acceptodds
Under review as a conference paper at ICLR 2027

Part-Level Self-Blending: Region Ground Truth by Construction for Deepfake Detection

Abstract

Deepfake detectors produce a single real-or-fake score and therefore cannot indicate which part of a face was manipulated. We show that the score itself misses such edits. Each of eighteen published detectors is at or near chance when a manipulation is confined to a small facial region and at chance on a single eye (worse-eye AUC 0.48–0.55). We attribute the gap to the training signal, since self-blending, the most common method of synthesizing training fakes, blends the whole face and therefore provides no information about where the manipulation lies. We introduce Part-Level Self-Blending (P-SBI), which confines each synthetic blend to one named facial region, namely a single part, such as the left eye, or a small group of parts, such as the mouth, drawn from a fixed set of twelve that together cover every part of the face. Because the region is chosen at synthesis time, its mask is exact, and the same construction both trains a per-region prediction head and measures the region sensitivity of any detector. The verdict and the localization map need no anatomical input at test time, and only the per-region scores are read on a parser’s regions. With a DINOv3-H+ backbone adapted on under one percent of its weights, our detector, PARTAWARE, acquires a region sensitivity none of the eighteen has (coverage 0.847 over three seeds, against 0.587–0.676 for all eighteen) at no cost to the properties the field already measures, reaching crossdataset transfer 0.980 AUC and robustness to twenty-one corruption conditions (0.960 in distribution and 0.931 zero-shot, each at or above the strongest published detector. Removing P-SBI from the same backbone isolates the effect. Over three seeds, transfer falls only to 0.969 while coverage falls from 0.847 to 0.657, back into the published band. The lead also holds on diffusion-inpainted edits from two independent generators, whose region truth does not come from our construction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
deepfake detection, face forgery detection, self-blended images, region-level supervision, region sensitivity, forgery localization, foundation model adaptation, explainability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_aEKsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Part-Level Self-Blending (P-SBI), which generates synthetic facial manipulations with known region masks, and PARTAWARE, a detector trained to identify manipulated facial regions and localize edits. Experiments across 18 existing detectors, 21 corruption types, and independent manipulation benchmarks show strong cross-dataset transfer, robustness, and region-level detection. However, attribution results are partly replicated by a parameter-free baseline, dense localization remains imprecise, and transfer to real-world lip-sync forgeries is limited.

Critical strengths
  1. P-SBI provides known region masks, enabling systematic evaluation of region-level detection.
  2. Comparisons, ablations, corruption tests, and external benchmarks reveal both improvements and generalization limits.
Critical weaknesses
  1. A simple statistic with no learned parameters achieves 0.706 top-1 attribution accuracy, raising questions about whether the evaluation truly measures learned localization. Dense localization is also limited, with an IoU of just 0.192.
  2. Despite strong mouth-region coverage on synthetic manipulations, the model struggles to generalize to real AVLips forgeries, limiting claims about its ability to detect manipulations from real-world generators.
Questions for authors
  1. Which results best demonstrate that the model can identify manipulated regions from the edited image alone, without relying on pristine-image restoration or parameter-free baselines?
  2. Do the main performance gains hold across different training seeds and evaluation splits, especially for results reported using only a single seed?
4 comments

We thank the reviewer for the detailed comments. Following the reviewer's suggestions, we ran each requested analysis, and all of them are now in the rebuttal and paper.

Weakness 1:

The new analyses show that the 0.706 statistic needs the pristine frame, while our part head names the region without it. The statistic's top-1 comes from the Shapley intervention, which restores each part from the pristine frame. Restoring an unedited part changes nothing, so the edited part receives the whole score change (Appendix L.2 calls this oracle assistance).

  1. With the pristine frame, ours names the region more often (Shapley top-1 0.857 against 0.706). With blur, mean face color or noise as fillers, both fall well below their pristine values (ours 0.23–0.39, the statistic 0.17–0.25, chance 0.226). The attribution result therefore depends on the pristine frame, and our evidence for naming from the edited image is the part head.
  2. From the edited image alone, the part head names the region, at 0.693/0.808/0.816 over three seeds against 0.27–0.35 for the same model trained without P-SBI, and a training-free floor from the edited image reaches 0.38.

Table L.1 adds these rows, Section 5.2 states that 0.706 needs the pristine frame, and the floor is renamed from “parameter-free” to “training-free”, since its edited-image version fits per-part clean-frame statistics.

