acceptodds
Under review as a conference paper at ICLR 2027

Chain of Evidence: Localizing Video Forgeries via Image Forensic Evidence in Space and Time

Abstract

Diffusion-based video editing produces photorealistic forgeries that leave subtle forensic traces. These forgeries are often confined to small regions within a few frames, requiring precise localization across both space and time. Existing video forgery research has mainly focused on video- or frame-level detection, while joint spatial and temporal localization, which is required to verify footage used in news reporting or as legal evidence, has received less attention. Image forgery models provide spatial forensic evidence, but frame-wise inference cannot exploit temporal context to refine localization. We present Chain of Evidence, a parameter-efficient framework that extends perturbation-based diffusion forgery localization from images to video on a frozen SAM 3 backbone. The model learns spatial forensic evidence from images, and a lightweight 8.4M-parameter causal memory carries it across frames so that preceding context strengthens weak or ambiguous cues. Training on video updates only the causal memory and the presence head, leaving the image-trained forensic representation intact. We also propose a protocol that scores space and time jointly. Under it, chaining evidence across frames more than doubles spatiotemporal IoU over framewise inference on the out-of-distribution AINPAINT, from 0.162 to 0.420. On image benchmarks that no method was trained on, Chain of Evidence reaches an average localization IoU of 0.299, against 0.218 for the strongest baseline. On TVIL, it reaches a temporal F1 of 0.370, against 0.259 for the strongest baseline. Code is available at https://anonymous.4open.science/r/coe-iclr27.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video forgery localization, Image forgery localization, Diffusion-based editing, Spatiotemporal localization, Temporal forgery localization, Segment Anything Model, SAM 3, Foundation model adaptation, Parameter-efficient adaptation, Temporal memory, Perturbation-based forensics, Media forensics, Deepfake detection, Video inpainting 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_Gup1OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Chain of Evidence localizes forged content in images and videos using a frozen SAM 3 backbone. An evidence transformer extracts multiscale forensic cues from clean and noise-perturbed features, guiding feature refinement, mask prompting, and forgery detection. A lightweight causal memory then incorporates preceding frames, with video training updating only the memory and presence head. The paper also introduces a gated spatiotemporal evaluation protocol that penalizes missed forged frames and false-positive masks within forged videos. Evaluation across seven image and five video benchmarks shows strong image localization and substantial video-adaptation gains for local edits, while documenting weaker transfer to whole-segment generation.

Critical strengths
  1. A well-motivated, parameter-efficient image-to-video design. Learning forensic cues from diverse image data and freezing that representation during video adaptation directly addresses scarce video mask annotations (Section 3). The evidence-guided segmentation and compact causal memory form a coherent approach, with only 8.57M parameters trained in the video stage.

  2. Broad evaluation with substantial gains and transparent limitations. Average localization IoU reaches 0.299 versus 0.218 for the strongest baseline on four jointly unseen image datasets; video adaptation raises AINPAINT ST from 0.162 to 0.420 (Tables 2–3). The joint evaluation, corruption tests, and explicit failure analysis make the findings particularly useful.

Critical weaknesses
  1. The memory contribution is not fully isolated. The framewise comparison removes both memory and the video-retrained presence head (Section 4.4). A presence-head-only adaptation control would clarify how much of the gain specifically requires temporal context.

  2. The image-stage mechanism needs more targeted ablation. The results establish the full system’s effectiveness, but do not isolate the benefit of perturbation-derived evidence from clean-feature adaptation or the multiscale transformer (Section 3.2). These controls would strengthen the architectural claims.

Questions for authors
  1. Are results available for video adaptation of the presence head alone, without temporal memory?
  2. Do existing ablations compare perturbation-derived evidence with clean features alone under the same backbone and training data?
1 comment
AuthorsOpenReview ↗

Reply to Reviewer Gup1

We thank the reviewer for the positive assessment and for the two controls suggested. Both have now been run.

Weaknesses

W1: The memory contribution is not isolated. A presence-head-only control would show how much of the gain needs temporal context.

[AW1] Yes, and Table A in the general comment is exactly this control. Retraining the presence head on video without a memory leaves all three scores at the framewise level, and adding the memory brings the whole gain. The improvement does need temporal context. Table A will go into Section 4.4.

