acceptodds
Under review as a conference paper at ICLR 2027

PatchRecover: Image-Centric Diffusion Memorization Detection via Local Counterfactual Patch Recovery

Abstract

Diffusion models can memorize parts of training images, posing privacy and copyright risks. Many existing detectors rely on global, prompt-conditioned signals, which can obscure local memorization and limit detection when only generated images are available. We propose *PatchRecover*, a framework that extracts local evidence through counterfactual patch recovery and can operate without the original prompt, seed, or denoising trajectory. The core idea is that memorized content should recover more readily after a local edit than non-memorized content. We test this by replacing a patch with a counterfactual alternative, regenerating the image, and measuring recovery toward the original content. We aggregate the strongest patch-level recovery margins to obtain an image-level detection score. To explain why memorized patches recover more strongly, we introduce local recovery regions and derive a reconstruction bound under a local contraction assumption. Trajectory measurements show that memorized patches enter these regions more often and recover more strongly than non-memorized patches. On Stable Diffusion v1.4 and v2-base benchmarks that include non-memorized outputs from memorization-associated prompts, PatchRecover achieves mean image-level AUCs of 0.9980 and 0.9968, respectively. When both memorized and non-memorized images come from memorization-associated prompts, PatchRecover improves mean AUC from the strongest baseline's 0.9374 to 0.9782.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion model memorization, Diffusion regeneration, Memorization detection

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_T6cfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PatchRecover, a memorization detection metric for text-to-image diffusion models. PatchRecover perturbes one local patch at a time and regenerates edited images with diffusion model. It measures how strongly each edited image returns toward the no-edit reference, then aggregates patch scores into an detection score. The method supports both prompt-conditioned and prompt-free settings. The authors report strong detection results on benchmarks built with Stable Diffusion v1.4 and v2-base, and analyze denoising trajectories to investigate local recovery.

Critical strengths

The paper provides a quite straightforward while effective way to flag memorization, compared to more sophisticated metrics proposed by previous works depending on attention, probability-landscape sharpness or score anisotropy.

Critical weaknesses
  1. One major weaknesses of this paper is its limited novelty, both on the research question it considers and method it proposes. As far as I know, memorization detection has been extensively studied, with existing methods already achieving near-saturated performance on established benchmarks, which are also used by this paper. future research would benefit more from investigating novel memorization scenarios or model architectures than from introducing another detection method for well-studied settings.
  2. The contribution of theorectical explanation. The theorectical results presented in Section 3.2 regarding local contraction region is intriguing, but, lacks clear connection to the proposed metric. The effectiveness of the metric is primarily supported by empirical observations, leaving the theorectical results isolated. For example, how are ϵi\epsilon_i and λt\lambda_t determined and how do they guide the design of the metric?
Questions for authors
  1. I would suggest the authors justify the unique challenges in memorization detection and the novelty of this work.
  2. The authors should clarify the contribution and necessity of theoretical analysis on the methodolody.
4 comments

We thank the reviewer for recognizing our method as simple yet effective compared with more complex prior metrics, and for finding the recovery-region analysis intriguing. We address each concern below.

W1 / Q1. The novelty is limited. Memorization detection is well studied, and existing methods are near-saturated on established benchmarks. Please justify the unique challenges and the novelty of this work.

We agree that existing methods achieve near-saturated performance on standard benchmarks (Table 2, AUC > 0.98). However, these benchmarks overlook a more practical challenge: detecting memorization in individual images rather than their generation prompts. Our novelty lies in addressing this overlooked image-level setting. We identify this gap, reveal the limitations of existing methods in this setting, and propose an effective image-level detection approach based on local counterfactual recovery.

