acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Trajectory Representation and Consistency Estimation for AI-Generated Video Detection

Abstract

Recent advances in generative video models have enabled the synthesis of visually realistic content, posing significant challenges to synthetic video detection. Existing detectors often rely on appearance artifacts, semantic inconsistencies, and temporal patterns that may be generator-specific, limitating generalization to unseen synthesis models. We investigate whether responses to a pretrained generative model provide more transferable forensic cues. Our key observation is that real and AI-generated videos exhibit distinct velocity responses under a pretrained Flow Matching video model. This distinction persists when different pretrained video-generation backbones are used as probes, suggesting that velocity responses offer transferable forensic signals beyond visual artificts. Motivated by this observation, we propose TRACE (Trajectory Representation and Consistency Estimation), a generation-process-aware framework for AI-generated video detection. TRACE leverages a pretrained video DiT as a velocity-field probe to extract representations at multiple flow time points, and models cross-frame consistency through velocity differences between adjacent frames. We further introduce a Real-Centered Trajectory Optimization objective that encourages generator-invariant representation learning. Extensive experiments on AIGVDBench demonstrate that TRACE generalizes effectively across diverse generators, substantially outperforming prior state-of-the-art methods on unseen open- and closed-source video generation models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-generated video detection, video forensics, Flow Matching, trajectory representation

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

This paper proposes a framework for AI-generated video detection that leverages velocity responses from a pretrained Flow Matching video DiT rather than relying exclusively on appearance based artifacts. The method extracts representations at multiple noise timesteps, incorporates temporal consistency information through velocity differences, and introduces a real-centered optimization objective to improve generalization to unseen generators. Experiments on various dataset demonstrate substantial improvements over existing methods.

Critical strengths
  • Using the velocity response of pretrained generative video models as forensic representations is novel and provides a valid alternative to conventional appearance and artifact based detection.
  • TRACE achieves SOTA performance in various setting and benchmarks.
  • Combining noise-conditioned velocity representations with frame to frame differences is a reasonable way to incorporate both generative model responses and temporal consistency.
  • The paper includes multiple ablations to support the method.
Critical weaknesses

Although, I like the idea and simplicity of the method, I have few questions and concerns.

  • The choice of S is not sufficiently justified, particularly given the appendix results showing that velocity representation remain informative at larger noise timesteps. It is unclear whether the selected timesteps contribute complementary information or whether a single well-chosen timestep could achieve comparable performance.
  • Extracting velocity features at every selected prove point increases computational overhead and feature dimensionality. The paper does not provide a sufficiently detailed accuracy efficienty analysis.
Questions for authors

See Weaknesses

5 comments

W1. Timestep Selection and Complementary Information

We thank the reviewer for raising this important question regarding timestep selection and the complementary benefits of multi-timestep representations. Our choice of S={0,25,50,75,100}S=\{0,25,50,75,100\} is primarily motivated by the empirical analysis in Appendix D, which shows that the discriminability of velocity representations reaches its peak around timestep 100. This observation suggests that the selected range covers the most informative stages of the flow-matching trajectory. Furthermore, the original manuscript already includes timestep ablations in Sec. 4.4.1, comparing timesteps 0, 100, and 200, to investigate the effectiveness of extending the timestep to larger noise levels.

To further examine whether different timesteps provide complementary information or whether a single well-chosen timestep is sufficient, we conduct additional experiments comparing individual timesteps with multi-timestep configurations. Specifically, we evaluate five single-timestep variants, each using velocity features extracted at one of s∈{0,25,50,75,100}s \in \{0, 25, 50, 75, 100\}, as well as two multi-timestep variants: the selected configuration {0,25,50,75,100}\{0,25,50,75,100\} and an extended configuration {0,50,100,150,200}\{0,50,100,150,200\}. All variants are evaluated under otherwise identical experimental settings. We report AUC on both open-source and closed-source generator settings, as summarized in Table 1.

Table 1. AUC (%) of different timestep configurations on open-source and closed-source generator settings.

