acceptodds
Under review as a conference paper at ICLR 2027

SPATIAL PROBING FOR LOCALIZED AI-EDITED IMAGE DETECTION

Abstract

Frozen vision-foundation-model (VFM) features with a linear probe detect fully generated images accurately, but fail on localized editing: real photographs in which an AI generated, inserted, or re-rendered part of the content. We show most of this failure is an aggregation artifact: the pooled feature is an attention weighted average over spatial content, so on an edited image it aligns more with the unedited majority (cosine 0.233 to the all-patch mean) than with the edited patches (0.100). Moving the probe to the spatial level, as multi-scale windowed scoring or per-patch token probing, recovers much of the signal with the backbone frozen and no added training. With the released MetaCLIP2 probe (we reproduce its published benchmark tables to within 0.001 average accuracy), windowed scoring reaches 0.937 AUC on FLUX-Kontext instruction-edited images, 0.850 AUC cross-domain on BR-Gen, and stays above 0.89 AUC under JPEG and blur; JPEG re-compression even improves detection. Across the AIGI-Now matrix, windowed wins all 9/9 semantic splits (sign-test p=0.002); pooled wins 8/9 pristine pixel splits. Across three benchmarks, max aggregation captures the strongest local evidence and mean aggregation the global statistics; the operative choice is spatial isolation, not feature granularity, since whole-image patch statistics dilute a localized edit while windowed cropping isolates it. The windowed probe also estimates the manipulation’s spatial scale via its argmax crop. The paradigm’s boundary remains honest: patch-token probes cannot localize small edits (IoU near 0), whose signal lives in native-resolution noise statistics that downsampling destroys.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-foundation-model (VFM),AI-generated images(AIGI),Localized editing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hox5OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses localized AI-edit detection, where only part of a real image is AI-modified. The authors show that existing MetaCLIP2-based probing struggles with such edits because global pooling dilutes local signals. They propose a training-free windowed probing method using multi-scale crops and maximum-score aggregation. They also investigate patch-level probing and aggregation strategies. Experiments across four benchmarks show improvements in some settings but inconsistent gains overall.

Critical strengths
  1. The paper addresses the practical challenge of detecting localized AI edits, where existing detectors often struggle because most of the image remains authentic. This is an increasingly relevant problem given the growing use of AI-based image editing tools.
  2. The proposed windowed probing method is simple and computationally efficient. It builds on an existing detector without requiring additional training or modifying the backbone, making it easy to reproduce and apply.
Critical weaknesses
  1. The results on localized edits don’t really support the main claim. On FLUX-Kontext the gain is small and not significant (0.928 to 0.937), and on BR-Gen windowed scoring is actually worse (0.850 to 0.788). The abstract and TL;DR also overstate things: 0.850 on BR-Gen is the pooled baseline, and patch probing needs training and is close to chance.
  2. The aggregation-geometry principle feels circular. The paper decides that evidence is “concentrated” because windowed wins, and then uses that as proof of the principle. On BR-Gen, where masks are available, pooled wins even for very small edits. Also, the dilution analysis uses patches the probe scores as fake, not the real edited regions, and real images show the same pattern (Table 16).
  3. The novelty is limited. The method is basically four center crops of an existing probe, similar to multi-crop test-time augmentation. Since crops are resized up, the gain might come from higher resolution and not from spatial isolation. Almost everything is also tested on one probe (MetaCLIP2), so it’s hard to call this a general principle.
Questions for authors
  1. Can spatial concentration be measured independently (e.g., from edit masks) to predict the better aggregation before seeing results? Why does pooled scoring win on BR-Gen even for edits under 5% of the image?
  2. What does the AIGI-Now semantic split contain, and does the windowed gain hold when measured with AUC instead of accuracy?
  3. How much of the gain comes from upsampling the crops rather than isolating the edit? For example, how does full-image evaluation at higher resolution compare?
Reviewer_hxjoOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper builds on a released detector by using a linear probe on frozen MetaCLIP2 features, trained on fully generated images. It asks why this probe struggles with localized AI edits, meaning real photos where only part of the image was generated or changed. The authors argue that the pooled image feature averages over the whole image, so a small edited region is diluted. They propose two “spatial” probing methods that keep the backbone frozen. First is the windowed probing, which can score center crops at four scales (0.4, 0.6, 0.8, and 1.0) with the same probe and take the maximum. No training is needed. Second is patch probing, which uses a per-patch linear head on patch tokens and distances to banks of real and fake patches. From these, they propose the "aggregation-geometry principle." The max aggregation suits concentrated evidence, while mean aggregation suits diffuse evidence. Following this, they have the results with the released probe on some of these generators, FLUX-Kontext windowed 0.937 AUC vs. pooled 0.928 (not significant). BR-Gen, which is around pooled 0.850 vs. windowed 0.788 (windowed significantly worse) and AIGI-Now windowed wins accuracy on 9/9 "semantic" cells and pooled wins 8/9 "pixel" cells. The argmax crop scale correlates moderately with edit size (ρ = 0.30). Frozen patch heads cannot localize edits (IoU ≈ 0), and edited regions are smoother in high-frequency noise. The authors also evaluate robustness under JPEG and blur, as well as trajectory truncation, re-editing, and tool-switching experiments.