On dense localization we agree. The map is coarse on every seed (IoU 0.192, 0.341 and 0.361), and Section 6 now states this limit together with a confidence rule for when a part-head name can be trusted, which reaches 0.90 top-1 on the half of blends it accepts.

Weakness 2:

We agree, and the paper now scopes P-SBI accordingly. It targets local spatial edits, and mouth coverage on synthetic blends does not transfer to real lipsync forgeries. Our seeds and controls show that P-SBI brings no gain on lipsync, where the control is higher on AVLips (by 0.01 to 0.04) and on DeepSpeak in every seed (Appendix K). We think lipsync leaves its evidence in mouth motion and in generation artifacts, which a spatial blend does not model. Appendix K now treats lipsync as outside the scope of P-SBI and adds the seeds and controls. Section 6 and the Conclusion state that P-SBI targets local spatial edits, while detection of lipsync and whole-face forgeries comes from the rest of the training pipeline.

Question 1:

The part head is the direct evidence, since it receives only the edited image in one forward pass.

  • Blends: top-1 0.693/0.808/0.816 against 0.27–0.35 for the same model without P-SBI, chance 0.226 and a constant guess 0.331, at left/right accuracy 0.92–0.95.
  • No pristine parse: with the part mask parsed from the edited image, 0.674/0.791/0.819 on 2168 blends, against 0.699/0.809/0.821 with the pristine parse on the same blends.
  • Unseen manipulation: on diffusion inpainting with each face re-detected and parsed on the edited image, 0.64–0.75 over seeds and generators (the same range as with the pristine parse), against 0.34–0.54 for the controls, scored with the pristine parse.
  • When confident: 0.90 held out on the half of blends the confidence rule accepts (reported seed).

These are in Table L.1 and Appendices L.4 and L.5.

Question 2:

We agree with the reviewer. The region gains hold on every seed, and the paper also states plainly where P-SBI gives no gain. New Appendix F.6 (Table F.6) repeats coverage, paired shift, part-head naming, localization IoU, diffusion inpainting and the in-the-wild benchmarks with three seeds and three no-P-SBI controls. The Shapley and occlusion readouts, the confidence rule and the case figures use the reported seed.

  • Every seed beats its control on paired shift (+1.81 to +2.00 SD against +0.71 to +1.19), part-head naming (0.69–0.82 against 0.27–0.35), localization IoU (0.19–0.36 against 0.10–0.11), and coverage and naming on both inpainting generators.
  • Three-seed means in the main text: new Table 3 reports each ability as the three-seed mean with its range (coverage 0.847, naming 0.772, IoU 0.298), and mean transfer AUC is 0.980, 0.982 and 0.980 on the three seeds.
  • Splits: intervals resample source videos (videos on Deepfake-Eval-2024), the confidence rule is checked on 200 held-out video-group splits, and leave-one-benchmark-out selection picks the reported model in all four folds.
  • WildDeepFake: P-SBI beats its control on every seed, by +0.011, +0.034 and +0.020. The seed mean (+0.022 [0.002,0.042][0.002, 0.042]) and the +0.034 seed are significant.
  • Scope: P-SBI targets local spatial edits, so it brings no gain on lipsync and on the in-the-wild Deepfake-Eval-2024 set (reply to W2, Appendix F.6).

In sum, region sensitivity and naming hold on every seed and on an unseen manipulation, and transfer is on par with the strongest published detectors. The limits of P-SBI are now stated in the main text.

AuthorsOpenReview ↗

Thank you, and a brief follow-up

We thank the reviewer again for the careful review. To briefly summarize our reply, from the edited image alone the part head names the edited region at 0.69–0.82 over three seeds, about twice the 0.27–0.35 reached without P-SBI, and every seed beats its control. Since the discussion period ends soon, we would be grateful to hear whether our reply addresses the reviewer's concerns, and we would be glad to clarify any remaining point.

Thanks for the response. My concerns are addressed. I will increase my score to 3 (Weak accept).

We thank the reviewer for taking the time to review our comments and we appreciate their response about increasing our score.

Reviewer_B9SsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary
  1. This paper investigates the ability of Deepfake detectors to perceive and localize localized facial manipulations. The authors first propose Part-Level Self-Blending (P-SBI), which generates training samples with explicit regional labels and pixel-level masks by restricting self-blending operations to predefined facial regions. Based on this, PARTAWARE is constructed, which utilizes the DINOv3 backbone network with efficient parameter fine-tuning to jointly achieve image-level detection, region classification, and pixel-level localization.