W2: The image stage needs targeted ablation separating perturbation evidence from clean-feature adaptation and the multiscale transformer.

[AW2] Table C in the general comment isolates each piece on the same backbone and training data. Every removal lowers the mean LL. No variant beats the full model on any test set. Clean features alone cost the most (see Q2).

Questions

Q1: Are results available for adapting the presence head alone, without temporal memory?

[AQ1] Yes. Row 2 of Table A in the general comment is this setting: the same frozen image stage, the presence head retrained on the video data, and no memory. FakeParts ST\mathit{ST} goes from .099 to .113, AINPAINT ST\mathit{ST} from .162 to .160 and TVIL TT from .133 to .121, so the head alone stays at the framewise level. Adding the memory (row 3) brings ST\mathit{ST} to .425 and .420 and TVIL TT to .370.

Q2: Do existing ablations compare perturbation evidence with clean features under the same backbone and training data?

[AQ2] Not in the submission. Table C in the general comment adds this comparison. The clean-only row keeps the backbone, the adapters, the training dataset, the hyperparameters and the seed, and replaces the perturbed view with a second copy of the clean image, so the difference channels are zero and the model is a plain SAM 3 adaptation. Its LL is lower on all four test sets, by .023 on BBC-PAIR, .029 on SID-Set, .012 on DiffSeg30k and .046 on BR-Gen, and the mean drops from .457 to .430. The revision will add Table C to Section 4.4.

We hope this addresses the reviewer's points and would be glad to answer anything further.

Reviewer_eq4WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Chain of Evidence, a parameter-efficient framework designed for localizing forged regions in images and videos. To address the scarcity of annotated data for video forgeries, the authors first leverage a frozen SAM to extract feature discrepancies between original images and their Gaussian-noise-perturbed counterparts, enabling forgery localization through multi-scale evidence extraction, feature enhancement, and the generation of mask prompts. Subsequently, a causal memory module is incorporated to propagate forensic information from preceding frames to the current frame, thereby improving both spatial localization and temporal detection in videos. Furthermore, the authors propose a spatiotemporal evaluation protocol that combines frame-level detection with pixel-level masking, and they validate the effectiveness of the proposed method across multiple image and video forgery datasets.

Critical strengths
  1. This paper leverages a relatively abundant dataset of forged images to learn forensic features. During the video-based stage, the image representations are frozen, and only a lightweight memory module and an existence prediction head are trained. This design directly addresses the scarcity of pixel-level forgery annotations for videos and offers a viable approach to reducing the number of trainable parameters during the video-processing phase.

  2. The experiment achieved an average localization IoU of 0.299 on four image datasets that were not used for training with any comparison method, which is higher than the strongest baseline of 0.218. As shown in Table 3, incorporating the video-stage training for the AINPAINT dataset improved the ST score from 0.162 to 0.420. Furthermore, the paper analyzes failure cases, specifically those involving image degradation and cross-dataset transfer, thereby enhancing the practical value of the experimental findings.

Critical weaknesses
  1. Section 4.4 compares the full video model with the frame-by-frame model, but the two differ not only in memory modules but also in the retraining of the presence head. Consequently, the performance gains reported in Table 3 cannot be attributed solely to the causal memory mechanism. Appendix G acknowledges that the current experiments do not disentangle these factors. Furthermore, the lack of corresponding component ablation in the image-stage evidence extraction Transformer, evidence-guided attention, and cue generator makes the specific advantages of the proposed evidence learning mechanism over existing SAM-based methods unclear.
  2. Table 3 shows that the memory-augmented model, when trained on FakeParts, achieves an ST score of 0.420 on the AINPAINT dataset. While this significantly outperforms its own frame-by-frame counterpart, it still falls short of the performance of DINOv3-IML, which does not rely on video-based temporal modeling. Moreover, the same memory module causes the temporal F1 score of TASLE to drop from 0.312 to 0.108, suggesting that its effectiveness is heavily dependent on the specific types of forgeries and the distribution of the training data. For TVIL, although the authors reported good cross-dataset results, Appendix A points out that it shares some data sources and the E2FGVI editor with FakeParts, making it difficult to use as completely independent evidence of cross-domain generalization.
Questions for authors
  1. Regarding the ablation of memory modules in Section 4.4, can you provide control results showing only retraining the presence head without introducing memory modules? This would help disentangle the contributions of video-based training from those of the causal memory mechanism itself.

