acceptodds
Under review as a conference paper at ICLR 2027

Does AI-Generated Video Detection Really Need Temporal Modeling?

Abstract

Temporal information is a natural source of evidence for AI-generated video detection (AIGVD), yet under shortcut-controlled evaluation on GenVideo, a simple spatial-only detector reaches about 0.95 AUC, substantially outperforming existing temporal detectors at around 0.80. Rather than concluding that temporal modeling is unnecessary, we investigate why current methods fail to exploit temporal evidence effectively. Our analysis suggests a decomposition hypothesis: observed inter-frame variation entangles generation-related evidence with substantial non-generative variation from preprocessing, camera motion, and object motion. We examine two consequences of this entanglement. First, scalar summaries commonly used by current methods can map generation-related and non-generative changes to overlapping values, whereas element-wise differences preserve richer structure and substantially improve detection. Second, progressively reducing major non-generative components yields further consistent gains, making temporal evidence competitive with spatial evidence on GenVideo and stronger in AUC on the unseen GenVidBench benchmark. Moreover, the resulting debiased temporal evidence complements spatial evidence, with their fusion further improving performance. Across controlled representation analyses, progressive reduction of non-generative variation, and cross-benchmark evaluation, our results consistently show that properly modeled temporal evidence remains highly discriminative and valuable for AIGVD.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Video Detection, Deepfake Detection, AI Safety

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UQ8mOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

AI-generated video detection (AIGVD) is an important research topic today. However, existing methods may exploit shortcuts introduced by preprocessing, such as duplicated frames caused by temporal resampling. This raises concerns about whether these methods capture the intrinsic characteristics of AI-generated videos. To investigate whether temporal information contributes to AIGVD, the authors conduct targeted controlled experiments, revisit the role of temporal modeling, and propose a decomposition hypothesis to reassess temporal detection methods.

Critical strengths

1 The paper identifies a key phenomenon in AIGVD and raises a thought-provoking research question that offers valuable insights. The topic is both novel and practically useful.

2 The paper presents extensive, targeted experiments that support its central arguments, making the study coherent, comprehensive and interesting.

Critical weaknesses

The experimental details in the current version require further refinement. Only a small number of videos remain after filtering, raising concerns about the generalizability of the findings and potential selection bias. Could excluding very short clips affect the results, given that their stylistic characteristics may differ substantially from those of longer videos? These concerns limit the justification for a higher rating.

Questions for authors

The reduction in sample size after filtering raises concerns about potential selection bias. Evaluating the proposed approach on the excluded short videos would help assess its applicability to real-world data. If such an evaluation is infeasible, could the authors explain the constraints and their implications?

3 comments
AuthorsOpenReview ↗

Response to Reviewer UQ8m

W&Q: The reviewer is concerned that filtering out short videos may introduce selection bias and limit the generalizability of the findings, and suggests evaluating the approach on the excluded short videos.

Response: We thank the reviewer for raising this concern and agree that evaluating the excluded short videos is important for assessing the generalizability of our findings. Our original experiments follow the shortcut-controlled protocol of Michels et al. (2026). Note that the MSR-VTT real videos released with GenVideo are only 3 FPS; we replaced them with their original high-frame-rate source videos, so real videos are not removed due to frame rate. Under this protocol, 14.45% of the GenVideo test videos are excluded.

To examine whether this filtering affects our conclusions, we further adopt a relaxed 1-second, 8 FPS protocol, which keeps the frame interval unchanged and only shortens the duration. Under this protocol, only 0.09% of the videos are excluded, covering nearly all previously excluded short clips, including HotShot, a generator entirely excluded under the original protocol due to its short video duration.

As shown below, each step of the progressive analysis consistently improves AUC on all eight generators, raising the overall AUC from 0.799 to 0.942, comparable to 0.949 under the original protocol. This indicates that our main findings do not depend on the original filtering condition and remain valid when short videos are included.

Moreover, HotShot directly corresponds to the case raised by the reviewer: as its videos are all shorter than 2 seconds, it was entirely excluded under the original protocol and may differ stylistically from longer videos. Even on HotShot, each step of the progressive analysis consistently improves AUC, from 0.593 to 0.859, further confirming the applicability of our findings to short videos. The full results have been added to the manuscript (Appendix D.8, marked in blue).