  2. Furthermore, the paper introduces a region-sensitivity evaluation protocol based on Region-AUC. This protocol is used to benchmark 18 publicly available detectors. The effectiveness of the proposed method is validated through cross-dataset testing, robustness tests against image degradations, component ablation studies, and experiments involving independent generation models.

Critical strengths
  1. This paper evaluates 18 publicly available detectors using a unified Region-AUC protocol, revealing the shortcomings of existing methods in detecting facial manipulations that cover only small areas. In particular, the analysis of regions such as monocular vision provides valuable evidence for understanding the local perception limitations of Deepfake detectors.
  2. The authors conducted controlled experiments involving fixed backbone networks, whole-face self-blending, the removal of P-SBI, and the use of multiple random seeds. The results demonstrate that P-SBI effectively enhances regional sensitivity. Furthermore, tests using independent diffusion-based inpainting confirm, to some extent, that this performance gain is not solely dependent on the original self-blending operation.
Critical weaknesses
  1. Section 3.3 defines Region-AUC as measuring the response of image-level detection scores to different types of tampering, rather than whether the model can accurately predict the location of the tampering. Although the authors have designed additional region classification and pixel localization branches, Appendix L shows that the Top-1 accuracy of region classification is only 0.693, and the [email protected] of dense localization is only 0.192. Therefore, the current results mainly prove that P-SBI can enhance the detection sensitivity to local tampering, but the support for accurate identification and localization of the tampered regions is relatively limited. It is recommended to clearly distinguish the three abilities of region sensitivity, region classification, and pixel localization, and make the main conclusion consistent with the actual verification capabilities.

  2. The core Region-AUC evaluation uses the same synthetic mechanism as the P-SBI training, while the other 18 detectors do not adopt this training strategy. Although the authors alleviated this problem through independent diffusion repair tests, the performance advantage has narrowed, and it cannot fully prove the broad generalization ability for real local tampering. Appendix J further shows that on Deepfake-Eval-2024, PARTAWARE's AUC is only 0.624, and the advantage over the strongest baseline does not reach statistical significance, and only about 40% of the videos can pass the preprocessing process. Meanwhile, Appendix F.3 indicates that four cross-dataset benchmarks participated in the backbone network selection, making the main cross-dataset results not considered as completely independent final tests.

Questions for authors
  1. Region-AUC primarily measures the detectability of local manipulations rather than the accuracy of their localization. Could the authors further elaborate on the relationship between this metric, region-classification Top-1 accuracy, and pixel-level localization IoU? Additionally, to what extent do the current results support the method's capability for regional localization?

  2. Experiments involving independent diffusion-based inpainting have demonstrated that certain performance gains can generalize across different synthesis mechanisms. Please further explain the differences between your results and the performance on the real Deepfake-Eval-2024 and Lipsync datasets, and how these differences affect the applicability of your method.

  3. Appendix F.3 mentions four cross-dataset benchmarks used for model selection. Could the authors specify which architectural choices or hyperparameter settings were determined based on these benchmarks, and clarify which experiments can be considered as truly independent evaluations of generalization?

1 comment

We thank the reviewer for the detailed comments. Following the reviewer's suggestions, we ran each requested analysis, and all of them are now in the paper. For clarity, Section 3.3 now defines the three abilities, and Section 6 states the limits of P-SBI.

Weakness 1:

We agree, and as the reviewer asked, the main text now reports the three abilities separately in new Table 3 (Section 5.2), per seed in Appendix L.4.

AbilityMetricPartAware (3 seeds)Without P-SBI (3 seeds)Chance
Sensitivitycoverage0.795/0.877/0.8680.682/0.641/0.6490.50
Namingpart-head top-10.693/0.808/0.8160.354/0.279/0.2650.226
Pixels[email protected]0.192/0.341/0.3610.111/0.103/0.1020.122 (random region)

P-SBI adds region sensitivity and naming on every seed, while pixel localization stays coarse, and the Conclusion now claims exactly this. Naming also holds without the two large recipes (0.64–0.78 against 0.13–0.23, chance 0.107). Chance is now computed from the recipe frequencies (0.226) instead of 1/12. Always naming the right eye reaches 0.331, and P-SBI stays well above both.