  2. To what extent are the performance gains during the image-based phase attributable to the extraction of perturbed evidence, as opposed to the pre-trained representations from SAM 3 and the learned mask prompts? Are there any existing ablation studies for these components that could support this assessment?

  3. Regarding cross-dataset generalization, it would be beneficial to further clarify the degree of overlap between the FakeParts and TVIL datasets at the raw video level. Additionally, please explain why the memory model exhibits such distinct performance differences when evaluated on AINPAINT versus TASLE.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer eq4W (1/2)

We thank the reviewer for the detailed reading and the precise questions.

Weaknesses

W1: The video gains in Table 3 cannot be attributed to the memory alone, since the presence head is also retrained. The image stage also lacks component ablations for the evidence transformer, the evidence-guided attention and the cue generator.

[AW1] We agree on both points and ran the two controls the reviewer asks for. Table A in the general comment separates the presence head from the memory. Retraining the head on video without a memory barely moves the three scores. Adding the memory is what raises them.

Table C in the general comment removes one architectural component at a time. Every row is trained on the same training dataset with the same hyperparameters, with the same frozen backbone and decoder.

Every removal lowers the mean LL. The full model reaches .457. Without the perturbation evidence it is .430. Without the cue generator it is .443. The MLP adapter gives .448 and single-scale evidence gives .449. No variant beats the full model on any test set.

W2: On AINPAINT the memory model falls short of DINOv3-IML, on TASLE it drops TT from .312 to .108, and TVIL shares sources and an editor with FakeParts. Is the memory's effectiveness tied to the training distribution?

[AW2] The video model is trained on one forgery family, FakeParts inpainting, and the presence head learns its operating point from that data. The memory still helps on the two test sets whose videos it never saw, AINPAINT and TVIL (Table B in the general comment). The two temporal-only sets, TASLE and ActivityForensics, are the exception (Table G in the general comment). We could not train on a second family because the only video sets with frame-level spatial masks available to us, FakeParts-ROVI and AINPAINT, are both inpainting, while TASLE, TVIL and ActivityForensics carry temporal labels only. We will state this as a limitation, next to the future work on diffusion-edited video with spatial masks that the paper already lists. On the three observations:

  1. DINOv3-IML is a frame-level model, so this compares image stages, not memories. Our memory's contribution is the gain over our own framewise model (Table B). DINOv3-IML cannot place a forgery in time (Table D).
  2. The TASLE drop is a calibration effect. The FakeParts-trained memory flags too few frames, and so does SAM 3's memory trained on the same data, on TASLE and on ActivityForensics (Tables E and G). The same memory trained on AINPAINT does not drop (Table E).
  3. TVIL and FakeParts share no video (Table F). Without the clips from the shared editor, our lead on TVIL remains (Table D).
AuthorsOpenReview ↗

Reply to Reviewer eq4W (2/2)

Continued from our reply above.

Questions

Q1: Can you show results for retraining only the presence head, without the memory?

[AQ1] Yes, row 2 of Table A.

Q2: How much of the image-stage gain comes from the perturbed evidence rather than from SAM 3's pretrained representations and the learned mask prompts? Are there ablations for these components?

[AQ2] Table C answers this. The clean-only row is SAM 3's representations alone and its mean LL is .430. Removing only the learned mask prompt gives .443. The full model reaches .457. Both components add to SAM 3's representations.

Q3: What is the raw-video overlap between FakeParts and TVIL, and why does the memory behave so differently on AINPAINT and TASLE?

[AQ3] No video is shared. Both sets draw on YouTube-VOS, so we checked video by video. None of the 508 TVIL test clips comes from any of the 5,169 source videos behind our training clips (matched by video ID and perceptual hash). The only shared element is the E2FGVI editor, which produced 110 of the 389 TVIL fakes. Those clips are easier for every method, and removing them does not change the ranking (Table D). Appendix A will be corrected.

