Instead of modifying the upstream protection mechanism, the authors propose SPARR, a downstream response-interpretation framework that converts manipulation-sensitive responses into pixel-level tampering masks. The method is evaluated on five conventional tampering benchmarks and four generative-AI editing settings under a fixed upstream protection mechanism. The experiments show substantial improvements over the original Blank Canvas localization outputs on several conventional benchmarks, and the progressive ablation indicates that both PRSR and especially SARD improve F1, IoU, and boundary F1.

Critical strengths
  • The paper identifies and clearly formulates a meaningful observability-to-localization gap in proactive image protection.

Existing proactive approaches may successfully expose manipulation-sensitive responses without necessarily producing spatially accurate tampering masks.The paper cleanly separates these two objectives and shows that substantial localization gains can be obtained purely through stronger downstream interpretation while keeping the upstream protection mechanism fixed.

  • The progressive design is well motivated and empirically effective.
Critical weaknesses
  • The contribution of the SARD acceptance mechanism is not sufficiently isolated from candidate generation and candidate selection.

This raises the question of whether most of the gain is actually due to selecting among Morph/BAC/Wide candidates rather than the proposed accept-or-fallback decision itself.

  • The overall pipeline contains many manually specified design choices and thresholds, but their robustness is not sufficiently characterized.
Questions for authors

Q1. Could the authors disentangle the effect of candidate selection from the effect of the SARD acceptance/fallback gate?

Q2. How sensitive is SPARR to the manually specified refinement hyperparameters and thresholds?

In particular, could the authors report sensitivity to the PRSR boundary width / percentile thresholds / area-change threshold and the SARD acceptance thresholds?

Q3. Please clarify the training and transfer protocol of the learned AMSP and SARD selectors across datasets.

2 comments
AuthorsOpenReview ↗

Part1/2—Response to Reviewer Xzws

Dear Reviewer Xzws,

Thank you for your careful review of Submission 31996 and for recognizing the distinction between manipulation observability and fine-grained localization. We appreciate your suggestions about the SARD gate, parameter sensitivity, and selector training.

W1 / Q1. Separating candidate generation, selection, and gate acceptance

We appreciate the suggestion to separate the effect of ranking from that of the acceptance rule. SARD generates Morph, BAC, and Wide candidates from the PRSR mask, ranks them by predicted sample-specific utility, and applies an acceptance rule that can fall back to PRSR. To isolate selection and acceptance, we held the pool fixed at 18,863 candidate records spanning 24 types on the 1,010 held-out IMD2020 images.

Selection policyMacro F1Macro IoUBoundary F1
PRSR baseline0.76300.66580.3013
Random selection, mean of 100 seeds0.78770.70010.3402
Best retrospective fixed policy0.83460.76260.4408
Ridge ranking without the gate0.84020.77400.5324
Full SARD with the frozen gate0.84010.77390.5306
Ground-truth F1 oracle0.85070.78590.5521

The fixed policy was selected retrospectively using test F1. The oracle uses ground truth to estimate the F1 selection ceiling. Adding the frozen gate changes mean F1 from 0.840198 to 0.840147, and the gate accepts refinements on 991 of 1,010 images. These results suggest that sample-wise ranking accounts for most of the mean improvement in this experiment. The acceptance rule provides a configurable fallback to PRSR.

Changes in the revised manuscript. Section 4.3 (p. 9) summarizes the ranking and gate results. Appendix F, Table 11 (p. 18), presents the comparisons while holding the candidate pool fixed, including the rounded gate-off and gate-on F1 values of 0.8402 and 0.8401.

AuthorsOpenReview ↗

Part2/2—Response to Reviewer Xzws

W2 / Q2. Sensitivity to PRSR and SARD settings

Thank you for identifying the settings that warranted closer examination. The one-factor-at-a-time PRSR study uses 1,010 images and 12 settings, including the frozen reference. That reference uses a 5-pixel boundary band, lower and upper residual quantiles of 0.20 and 0.80, and a maximum relative mask-area change of 0.12. Each row below varies only the parameter shown.