Dataset[0][25][50][75][100][0, 25, 50, 75, 100][0, 50, 100, 150, 200]
Open-source96.7197.5294.9596.4197.8897.9597.21
Closed-source92.3891.0693.8392.7093.8995.1193.86

As shown in Table 1, the individual timesteps exhibit different detection performance, suggesting that velocity responses at different stages of the flow-matching trajectory capture distinct forensic characteristics. More importantly, fusing the five selected timesteps achieves the best performance in both settings, reaching 97.95% AUC on open-source generators and 95.11% AUC on closed-source generators. Compared with the best single-timestep variant (i.e., s=100s=100), multi-timestep fusion improves AUC by 0.07 and 1.22 percentage points, respectively. This improvement is particularly evident in the closed-source setting, suggesting that aggregating velocity responses across different trajectory stages provides complementary forensic information, especially for generalization to unseen generators.

We also compare the selected configuration with the extended configuration {0,50,100,150,200}\{0,50,100,150,200\}, which incorporates larger noise timesteps. The selected configuration achieves higher AUC by 0.74 and 1.25 percentage points on open-source and closed-source settings, respectively. These results are consistent with the discriminability analysis in Appendix D, further supporting our choice to focus on timesteps up to 100 rather than extending the representation to larger noise levels.

The analysis in Appendix D, the existing timestep ablations in Sec. 4.4.1, and our additional single- versus multi-timestep comparisons consistently support the effectiveness of the selected timestep configuration. In particular, the results demonstrate that the selected five timesteps capture complementary forensic information and provide better overall performance than either individual timesteps or configurations extending to larger noise levels. We have incorporated these additional results and the corresponding discussion into Appendix H of the revised manuscript.

W2. Computational Efficiency and Accuracy-Efficiency Trade-off

We thank the reviewer for highlighting the importance of computational efficiency. We provide additional evaluations of TRACE in terms of computational complexity, inference latency, GPU memory consumption, and model parameters. We also clarify the distinction between the training and inference costs of our framework.

First, TRACE has a lightweight training process despite using a large pretrained video DiT. Specifically, we keep the pretrained DiT backbone completely frozen and optimize only the lightweight detection components. As shown in Table 2, the number of trainable parameters in TRACE is only 0.0008M for the single-timestep variant and 0.0041M for the five-timestep variant, compared with 14.415B total parameters in the Wan2.2-A14B-based model. This substantially reduces the number of parameters requiring optimization and avoids the substantial training cost of fine-tuning large-scale video generation models. Therefore, although TRACE leverages a large pretrained backbone, its task-specific training remains highly parameter-efficient.

Second, we acknowledge the inference overhead introduced by the large pretrained DiT. To provide a comprehensive assessment, we report the FLOPs, end-to-end inference latency, peak GPU memory consumption, total parameters, and trainable parameters of different methods and TRACE configurations in Table 2. All computational costs are measured for the complete detection pipeline, including velocity extraction and subsequent feature processing. The results show that the five-timestep TRACE configuration incurs considerable inference overhead, primarily due to repeated DiT forward passes. To explore a more efficient configuration, we evaluate a single-timestep variant using S={0}S=\{0\}. This reduces the number of DiT forward passes and substantially lowers computational complexity and inference latency, while achieving 96.71% AUC on open-source and 92.38% AUC on closed-source generator settings. Notably, it achieves the best performance among the single-timestep configurations, providing a more favorable accuracy-efficiency trade-off.

Finally, we explore the feasibility of reducing inference costs through smaller backbones. As shown in Table 2, TRACE can also operate with the Wan2.1-T2V-1.3B backbone, demonstrating that our framework is not inherently restricted to the 14B model. Although the smaller backbone offers a substantially more compact alternative, further reducing its inference cost remains an important direction. In future work, we will investigate knowledge distillation to transfer the forensic representations and discriminative capabilities of the large DiT into lightweight video models, together with model pruning and quantization to further reduce computational complexity, memory consumption, and energy usage. These techniques could enable TRACE to retain its forensic effectiveness while operating with smaller, more efficient backbones, facilitating low-power inference and deployment in resource-constrained environments.

We have incorporated these additional efficiency evaluations and discussions into Appendix G of the revised manuscript.