SettingCrafterGen-2HotShotLaVieMSE.MS.Show-1SoraOverall
Scalar (ℓ2\ell_2 distance)0.73740.90140.59280.71520.64810.61970.78450.60070.7988
Element-wise difference0.88450.96040.71530.82890.80600.80820.87620.76840.8594
+ Camera motion compensation0.94100.99070.83460.88530.92130.88650.89670.90080.9231
+ Object motion compensation0.96350.99360.85870.92130.94390.91140.90830.91160.9423

Thank you for the response. My main concerns have been addressed, and I maintain my positive rating.

Thank you for your positive feedback. We sincerely appreciate your constructive comments and support.

Reviewer_7gNSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper conducts an in-depth investigation into the core debate surrounding whether "temporal modeling is necessary" in AI-generated video detection, proposing a de-noising framework that decouples frame-to-frame changes and significantly improving cross-benchmark detection generalization through systematic multidimensional compensation and feature optimization, thereby demonstrating exceptional academic insight; however, the method still leaves room for further refinement regarding accumulated errors from traditional motion estimation, computational overhead for industrial real-time deployment, the diversity of backbone networks and comparison baselines, as well as the deep analysis of typical failure cases.

Critical strengths

1.This paper keenly highlights the core debate in the current field of AI-generated video detection regarding whether "temporal modeling is necessary," and through rigorous controlled experiments, reveals that the fundamental reason existing temporal detectors fail is interference from non-generative changes, demonstrating high academic insight.

  1. The authors innovatively propose the hypothesis of decoupling frame-to-frame changes, and systematically de-noising step-by-step from four dimensions: representation format (using element-wise differences instead of scalar summaries), preprocessing control, camera motion compensation, and target motion compensation, resulting in a progressive and logically rigorous methodology.

  2. The paper not only validates the effectiveness of de-noised temporal features on in-distribution datasets, but also showcases excellent generalization capabilities in cross-benchmark (GenVidBench) evaluations, proving the natural complementarity of spatio-temporal features through parameter-free fusion, which provides important guidance for future video forensics research.

Critical weaknesses

1.Despite emphasizing the importance of de-noising, the camera motion estimation (homography matrix) and target motion estimation (RAFT optical flow) employed rely on traditional computer vision algorithms, and these estimation errors may accumulate in long videos or complex scenes, thereby adversely affecting detection performance.

2.The paper lacks an evaluation of computational overhead (such as the memory footprint and inference latency introduced by optical flow computation and dense element-wise differences), leaving the practicality and efficiency of the method questionable when facing large-scale, real-time content moderation industrial scenarios.

3.The ablation studies and step-by-step de-noising rely primarily on a few fixed feature encoders (such as DINOv3), failing to thoroughly validate the universality and stability of this decoupling framework across diverse or lightweight backbone networks.

4.The methods compared on GenVidBench are relatively few, lacking a comprehensive benchmark against current mainstream or state-of-the-art (SOTA) spatio-temporal detection schemes, which is thus insufficient to robustly support the absolute advancement of the proposed method in cross-benchmark generalization.

Questions for authors

See the weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer 7gNS (Part 1/2)

We sincerely thank the reviewer 7gNS for the positive assessment and constructive suggestions. We address each comment below, with all revisions marked in blue in the manuscript.


W1: The reviewer is concerned that the traditional motion estimation methods (homography and RAFT) may introduce estimation errors in complex scenes and accumulate errors in long videos, both of which may adversely affect detection performance.

Response: We agree that traditional motion estimation methods inevitably introduce errors in complex scenes, as already discussed in Section 3.3 and Appendices B.2–B.3. That said, our work is an analytical study rather than an attempt to develop a perfect motion compensation algorithm. We employ Homography and RAFT as interpretable, approximate interventions to investigate whether reducing non-generative motion variation improves the discriminative power of temporal evidence. Even with these imperfect off-the-shelf methods, motion compensation improves AUC on all seven generators, raising the overall AUC from 0.883 to 0.949 and supporting our core hypothesis. We expect that more accurate and robust motion estimation could further strengthen temporal evidence, which is a natural extension of our framework.

