acceptodds
Under review as a conference paper at ICLR 2027

DELTA: Detecting and Localizing Generated Video Content through Generative Latent Differences

Abstract

Partially manipulated videos pose a growing forensic challenge, as generated content can be inserted into short intervals while the rest of the video remains authentic. Existing approaches mainly derive temporal cues from frame-level discriminative representations. We argue that focusing on individual representation states can overlook how video content evolves over time. Based on this view, we introduce DELTA, a training-free, multi-scale descriptor that characterizes local transitions along generative latent trajectories. By modeling local evolution rather than isolated latent states, DELTA captures temporal transition structure beyond frame-level discriminative representations. For full-video detection, we map continuous spatial transitions into reusable discrete patterns for robustness. We further introduce a benchmark that evaluates temporal localization of partial forgeries using paired real sources and exact interval annotations. DELTA achieves strong performance across both temporal localization and full-video detection. In particular, on the unseen-generator evaluation, replacing CLIP features with DELTA under the same ActionFormer localizer improves mean AP from 77.68 to 86.31.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated video detection, video forensics, latent dynamics

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_JKHeOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DELTA adopts a frozen, frame-wise Wan2.1 VAE to extract signed latent differences. Multi-horizon spatial statistics are aggregated into 256-dimensional descriptors for an ActionFormer localizer, while a separate FSQ–CNN branch classifies videos from spatial difference maps. The paper also introduces a source-disjoint, paired partial-forgery benchmark. Its main contribution is an effective representation for temporal localization, supported by broader AIGV detection experiments.

Critical strengths
  1. Substantial gains in boundary precision with the localizer itself held fixed, which cleanly isolates the contribution of the proposed representation.
  2. The paired benchmark provides useful experimental control: matched intervals enable more meaningful cross-pipeline comparisons than evaluations over unrelated real and synthetic videos.
Critical weaknesses
  1. Incremental novelty & missing related work. Differencing frame representations is a well-established idea (DeCoF, D3, ReStraV, as well as temporal difference networks in action recognition). The choice of VAE latents, per-channel statistics, and multiple temporal horizons are incremental. ReStraV and DeCoF are neither cited nor compared against. Moreover, all reported models require training (an ActionFormer or FSQ–CNN head), so the "training-free" characterization is somewhat misleading.
  2. Three of the five seen generators in ActivityForensics decode through the same VAE family that DELTA uses as its encoder. Wan and VACE use the Wan2.1 VAE, and Sci-Fi uses the CogVideoX VAE. The encoder ranking in Table 11 (Wan2.1 84.04 > CogVideoX 78.72 > SD 74.21) is consistent with this degree of overlap. Furthermore, on the authors' own benchmark which contains no Wan-family forgeries, DELTA underperforms ActionFormer on SVD-XT. These observations suggest that DELTA may be detecting encoder–decoder pairing rather than generic generative latent trajectories.
  3. The GenVideo evaluation is likely biased by a sampling shortcut. Michels et al. show that under this protocol the frame sampling interval of real GenVideo videos is over 5× that of fake videos. Because the paper samples frames uniformly over the full video duration, δ_t may encode video length and motion magnitude rather than generative artifacts. In addition, the reported test split (0.5k real, 4.5k fake) is much smaller than the official GenVideo split (10K real, 8.6K fake), with the real portion particularly reduced.
  4. The benchmark contribution is weak. (i) The Framer split contains only a single fixed-length forged interval, so an in-domain localizer can easily learn a duration prior. (ii) The dataset is small, and two generators (Framer and RIFE) are closer to frame-interpolation methods than to genuine localized video editing, which limits the benchmark's validity for the target task.

ReStraV: AI-generated video detection via perceptual straightening, NeurIPS 2025
DeCoF: Generated Video Detection via Frame Consistency: The First Benchmark Dataset

Questions for authors
  1. Can you provide per-generator results on ActivityForensics, or results under the official cross-domain protocol (train on A → test on B, and vice versa)?
  2. Can you rerun the evaluation on the full GenVideo set and on FakeParts with fixed-FPS sampling and a duration-matched real subset? Results on additional AIGV benchmarks with better real–fake matching (e.g., GenBuster++, ViFBench, MagicVideo) would further strengthen the claims.
  3. Please discuss and compare against ReStraV and WaveRep, and provide comparisons with these methods as well as D3 on GenVideo.

GenBuster++: BusterX++: towards unified cross-modal AI-generated content detection and explanation with MLLM
MagicVideo: Preserving forgery artifacts: AI-generated video detection at native scale, ICLR 2026
ViFBench: Skyra: AI-generated video detection via grounded artifact reasoning, CVPR 2026

5 comments
AuthorsOpenReview ↗

Response to Reviewer JKHe (1/4)

We thank the reviewer for the careful reading and constructive feedback, and we are glad that the reviewer finds value in our work. We have uploaded a revised PDF with all changes marked in blue, and we address each concern below.