Table 2. Computational efficiency comparison of different methods and TRACE configurations.

MethodFLOPs (G)Latency (s/video)Peak GPU Memory (GB)Total Params (B)Trainable Params (M)
Effort1,296.40.0221.970.5050.199
ForgeLens169.70.0201.690.4447.097
V-PVP361.00.0150.390.0870.493
Qwen2.5-ViT13,878.80.2921.730.669668.686
TRACE (Wan2.1-T2V-1.3B, S={0}S=\{0\})67,500.00.8903.491.5460.0008
TRACE (Wan2.2-A14B, S={0}S=\{0\})438,200.03.05229.2414.4150.0008
TRACE (Wan2.2-A14B, S={0,25,50,75,100}S=\{0,25,50,75,100\})2,191,000.013.69929.2414.4150.0041

Thank you for your response.

First Concern

Resolved. Given the differences between the two methods and the improved performance.

Second Concern

Remains unresolved. Although the number of trainable parameters is small, the substantially higher inference latency of TRACE, particularly as the number of timesteps increases (which is an important factor in achieving better performance), may significantly limit its practical applicability.

As another reviewer JYDM also raised a similar concern, I would be interested in hearing their perspective on this issue before finalizing my rating.

Reviewer_m8j1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on AI-generated video detection, a practically relevant and important research problem. The authors raise a critical question in this field: whether existing detection methods that rely on conventional cues such as texture artifacts will become obsolete as video generation models continue to evolve rapidly. To address this challenge, the authors present an interesting observation that real videos tend to exhibit smaller velocity responses than AI-generated videos, suggesting a potentially robust and discriminative cue for detection. Building on this insight, the authors propose the TRACE framework and conduct extensive cross-generator experiments to evaluate its effectiveness. The experimental results demonstrate the effectiveness of the proposed approach.

Critical strengths

The paper’s main strength lies in its intriguing motivation and new perspective: real and AI-generated videos may exhibit consistent differences in their responses within the velocity field of a pretrained generative model. The paper builds coherently on this observation and develops TRACE around this central motivation.

Critical weaknesses

1 The paper’s motivation rests on the implicit assumption that conventional cues for detecting AI-generated videos such as temporal inconsistencies between frames may diminish as generative models improve, whereas differences in velocity-field responses may provide more persistent cues. However, are these response differences truly more persistent than conventional cues? Could future generative models also be explicitly optimized to reduce or eliminate them?

2 The observation that real videos exhibit smaller velocity responses than AI-generated videos remains insufficiently explained. The paper would benefit from a more detailed investigation into the underlying causes of this difference.

Questions for authors

As discussed in the weaknesses, my concerns about the paper center on two points: (1) whether the proposed signal is truly more persistent than conventional forensic cues, and (2) whether the authors can provide a more thorough explanation of the observed response differences underlying the paper’s core motivation.

3 comments

W1 & Q1. Persistence and Robustness of Velocity-Based Forensic Signals

We thank the reviewer for this insightful question. We would first like to clarify that, as stated in the Abstract and Introduction, TRACE extracts velocity-field responses predicted by a pretrained Flow Matching video DiT as forensic representations, rather than relying solely on conventional visual artifacts or temporal inconsistencies in generated videos. Our key motivation is to exploit generative models themselves as forensic probes, providing a complementary perspective beyond conventional output-level cues.

Fundamentally, diffusion and flow-matching models learn to transform noise into data through progressive state evolution, characterized by learned denoising directions or velocity fields. By leveraging a pretrained generative model, TRACE captures the internal responses associated with this generative process, which are fundamentally different from superficial output-level artifacts. Since these generative principles are shared across a broad family of generation models, the extracted responses may provide more transferable forensic cues across different generators. Furthermore, TRACE exploits both velocity magnitudes and adjacent-frame velocity differences at the same noise level to capture complementary response characteristics. We also provide a robustness evaluation in Sec. 4.5 to further investigate the reliability of the proposed representations under different input perturbations.

To further substantiate the generalization capability of the proposed forensic cues, we summarize our experimental results from two complementary perspectives.