Weakness 2:

We agree. Selection is answered under Q3.

Independent manipulation. On Kandinsky inpainting, a manipulation never seen in training, every seed beats its control in coverage (0.762–0.789 against 0.624–0.658) and naming (0.639–0.704 against 0.358–0.542), and SDXL agrees (Table F.6). As the reviewer notes, the margin over the best published detector is smaller than on blends (0.068 against 0.119), yet it holds on every seed. The paper also checks for shortcuts from our blend synthesis: a model trained without Poisson compositing still detects Poisson blends (0.741, against 0.720 with it), so the gain is not from compositing, and at JPEG quality 75 PartAware still leads all 18 detectors (0.751 against at most 0.637).

Deepfake-Eval-2024. Only 40% of the videos were evaluated because of a face-tracking setting. The tracker counted gaps between linked detections in source frames, not sampled frames, so faces in long videos were not linked. It now counts sampled frames. 1850 of 2036 videos (91% of real and of fake) are now evaluated, and every detector is rescored on every benchmark. Our three-seed mean is 0.633, significantly above 14 of the 18 published detectors and not significantly different from Effort, GenD CLIP-L and GenD PE-L. Unlocking-CLIP is 0.019 higher. The Table 1 mean and margin are unchanged (0.9797, 0.008 above ForensicsAdapter), and PartAware still leads Cross-AUC (0.982).

Question 1:

Coverage measures how well edited frames rank above clean frames, and the paired analysis measures the score change against each frame's own clean version. Naming asks which region was edited, and IoU which pixels. The controls show that naming does not follow from sensitivity. The same model trained without P-SBI, whose part head learns from FaceForensics++ difference masks, still detects edits (coverage 0.64–0.68), yet names the region at 0.27–0.35, the level of the constant guess. What P-SBI adds is naming from the edited image alone. Naming reaches 0.69–0.82 (0.90 on the half of blends a confidence rule accepts), while pixel localization stays coarse (IoU 0.19–0.36).

Question 2:

The two results differ because the edits differ, which marks the scope of P-SBI. Inpainting is a local spatial edit, so the gain transfers. Deepfake-Eval-2024 mixes in-the-wild lipsync, face swap and diffusion generation. Our controls show no gain from P-SBI there, and on AVLips the control is 0.01 to 0.04 higher (Appendices F.6 and K). WildDeepFake supports this reading: its face swaps are blended into the frame, and P-SBI gains on every seed (+0.022 [0.002,0.042][0.002, 0.042] over its control). P-SBI therefore applies to local face edits and to naming the edited region, and detection of whole-face and temporal forgeries comes from the rest of the training pipeline.

Question 3:

Appendix F.7 groups each choice by the data that set it.

  • Set before training. The residual cutoff (0.12), the P-SBI probability (0.5) and both loss weights (0.5) were fixed in advance.
  • Chosen on the four transfer benchmarks. Backbone family and scale (Appendix F.3), and LoRA rank 32 over 16, judged on transfer and coverage (Appendix F.4). Region coverage on FaceForensics++ also served to reject P-SBI variants, and AVLips and FakeAVCeleb were used to judge a cross-frame mouth variant.
  • Outside model selection. WildDeepFake (face padding set without detector scores), Deepfake-Eval-2024, both inpainting sets and DeepSpeak. Leave-one-benchmark-out selection picks the reported model in all four folds (details in Appendix F.7 and in our reply to Reviewer nRfY, Weakness 2).

Region sensitivity and naming hold on every seed and on unseen inpainting, and transfer is on par with the strongest published detectors.

Reviewer_nRfYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Part-Level Self-Blending (P-SBI), which confines synthetic manipulations to named facial regions and records supervision masks during synthesis. It also proposes a region-coverage protocol and PARTAWARE, a detector with frame, part, and localization heads. Experiments identify poor sensitivity to small local edits across eighteen released detectors and show that P-SBI improves this capability while retaining strong conventional detection performance.

Critical strengths

The paper measures a useful capability that aggregate detection AUC can conceal. The region-level evaluation reveals substantial variation across facial parts, particularly near-chance performance on single-eye edits. Reporting worst-region performance and operating points alongside mean coverage makes the evaluation more informative.

The experimental controls are extensive. Same-backbone comparisons, whole-face blending controls, held-out compositing operators, recompression, and three training seeds support the contribution of region-confined training. Coverage improves from 0.657 without P-SBI to 0.847 with it, averaged over three seeds. Independent Kandinsky and SDXL inpainting evaluations further support transfer beyond the training synthesis.