Regarding error accumulation, motion is estimated independently for each pair of adjacent frames, so errors do not accumulate over long videos.

W2: The reviewer suggests evaluating the computational overhead of the method, such as its inference latency and memory footprint.

Response: Following your suggestion, We have added a computational overhead analysis in the manuscript (Appendix D.5, marked in blue), reporting the per-video inference time and peak GPU memory of each configuration in our progressive analysis (16 frames per video, batch size 1, on an RTX PRO 6000).

ConfigurationTime (s)GPU Memory (GB)
Spatial-only0.16491.722
Temporal (Scalar)0.16511.723
Temporal (Element-wise)0.16801.757
Spatial + Temporal0.33323.022
Ours0.42313.093

The results reveal two important observations. First, replacing scalar temporal statistics with element-wise differences introduces negligible computational overhead, increasing inference time by only 1.8% (0.1651 s to 0.1680 s) and GPU memory by 2.0% (1.723 GB to 1.757 GB). This suggests that preserving richer temporal variation does not necessarily require substantial additional computation.

Second, compared with the spatial-temporal configuration (w/o motion compensation), our pipeline introduces a 27.0% increase in inference time (0.3332 s to 0.4231 s) and only a 2.3% increase in GPU memory (3.022 GB to 3.093 GB), representing a reasonable computational overhead given the performance improvements. Importantly, our work is primarily an analytical study rather than a framework specifically optimized for real-time industrial deployment. We employ existing motion estimation techniques as interpretable interventions to investigate the discriminative value of temporal evidence. Improving inference efficiency through lightweight motion estimation or more efficient implementations is a promising direction for future work.

W3: The reviewer suggests validating the decoupling framework on more diverse or lightweight backbones.

Response: We agree that validating the framework across diverse backbones further strengthens our analysis. We have additionally repeated the progressive analysis with a lightweight CNN encoder, consisting of a 3×3 convolution (32 channels), two downsampling stages (64 and 128 channels) each followed by a residual block, global average pooling, and a linear layer producing 256-dimensional features. As shown below, each debiasing step consistently improves AUC on all seven generators, raising the overall AUC from 0.603 to 0.882. This trend mirrors that observed with DINOv3, demonstrating the universality and stability of our decoupling framework across different backbones. The full results have been added to the manuscript (Appendix D.7, marked in blue).

SettingCrafterGen-2LaVieMSE.MS.Show-1SoraOverall
Scalar (ℓ2\ell_2 distance)0.6080.7650.5120.6020.5750.6950.5540.603
Element-wise difference0.7270.8830.5730.7220.6280.8040.6040.722
+ Camera motion compensation0.9000.9230.7350.8320.8420.8330.7960.833
+ Object motion compensation0.9400.9710.8040.8830.8730.8570.8470.882
AuthorsOpenReview ↗

Response to Reviewer 7gNS (Part 2/2)

W4: The reviewer suggests comparing with more state-of-the-art spatio-temporal methods on GenVidBench.

Response: We agree that a more comprehensive comparison strengthens our cross-benchmark evaluation. We have therefore additionally evaluated three state-of-the-art spatio-temporal detectors, STALL (CVPR’26), NSG-VD (NeurIPS’25), and DeMamba, on GenVidBench under the same setting. As shown below, all three methods fall well below our debiased temporal detector and fusion on both metrics. This further confirms that debiased temporal evidence generalizes better under benchmark shift than existing spatio-temporal modeling, and that it effectively complements spatial evidence. The full per-category results have been added to the manuscript (Appendix D.6, marked in blue).

MethodAPAUC
NSG-VD0.7550.642
STALL0.7740.660
Demamba0.8420.772
Debiased temporal0.8770.791
Fusion0.9240.842
Reviewer_48RZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the role of temporal modeling in AI-generated video detection (AIGVD). The authors show that a simple spatial-only detector substantially outperforms existing temporal methods under shortcut-controlled evaluation. To explain this, they propose a decomposition hypothesis that inter-frame variations contain both generation-related evidence and non-generative changes caused by preprocessing, camera motion, and object motion. By preserving richer temporal representations and progressively reducing non-generative variations, the proposed approach significantly improves temporal detection performance. Experiments on GenVideo and GenVidBench demonstrate that debiased temporal evidence is competitive with spatial evidence, while their fusion further improves detection performance.