PRSR settingMacro F1Boundary F1
Frozen reference0.763150.30194
Boundary width 2 px0.755480.29853
Boundary width 3 px0.759820.30058
Boundary width 7 px0.762370.30527
Boundary width 9 px0.760930.31246
Lower quantile 0.100.761240.28964
Lower quantile 0.300.763980.31768
Upper quantile 0.700.757710.28389
Upper quantile 0.900.757330.32328
Area-change limit 0.060.755470.30869
Area-change limit 0.180.765480.29985
Area-change limit 0.240.766700.29548

An upper quantile of 0.90 lowers region F1 and increases Boundary F1, while relaxing the area-change limit to 0.24 has the opposite effect. We retained the frozen settings throughout the held-out evaluation.

For SARD, the frozen thresholds are 0.002 for predicted utility gain and −0.003 for predicted F1 gain. The utility gain threshold is 0.010 for Wide candidates. Varying only the predicted F1 gain threshold gives the following results.

Predicted F1 gain thresholdAccepted / 1,010Macro F1
−0.0101,0000.840183
−0.003 (frozen)9910.840147
09770.840134
0.00159680.840028

The archived gate study covers 44 configurations, including the frozen reference, one-factor changes, and joint changes. The one-factor configurations give mean F1 between 0.839683 and 0.840198. The joint changes give mean F1 between 0.839821 and 0.840183. The four settings above show that acceptance counts can change while mean F1 remains close to the frozen result.

Changes in the revised manuscript. Appendix F (p. 18) reports the frozen PRSR setting, the effects of raising the upper quantile to 0.90 and the area-change limit to 0.24, and all four predicted F1 gain thresholds shown above. Appendix B (p. 14) documents the frozen acceptance thresholds, including the stricter utility threshold for Wide. The full 12-setting PRSR grid and the broader gate study are additional detail in this response.

W3 / Q3. Selector training and transfer across datasets

We appreciate the request for a clearer account of the training and transfer protocol. For IMD2020, Dev1000 is used to construct training candidates and fit the downstream predictors. AMSP uses ground-truth-based ranking supervision and validation-based LambdaMART early stopping with patience 40. SARD uses 19,515 development candidate records, 23 features, and a five-fold Ridge ensemble with α = 10 to predict utility and F1 gains.

The 1,010 held-out images are evaluated with frozen selectors and decision thresholds. Prediction uses image and candidate features, and ground truth is used for scoring. The full IMD2020 score combines 1,000 out-of-fold development predictions with the 1,010 held-out predictions. BC protection and SAM remain fixed in these analyses.

Changes in the revised manuscript. Section 4.1 (p. 7) now specifies the training targets, early stopping, Ridge fitting, and frozen inference on IMD2020. Appendix A.1–A.2, Tables 4–5 (p. 13), document the split and use of ground truth. Section 4.2 (p. 8) specifies frozen SPARR predictions for the 1,000-image LaMa evaluation.

The checkpoint and fitting details for each external benchmark remain open. A full account of cross-dataset transfer would identify the checkpoints and any target-dataset fitting or threshold changes for each benchmark.

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

This paper studies Image Forgery Localization (IFL) in proactive image protection scene. Upstream protection uses Blank Canvas (BC) to add perturbation to original images, which means when the protected image is edited, SAM will respond anomalously on that region. The authors claim that this tampering response is too coarse and inconsistent to serve as a pixel-accurate mask. They call it the observability-to-localization gap. Hence, the authors manage to leave the protection image and SAM unchanged, and only study the downstream interpretation part. The contributions can be listed as follow:

  1. Proposing the gap between observability and localization.
  2. This paper proposes a three-stage pipeline to address prompt-dependent ambiguity, sample-dependent refinement risk and semantic-structural mismatch, and does no harm to the upstream protection.
  3. Providing the controlled same-protection evidence and evaluations on five traditional manipulation datasets and four generative editing benchmarks.
Critical strengths
  1. The problem is well-defined. The pipeline is lightweight due to the separation of observable tampering from interpretable responses. It is easy to apply after any responses-producing protection mechanism.
  2. SPARR includes explicit fallback rules. PRSR reverts to the semantic proposal if the mask area changes by more than 12%. SARD accepts a refinement only when its predicted gain passes a fixed threshold. Otherwise, it keeps the PRSR result.
Critical weaknesses
  1. The generalization comparison may not fully isolate the effects of the downstream interpretation and the so-called controlled same-protection is doubtful. Since the baseline results in Table 1 and 2 are directly copied from the Blank Canvas paper, it was not re-test on the same pairs. It may introduce unfairness. Additionally, the Blank Canvas (BC) is a training-free method, while SPARR involves supervised LambdaMART and Ridge models. The latency and computational cost comparison is not discussed in the paper.
  2. The claim of "sample-adaptive" in SARD lacks sufficient evidence. In fact, SARD accepts a non-baseline refinement input for 991 out of 1010 samples, the gate is supposed to rarely reject. There should be comparison with using fixed, random or oracle candidate selection. Also, there are many hyperparameters in three stages like grid size, band width, quantiles etc., but it is not stated in the paper how were they selected. Also, there are no ablation studies on the important selections in SARD and AMSP stages.
Questions for authors
  1. What data were the LambdaMART and Ridge models trained on, and how were the IMD2020 1,000/1,010 split and early stopping done?
  2. How were I^P constructed for the conventional PS datasets like CASIA1? And please specify the definition and experimental settings of "basic Blank Canvas".
  3. Compared to Blank Canvas, why were the evaluation results on the datasets DSO and Lama omitted here? Please add these two datasets in Table 1 and 2. Suggestions:
  4. How would results change if other candidate selection ways are used in SARD stage? Please compare SARD with different candidate selection strategies like fixed, random and oracle candidate selection with threshold sensitivity.
  5. According to Blank Canvas, they deeply understand that the perturbation is supposed to be vulnerable and any degradation like JPEG compression will affect the localization performance. What is more, for this paper, I^T and I^P are suspected to be almost identical on untampered regions in PRSR and AMSP stages. While after distribution to social media, this assumption may fail. Please include robustness analysis on common attacks like resizing, gaussian blur, gaussian noise and JPEG compression. It would be better to include online social network (OSN) attack like the one studied in IF-OSN.
  6. Please add trivial baselines using |I^T − I^P|, and re-run Blank Canvas on the same 1,010 held-out samples.
3 comments
AuthorsOpenReview ↗

Part1/3—Response to Reviewer QauM

Dear Reviewer QauM,

Thank you for your careful review of Submission 31996 and for highlighting the role of SPARR's fallback rules. Your suggestions helped us clarify what the comparisons measure and identify where additional evaluation would be valuable.

W1. Comparison fairness, supervision, and computational cost

We agree that the published scores in Tables 1 and 2 of the submission provide task-level context. The same-input analyses in this response use fixed Blank Canvas (BC) protection and a fixed pretrained SAM. On IMD2020 Remaining1010, the progressive and fixed-morphology controls measure the downstream contribution on matched pairs. The same-pair rerun of the published BC localization pipeline is addressed in Q6.

AMSP's LambdaMART ranker and SARD's Ridge models use offline supervision. At inference, they use image and candidate features. For runtime, the measured PRSR CPU median is 29.42 ms/image (95th percentile 566.12 ms), and Ridge ranking takes approximately 0.054 ms/image after feature construction. An end-to-end comparison with BC would also need to measure SAM inference, candidate generation, and feature construction. Offline training cost would need to be reported separately.

Changes in the revised manuscript. Section 4.1 (p. 7) clarifies supervision and frozen inference. Section 4.2 (p. 8) identifies the published scores as task-level references. Section 4.3 (p. 9) and Appendix F, Table 10 (p. 18), present the matched-input controls. Appendix F also reports the measured component timings and the scope of those measurements.

W2. Sample adaptivity, ablations, and parameter choices

We appreciate the suggestion to separate candidate selection from gate acceptance. The frozen gate accepts refinements on 991 of 1,010 images and falls back to PRSR on 19. As shown in Q4, the learned ranker improves over random and fixed selection within the shared candidate pool. Adding the gate changes mean F1 from 0.840198 to 0.840147. This suggests that sample-wise candidate ranking accounts for most of the mean gain, with the gate providing a fallback.

The PRSR sensitivity analysis varies the boundary width, residual quantiles, and area-change limit individually. The frozen reference uses a 5-pixel band, residual quantiles of 0.20 and 0.80, and a 0.12 area-change limit. For example, raising the upper quantile to 0.90 changes F1 from 0.76315 to 0.75733 and Boundary F1 from 0.30194 to 0.32328. The gate results in Q4 further show how the acceptance threshold changes the number of refinements. AMSP uses a fixed 4 × 4 prompt grid. A grid-size sweep and an isolated ablation of the learned AMSP selector would complement the existing stage-wise comparison. We also appreciate the request to explain how the initial defaults were chosen. The current analyses quantify sensitivity around those defaults, while the original parameter-selection criteria need further clarification.