  1. A more practical and overlooked problem. Existing methods mainly focus on prompt-level memorization, but a memorized prompt does not always produce a memorized image. Moreover, the original generation seeds are often not available in real-world cases, making existing prompt- and seed-based methods less practical.
  2. A more challenging benchmark revealing the limitations of existing methods. We introduce a larger prompt-image mismatch benchmark (112 SD1 and 182 SD2 prompts), including non-memorized images generated from memorized prompts. The table below summarizes TPR@5%FPR results from Tables 14 and 20, where "Best baseline" refers to the best of 12 variants across six baselines. Despite their strong performance on standard benchmarks, existing methods struggle in this setting, with AUC dropping to near chance (0.40–0.62) when the original seed is unavailable (see pyGY, W4). In contrast, PatchRecover directly evaluates the candidate image and achieves substantially better performance.
Method (what it needs)SD1SD2
PatchRecover-PF (image)0.540.96
PatchRecover (image + prompt)0.330.98
Best baseline (prompt + seed)0.440.52
Best baseline with DDIM inversion (image + prompt)0.100.17
  1. Local memorization evidence without optimization. PatchRecover edits individual patches and measures their recovery toward the original content, providing local evidence of memorization without per-image optimization. It is also approximately 120× faster than InvMM.

We have clarified the novelty of our work and contributions in the revised Introduction. We hope they address the concern about novelty.

W2 / Q2. The theoretical results in Sec. 3.2 lack a clear connection to the metric. How are ρ\rho and ϵ\epsilon determined, and how do they guide the design of the metric? Please clarify the contribution and necessity of the analysis.

How ρ and ε are determined. Both parameters are introduced only for theoretical analysis and are not used to define or compute the detection metric. Specifically, ρ\rho defines the recovery region and is estimated from recorded trajectories, while ϵ\epsilon represents the residual recovery error for each patch. Their purpose is to explain the local recovery behavior underlying PatchRecover, rather than directly determine its metric.

How the analysis guides the design. The analysis explains why memorized patches are more likely to recover their original content and motivates four design choices, supported by ablation studies below:

  1. Regeneration strength. Sufficient denoising is needed for recovery, but excessive noise makes recovery harder. Memorized patches enter the recovery region with about 16 steps left, non-memorized ones with about 8 (Table 23). Too much noise, however, starts the trajectory far from the region. Accordingly, AUC on SD1 is 0.74 at strength 0.1, 0.999 at 0.5 and 0.956 at 0.9 (Table 10).
  2. Top-k aggregation. Since memorization may occur only in certain patches, we focus on patches with the strongest recovery signals. Using all patches reduces SD1 AUC from 0.9990 to 0.9592 (Table 9).
  3. Recovery margin. The margin dedit−drefd_{\text{edit}}-d_{\text{ref}} measures recovery toward the original content relative to the edited content. It achieves higher TPR@1%FPR (0.64) than using either distance alone (0.41 and 0.05), while reconstruction without editing performs poorly (AUC 0.59, Table 22).
  4. Meaningful patch edits. Recovery requires a noticeable change to the original content. Gray, Neighbor, Flip, and Blur edits work well, while JPEG compression, which largely preserves the content, performs poorly (Table 8).

Overall, the analysis explains and motivates our metric design, supported by the ablation results. We clarify this in the revised Sec. 3.2.

We hope these clarifications adequately address your concerns and welcome any further questions.

Dear Reviewer,

We’re nearing the end of the discussion period. This is a friendly reminder to leave a comment for the authors.

Thank you

~ AC

Thank you again for your review, and thanks to the AC for the reminder. In our response, we clarified the challenge this work targets: existing detectors are near-saturated when they use the prompt and its generation seed, but they fall close to chance when they receive only the image. We also explained how the recovery analysis motivates the design choices of the metric.

We would appreciate it if you could let us know whether this addresses your concerns. If so, we hope you might consider updating your assessment. With the discussion period closing soon, we are happy to clarify anything further in the time left.

Reviewer_QRapOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PatchRecover, a method for detecting memorization in diffusion-generated images through local counterfactual patch recovery. It replaces image patches with alternative content and measures their recovery after regeneration. The resulting patch-level scores are aggregated for image-level detection. The authors also introduce local recovery regions and derive a reconstruction bound under a contraction assumption. Experiments on Stable Diffusion v1.4 and v2-base demonstrate competitive detection performance. The paper further evaluates a prompt-free variant and examines memorized and non-memorized images generated from memorization-associated prompts.