Critical strengths
  1. This paper presents a well-founded explanation for evaluating the role of temporal evidence in AIGC video detection tasks. The experimental results show that biased temporal evidence outperforms spatial evidence in terms of the AUC metric. This finding is significant because it redefines how the academic community should design and evaluate temporal evidence.
  2. The analysis and design were rigorous. Factors such as univariate control comparisons, stepwise interventions with consistent gains per generator, cross-method validation, and cross-benchmark evaluation by action category collectively ensured the credibility of the evidence chain.
Critical weaknesses
  1. Lack of key benchmarks. The argument that the performance of existing temporal methods plateaus around 0.80 is based solely on the models D3, ReStraV, Over-Coherence, NSG-VD, and STALL. GenVideo benchmarks state-of-the-art methods such as DeMaamba and recent MLLM-based detectors—BusterX, VidGuard-R1, and Qwen2.5-ViT—which are cited in related work but were not evaluated under the protocol described in this paper. If these methods perform significantly better than those mentioned above, the premise that “spatial methods outperform temporal methods” would be undermined.
  2. The source of the spatial-only detector’s advantage has not yet been analyzed, and the debiasing process may introduce new hidden shortcuts. The fact that the spatial detector alone achieves an AUC of >0.95 indicates overwhelming single-frame evidence; however, it remains to be verified whether this reflects genuine generative artifacts or domain differences in the training data sources (between Youku-mPLUG and Pika). The sharp drop in the spatial AUC value on GenVidBench (from 0.95+ to 0.767) corroborates these concerns regarding domain differences. Furthermore, SIFT/RANSAC and RAFT themselves may fail on generative videos; “failed motion estimation” may serve as an implicit detection cue, and the argument attributing the debiasing gains to this is based solely on three case studies.
  3. Its “bias-free temporal detector” is not a new end-to-end method, but rather an analysis pipeline assembled from existing components: a frozen DINOv3 + LoRA + SIFT/RANSAC monostatic matching + RAFT optical flow + element-level differences + a lightweight MLP head. Each component is off-the-shelf. It does not propose a new network architecture, a new loss function, or a new training paradigm, nor does it achieve a state-of-the-art (SOTA) ranking that challenges the GenVideo benchmark. It is better categorized as a “high-quality analytical paper” rather than a “paper on a new method.”
Questions for authors
  1. How did GenVideo evaluate the performance of transformer-based temporal methods and MLLM-based detectors under your “Shortcuted-Control” protocol? Could you please include at least one such comparison in the appendix?
  2. What do detectors based solely on spatial features actually rely on? Can you provide feature attribution or simple diagnostic information (e.g., frequency domain, texture, color statistics) to rule out the possibility that differences in the source domains of the training data are the primary reason for the AUC value exceeding 0.95?
  3. Do motion-compensated pipelines (SIFT+RANSAC, RAFT) produce similar residual patterns on real-world video? How can we rule out the possibility that “failed motion estimation on generated video” is merely a new implicit shortcut, rather than evidence of reduced non-generative variability?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 48RZ (Part 1/3)

We sincerely thank the reviewer 48RZ for the time and valuable comments. We address each concern below, with all revisions marked in blue in the manuscript.


W1&Q1: The reviewer is concerned that our spatial-only detector is not compared with DeMamba and MLLM-based methods (BusterX, VidGuard-R1, and Qwen2.5-ViT), whose potentially stronger performance could undermine our finding that a simple spatial-only detector outperforms existing temporal methods.

Response: The reviewer's concern stems from a distinction central to our study: we do not claim that spatial-only detection outperforms all existing video detectors. Our finding is that spatial-only detection substantially outperforms existing methods that rely exclusively on temporal evidence, once preprocessing shortcuts are controlled. This observation motivates our investigation into why temporal information, despite its apparent importance for video authenticity, is poorly exploited by current approaches.

The suggested methods do not provide direct counterexamples to this finding. DeMamba jointly exploits spatial and temporal representations, while BusterX, VidGuard-R1, and Qwen2.5-ViT incorporate additional visual or semantic information rather than relying exclusively on temporal evidence. Strong performance from these methods does not establish the effectiveness of temporal modeling alone. This scope is explicitly defined in Section 2 (Analysis Setup) and Appendix D.2. Nevertheless, we have added comparisons with DeMamba and BusterX++, which further reinforce our findings.