Changes in the revised manuscript. Section 4.3 (p. 9) adds the gate-on versus gate-off comparison and the 991 acceptance count. Appendix F, Table 11 (p. 18), gives the shared-pool comparison. The sensitivity paragraph on the same page gives the frozen PRSR values, selected one-factor perturbations, and the predicted F1 gain threshold sweep for SARD.

AuthorsOpenReview ↗

Part2/3—Response to Reviewer QauM

Q1. Training data, the 1,000/1,010 split, and early stopping

We use the 1,000-image IMD2020 development split (Dev1000) to construct candidates and fit the downstream predictors. Ground truth supplies the ranking targets for AMSP's LambdaMART model and the candidate-level F1 and utility gain targets for SARD's Ridge models. AMSP uses validation-based early stopping with patience 40. SARD uses 19,515 development candidate records, 23 features, Ridge regularization with α = 10, and a five-fold ensemble.

The 1,010 held-out images (Remaining1010) are evaluated with the selectors and thresholds frozen. In the frozen evaluation, ground truth is used for scoring. The full IMD2020 result combines 1,000 five-fold out-of-fold development predictions and 1,010 held-out predictions. The manuscript documents these split and validation roles. The sample assignment rule and fold-level validation details require a fuller account.

Changes in the revised manuscript. Section 4.1 (p. 7) now specifies the training targets, early-stopping patience, SARD training records and features, Ridge settings, and ground-truth use. Appendix A.1–A.2, Tables 4–5 (p. 13), document the split, out-of-fold predictions, and when ground truth is used.

Q2. Construction of the protected reference and the definition of basic BC

Thank you for asking about the paired-image construction. In the controlled hard-composite protocol, protection is first applied to the reference to obtain IPI^P. We then insert the edited pixels under mask MM, giving IT=(1−M)⊙IP+M⊙IeditI^T = (1-M)\odot I^P + M\odot I^{\mathrm{edit}}. This construction preserves the protected reference pixels outside the edit mask and helps explain why direct RGB differencing can perform well on these pairs. CASIA1-specific reference selection and pairing require a separate description, which the general equation alone does not provide.

The request for a “basic BC” comparison concerns the original BC localization pipeline on the same pairs. Our controlled comparisons use BC protection and SAM as fixed upstream components, with AMSP as the first SPARR stage. Published BC scores are reported as task-level references. Q6 explains the status of the same-pair BC rerun.

Changes in the revised manuscript. Section 4.3 (p. 9) describes the outside-mask pixel preservation in the aligned hard-composite controls and records that a faithful same-pair BC localization rerun remains unverified. CASIA1-specific reference selection and pairing are still open in the manuscript.

Q3. Additional evaluation on DSO and LaMa

Thank you for the suggestion. The revised manuscript adds results for DSO in Table 1 and LaMa in Table 2 (p. 8).

Q4. Fixed, random, oracle, and threshold controls for SARD

We evaluated the selection policies on the same 1,010 held-out images using a shared pool of 18,863 candidate records spanning 24 candidate types. The fixed policy was chosen retrospectively using test F1 and serves as a descriptive reference. The oracle uses ground truth to estimate the selection ceiling in F1.

Selection policyMacro F1Macro IoUBoundary F1
PRSR baseline0.76300.66580.3013
Random selection, mean of 100 seeds0.78770.70010.3402
Best retrospective fixed policy0.83460.76260.4408
Ridge ranking without the gate0.84020.77400.5324
Full SARD with the frozen gate0.84010.77390.5306
Ground-truth F1 oracle0.85070.78590.5521

The frozen predicted F1 gain threshold is −0.003. Varying this threshold alone changes how often the gate accepts a refinement.

Predicted F1 gain thresholdAccepted / 1,010Macro F1
−0.0101,0000.840183
−0.003 (frozen)9910.840147
09770.840134
0.00159680.840028

Changes in the revised manuscript. Appendix F, Table 11 (p. 18), reports the six shared-pool policies. The threshold sweep is reported in the parameter-sensitivity paragraph on the same page, and Section 4.3 (p. 9) summarizes the ranking and acceptance findings.