Critical strengths

1.The paper considers local memorization rather than relying only on a single image-level signal. The idea of editing individual patches and measuring their recovery is reasonable, especially when memorized content may appear only in certain parts of an image.

2.The prompt-image mismatch experiments in Section 4.2 are a useful part of the evaluation. By including non-memorized images generated from memorization-associated prompts, the experiments help examine whether the detector responds to the actual image content rather than only to the prompt.

3.The experimental section includes comparisons with several existing detectors and ablations on different editing operations, distance measures, and patch aggregation strategies. The prompt-free variant is also worth investigating because it does not require access to the original generation prompt.

Critical weaknesses

1.The use of reconstruction behavior as a memorization signal has some similarities to existing methods. For example, Jiang et al.(2025), Image-Level Memorization Detection via Inversion-Based Inference Perturbation, and Ma et al.(2025), An Inversion-Based Measure of Memorization for Diffusion Models, both investigate memorization using inversion-related signals. PatchRecover differs by applying local counterfactual edits, but the paper could better explain what additional information the proposed patch recovery margin provides compared with these approaches.

2.The theoretical analysis relies on the local contraction assumption in Eq.(4), but the empirical experiments do not directly establish that this condition holds during denoising. The trajectory analysis shows that memorized patches tend to recover more strongly, but this is not necessarily equivalent to verifying the contraction property. If the authors could provide more direct measurements of the contraction behavior, the theoretical explanation would be more convincing.

3.In Section 4.4, the patch-level evaluation assigns the image-level label to every patch. This may introduce considerable label noise because not all regions of a memorized image are necessarily memorized. Therefore, the patch-level results do not fully demonstrate whether the method can accurately locate memorized regions.

Questions for authors

1.Have the authors examined whether the actual denoising trajectories satisfy the local contraction condition in Eq.(4)? In particular, does the patch distance consistently decrease after entering a recovery region?

2.Have the authors checked whether the highly reproducible images used as memorized samples have close matches in the training dataset? This could help distinguish training-data memorization from generation consistency.

5 comments

We thank the reviewer for recognizing our patch-level approach, the mismatch evaluation, and the thorough comparisons and ablations. The constructive comments led to a direct test of the contraction assumption and a clearer comparison with inversion-based methods. We address each concern below.

W1. Using reconstruction behavior as a memorization signal is similar to IIP and InvMM. What additional information does the patch recovery margin provide?

The key difference is that IIP and InvMM measure image-level reconstruction behavior through per-image optimization, while PatchRecover measures local recovery from counterfactual edits without optimization.

IIPInvMMPatchRecover
Per-image optimizationYes (a prompt embedding, after DDIM inversion)Yes (noise distribution and soft prompt, up to 2,000 steps)No
EvidenceOne score per imageOne score per imageOne score per patch
What is measuredResponse to the optimized perturbationKL to the prior needed to reproduce the imageWhether an edited patch returns to its original content

To quantify the benefit of the recovery margin, we compared it with its individual components (dissimilarity to the edited and reference patches) and reconstruction without editing (Table 22 in the revised paper, SD1 mismatch test, A18 vs. B, AUC / TPR@1%FPR). All scores use the fixed Gray/L2 configuration with top-2 aggregation:

ScorePatchRecoverPatchRecover-PF
Margin d_edit − d_ref0.9680 / 0.64430.8458 / 0.4403
d_edit only0.9552 / 0.40780.8842 / 0.3050
−d_ref only0.8196 / 0.04720.7369 / 0.0000
Reconstruction without an edit0.5908 / 0.02590.6113 / 0.0528

The results show that: (1) Counterfactual edits provide the key signal. Reconstruction without editing performs poorly (AUC ≈ 0.6), whereas the recovery margin achieves substantially higher AUC. (2) The margin improves low-FPR detection. It achieves the highest TPR@1%FPR for both variants, outperforming either component alone.