DeMamba. We examine DeMamba's architecture to isolate the contribution of temporal modeling. Its CLIP-based variant combines a Mamba-based temporal branch with a global representation derived from independently encoded frames. By removing this global representation and freezing the CLIP encoder, we evaluate its temporal branch alone under a setting comparable to our temporal analysis. The temporal branch achieves only 0.765 AUC, consistent with existing temporal-only methods (0.75–0.80) and substantially below our debiased temporal detector (0.949). The full CLIP-based configuration reaches 0.822 AUC, while its strongest official spatio-temporal configuration (XCLIP-based, full-parameter training) achieves 0.943, still below our spatial-only detector (0.954) and fusion (0.977). These results provide additional evidence that the limited performance of existing temporal-only approaches is not specific to our original baseline selection.

MLLM-based detectors. BusterX and VidGuard-R1 follow a different detection paradigm, incorporating semantic knowledge and reasoning from large-scale vision-language pretraining rather than relying exclusively on temporal evidence. As their weights are not publicly available, we additionally evaluate BusterX++, the Qwen-based successor of BusterX, under zero-shot evaluation. It achieves 0.895 AUC, below our spatial-only detector (0.954) and debiased temporal detector (0.949).

Importantly, these additional comparisons reinforce rather than undermine our central finding. Existing temporal-only detectors remain substantially weaker than a simple spatial-only baseline after shortcut control. However, our progressive analysis demonstrates that this weakness is not inherent to temporal evidence: preserving richer inter-frame variation and reducing non-generative interference raises temporal detection to 0.949 AUC, with spatial-temporal fusion further improving performance to 0.977.

The full per-generator comparisons are provided below and incorporated into Table 3 and Appendix D.1 of the revised manuscript.

MethodCrafterGen-2LaVieMSE.MS.Show-1SoraOverall
DeMamba (CLIP, frozen), temporal branch0.6910.8530.7360.7750.7480.7490.7920.765
DeMamba (CLIP, frozen), full0.7960.8740.7930.8820.8130.7960.8420.822
DeMamba (XCLIP, full-parameter)0.9620.9910.9320.9540.9800.8640.9030.943
BusterX++ (MLLM)0.8660.9610.8500.9170.9050.8290.8680.895
Spatial-only (DINOv3)1.0000.9950.9270.9910.9960.8390.9630.954
Debiased temporal (ours)0.9630.9960.9250.9500.9320.9150.9230.949
Fusion (ours)0.9990.9990.9620.9940.9840.9350.9550.977
AuthorsOpenReview ↗

Response to Reviewer 48RZ (Part 2/3)

W2: The reviewer raises two concerns: (a) the spatial-only detector's advantage may stem from training-source differences rather than genuine generation artifacts, as suggested by its drop on GenVidBench; and (b) motion estimation failures may introduce new detection shortcuts, with the evidence limited to three case studies.

Response: (a) The GenVideo evaluation protocol already provides substantial evidence against the suggested training-source shortcut. Both the real source (MSR-VTT) and all seven generators in the test set are unseen during training, which uses only Youku-mPLUG and Pika. The spatial-only detector nevertheless achieves 0.954 AUC, demonstrating that its advantage is not limited to distinguishing the two training datasets, although transferable domain-related cues cannot be entirely excluded.

The drop on GenVidBench does not invalidate this observation. Both the real source and generators are replaced, and all methods degrade, including temporal methods based on different designs (second-order variation, angular relationships, and temporal coherence). This indicates a broader benchmark shift rather than a failure unique to spatial detection. Importantly, the spatial-only detector still outperforms all three original temporal-only baselines (0.767 vs. 0.669-0.758 AUC).

More fundamentally, our central contribution does not depend on identifying the exact spatial cues. The spatial-only result motivates our investigation into why temporal-only methods perform poorly. All variants in our progressive temporal analysis use the same training sources and frozen encoder, with controlled interventions targeting non-generative variation. Thus, the consistent improvements cannot be attributed simply to differences in training sources or backbone representations.