Source videosEditor(s)Public release
TVIL (test, 508 clips)YouTube-VOS 2018E2FGVI, FGT, FuseFormer, STTNgithub.com/ymhzyj/UMMAFormer (Zhang et al., ACM MM 2023)
FakeParts-ROVI (our video training set)ROVI: Refer-YouTube-VOS and A2D-Sentences (Wu et al., 2024)E2FGVIhuggingface.co/datasets/hi-paris/FakeParts (Liu et al., 2025)

Table F: Provenance of TVIL and FakeParts-ROVI.

MethodTVIL TT, all clipsTVIL TT, without E2FGVI
Ours (memory).370.353
AINPAINT-ST.259.238
Ours (framewise).133.126
RelayFormer.112.102
DINOv3-IML.107.099
Flag every frame.092.084

Table D: TVIL TT with and without the E2FGVI fakes.

On AINPAINT versus TASLE, Table E explains the TASLE drop. The FakeParts-trained memory flags 8% of frames, against 51% for the framewise model, so TT falls while it clears far more real clips. The same memory trained on AINPAINT flags 83% and returns to the framewise level, while SAM 3's memory trained on FakeParts is more cautious still. The drop is a cautious presence decision learned from FakeParts, not a failure of the memory. AINPAINT is inpainting like FakeParts, so that decision transfers there (Table B). We will state the single-family training data as a limitation (see W2).

TASLE, 1,293 clips, 15% of frames forgedTTShare of all frames flaggedReal clips cleared, of 662
Flag every frame.257100%0
Ours (framewise).31251%17
Ours (memory, trained on FakeParts).1088%224
SAM 3 memory, trained on FakeParts.0444%432
Ours (memory, trained on AINPAINT).26983%1

Table E: TASLE. TT follows how many frames a model flags; both FakeParts-trained memories are cautious, the AINPAINT-trained one is not.

We hope this addresses the reviewer's points and would be glad to answer anything further.

References

  • Wu et al. Towards Language-Driven Video Inpainting via Multimodal Large Language Models (ROVI dataset). CVPR 2024. arXiv:2401.10226.
  • Liu et al. FakeParts: A New Family of AI-Generated Video Forgeries. arXiv:2508.21052, 2025.
  • Zhang et al. UMMAFormer: A Universal Multimodal-adaptive Transformer Framework for Temporal Forgery Localization (TVIL dataset). ACM MM 2023. arXiv:2308.14395.
Reviewer_bEB3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Chain of Evidence, a framework for joint spatial and temporal localization of video forgeries. The method builds upon a frozen SAM 3 backbone and extracts forensic cues by comparing features from original and Gaussian-noise-perturbed images. An evidence transformer aggregates these cues across multiple scales to guide feature adaptation, mask prompting, and forgery presence prediction. To extend image-based forgery localization to videos, the authors introduce a lightweight causal memory bank that propagates forensic information across preceding frames. Experiments on seven image and five video benchmarks demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The proposed two-stage training strategy addresses the scarcity of pixel-level annotations for video forgeries by leveraging more abundant image forgery datasets. The paper is generally well organized and provides additional robustness evaluations and failure-case analyses.
  2. Experiments on seven image and five video benchmarks demonstrate improved out-of-distribution image localization and substantial gains in spatiotemporal localization over the framewise variant of the proposed method.
Critical weaknesses
  1. The methodological novelty of this work appears limited. My primary concern is the methodological novelty of the proposed framework. The method appears to combine existing perturbation-based image forensic localization techniques with established memory-based temporal propagation mechanisms from video segmentation. In particular, prior works such as SAM2Long (Ding et al., ICCV 2025) have already explored improving temporal memory and propagation within SAM-based video segmentation frameworks. While the proposed combination is reasonable, it remains unclear what fundamentally distinguishes the proposed "Chain of Evidence" from a straightforward integration of an existing SAM-based image forgery localizer with an established video memory mechanism.
  2. Insufficient comparisons with closely related SAM-Based temporal methods. Although the paper compares against several video forgery localization methods, including TruVIL, RelayFormer, and AINPAINT-ST, it lacks direct comparisons with existing SAM-based video segmentation and temporal propagation approaches. This omission is particularly concerning because SAM 3 itself already incorporates a memory-based video tracker. It is unclear why the authors introduce a separate causal memory bank instead of adapting the existing SAM 3 tracking mechanism for forensic evidence propagation.

Ding et al., SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree, ICCV 2025.