Compared with InvMM on the matched 72-image SD1 subset (App. C.3), PatchRecover also achieves higher AUC on A18 vs. B (0.9946 vs. 0.7531; Table 6) with approximately 120× lower runtime (18.4 s vs. 2,253 s per image; Table 21). Since IIP's code is unavailable, we compare it conceptually. We clarify these differences in the revised Sec. 2.

W2 / Q1. The analysis relies on the local contraction assumption (Eq. 4), but the experiments do not show directly. Do the actual trajectories satisfy it? Does the patch distance consistently decrease after entering a recovery region?

We directly evaluated Eq. 4 on 92,032 denoising trajectories (16 patches × 5,752 images, 25 steps each) using deterministic PNDM sampling. Table 23 of the revised paper reports the results (M = patches of memorized images, N = patches of non-memorized images):

SD1 MSD1 NSD2 MSD2 N
Enter the recovery region78.5%10.6%73.0%12.2%
Steps left at entry (median)168157
Steps in the region that satisfy the fitted Eq. 461.6%26.3%68.1%3.1%
Final distance below the distance at entry86.7%95.7%98.8%95.9%
Final-to-entry distance (median)0.6550.9170.4810.925
Distance never increases after entry33.5%40.2%57.2%37.5%

Does the distance decrease after entry? Yes. Memorized patches enter the recovery region more frequently and recover more strongly. On SD1, 78.5% of memorized patches enter the region, compared with 10.6% of non-memorized patches. On SD2, the rates are 73.0% and 12.2%. After entry, the median final-to-entry distance ratios are 0.655 (SD1) and 0.481 (SD2) for memorized patches, substantially lower than 0.917 (SD1) and 0.925 (SD2) for non-memorized patches. Moreover, 86.7% and 98.8% of memorized patches finish closer to their original content.

Does the distance decrease strictly at every step? Not always. Eq. 4 holds for 61.6% (SD1) and 68.1% (SD2) of in-region steps for memorized patches. Importantly, Proposition 1 bounds the expected final distance under stepwise contraction in expectation, rather than requiring every individual trajectory to decrease monotonically. Thus, temporary distance increases, such as the late-stage rebound on SD1 (Fig. 2(a)), do not necessarily violate the assumption.

Overall, the experiments support the predicted net recovery behavior, while showing that the local contraction assumption is an approximation rather than a universal property. We clarify the conditional nature of Proposition 1 in Sec. 3.2 and report these measurements in App. G.6.

W3. In Sec. 4.4, every patch receives the label of its image. This is noisy, because not all regions of a memorized image are memorized, so the patch-level results do not show that the method can locate memorized regions.

We agree. Our method focuses on patch-level recovery behavior rather than memorized-region localization, and we do not claim localization accuracy. Table 3 evaluates patch-wise separation using image-level labels, which we have clarified in the revised Secs. 3.1 and 4.4.

To further validate the recovery signal, we conducted an additional experiment using ground-truth copied-region masks from Webster (2023), also used by Kim & Lee (2026), covering 40 memorized images (28 from SD1 and 12 from SD2). We classified patches as copied (≥50% mask overlap) or uncopied (≤10% overlap) and measured their recovery rates (Table 24 in the revised paper). Of the 40 images, 32 contain both patch types, allowing direct comparisons.

Entry rateSD1SD2
copied87%81%
uncopied39%4%

Using ground-truth copied-region labels reduces the noise introduced by image-inherited patch labels and yields clearer separation (Table 24). This suggests that the original patch-level results are conservative rather than inflated, as the recovery signal becomes stronger with more accurate region-level labels. Copied patches also enter more often than uncopied ones in every image that contains both (32 of 40 images), and Fig. 8 shows qualitative examples. Thus the patch-level recovery signal concentrates in copied regions, which suggests that PatchRecover can also provide localization evidence, although localization was not our goal. Because this check is small, these results provide preliminary evidence for the patch recovery signal. A full localization evaluation remains further evaluation in the future.

Q2. Do the highly reproducible images used as memorized samples have close matches in the training data? This would distinguish training-data memorization from generation consistency.