(1) Performance across open-source and closed-source generation settings. Table 1 reports the detection performance across 20 open-source and 11 closed-source generation settings. TRACE achieves 94.27% and 92.03% ACC@5% FPR on open-source and closed-source settings, respectively, substantially outperforming the compared methods. In particular, TRACE consistently achieves higher accuracy than the strongest competing methods in both settings, demonstrating its effectiveness across diverse generator conditions, including unseen closed-source generators.

Table 1. Performance across open-source and closed-source generation settings (ACC@5% FPR, %).

MethodOpen-sourceClosed-source
UnivFD68.8562.36
D370.6661.91
DeCoF71.6868.76
V-PVP68.6768.52
TRACE94.2792.03

(2) Cross-dataset generalization. We further evaluate TRACE on six different datasets, with the results presented in Table 2. TRACE achieves the best performance on four datasets, namely VideoFeedback, ComGenVid, GVD, and GVF, with ACC@5% FPR scores of 79.87%, 88.13%, 94.31%, and 95.39%, respectively. On GenVideo, TRACE achieves 87.93%, ranking second among the compared methods. Although its performance on Magic Videos (67.22%) is lower than that of some competing methods, TRACE achieves the highest average ACC@5% FPR of 85.48% across all six datasets, outperforming the second-best method, V-PVP (76.48%), by 9.00 percentage points. These results demonstrate that TRACE achieves strong overall cross-dataset generalization while maintaining competitive performance under diverse data distributions.

Table 2. Cross-dataset generalization performance (ACC@5% FPR, %).

MethodVideoFeedbackGenVideoComGenVidMagic VideosGVDGVFAverage
DDA66.1878.2770.0668.7074.2391.5374.83
DeMamba65.4175.2667.8066.6375.5691.8373.75
DeCoF61.8274.6570.5370.2278.9192.6474.80
STALL68.5990.5067.2371.3474.8984.3376.15
V-PVP66.1875.4972.5368.8583.6892.1676.48
TRACE79.8787.9388.1367.2294.3195.3985.48

Taken together, the results across diverse generation settings and datasets demonstrate the strong generalization capability of TRACE. Unlike conventional detectors that primarily rely on generator-specific visual artifacts, TRACE exploits the internal velocity responses of a pretrained generative model, capturing forensic characteristics rooted in learned generative dynamics. Such representations are not tied to the specific artifacts of individual generators, making TRACE more adaptable to the continuously evolving landscape of video generation models. We therefore expect TRACE to maintain stronger generalization to emerging synthetic videos and unseen generators, providing a promising direction for robust AI-generated video detection.

W2 & Q2. Underlying Mechanism of Velocity-Field Differences

We thank the reviewer for highlighting the need to further investigate the underlying mechanism of the observed velocity-response differences. We attribute this phenomenon to the distributional discrepancy between real and AI-generated videos with respect to the generative dynamics learned by the pretrained video DiT. Specifically, through large-scale training on real-video data, the pretrained DiT learns the underlying spatiotemporal statistics of real videos. Consequently, real videos tend to exhibit responses more consistent with the learned generative dynamics, resulting in relatively smaller velocity magnitudes. In contrast, despite their perceptual realism, AI-generated videos may still deviate from the learned real-video distribution in their underlying spatiotemporal representations, leading to larger velocity responses. Beyond response magnitudes, adjacent-frame velocity differences at the same noise level further characterize how the pretrained DiT responds to temporal variations, providing complementary forensic evidence.

This distributional discrepancy is also supported by the t-SNE visualization in Appendix H, where real and generated videos exhibit distinguishable distributions in the velocity-response feature space. Furthermore, motivated by this observation, we introduce a triplet-based constraint in TRACE to explicitly pull real-video representations toward the real-data center while pushing generated-video representations away from it. This objective further enhances feature discriminability, connecting the observed distributional differences with our optimization strategy.