(b) The key distinction is between ordinary motion-estimation errors and generation-related temporal inconsistencies that cannot be explained by motion. Errors caused by occlusions, motion boundaries, and imperfect correspondence estimation occur in both real and generated videos, typically leaving residuals along object contours. In contrast, generated videos may exhibit implausible appearance or structural changes inside objects or otherwise stable regions, where no physically consistent cross-frame correspondence exists. In such cases, the residual reflects the generation-related inconsistency itself, rather than merely an arbitrary estimator failure (Appendix B.3). Our motion compensation is therefore an analytical intervention to suppress explainable motion variation, not an attempt to obtain a perfectly artifact-free residual. The same pretrained estimators are applied to both real and generated videos.

To address the concern about limited case studies, we have added further generated examples together with real-video references (Appendix D.4, marked in blue). Consistent with Figure 4, real-video residuals predominantly follow object and human contours, whereas generated-video residuals concentrate in localized regions of implausible appearance changes. Importantly, our evidence is not limited to these visualizations: motion compensation consistently improves AUC across all seven unseen generators (0.883 to 0.949, Tables 1 and 3). The resulting temporal detector also generalizes to GenVidBench without retraining, achieving 0.791 AUC, surpassing the spatial-only detector (0.767).

These quantitative and qualitative results support our central hypothesis that reducing non-generative motion variation systematically strengthens temporal evidence, even with imperfect motion estimators, without requiring every residual to be free of estimation errors.

AuthorsOpenReview ↗

Response to Reviewer 48RZ (Part 3/3)

W3: The reviewer notes that the debiased temporal detector is assembled from off-the-shelf components rather than being a new method, and suggests that the paper is better categorized as a “high-quality analytical paper” rather than a “paper on a new method.”

Response: We appreciate the reviewer's recognition of our work as a ``high-quality analytical paper,'' which is precisely its intended contribution. Our paper does not claim to introduce a new detection architecture or training paradigm. As clearly established in the title, abstract, and contributions, our objective is to investigate why existing temporal detectors underperform after shortcut control and whether temporal evidence remains valuable for AIGVD. Our contribution is a decomposition hypothesis, supported by systematic controlled experiments, rather than a new network architecture. The debiased temporal detector is the resulting configuration of these analyses, not a separately proposed method. Notably, its encoder is fully frozen; LoRA is used only in the spatial-only baseline.

The use of off-the-shelf components is deliberate. By progressively examining representation form, sampling and compression, camera motion, and object motion while keeping the underlying encoder and training sources fixed, we can investigate the contribution of each factor without confounding the analysis with novel architectures or training objectives. Introducing additional network designs would obscure rather than strengthen these findings.

More importantly, our study provides insights beyond assembling existing components: scalar temporal summaries discard discriminative variation structure; non-generative motion can obscure useful temporal evidence; and reducing such variation makes temporal evidence highly discriminative and complementary to spatial evidence. These findings explain the surprising weakness of existing temporal-only detectors and establish why temporal modeling remains valuable. Although SOTA performance is not our objective, the resulting fusion achieves 0.977 AUC, outperforming all compared methods under the shortcut-controlled protocol, including DeMamba (see W1). The scientific contribution lies in explaining these observations, not in introducing another detection architecture.

Q2: The reviewer asks what cues the spatial-only detector actually relies on, and whether diagnostic evidence can rule out training-source differences as the main reason for its $>$0.95 AUC.

Response: The GenVideo cross-source evaluation already provides substantial evidence against the suggested training-source shortcut. Both the real source and all seven generators are unseen during training, yet the spatial-only detector achieves 0.954 AUC. This demonstrates that its performance is not limited to distinguishing Youku-mPLUG from Pika. Furthermore, performance remains nearly unchanged with a frozen encoder (0.952, Appendix D.3) and is consistently strong across CLIP, DINOv2, and DINOv3, indicating that discriminative frame-level information is already present in generic visual representations.

Importantly, identifying the exact spatial cues and understanding why temporal-only detectors underperform are two distinct research questions. Our paper investigates the latter, using the strong spatial-only result as motivation rather than claiming to explain its underlying artifacts. While identifying specific frequency, texture, or color cues requires a separate attribution study, our central findings do not depend on such attribution, as they are established through controlled temporal analyses using the same training sources and frozen encoder.