Yes. For each memorized image, we computed SSCD to the training images released with Webster (2023) for its prompt and took the best match, using the prompts whose released training images were all available (Table 25 in the revised paper):

ModelPromptsImagesBest SSCD, mean ± SDSSCD ≥ 0.5
SD1233 of 4829320.760 ± 0.16890.2%
SD2188 of 2167520.695 ± 0.13590.6%

About 90% of the memorized images match a training image above the standard 0.5 threshold. Their reproducibility therefore reflects training-data memorization, not only generation consistency.

We hope these experiments and clarifications address your concerns, and we would be happy to answer further questions.

Reviewer_pyGYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PatchRecover, which is a memorization detection technique for text-to-image diffusion models. The method does not necessarily require prompts for its detection, unlike most prior work. They do it by patching part of an image with a counterfactual, for instance flip, blur, etc. and then regenerating both the edited image and the unedited image with img2img. Finally, the score depends on the difference of the edited patch and unedited reference. Experiments on SDv1 and v2 show strong performance.

Post-rebuttal:

The rebuttal addresses my main concerns. In particular, the additional experimental evidence and the clearer presentation of the results make the proposed approach a solid contribution. Having reviewed the other comments, I see that the authors have addressed several concerns. For this reason, I am raising my score to 'weak accept'. The final version should keep all new additions.

Critical strengths
  • The paper is well written and well motivated. Specifically, framing the memorization problem as image-level and not prompt-level is intuitive and practical.
  • The method can run without the prompt (PatchRecover-PF) and without per-image optimisation, unlike prior detectors. It only needs an img2img pipeline and plausibly extends to other latent diffusion models. The prompt-free variant is, however, weaker than most baselines (0.9712 AUC on SD1, Table 2).
  • The method outperforms strong baselines in memorization detection. Even if the gains are sometimes small, they seem to be somewhat consistent.
  • The ablations such as steps, strength, grid size, query budget are well-thought and thorough.
Critical weaknesses
  • One of the main weaknesses of this paper is that the prompt–image mismatch experiment (Table 1) rests on very little data. For example, the group B contains 18 prompts on SDv1 and only 3 prompts on SDv2. Thus, effectively, the key claim is SD1-only. Another example would be that with 72 negatives, TPR@1%FPR means zero false positives and is extremely noisy.
  • The method proposed requires about 17 img2img passes with 25 steps, resulting in about 850 UNet passes, which is far more than prior prompt-based detectors. The paper does not compare the wall clock time with other approaches.
  • The introduction mentions localisation of memorization but in the patch-level experiment , patch labels are taken from the image. Thus, Table 3 measures patch-wise separation and not localization against region-level ground truth. A localization metric against annotated masks (for example, local-memorization annotations in the style of Chen et al., 2025) would strengthen the claims.
  • The baselines score the prompt with a query seed and never see the candidate image (Sec. 4.2 and Table 5 in App. C.2). They can therefore separate A from B only through differences between the prompts, which favours the proposed method in the mismatch test. A fairer comparison would use image-only adaptations of the baselines, for example running them on a DDIM-inverted latent of the image.
  • The edit and distance are selected on the evaluated image cohorts, which the paper itself calls retrospective (Sec. 4.1). The chosen configurations differ by model and variant (App. C.1), and each is the best cell of the sweep in Table 8. The patch-level configuration is also selected on the 20 evaluation draws (App. E). Each baseline, in contrast, runs one fixed configuration, and the choice matters. JPEG reverses the score ordering.
Questions for authors
  • Does the method transfer across models (an image generated by model A, scored with model B) or to newer architectures?
  • What is the computational cost of the approach, and how does it compare with previous methods?
  • How does the method perform in a localization experiment, such as in Chen et al. ?
  • Can you report results with the edit, distance and aggregation selected on prompts disjoint from all evaluated cohorts?
8 comments

We thank the reviewer for recognizing our image-level framing, the practicality of our method, its consistent gains and our thorough ablations. Below, we address each concern.