Questions for authors

Please refer to weakness part.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer bEB3

We thank the reviewer for recognising the two-stage training strategy and the robustness and failure analyses.

Weaknesses

W1: The novelty appears limited. What distinguishes Chain of Evidence from a straightforward integration of a SAM-based image forgery localiser with an existing video memory such as SAM2Long?

[AW1] We thank the reviewer for pointing us to SAM2Long and will add it to the related work. SAM2Long keeps SAM 2's memory and only changes which stored frames the tracker reads. It is built to track a prompted object, whereas forgery localisation gives it neither a prompt nor an object to track.

We also ran the direct integration the reviewer has in mind, and it turned out to be a useful test (Table B in the general comment). We kept our image stage and attached SAM 3's own memory to it pretrained, pretrained with SAM2Long, and trained in place of ours. The first two stay at the framewise level. The third, trained on the evidence our image stage produces, works as well as our memory and on some sets better.

This clarifies where the contribution lies. It is the evidence produced by the frozen image stage. Any memory trained on that evidence, ours or SAM 3's, lifts ST\mathit{ST} from below .17 to above .42 on AINPAINT and FakeParts (Table B), and the memory block can be either one. We will say this in the introduction and report both memory variants in the revision.

W2: Why introduce a separate causal memory bank instead of adapting SAM 3's own tracker memory, and how does the method compare with SAM-based temporal propagation?

[AW2] We did start with SAM 3's own memory. In our first video experiments we attached it to an early version of our image stage. Without training, it left the per-frame masks unchanged. Fine-tuned on FakeParts, it became unstable. On some clips it stopped predicting anything after the first frame, and on others it marked most of the video as forged. We traced this to the way a tracking memory works. It stores the predicted mask and reads it back on the next frame, so one wrong mask feeds on itself, and a forgery gives it no object to hold on to. We built our own memory to avoid this. It stores the evidence features from the image stage rather than masks, each location reads only its own history, and it starts out identical to the per-frame detector.

Table B in the general comment repeats the comparison with the paper's model, and two things changed since that early test. The image stage is now the frozen evidence model of the paper. Both memories are trained with the same training dataset, hyperparameters, seed and retrained presence head (7.5M against 8.4M trained parameters), and AINPAINT and TVIL are not seen in training. Under this protocol SAM 3's memory no longer collapses. It now receives stable evidence features, so the mask it writes back is mostly right and the feedback helps rather than hurts. It also benefits from SAM 3's video pretraining, which our memory does not have. We thank the reviewer for pointing us to this improvement, which we would otherwise have missed.

Table G in the general comment adds the two temporal-only test sets. There both trained memories fall below the framewise model, and SAM 3's falls further. Both flag far fewer frames than the framewise model. Two different memories trained on the same data behave the same way, so this caution comes from the FakeParts training data, not from the memory design.

Table B shows that the gain comes from training a memory on the evidence of the image stage, not from the design of the memory block. SAM 3's pretrained memory stays at the framewise level. Trained on the evidence, both memories raise ST\mathit{ST} and TT on all three test sets. SAM 3's memory localises better, and ours detects more reliably on the two unseen sets. We keep ours as the paper model and report the SAM 3 variant next to it. The revision will state this in Section 3, add Table B to Section 4.4, and add the SAM 3 variant to Table 3.

References

  • Ding et al. SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree. ICCV 2025. arXiv:2410.16268.
AuthorsOpenReview ↗

Follow-up for Reviewer bEB3

Dear Reviewer bEB3,

As the discussion period is still open, we wanted to check whether our reply addresses your concerns. We would be glad to answer any remaining questions.

Best regards, The Authors

Reviewer_bEB3OpenReview ↗

Still unclear what is novel beyond existing temporal methods for video segmentation

Thank you for adding the comparisons in Table B. My original concern was what distinguishes Chain of Evidence from a straightforward integration of an existing SAM-based image forgery localizer with an established video memory mechanism. Table B addresses the temporal part of this question. With the same frozen image stage and matched training conditions, the fine-tuned SAM 3 memory achieves higher spatiotemporal IoU than the proposed causal memory on both AINPAINT and FakeParts, as well as higher temporal F1 on TVIL. I understand the architectural difference: the proposed memory stores forensic evidence features and allows each spatial location to read its own history. Given that an existing video memory also makes effective use of these features, the methodological contribution of the temporal design remains unclear to me.