Critical strengths
  • The authors reproduce the released probe’s published tables closely (Table 20). Every pooled-vs-windowed comparison then uses the same probe, so the differences come only from how scores are aggregated. They report paired significance tests (Table 17) and state results that go against their method: windowed is significantly worse on BR-Gen (p < 10⁻⁴), the FLUX-Kontext gain is not significant, localization fails (IoU ≈ 0), and VAE reconstruction is a blind spot (DDA-COCO).

  • Several side results are informative on their own: Frozen final-layer patch tokens cannot localize edits even with in-domain mask supervision (Table 4). Early layers carry more of the edit signal than late layers (Table 13). Edited regions have lower high-frequency energy, so classical residual detectors score below chance (Table 14). Switching the editing tool sharply lowers the probe’s score (App. L).

Critical weaknesses

1. BR-GEN in the abstract (and the conclusion) credits windowed scoring with "0.850 cross-domain on BR-Gen." That number is the pooled baseline. Windowed gets 0.788, which is significantly worse (Table 17). Table 14 and (Section 5.2) also list "Windowed frozen-VFM probe (ours)" at 0.850 on BR-Gen LaMa, but Table 7 gives windowed 0.779 there (pooled 0.834). The FLUX-Kontext which shows the gain as 0.937 vs. 0.928 and not significant (p = 0.52) as (Section 4.3) says so, but the abstract, introduction and conclusion present 0.937 as the signal the method "recovers." The abstract says AUC "stays above 0.89 under JPEG and blur." Table 9 shows 0.886 (FLUX-Kontext windowed, blur) and 0.787-0.788 on BR-Gen. (Section 4.6) says ">=0.91 windowed," which is also contradicted by 0.886. "Re-compression even helps" holds only on FLUX-Kontext. On BR-Gen, pooled AUC drops from 0.850 to 0.837 under JPEG80.

2. (Section 5.4) and App. K say JPEG/blur "purification" closes the detection window, which contradicts (Section 4.6) and contribution (v). The baselines shown in Table 3, UnivFD beats both pooled and windowed on BR-Gen (0.893 vs. 0.850 and 0.788) and on the AIGI-Now pixel split (0.951). The text says UnivFD is "overtaken on the localized-edit benchmarks where spatial isolation matters most." That leaves out BR-Gen, the only benchmark with real edit masks. The introduction also says CLIP linear probes "cluster near chance" on localized edits, while UnivFD reaches 0.906 and 0.893.