W1. The mismatch experiment (Table 1) rests on little data. Group B has 18 prompts on SD1 and 3 on SD2. TPR@1%FPR with 72 negatives is noisy.

We agree. To address this problem, we built a larger mismatch benchmark on both models. Group B (non-memorized images from memorized prompts) now has 112 prompts on SD1 and 182 on SD2. For each memorized prompt, we selected one generation that meets two criteria: (1) its SSCD to each of the prompt’s four memorized generations (Group A, produced with seeds that reproduce training data) is below 0.5, the standard copy threshold; and (2) its SSCD to every released training image of the prompt is also below 0.5, so it does not copy any training image. No detector output was used in the selection. This procedure enlarges the benchmark while keeping the selection independent of every detector, so the comparison is fair.

Instead of TPR@1%FPR, we now report TPR at 5% and 10% FPR, which are more reliable metrics at this sample size. The results are in Table 14 of the revised paper (A vs B, AUC / TPR@5%FPR / TPR@10%FPR):

MethodSD1SD2
Ren et al.0.645 / 0.16 / 0.170.772 / 0.45 / 0.57
Wen et al.0.545 / 0.15 / 0.210.211 / 0.10 / 0.11
Jeon et al.0.712 / 0.01 / 0.380.659 / 0.52 / 0.54
Chen et al.0.566 / 0.15 / 0.200.258 / 0.09 / 0.13
Asthana & Belagiannis0.766 / 0.06 / 0.580.330 / 0.18 / 0.21
Kim & Lee0.834 / 0.08 / 0.440.094 / 0.01 / 0.02
PatchRecover0.806 / 0.33 / 0.460.985 / 0.98 / 0.98
PatchRecover-PF0.881 / 0.54 / 0.610.963 / 0.96 / 0.96

PatchRecover is the only method that performs well on both models. On SD2, both variants lead by a wide margin on all three metrics, while 4/6 baselines fall below 0.5 AUC. SD1 is now harder for all methods than in Table 1, because many of its Group-B images lie close to the copy threshold, where most methods lose accuracy (Table 15). Even so, PatchRecover-PF is best on all three metrics, and PatchRecover is slightly below Kim & Lee in AUC (0.806 vs 0.834) but well above it at 5% FPR (0.33 vs 0.08). Overall, each baseline performs poorly on at least one model, whereas both PatchRecover variants stay above 0.80 AUC on both. These results confirm that PatchRecover’s advantage holds across models.

W2 / Q2. The paper does not compare wall-clock time. What is the cost compared with previous methods?

We add the cost comparison with previous methods in Table 21 (App. G.4), which summarizes the UNet passes and the average wall-clock time per image on a single RTX 3090:

MethodUNet passesSD1 s/imageSD2 s/image
Asthana & Belagiannis40.100.10
Jeon et al.4 (+ 2 JVP)1.991.83
Wen et al., Chen et al., Ren et al.1002.29–2.362.14–2.16
Kim & Lee100 (+ 16 VJP)3.313.00
The six baselines with DDIM inversion (see W4)104–2006.3–11.15.6–10.4
PatchRecover / -PF, 4 regions2605.65.2
PatchRecover / -PF, 16 regions (paper)88418.417.1
InvMM (SD1)47,833 on average2,253—

PatchRecover’s cost is adjustable through the number of queried patches. With 16 patches, it is slower than the other baselines but about 120× faster than InvMM. With 4 patches, it runs faster than every inverted baseline in wall-clock time, while AUC stays nearly unchanged (0.971 on SD1 and 0.992 on SD2, vs 0.999 and 0.997 with 16 patches; Table 12). PatchRecover is therefore simple, needing only an img2img pipeline, flexible in cost, and effective even at a low budget.

W3 / Q3. Table 3 inherits patch labels from the image, so it measures patch-wise separation, not localization. How does the method perform in a localization experiment, as in Chen et al.?

We agree, and we did not intend Table 3 as a localization result. The paper studies local edits and whether each patch recovers its original content, and it does not claim localization accuracy. App.Sec.E notes that assessing localization accuracy would require independent patch annotations. To avoid ambiguity, the revised Sec. 3.1 and Sec. 4.4 now describe Table 3 as patch-wise separation.