Q3: Do motion-compensated pipelines produce similar residual patterns on real videos, and how can failed motion estimation on generated videos be ruled out as a new implicit shortcut?

Response: No, the residual patterns on real and generated videos are different. Following the reviewer's suggestion, we have added residual visualizations on real videos, together with more generated examples (Appendix D.4, marked in blue). On real videos, debiasing suppresses most motion-induced structure, leaving residuals mainly along object and human contours, where motion estimation errors typically occur. On generated videos, residuals frequently concentrate within objects or stable regions, corresponding to implausible changes in eye-corner blood vessels, mouth shape, or clothing color. As explained in W2, ordinary estimation errors occur in both real and generated videos, whereas generation-related appearance changes may admit no physically consistent cross-frame correspondence. In such cases, the residual reflects the temporal inconsistency itself rather than merely an arbitrary estimator failure. Together with the quantitative evidence in W2, these observations support our interpretation that debiasing suppresses non-generative variation rather than simply exploiting estimation failures.

Comments on the paper

2 comments

Dear Reviewers,

Thank you all for your valuable and constructive feedback. We have carefully addressed the concerns raised through detailed clarifications and additional experimental results. As the discussion phase is nearing its conclusion, we believe our responses have addressed the main concerns. If there are any remaining questions or concerns, we would be happy to provide further clarification.

Sincerely,

Authors of Submission 1199

Summary of Rebuttal and Discussion

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their constructive feedback. To facilitate the discussion, we briefly summarize our responses below. All revisions are marked in blue in the updated manuscript.

Clarification of the scope of our finding

Reviewer 48RZ questioned whether potentially stronger performance from DeMamba and MLLM-based detectors would challenge our motivating observation. We clarified that our finding, as defined in Section 2, is that a simple spatial-only detector outperforms methods specifically designed to exploit temporal evidence alone, not that it outperforms all methods. Both DeMamba and MLLM-based detectors jointly exploit spatial, temporal, or even semantic information, so neither falls within this scope. Nevertheless, as suggested, we added DeMamba and BusterX++: DeMamba’s temporal branch alone reaches only 0.765 AUC, on par with existing temporal methods, and both full models remain below our spatial-only detector.

Clarification of the analytical positioning

As Reviewer 48RZ noted that our work is better categorized as a "high-quality analytical paper," we clarified that this is precisely our intended positioning. We do not claim a new detection architecture; our contribution is a decomposition hypothesis supported by controlled experiments, and off-the-shelf components are used deliberately so that each intervention can be cleanly attributed to a single factor.

Training-source bias and motion-estimation shortcuts

In response to Reviewer 48RZ, we showed that the GenVideo test set uses entirely unseen real sources and generators, providing strong evidence against a training-source shortcut. We also added case studies on real videos, showing that their residuals mainly follow object contours, whereas residuals on generated videos concentrate in localized regions of implausible change.

Changing backbone for comparison

As Reviewer 7gNS suggested, we repeated the progressive analysis with a lightweight CNN backbone, where each step consistently improves AUC on all seven generators, confirming that our decoupling framework generalizes across backbones of very different capacities.

Additional experiments

To fully address the reviewers' suggestions, we added further experiments: (1) computational overhead analysis, showing that debiasing introduces moderate overhead, with nearly unchanged GPU memory and an inference time well below the clip duration; (2) comparisons with more state-of-the-art spatio-temporal baselines on GenVidBench (STALL, NSG-VD, and DeMamba), all of which fall well below our debiased temporal detector and fusion; and (3) evaluations on short videos: we relaxed the protocol from 2-second to 1-second clips, retaining 99.91% of the test videos, including the previously excluded HotShot generator, and our progressive analysis still yields consistent gains on all eight generators.


Summary and discussion status

Reviewer UQ8m has confirmed that our responses addressed their concerns. We have not yet received follow-up from Reviewers 48RZ and 7gNS. We believe the main concerns of Reviewer 48RZ largely stem from the scope and positioning of our work, which we have clarified with supporting experiments, while the remaining concerns mainly relate to experimental completeness and have been addressed with substantial additional experiments. We remain happy to provide any further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.