3. The evidence for the "aggregation artifact" and the aggregation-geometry principle does not isolate spatial localization, as the main statistical evidence comes from a split that is not localized editing. The 9/9 result (sign test p = 0.002) is on the AIGI-Now semantic split, which the paper calls "spatially localized anomalies" (Section 4.4) and "degraded edited content" (Fig. 3). The AIGI-Now paper describes this subset as real images paired with surreal, heavily post-processed generated images, which is semantic-level fully generated content. So this result says little about localized edits. It is accuracy at a fixed threshold, not AUC. App. J states that max-over-crops "adds a positive score offset to all images." On degraded fakes that the pooled probe under-scores, shifting every score upward raises fake accuracy without improving ranking. On the clean pixel split, the same shift costs real-image accuracy. That alone predicts the "mirror image" pattern, with no need for a spatial explanation. AUC per cell is needed. Zoom is a confound, because the backbone runs at 224 px, so a 0.4 crop resized to 224 px is a 2.5x zoom and the windowed gains may come from seeing finer detail, which the paper itself says downsampling destroys (Section 5.2), rather than from isolating the edit. There is no control: no same-size crops away from the edit, no crops that exclude the BR-Gen mask region, and no higher-resolution pooled input. The dilution measurement does not separate edited from real images. Table 16 reports cosine 0.233 vs. 0.100 for edited images. Real images show the same pattern with an even larger gap (0.198 vs. 0.056; gap +0.142 vs. +0.133). The text (Section 5.1, abstract) cites only the edited row. In addition, the "top-15% fake-scored patches" are not edit masks.

4. BR-Gen edits are not "small and scattered." because we see that BR-Gen is used as the example where mean should win because edits are small and scattered. But Table 18 shows 184 of 500 forged images have masks covering >=75% of the image and 276 cover >=50%, and pooled still wins every bin. Large edits are exactly where the principle predicts max should help. The center-only crops (Limitation iii) could explain BR-Gen better than "scattered" edits do. As stated, the principle fits whichever aggregator wins, since there is no independent measure of how concentrated the evidence is.

The following are some minor issues that i have noticed as well that the authors can clarify and modify in their menuscript.

  • BR-Gen windowed is 0.799 in Table 9 and 0.788 elsewhere.
  • Generator names are inconsistent or garbled ("flux-kera", "keling", "stable diffusion v.1.4", Fig. 3 axis labels), and "(4.5)" text missing section name.
Questions for authors

1. Are the real images in the FLUX-Kontext and BR-Gen tests the unedited sources of the edited images, or a separate pool? Table 3 says reals come from "AIGI-Now flux-dev pixel." If they are unpaired, differences in content, resolution or format could drive the scores.

2. Please report AUC, not only fixed-threshold accuracy, for all 18 AIGI-Now cells. Please also report windowed accuracy with a recalibrated threshold. Given the score offset in App. J, does the 9/9 semantic-split result hold?

3. What is the evidence that the AIGI-Now semantic split contains spatially localized anomalies? The AIGI-Now paper describes it as surreal, heavily post-processed generated images.

4. Can you control for the zoom effect? For example: same-size crops placed away from the edit, crops that exclude the BR-Gen mask region, or the pooled probe on higher-resolution input or tiled inputs.

5. Which numbers are correct for windowed on BR-Gen: 0.788 (Tables 2, 7, 12, 15), 0.799 (Table 9), or 0.850 (Table 14, Section 5.2)? Was the released probe trained on SD1.4 or on flux-dev?

6. Table 16 shows the same pooled-vs-patch gap for real images (+0.142) as for edited ones (+0.133). How does this support edit-specific dilution? Can you repeat the analysis with BR-Gen ground-truth masks?

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

This paper studies detection of localized AI-edited images: real photographs in which only part of the content has been generated, inserted, or re-rendered by AI. The authors argue that frozen vision foundation model (VFM) features with a linear probe fail on such images mainly because image-level pooling dilutes the local edit signal. The authors propose two spatial probes that keep the backbone frozen and need no additional training.

Critical strengths

The overall experiment results provide empirical evidence and metric improvements.

Critical weaknesses
  1. The conclusions are close to self-evident. Windowed scoring is essentially multi-crop test-time augmentation with max aggregation.
  2. At test time one cannot know whether an edit is concentrated or diffuse, and the paper offers no adaptive selection or fusion mechanism. So the principle can only explain results after the fact; it cannot guide actual detection. Deploying windowed scoring by default, for example, loses 6 points on BR-Gen.
  3. All crops are center crops, which implicitly assumes edits lie near the image center. The paper has no comparison with sliding-window or grid crops and no analysis stratified by edit location, so it cannot rule out that the gains of windowed scoring come from a center bias in the datasets rather than from the method isolating the edited region.
  4. Edit-scale estimation, presented as a new capability, reaches only ρ = 0.302.
  5. All the issues above indicate this paper shows no novel insights and mininal technical contributions. So i recommand clear-reject recommandation.
Questions for authors

I have no further questions for the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.