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

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

TRACE employs a frozen Wan2.2-A14B video DiT as a forensic probe, extracting responses at five flow-time points and combining them with difference features. A lightweight classifier is trained with binary cross-entropy and Real-Centered Trajectory Optimization (RCTO), which attracts real representations toward an EMA center while repelling generated ones. Training uses 14K real videos and 14K Open-Sora videos. The potentially valuable contribution lies in adapting generative-model probing to flow-matching video representations and temporal consistency.

Critical strengths
  1. A useful empirical finding. A frozen large video generator is a very strong forensic feature extractor under single-source training. This finding is worth publishing.
  2. The trainable design is minimal, and the shared-noise choice is well judged. Sharing one noise realization across frames makes frame-to-frame velocity differences content-dominated rather than noise-dominated.
Critical weaknesses
  1. Inconsistent definition of the final difference-feature set. The Introduction describes residuals taken along flow times, whereas Eqs. (8)–(9) define Δv\Delta v for all N−1 adjacent frame pairs at every probe level. In contrast, Appendix A states that the fused representation contains adjacent flow-time differences and two temporal differences, both at the clean level (first–middle and middle–last). Moreover, Section 3 suggests that v denotes the DiT-predicted velocity field of shape (B, 16, F, H, W), whereas Section 4.1 refers to "final-layer DiT features." These descriptions need to be reconciled, otherwise the method cannot be reproduced as written.
  2. Strong ablation baseline that is not discussed. The variant without noise perturbation and RCTO already achieves 96.3%/91.7% AUC, compared with 97.9%/95.1% for the full method. This indicates that the base representation is already highly discriminative, and at s = 0 the remaining signal consists solely of the video's latent features rather than a smaller velocity response for real videos. This weakens the link between the paper's core claim and the experimental results.
  3. Unfair baseline comparisons. (i) The strongest methods in the original AIGVDBench, Effort and ForgeLens are not compared. Under a matched protocol, the claimed closed-source margin are largely limited. (ii) The results reported in the V-PVP paper are substantially higher than the authors' reproduced numbers .(iii) WaveRep and the frozen Qwen2.5-ViT achieve only 50-60 AUC on the training domain, indicating unsuccessful training. These baseline results undermine the fairness of the comparison.
  4. Missing computational analysis. Although only the classifier head is tuned, inference requires five Wan2.2 14B DiT forward passes per video, incurring substantial overhead. The paper lacks FLOPs, latency, and memory comparisons against baselines, which limits assessment of practical deployability.
  5. Missing fixed-threshold metrics. The paper reports AUC, AP, and ACC@5% FPR, which adjusts the threshold on the real videos of each test split. These metrics capture ranking ability and are insensitive to common cross-generator or cross-postprocessing score drift. Fixed-threshold accuracy should be reported for TRACE and the baselines to enable a more comprehensive evaluation.
Questions for authors
  1. Please unify Figure 2, Eqs. (8)–(10), and Appendix A: state unambiguously whether velocity fields or DiT latent features are used, whether both types of temporal differences are employed, and give the shape of the final fused tensor.
  2. Please provide computational-cost comarision and analysis, e.g., per-video FLOPs, latency, and GPU memory.
  3. Please describe in greater detail the settings of retrained baseline methods such as WaveRep, Qwen2.5-ViT, and V-PNP. Can the authors additionally report results using their pretrained backbones?
5 comments

W1 & Q1. Clarification of Velocity Representations and Temporal Differences

We thank the reviewer for identifying these inconsistencies. We clarify that TRACE probes the pretrained DiT along multiple flow timesteps, while the residuals are specifically computed as temporal differences between adjacent video frames at the same flow timestep, rather than differences between adjacent flow timesteps. Specifically, for each probe timestep sks_k, we extract the DiT-predicted velocity representations of all video frames and compute:

rsk(n)=vsk(n+1)−vsk(n),n=1,…,N−1.\mathbf{r}{s_k}^{(n)} = \mathbf{v}{s_k}^{(n+1)} - \mathbf{v}_{s_k}^{(n)}, \quad n=1,\ldots,N-1.

This operation is independently performed at each selected timestep sk∈{0,25,50,75,100}s_k \in \{0,25,50,75,100\}. Thus, the multi-timestep representations characterize the DiT responses along different stages of the flow-matching trajectory, while the corresponding residuals capture temporal variations in these responses across adjacent video frames. We also use the same noise realization across all frames at each probe timestep to avoid introducing differences caused by independently sampled noise.