The evaluation is broad and transparent. Released checkpoints receive identical crops and corrupted pixels, preprocessing effects are disclosed, and direct localization is distinguished from attribution requiring a pristine reference. Maintaining approximately 0.980 transfer AUC and 0.931 zero-shot corrupted AUC while adding region sensitivity is a useful empirical result.

Critical weaknesses

Interpreting coverage as whether a detector’s score “responds” to an edit is stronger than the pooled estimator establishes. Section 3.3 compares edited frames against all clean frames rather than their own pristine counterparts. Low coverage can therefore reflect large between-image score variation despite a consistent response to editing, as Appendix I discusses for ProDet and FSFM. A complementary paired score-change or paired win-rate analysis would distinguish intervention sensitivity from cross-image separability.

The four headline transfer benchmarks also guide backbone selection. Consequently, the 0.008 advantage over ForensicsAdapter and its bootstrap interval do not account for model-selection optimism. The independent Deepfake-Eval-2024 evaluation is valuable, but its advantage is statistically inconclusive. A selection-independent comparison, or narrower wording around the headline transfer ranking, would strengthen the generalization claim.

Direct localization remains less reliable than the coverage results might suggest. The part head identifies the edited region in 0.693 of synthetic cases, while the dense map achieves mean IoU 0.192 and pointing accuracy 0.474. On independent inpainting, single-eye region naming falls to 0.37–0.42. These results demonstrate meaningful learning, but limit the reliability of explanations for individual images. These limitations should be more prominent, with representative failure cases illustrating when localization can be trusted.

Questions for authors

Please address the points raised in the weaknesses above.

2 comments

Thank you for a precise review. We ran each requested analysis.

Weakness 1:

We agree, and new Appendix E.4 gives the paired analysis the reviewer suggested. Each blend is compared with its own clean frame, and we report the paired win rate and the standardized shift (mean blend-minus-clean score change in units of the clean-score SD), with intervals resampling source videos.

Win rateShift (SD)Worse-eye winWorse-eye shift (SD)
PartAware, 3 seeds (reported first)0.928/0.946/0.942+1.98/+2.00/+1.810.873/0.919/0.906+2.07/+2.03/+1.68
Same backbone without P-SBI, 3 seeds0.887/0.776/0.791+1.19/+0.84/+0.710.803/0.569/0.578+0.17/+0.04/+0.03
18 published detectors0.731–0.907+0.30 to +0.930.566–0.890+0.04 to +0.25

Higher is better in every column, and PartAware leads, which confirms the coverage ranking. No published detector reaches any PartAware seed on win rate, shift or worse-eye shift, and on worse-eye win rate the PartAware mean (0.899) is above every published detector. PartAware moves the score by about 2 SD, about twice the largest published shift, and each of the 12 regions moves by 0.67 SD or more on every seed. As the reviewer suggested, published detectors' scores also move toward fake in most pairs, so part of their low coverage is between-image spread, but the move is small (at most 0.93 SD, and 0.25 SD on the worse eye). The shift separates detectors best, as the win rate ignores the size of the move.

Section 3.3 now calls coverage region sensitivity (edited frames scoring above clean ones) and adds:

“Pairing each blend with its own clean frame shows that the score responds to the edit itself, as PartAware moves it by 1.8–2.0 clean-score standard deviations against 0.30–0.93 for the published detectors.”

Weakness 2:

We agree and have narrowed the transfer claim. Section 5.1 now reads the 0.008 margin as parity with the strongest published detectors (Table 1's mean, 0.9797, is still the highest), and the Abstract says “on par with the strongest published detectors”. Two new checks support it.

Leave-one-benchmark-out selection (Appendix F.7). Holding out each benchmark in turn, we pick among our 17 trained configurations scored on transfer, and among the 18 published detectors, by mean AUC on the other three. The reported model is picked in all four folds. On the held-out benchmark, PartAware is ahead of the published detector picked the same way by +0.019 AUC, and of the best published detector on that benchmark by +0.004 on average, where no margin is significant (DFD and UADFV are at ceiling). So no single benchmark drives the pick, and WildDeepFake gives the check outside model selection.