AuthorsOpenReview ↗

Part3/3—Response to Reviewer QauM

Q5. Robustness to degradations and online social-network processing

We agree that JPEG compression, resizing, Gaussian blur and noise, and online social-network processing matter for practical use. These transformations can change SAM responses, disturb alignment, and create residuals outside the manipulated region. The current results are for aligned protected-reference pairs. Evaluating each stage across degradation types and strengths remains future work.

Change in the revised manuscript. Section 5 (p. 10) adds the alignment limitation and discusses how these degradations can affect SAM responses and introduce residuals outside the manipulated region. It identifies stage-by-stage measurements as future work.

Q6. Direct RGB baselines and a same-pair BC rerun

Thank you for suggesting a direct baseline. We evaluated pixelwise RGB differencing with per-image Otsu thresholding on the 1,010 held-out pairs. This baseline requires no training and chooses the threshold from each image pair. It achieves macro F1 0.7445, IoU 0.6141, and Boundary F1 0.4364. Full SPARR reaches 0.8401, 0.7739, and 0.5306 on the same comparison. Alignment and the preservation of untampered pixels make this hard-composite setting favorable to direct RGB differencing.

Figure 11 in Appendix F.1 of the revised manuscript shows four post-hoc examples where RGB+Otsu misses parts of the tampered region. In sample 1710, SPARR increases F1 from 0.3517 to 0.8878 and reduces false negatives from 81,374 to 1,236 pixels, with additional false positives visible in the overlay. The Otsu masks were reconstructed from the archived evaluation procedure and checked against the recorded F1 scores. The four cases illustrate a specific failure mode of automatic thresholding, and the aggregate comparison above gives the dataset-level results.

We appreciate the request to rerun BC on the same 1,010 pairs. The public materials left several settings required for a faithful reproduction of the complete localization pipeline unverified. We therefore report the published BC scores as task-level context and present the matched-pair stage and RGB analyses as the controlled comparisons completed here.

Changes in the revised manuscript. Section 4.3 (p. 9) explains the RGB control results and states the status of the BC rerun. Appendix F, Table 10 (p. 18), reports Macro F1 and Boundary F1 for the RGB and morphology controls. The IoU figures above are additional detail in this response. Appendix F.1, Figure 11 (p. 19), gives the four qualitative examples.

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

This paper studies fine-grained image tampering localization under a proactive protection setting, where both a protected reference image and its manipulated counterpart are available. Building on the perception-based protection method Blank Canvas, which suppresses the normal responses of a pretrained SAM so that subsequent manipulation exposes anomalous responses, the authors argue that such exposed responses are too coarse and spatially inconsistent to serve directly as pixel-accurate masks. They refer to this as the "observability-to-localization gap". To address it, they propose SPARR, a three-stage downstream response-interpretation framework that keeps the upstream protection mechanism and SAM fixed.

Critical strengths

The paper explicitly distinguishes making tampering observable from converting the exposed evidence into pixel-accurate localization, and decomposes the latter into three failure modes: prompt-dependent ambiguity, semantic–structural mismatch, and sample-dependent refinement risk. The three stages map directly onto these failure modes, which makes the overall design easy to follow.

Detailed analysis and honest discussion of limitations. The appendix provides per-sample statistics (e.g., how many samples improve or degrade after PRSR and SARD), a breakdown of SARD decisions by candidate family, module-wise qualitative examples, and representative failure cases. The authors openly acknowledge that refinement-risk estimation is not sample-wise monotonic, with 99 held-out samples degraded relative to PRSR.

Critical weaknesses

1、The paper addresses a narrow setting. The upstream proactive protection mechanism (Blank Canvas) and the pretrained SAM are kept fixed, both the protected reference and the manipulated image are assumed available at test time, and the task is reduced to post-hoc refinement of responses that already indicate the tampered regions. Under these assumptions, the core localization capability comes from the existing protection mechanism, and SPARR mainly converts its coarse outputs into finer masks. The three stages themselves are largely combinations of established techniques: median fusion of multi-prompt SAM responses with LambdaMART-based subset ranking, RGB-residual-guided boundary-band correction, and Ridge-regression-based selection among morphological and boundary refinement candidates. The work therefore reads more as an engineering post-processing pipeline tailored to one specific upstream method than as a general methodological contribution. The authors should clarify what is conceptually new beyond refining Blank Canvas outputs, and whether SPARR is applicable to other proactive protection schemes (e.g., EditGuard, OmniGuard, MaLP). Otherwise, the significance of the contribution for the ICLR community appears limited.