W1 & Q3 (a): Novelty and comparison with the suggested baselines

On the suggested related work. We thank the reviewer for the suggested related work. D3, ReStraV, and DeCoF all exploit how frame-level features change over time and ask whether this change is regular: D3 takes the standard deviation of temporal differences of adjacent-frame embedding distances, ReStraV uses step lengths and curvature of the DINOv2 feature sequence, and DeCoF feeds frozen CLIP frame features to a Transformer trained with video-level labels. All three compute temporal change in a semantic or perceptual feature space and output a single decision for the whole video. WaveRep instead uses a VAE only to construct training data: the Pyramid Flow VAE reconstructs real videos into additional fake samples, detection is performed by a fully fine-tuned DINOv2, and the VAE is not involved at inference. In contrast, DELTA differs from these methods in its motivation and design.

DELTA is motivated by partial forgery. Real and generated content alternate within the same video, so one needs to determine not only whether a video is forged but also which segment is forged. We therefore view a video as a trajectory in the latent space of a generative model, and use the latent motion between adjacent time steps along this trajectory to characterize the temporal evolution of appearance.

DELTA is distinct in three aspects.

  1. Temporal change is computed in a generative latent space. Generative latents retain richer spatial and appearance information than semantic features, so generation traces show up in fine-grained temporal changes.
  2. The change is recorded frame by frame. Instead of compressing the whole video into one score, we record the direction and magnitude of the latent change at each frame, so the sequence shows when a change occurs.
  3. The VAE is the inference-time representation. Unlike WaveRep, which uses the VAE only to build training data, DELTA uses the frozen VAE encoder as the representation at inference.

The same representation serves both tasks: spatial transition maps with an FSQ–CNN for video-level detection, and a temporally aligned descriptor sequence for localization.

Experiments support these design choices. Wan-latent transitions reach 84.04 mAP vs. 44.78 for CLIP (Table 5), and removing the signed mean lowers mAP to 75.10 (Table 15). On fixed-FPS GenVideo, DELTA reaches 93.50 AUROC vs. 83.23 (D3) and 79.55 (ReStraV) (see our response to W3 / Q2).

We added comparisons with all four suggested methods. Results for ReStraV, D3, WaveRep, and DeCoF on the benchmarks are given in our response to W3 / Q2 and in Appendix J of the revised PDF. We have also added citations to and discussion of DeCoF, ReStraV, and WaveRep in Sec. 2.1 and Appendix H.1.

W1 & Q3 (b): About the training-free claim

We thank the reviewer for pointing out this ambiguity. In the paper, "training-free" refers to the DELTA descriptor rather than the full detection pipeline. Specifically, the descriptor is built on a pretrained and frozen Wan2.1 VAE encoder, its construction (Eqs. 1–3) involves only multi-horizon differencing and four spatial statistics with no learnable parameters (Sec. 4.1), and FSQ uses no learned codebook. Therefore, extracting the descriptor requires no labels or target data.

In contrast, the task heads are trained with supervision, namely ActionFormer for localization and a 0.206M classifier for video-level detection (Secs. 4.2–4.3, Appendix D).

To avoid ambiguity, we have revised the abstract, introduction, and contribution statement to describe this as training-free descriptor extraction using a frozen generative encoder, and to state that the task heads are supervised.

AuthorsOpenReview ↗

Response to Reviewer JKHe (2/4)

W2 & Q1: About encoder–decoder pairing

We thank the reviewer for raising the question of VAE encoder–decoder family effects. We first clarify that DELTA uses only the Wan2.1 VAE encoder, with no decoder or video reconstruction. We agree, however, that even with the encoder alone, same-family models may share correlated features that affect detection. We therefore added a per-generator analysis and the official cross-domain evaluation.

Per-generator results. The table reports mAP on the six generators of ActivityForensics. Wan and VACE use the Wan2.1 VAE, Sci-Fi uses the CogVideoX VAE, and FCVG, LTX, and Vidu belong to neither family.

MethodWanSci-FiVACEFCVGLTXVidu
ActionFormer57.3750.7774.4074.4088.2977.68
DELTA (SD1.5 VAE)39.3176.6484.3581.3792.3374.69
DELTA (CogVideoX1.5 VAE)49.4178.5087.6485.5394.4279.77
DELTA (Wan2.1 VAE)62.8081.6593.3789.2696.8086.31
  1. On Sci-Fi, which is decoded by the CogVideoX VAE, the CogVideoX encoder (78.50) still trails the Wan encoder (81.65).
  2. Among the six generators, Wan2.1 DELTA obtains its lowest mAP on the Wan generator itself (62.80).
  3. The ordering Wan2.1 > CogVideoX1.5 > SD1.5 holds on all six generators, including FCVG and LTX, which belong to neither VAE family, and the unseen generator Vidu.

The advantage of the Wan encoder is thus not confined to same-family generators, and encoder–decoder pairing alone is unlikely to explain the full performance difference. We have added this analysis to Appendix K.1.