To answer Q3, we conducted an initial evaluation using region-level ground truth. Chen et al. (2025) provide only qualitative localization masks without ground-truth validation. We therefore used the copied-region masks released by Webster (2023), also used by Kim & Lee (2026), covering 28 memorized images from SD1 and 12 from SD2 (40 images in total).

We classified patches as copied (≥50% overlap with the ground-truth mask) or uncopied (≤10% overlap), and measured how often an edited patch enters the recovery region defined in Sec. 3.2, where its regenerated content closely matches the original. Results are reported in Table 24 of the revised paper.

Entry rateSD1SD2
copied87%81%
uncopied39%4%

Using ground-truth copied-region labels reduces the noise from image-inherited patch labels and reveals stronger separation (Table 24). This suggests that the original patch-level results are conservative rather than inflated. Copied patches also enter more often than uncopied ones in every image that contains both (32 of 40 images), and Fig. 8 shows qualitative examples. Thus the patch-level recovery signal concentrates in copied regions, which suggests that PatchRecover can also provide localization evidence, although localization was not our goal. Because this check is small, we present it as preliminary, and a full localization benchmark remains future work.

W4. The baselines never see the candidate image but only prompt and seed, which favours the proposed method. A fairer comparison would run them on a DDIM-inverted latent of the image.

Thank you for this suggestion. This leads us to a fairer comparison, which is now in the revised paper (Sec. 4.2 and App. G.3).

Experiment. We DDIM-inverted each candidate image (50 steps, with its prompt) and ran all six baselines from the inverted latent instead of the seed. Table 19 in the revised paper reports the SD1 mismatch test of Table 1 (A18 vs B, AUC / TPR@1%FPR):

BaselinePrompt + seed (as in the paper)Prompt + DDIM-inverted image
Ren et al.0.8577 / 0.35350.7671 / 0.1792
Wen et al.0.9170 / 0.43680.5204 / 0.0632
Jeon et al.0.8616 / 0.12360.4574 / 0.0000
Chen et al.0.9374 / 0.45970.5104 / 0.0215
Asthana & Belagiannis0.8922 / 0.27640.5149 / 0.0167
Kim & Lee0.9003 / 0.24510.3753 / 0.0125

PatchRecover reaches 0.9782 / 0.6495 on this test (Table 1). Our method does not require source seed or inversion.

With the inverted image, 5 out of 6 baselines fall to chance level or below. The strongest, Ren et al., drops to 0.77 AUC. The same holds on the larger mismatch benchmark, where no inverted baseline exceeds 0.57 AUC on SD1 or 0.62 on SD2 (see Table 20).

Why. In fact, our original experiment protocal favours the baselines. Each baseline receives the image’s generation seed and re-runs the exact trajectory that produced it. Without original seed, DDIM inversion only approximates that trajectory, and these detectors depend on it. However, in a real audit, the seed is not available. PatchRecover does not rely on this seed, as its regeneration seeds are independent of the original one. On the larger benchmark, PatchRecover's AUC remains nearly unchanged, while the baselines suffer substantial drops. These results further demonstrate the effectiveness of PatchRecover, whose advantage becomes even greater when baselines use the candidate image instead of the original seed.

Q1. Does the method transfer across models (an image generated by model A, scored with model B), or to newer architectures?

To evaluate cross-model transferability, we generated images using SD1.4 and scored them using SD1.4, SD1.5, and SD2. SD1.4 and SD1.5 share the same initialization and LAION training subset, while SD2 was trained on a different LAION subset. We report results from the first draw of the SD1 mismatch test in Table 26 of the revised paper (A18 vs. B, AUC / TPR@1%FPR):