2、The claimed "observability-to-localization gap" is not quantitatively established. The paper's central motivation is that the responses exposed by proactive protection are "too coarse and spatially inconsistent" for fine-grained localization. However, this gap is never measured. The paper does not report, on the same set of protected images, how far the original Blank Canvas output is from the ground-truth masks under the same metrics, nor how this gap varies across datasets and manipulation types. In fact, Table 2 shows that Blank Canvas already achieves F1 ≈ 0.96–0.97 on all four generative-AI editing settings, and SPARR improves this by only about 0.001–0.010 (and is slightly lower on ControlNet, 0.972 vs. 0.973). This suggests that, at least in these settings, the gap is marginal. The authors should provide a controlled, quantitative characterization of the gap to demonstrate that the gap is substantial and widespread enough to justify a dedicated research problem.

3、SPARR is a downstream refinement module: it operates on the responses produced by a fixed upstream protection method (Blank Canvas) and additionally exploits the paired reference image. However, Tables 1 and 2 compare it with passive forensic methods (e.g., ManTraNet, MVSS-Net, CAT-Net, PSCC-Net, FakeShield) and standalone proactive methods (e.g., EditGuard, OmniGuard). These are complete, self-contained localization methods that do not depend on the outputs of any other method. Such baselines are not appro priate competitors for a refinement approach.The resulting comparison is confounded: SPARR's reported performance mixes the localization capability inherited from the upstream protection mechanism, the additional information from the clean reference image, and the contribution of the refinement itself. It is therefore impossible to tell from Tables 1 and 2 how much of the improvement is actually attributable to SPARR. The authors themselves acknowledge that passive methods serve only as "task-level references rather than strictly identical-input competitors", which further shows that these comparisons cannot support the claimed superiority. Since the paper's stated contribution is downstream response interpretation, the meaningful comparison is against other refinement or post-processing methods applied to the same Blank Canvas outputs, for example: direct thresholding of the raw responses; CRF or morphological post-processing; SAM-based or other learned mask refinement methods.

4、The experiments do not evaluate the robustness of the proposed method under common degradations such as JPEG compression, resizing, Gaussian noise, and blurring. This is concerning because PRSR relies directly on the pixel-wise RGB residual. Any global degradation applied to the manipulated image (e.g., a single JPEG re-compression or resizing during social-media distribution) would produce non-zero residuals over the entire image, which could severely corrupt the residual-guided correction and the residual-based features used by SARD. The authors should report results under a range of degradation types and strengths, and analyze how each stage (AMSP, PRSR, SARD) is affected. Without this, the practical contribution of the method cannot be properly assessed.

Questions for authors

see the above sections

2 comments
AuthorsOpenReview ↗

Part1/2 — Response to Reviewer HuBt

Dear Reviewer HuBt,

Thank you for your careful review of Submission 31996. We appreciate your recognition of the distinction between making tampering observable and localizing it accurately, as well as your comments on the sample-level failure analysis.

W1. Conceptual contribution and applicability beyond Blank Canvas

We appreciate the opportunity to clarify the contribution. SPARR studies a downstream question in the paired proactive setting: how can manipulation-sensitive responses be converted into pixel-level masks when prompt ambiguity, fragmented regions, or boundary errors remain? Each stage addresses a distinct failure mode. AMSP ranks and fuses prompt responses to improve proposal consistency. PRSR uses local residual evidence to correct spatial mismatches and applies an area-change safeguard. SARD ranks candidate refinements by predicted utility for each image and can retain the PRSR mask when the acceptance criteria are unmet.

In the shared-candidate-pool ablation, Ridge ranking without the gate achieves F1 0.8402, compared with 0.7877 for random selection averaged over 100 seeds and 0.8346 for the fixed policy selected retrospectively by test F1. Full SARD with the frozen gate reaches 0.8401. These results suggest that sample-wise ranking accounts for most of the observed gain, while the gate provides a fallback when predicted utility is low.

Blank Canvas (BC) and the pretrained SAM remain fixed in our controlled experiments. This setting lets us measure the downstream contribution on the same aligned protected-reference and edited-image pairs. We appreciate the suggestion to examine EditGuard, OmniGuard, and MaLP. Transfer would require suitable spatial evidence and compatible paired inputs for each method. Its effectiveness remains to be tested.