Regarding the Introduction, our original description was intended to convey the overall idea of probing the pretrained DiT across its flow-matching trajectory. However, its wording could be interpreted as computing residuals between adjacent flow timesteps. We have revised the description to distinguish multi-timestep trajectory probing from the actual temporal residual computation.

We also acknowledge that the description in Appendix A of adjacent flow-time differences and two temporal differences computed only at the clean level (first–middle and middle–last) does not accurately reflect our implementation. We have corrected Appendix A and revised Eqs. (8)–(10) to consistently describe the adjacent-frame differences computed independently at each probe timestep.

Finally, we clarify that vsk(n)\mathbf{v}_{s_k}^{(n)} denotes the velocity field predicted by the DiT, rather than intermediate hidden features extracted from its final layer. The predicted velocity field has the same shape as the VAE latent representation, namely (B,16,F,H,W)(B,16,F,H,W), and is extracted before spatial pooling. Under our current experimental setting, each video is represented by a tensor of shape (1,16,9,60,104)(1,16,9,60,104), where the dimensions correspond to the batch size, latent channels, number of video frames, and spatial dimensions, respectively. We will replace the ambiguous term "final-layer DiT features" in Sec. 4.1 with "DiT-predicted velocity representations" and clarify the feature construction and fusion procedure.

Accordingly, the final fused representation consists of two complementary components: (1) the DiT-predicted velocity representations at the selected flow timesteps, capturing the model's responses along the flow-matching trajectory; and (2) the corresponding adjacent-frame temporal residuals at each timestep, capturing temporal variations in these velocity responses. We will unify the descriptions and mathematical definitions throughout the revised manuscript to facilitate reproducibility. For additional implementation details and reproducibility, we provide an anonymous code repository: https://anonymous.4open.science/r/TRACE-D99D/.

W2. Forensic Capability of Single-Timestep Representations

We thank the reviewer for this insightful comment. We agree that the strong performance of the variant without noise perturbation and RCTO deserves further clarification. We would like to emphasize that a key finding of TRACE is that a pretrained generative video model can serve not only as a video generator but also as a powerful forensic probe, as its internal velocity responses encode discriminative information between real and AI-generated videos. This capability is already evident at the clean endpoint (s=0s=0), without additional noise perturbation.

As shown in Appendix D, the velocity responses at s=0s=0 exhibit relatively smaller differences between real and AI-generated videos compared with those at certain noisy timesteps. Nevertheless, these differences remain sufficiently discriminative for distinguishing real from synthetic videos, demonstrating that the pretrained DiT's velocity responses at the clean input level already contain useful forensic information.

Building upon this observation, multi-timestep probing and RCTO further enhance the intrinsic forensic capability of pretrained generative models. The variant using only the clean-endpoint velocity response, without additional noise perturbation or RCTO, already achieves AUCs of 96.3% and 91.7% on open-source and closed-source generator settings, respectively. Building upon this strong representation, the full TRACE further improves the AUCs to 97.9% and 95.1%, corresponding to gains of 1.6 and 3.4 percentage points, respectively. Notably, the improvement on closed-source settings is more than twice that on open-source settings, suggesting that the additional components are particularly beneficial for generalization to diverse generators. These results indicate that multi-timestep probing and RCTO do not create forensic discriminability from scratch; instead, they exploit complementary evidence across different flow stages and further enhance the separation between real and generated representations.

To further examine whether these benefits extend beyond AIGVDBench, we conduct additional cross-dataset evaluations. As shown in Table 1, we compare the S={0}S=\{0\} variant with the full TRACE on six additional datasets. The full TRACE consistently outperforms the single-timestep variant across all six datasets, with improvements of 4.91, 6.42, 7.37, 0.71, 6.68, and 5.35 percentage points on VideoFeedback, GenVideo, ComGenVid, Magic Videos, GVD, and GVF, respectively. In particular, substantial improvements are observed on VideoFeedback, GenVideo, ComGenVid, GVD, and GVF. These results further demonstrate that the benefits of multi-timestep trajectory probing and RCTO are not limited to AIGVDBench but also extend to diverse evaluation datasets. Although the improvement on Magic Videos is relatively modest, the full TRACE still consistently surpasses the S={0}S=\{0\} variant.