MethodScored by SD1.4Scored by SD1.5Scored by SD2
Ren et al.0.8302 / 0.44440.7284 / 0.38890.5710 / 0.0000
Wen et al.0.8889 / 0.33330.8426 / 0.38890.6296 / 0.2778
Jeon et al.0.8086 / 0.22220.7685 / 0.61110.4630 / 0.0000
Chen et al.0.8920 / 0.27780.8179 / 0.44440.6327 / 0.3889
Asthana & Belagiannis0.7963 / 0.50000.7006 / 0.44440.5216 / 0.0556
Kim & Lee0.8642 / 0.16670.7747 / 0.05560.5772 / 0.0000
PatchRecover0.9743 / 0.55560.9929 / 0.93060.6630 / 0.1806
PatchRecover-PF0.8142 / 0.43060.8756 / 0.48610.6287 / 0.0972

The results show two key observations: (1). Strong transfer to SD1.5. PatchRecover achieves an AUC of 0.9929 and TPR@1%FPR of 0.9306, outperforming its SD1.4 results (0.9743 / 0.5556), while all six baselines show AUC drops of 0.04–0.10. (2). Limited transfer to SD2. All methods show performance drops, likely due to differences in training data.

These results suggest that cross-model transfer is stronger between models sharing training data, consistent with memorization being model-dependent. For newer architectures (e.g., SDXL and DiT), suitable memorization benchmarks remain scarce. We acknowledge this limitation in App. H.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for their careful and constructive feedback. In response, we have revised the paper and conducted additional experiments to address the key concerns. All changes are highlighted in blue, with the original table, figure, and section numbering preserved. New experimental results are provided in Appendix G (Tables 13–26 and Fig. 8). Below, we summarize the main findings, with detailed responses provided to each reviewer.

  • A larger mismatch benchmark on both models (pyGY, W1). We enlarged the benchmark from 18 to 112 prompts on SD1 and from 3 to 182 on SD2. PatchRecover reaches 0.985 AUC against 0.772 for the strongest baseline, and detects 98% of memorized images at 5% FPR against 52% on SD2. On SD1, which is more challenging for all methods, PatchRecover-PF achieves the best AUC (0.881). Our method is the most effective in these comparisons.
  • Adapting baselines for fair image-based comparisons (pyGY, W4). When the baselines see the image through DDIM inversion instead of the source seed, their performance drops substantially, most to near chance on the enlarged benchmark. Our method’s advantage grows even larger.
  • Configuration selection (pyGY, W5). We tested disjoint-prompt selection, a fixed configuration, and all 12 baseline variants. The results confirm PatchRecover’s robustness to configuration choices and its strong performance against the baselines.
  • Cross-model transfer (pyGY, Q1). Cross-model experiments show strong transfer between SD1.4 and SD1.5, which share training data (PatchRecover: 0.993 AUC), but substantial performance drops for all methods on SD2, which uses different training data. This suggests that transferability depends on shared training data.
  • Time cost (pyGY, W2). With 16 regions, PatchRecover is about 120× faster than InvMM but slower than the other baselines. With 4 regions, it cuts cost by 70%, making it faster than all the DDIM-inverted baselines, while keeping AUC at 0.97 (SD1) and 0.99 (SD2).
  • Localization (pyGY, W3 and QRap, W3). Table 3 measures patch-wise separation, and localization was not the focus of this work. We now state this more clearly. Even so, a preliminary check with ground-truth copied-region labels shows that the recovery signal concentrates in copied regions and separates more sharply than under image-inherited labels. This suggests that PatchRecover can also provide localization evidence.
  • A direct test of the contraction assumption (QRap, W2). We measured contraction directly on 92,032 trajectories using entry rate, per-step distance ratio and final-to-entry distance. 87% (SD1) and 99% (SD2) of memorized patches end closer to their original content after entering the recovery region, supporting the contraction hypothesis in Sec. 3.2.
  • Novelty and the role of the analysis (T6cf). We highlight the overlooked yet important challenge of image-level memorization detection, where existing methods struggle despite near-saturated benchmark performance. Our theoretical analysis explains the patch recovery mechanism and motivates the key design choices, supported by extensive ablation studies.

We believe these additional experiments and revisions substantially address the reviewers’ concerns and strengthen the paper. We hope the reviewers will reconsider their assessments in light of the new evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.