The revised tables report OmniGuard as a standalone method. Applying the SPARR stages to spatial evidence from OmniGuard, EditGuard, or MaLP remains a separate experiment.

Changes in the revised manuscript. Section 4.3 (p. 9) explains how much of the gain comes from ranking within the shared candidate pool. Appendix F, Table 11 (p. 18), reports the fixed, random, learned, and oracle policies. Section 5 (p. 10) describes the dependence on the upstream responses and identifies evaluation under broader protection settings as future work.

W2. Quantifying the observability-to-localization gap

We agree that the magnitude of the gap should be quantified. On the same 1,010 held-out IMD2020 images, with the upstream protection and SAM fixed, the stage-wise evaluation gives the following macro scores.

StageMacro F1Macro IoUBoundary F1
AMSP0.72770.61360.3023
AMSP + PRSR0.76310.66600.3286
Full SPARR0.84020.77410.5577

From AMSP to full SPARR, F1 increases by 0.1125 and Boundary F1 by 0.2554. These values quantify the improvement from the initial AMSP proposal to the final mask. A faithful same-pair evaluation of the original BC localization pipeline remains unverified.

We appreciate the point about generative editing. The published task-level comparisons show small F1 differences in these settings. On ControlNet, SPARR reports 0.972 and BC reports 0.973. We therefore limit our conclusion to the downstream gains measured under the controlled IMD2020 protocol. The size of the gap across manipulation types remains an open question.

Changes in the revised manuscript. Section 4.2 (p. 8) now states the ControlNet scores explicitly and describes the published scores as task-level references. Section 4.3, Table 3 (p. 9), presents the progressive comparison, with the four-decimal means in Appendix C.1, Table 7 (p. 14). The introduction to Appendix F (p. 18) identifies the Boundary F1 values reported for each comparison.

AuthorsOpenReview ↗

Part2/2 — Response to Reviewer HuBt

W3. Same-input comparisons for downstream refinement

We agree that the published passive and standalone proactive scores provide task-level context. To evaluate downstream refinement on matched inputs, we compared fixed morphology and full SPARR on the same 1,010 held-out images in the morphology control study.

MethodMacro F1Boundary F1
AMSP0.72760.2863
AMSP + fixed morphology (5 × 5)0.73450.2803
PRSR0.76300.3013
PRSR + fixed morphology (5 × 5)0.76710.3080
Full SPARR0.84010.5306

In the paired RGB comparison, direct RGB differencing with per-image Otsu thresholding reaches macro F1 0.7445, and full SPARR reaches 0.8401. Alignment and the preservation of untampered pixels make this hard-composite setting favorable to direct RGB differencing. We appreciate the additional suggestions. Direct thresholding of raw BC responses, CRF, and SAM-based or other learned refinement remain useful same-input comparisons for future work.

Changes in the revised manuscript. Section 4.3 adds a matched-input controls paragraph that explains why aligned hard composites favor direct differencing and states the status of the same-pair BC rerun. The fixed morphology and RGB controls appear in Appendix F, Table 10 (p. 18). Figure 11 (p. 19) shows selected RGB+Otsu failure cases.

W4. Robustness to common degradations

We agree that robustness under JPEG compression, resizing, Gaussian noise, and blur matters for practical use. These transformations may change or misalign the SAM responses used by AMSP, introduce RGB residuals in untampered areas that affect PRSR, and alter the residual-based features used by SARD to rank and accept refinements. Similar issues may arise after social-media processing.

Our current experiments use aligned protected-reference pairs. We acknowledge the need for a controlled evaluation across degradation types and strengths, with results reported separately for AMSP, PRSR, and SARD. Such an evaluation would help establish which stages are most sensitive and under what conditions.

Change in the revised manuscript. Section 5 (p. 10) explicitly states the alignment assumption, explains how JPEG compression, resizing, blur, noise, and social-network processing can affect SAM responses and RGB residuals, and identifies stage-by-stage robustness measurement as future work.

Comments on the paper

1 comment

Dear reviewers,

As the discussion period approached the end, please take a moment to read the authors’ responses and share your thoughts on any remaining points of discussion. If the discussion has changed your assessment, please update your review accordingly.

Thank you for your time and contributions!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.