A benchmark outside model selection (Appendix M). On the WildDeepFake test split (801 face sequences from 157 source videos, identical for every detector), PartAware reaches AUC 0.854/0.852/0.836 (mean 0.847). With intervals resampling source videos, the mean is significantly above 13 of the 18 published detectors, not significantly different from the other five (FSFM 0.850, Unlocking-CLIP 0.846, Xception, GenD DINOv3-L and GenD CLIP-L) and significantly below none. The dataset's tight crops are padded before face detection, identically for every detector, with the padding set by face-detection count and without detector scores (Appendix M).

On Deepfake-Eval-2024, also outside model selection, our three-seed mean is significantly above 14 of the 18 published detectors (details in Appendix J and in our reply to Reviewer B9Ss, Weakness 2).

Weakness 3:

We agree, and a new paragraph in Section 6 now states these limits in the main text: IoU 0.19–0.36 over three seeds (mean 0.30, Table 3), and single-eye naming of 0.37–0.42 on inpainting.

To show when a name can be trusted, we added a confidence rule. The part head is trusted when its largest softmax weight clears a threshold set for top-1 0.9 on half of the video groups (reported seed). Over 200 random splits, it keeps on average 0.499 of the held-out blends at top-1 0.898 (central 95% of splits 0.828–0.960). On inpainting, which never set the threshold, it raises top-1 from 0.64 to 0.82 on Kandinsky and from 0.65 to 0.85 on SDXL, keeping 61.8% and 58.7% of edits. The rule raises reliability without guaranteeing a correct name, which the case figures in Appendix L.5 make visible.

Two figures with cases chosen by fixed rules show successes and failures (Appendix L.5). For each recipe, the first shows the trusted and the flagged blend at median confidence. The second shows the failure cases the reviewer asked for, the trusted blend with the lowest IoU for each recipe, and in 5 of these 12 worst cases the part head still names the right region. Naming also holds with the part mask parsed from the edited image (top-1 0.674/0.791/0.819, within 0.025 of the pristine parse).

AuthorsOpenReview ↗

Thank you, and a brief follow-up

We thank the reviewer again for a precise and constructive review. To briefly summarize our reply, the paired analysis confirms that coverage reflects a response to the edit: PartAware shifts its score by 1.8–2.0 clean-score standard deviations on the edited region, about twice the largest shift of the 18 published detectors (0.93), and on WildDeepFake, which played no part in selection, it is significantly above 13 of the 18 detectors and significantly below none. Since the discussion period ends soon, we would be grateful to hear whether our reply addresses the reviewer's concerns, and we would be glad to clarify any remaining point.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion for the Area, Program and Senior Area Chairs

We thank the Area, Program and Senior Area Chairs for overseeing the review of our paper, and the reviewers for their constructive feedback. Reviewers B9Ss and aEKs confirmed that our replies address their concerns. We hope our reply to Reviewer nRfY, whose suggestions shaped several of the new analyses, addresses their concerns as well. For convenience, we briefly summarize the main concerns of each reviewer and how we resolved them.

Reviewer nRfY

  • Does coverage reflect a response to the edit? Compared with each blend's own clean frame, PartAware shifts its score by 1.8–2.0 standard deviations, about twice the largest shift of the 18 published detectors.
  • Is transfer independent of model selection? Leave-one-benchmark-out selection picks the reported model in all four folds. On WildDeepFake, unused in selection, PartAware is significantly above 13 of 18 detectors and significantly below none.
  • When can a region name be trusted? A confidence rule keeps half of the blends at top-1 accuracy 0.90.

Reviewer B9Ss

  • Which abilities does the method show? We report region sensitivity, naming and localization separately over three seeds, with naming (top-1 0.772, chance 0.226) and localization (IoU 0.298, chance 0.122) well above chance.
  • Independent manipulations and Deepfake-Eval-2024. Every seed beats its control on unseen diffusion inpainting, and PartAware is significantly above 14 of 18 detectors on Deepfake-Eval-2024.

Reviewer aEKs

  • Does naming need the pristine frame? From the edited image alone, the part head names the region at 0.69–0.82, about twice the 0.27–0.35 reached without P-SBI.
  • Lipsync and seeds. P-SBI is scoped to local spatial edits, and every seed beats its control, with transfer AUC 0.980–0.982 on all three seeds.

The final version of the paper will include all of these analyses which we have discussed in our rebuttal. We thank all three reviewers, nRfY, B9Ss and aEKs, again for their time and for suggestions that strengthened our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.