Your response appears to identify the evidence produced by the frozen image stage as the main contribution. However, Detective SAM already uses features from original and perturbed images to guide forgery localization with a frozen SAM backbone. Could you clarify where you see the novelty of Chain of Evidence? In particular, what distinguishes the way your framework extracts or propagates forensic evidence from the straightforward combination described above?

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Shared Tables

We thank all reviewers for their careful reading and suggestions. Three concerns are shared across the reviews, and we addressed each with a new experiment:

  1. Is the video improvement due to the memory, or to the retrained presence head? We retrained the head alone, without memory (Table A).
  2. What does each part of the image stage contribute? We removed one component at a time and retrained (Table C).
  3. Is TVIL really independent of FakeParts? We compared the two sets video by video (Tables D and F in the eq4W thread).

Each reviewer thread covers the remaining review-specific points. All changes will go into the revised manuscript.

ExperimentAddresses
Presence head retrained without memory (Table A)bEB3 W1; eq4W W1, Q1; Gup1 W1, Q1
SAM 3 memory, pretrained (Table B)bEB3 W1, W2
SAM 3 memory with SAM2Long (Table B)bEB3 W1
SAM 3 memory, fine-tuned (Table B)bEB3 W2
Clean features only, no perturbation (Table C)eq4W W1, Q2; Gup1 W2, Q2
MLP adapter instead of evidence-guided attention (Table C)eq4W W1; Gup1 W2
Single-scale evidence (Table C)eq4W W1, Q2; Gup1 W2
No cue generator (Table C)eq4W W1, Q2; Gup1 W2
SAM 3 memory, fine-tuned, on TASLE and ActivityForensics (Tables G, E)bEB3 W2; eq4W W2
TVIL–FakeParts video overlap (Tables F, D)eq4W W2, Q3
TASLE frame statistics (Table E)eq4W W2, Q3

Shared tables

The replies cite the four tables below. Tables A, B and G use the video test sets and Table C the image test sets.

MemoryPresence headFakeParts ST\mathit{ST}AINPAINT ST\mathit{ST}TVIL TT
no (framewise)original.099.162.133
noretrained.113.160.121
yes (paper model)retrained.425.420.370

Table A: Memory and presence head on the same frozen image stage.

Temporal memoryAINPAINT DDAINPAINT TTAINPAINT SSAINPAINT ST\mathit{ST}TVIL DDTVIL TTFakeParts DDFakeParts TTFakeParts SSFakeParts ST\mathit{ST}
None (framewise).943.814.140.162.856.133.735.526.056.099
SAM 3, pretrained.925.773.160.202.815.134.706.454.045.095
SAM 3, pretrained + SAM2Long.899.780.158.190.784.127.692.458.045.093
Ours (paper model).943.877.309.420.891.370.855.917.352.425
SAM 3, fine-tuned.908.845.319.438.871.450.915.931.442.518

Table B: Temporal memories on the same frozen image stage; AINPAINT and TVIL unseen in training, FakeParts in-distribution; single seed.

Temporal memoryTASLE DDTASLE TTTASLE frames flaggedActivityForensics TTActivityForensics frames flagged
None (framewise).659.31251%.49255%
Ours (paper model).588.1088%.34816%
SAM 3, fine-tuned.425.0444%.2138%

Table G: The two temporal-only test sets, both unseen in training. 15% of TASLE's frames and 34% of ActivityForensics' frames are forged. ActivityForensics has forged clips only, so DD is not defined there.

VariantTrained parametersBBC-PAIRSID-SetDiffSeg30kBR-GenMean
Full model (paper checkpoint)7.62M.430.588.451.360.457
No perturbation (clean features only)7.62M.407.559.439.314.430
MLP adapter instead of evidence-guided attention7.42M.425.580.435.353.448
No cue generator (no learned mask prompt)7.62M.413.571.446.340.443
Single-scale evidence7.57M.417.573.447.360.449

Table C: Detection-gated LL per image test set (BBC-PAIR and SID-Set in-distribution, DiffSeg30k and BR-Gen out-of-distribution) and its mean over the four sets; single seed.

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.