Taken together, these findings highlight the intrinsic forensic capability of pretrained generative video models and demonstrate how TRACE further exploits this capability through multi-timestep trajectory probing and RCTO. We have clarified the distinction between the DiT-predicted velocity field and the input latent representation in Section~4.4.1 of the revised manuscript.

Table 1. Cross-dataset ACC@5% FPR (%) comparison between the single-timestep variant and the full TRACE.

MethodVideoFeedbackGenVideoComGenVidMagic VideosGVDGVF
TRACE (S={0}S=\{0\})74.9681.5180.7666.5187.6390.04
TRACE79.8787.9388.1367.2294.3195.39

W3 & Q3. Baseline Comparisons and Evaluation Protocols

We thank the reviewer for raising these important concerns regarding baseline comparisons and reproducibility. We agree that differences in training data and experimental protocols should be clearly specified to ensure a fair evaluation. To address these concerns, we have conducted additional experiments, including evaluating publicly available pretrained checkpoints and incorporating stronger baseline methods.

We first clarify the training protocol used in the main experiments. All methods trained under our experimental protocol, including TRACE and the originally reported baselines, use Open-Sora-generated videos as the sole synthetic training source, together with real videos, and are trained for 5 epochs. The remaining model configurations and hyperparameters follow the respective original papers. Detailed training settings are provided in Appendix B. This establishes a consistent training data protocol for comparing methods trained in our experiments.

We then clarify the evaluation of Qwen2.5-ViT and V-PVP. The results of these two methods reported in Table 2 are obtained by directly evaluating the checkpoints provided by their respective authors, rather than retraining them under our experimental protocol. In particular, the V-PVP checkpoint was trained using three synthetic video datasets: Open-Sora, CogVideoX1.5, and EasyAnimate. Therefore, its results reflect a broader synthetic training data setting than ours. We explicitly distinguish these author-provided checkpoint results from those of methods trained under our Open-Sora-only protocol to ensure transparency in the comparison.

Following the reviewer's suggestion, we further incorporate Effort and ForgeLens as additional baselines. Both methods are trained and evaluated under our experimental protocol, using the same real and Open-Sora-generated training videos for 5 epochs, with their remaining configurations following the respective original papers. As shown in Table 2, TRACE achieves the best performance among all compared methods across both open-source and closed-source generator settings, outperforming even these additional strong baselines. Specifically, TRACE achieves 97.95%/95.11% AUC, 94.27%/92.03% ACC@5% FPR, and 98.29%/95.25% AP on open-source/closed-source settings, respectively.

We have revised Appendix C to clarify the distinction between author-provided checkpoint evaluations and methods trained under our experimental protocol, and to provide the corresponding training details. We will incorporate the additional comparisons with Effort and ForgeLens into the main results table. These results further demonstrate the strong performance of TRACE under a single-source synthetic training setting.

Table 2. Comparison with baseline methods on AIGVDBench.

MethodAUC (Open) ↑AUC (Closed) ↑ACC@5% FPR (Open) ↑ACC@5% FPR (Closed) ↑AP (Open) ↑AP (Closed) ↑
Qwen2.5-ViT (Author checkpoint)68.5691.1264.7986.8570.5685.44
V-PVP (Author checkpoint)95.5688.4891.3080.5094.4184.71
Effort87.7682.4076.7271.0387.5173.62
ForgeLens96.7093.9991.7187.7893.9192.71
TRACE (Ours)97.9595.1194.2792.0398.2995.25

Note: Qwen2.5-ViT and V-PVP are evaluated using author-provided checkpoints. The V-PVP checkpoint was trained on Open-Sora, CogVideoX1.5, and EasyAnimate. All other methods in the table are trained using real videos and Open-Sora-generated videos for 5 epochs, following their respective original model configurations and hyperparameters.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.