Official cross-domain protocol. Following the reviewer's suggestion, we evaluated A→B and B→A under the official ActivityForensics cross-domain protocol (A = Wan / Sci-Fi / VACE, B = FCVG / LTX, with Vidu not part of this protocol).

MethodA→B mAPB→A mAPA→B [email protected]B→A [email protected]
ActionFormer61.1031.7832.6811.31
DELTA58.5733.8141.1621.39

At the strictest threshold, [email protected], DELTA is higher by 8.48 and 10.08 points in the two directions, consistent with its localization advantage at high tIoU in the main experiments. We have added these results to Sec. 6 and Appendix K.2.

Our benchmark and SVD-XT. None of the three domains of our benchmark involves a Wan-family generator. DELTA obtains 88.52 and 12.05 mAP on Framer and RIFE, above ActionFormer's 38.52 and 3.52, and 17.13 on SVD-XT, below ActionFormer's 21.79. Since this does not occur on Framer and RIFE, which are also non-family domains, we believe the gap on SVD-XT more likely stems from cross-manipulation transfer: the models are trained only on Framer, while SVD-XT generates conditioned only on the frame preceding the interval, so its boundary patterns may differ from those of Framer. We have added this discussion to Appendix L.3.

AuthorsOpenReview ↗

Response to Reviewer JKHe (3/4)

W3 & Q2: About the sampling shortcut on GenVideo and additional evaluations

We thank the reviewer for pointing out the potential sampling bias in the GenVideo evaluation. We agree with Michels et al. that uniform 8-frame sampling gives real and generated videos different frame intervals, so temporal differences may reflect video duration or motion magnitude rather than generation traces. The original sampling and test set (500 real and 4,551 generated videos) follow the official NSG-VD setting and were applied identically to all methods, but this does not remove the bias of the protocol itself. We therefore re-evaluated all methods with fixed-frame-rate sampling on a more complete GenVideo test set.

New protocol and test set. All methods use 8 consecutive frames at a fixed 8 FPS, with no repeated or interpolated frames. Real videos come from MSR-VTT at its original high frame rate and are shared by the ten generators. From the full GenVideo test set (9,984 real and 8,288 generated videos), we removed 691 generated videos that overlap with our validation set (the public NSG-VD validation list), leaving 9,984 real and 7,597 generated videos (17,581 in total) across ten generators.

Results on GenVideo. The table reports per-generator AUROC (%) under the fixed-FPS protocol, together with the trainable parameters of each method. DELTA performs strongly on GenVideo with a very small trainable head.

MethodParamsMSMorphMoonHotShotShow-1Gen-2CrafterLaVieSoraWildAvg
DeMamba125.71M95.7197.2298.8280.2585.4397.6694.3189.2793.8783.0391.56
NPR1.44M93.4296.3698.7471.5570.4196.9594.8683.0687.0782.4887.49
TALL86.92M88.9489.1397.3776.5887.1997.2993.4384.6489.6585.5288.97
STIL22.69M94.2097.5997.7582.5984.6598.1295.1689.2190.6385.7891.57
NSG-VD0.26M72.2972.6289.0261.6060.5979.1084.3758.1163.9071.0171.26
D3077.6979.5589.5479.9985.4290.5787.1584.3582.7475.2883.23
ReStraV0.0035M79.9270.8486.8576.8484.4893.1079.2878.0775.4470.6779.55
WaveRep86.53M98.2999.6499.9689.7093.8299.8999.9695.1988.7587.0095.22
DeCoF7.10M97.8998.3199.5695.0094.1699.8299.1092.9798.8493.0496.87
DELTA0.21M76.5098.0199.7097.2399.4098.0798.8498.6772.9695.6793.50

In particular, DELTA achieves the best AUROC on four of the ten generators (HotShot, Show-1, LaVie, and Wild).

Results on FakeParts. We also evaluated FakeParts at a fixed frame rate. Because no Interpolation or Inpainting samples satisfy the 8-frame input requirement at 8 FPS, we use 5 FPS × 8 frames to retain all eight tasks. AUROC (%) per task:

MethodTI2VT2VInterp.Extrap.Inpaint.Outpaint.FaceStyleAvg
NPR93.1888.1985.5844.1357.5567.3566.7184.0673.34
CNNDet94.5189.7169.5656.8855.5059.3882.6687.7174.49
DeMamba90.9287.2466.8948.2155.7257.7382.2289.4672.30
C2P-CLIP97.4089.9495.7953.2766.0265.7694.9691.7381.86
D363.9169.8851.8743.6950.8855.8654.1957.4255.96
WaveRep98.7688.6499.4382.7870.9086.9853.5281.0582.76
DeCoF96.5691.7561.9860.9955.0256.7178.6192.6674.28
DELTA90.4688.5590.7065.7857.2787.4754.5491.8078.32

DELTA is strong on tasks that change whole frames or temporal content (TI2V, T2V, Interpolation, and Style).

Results on ViFBench. To further examine cross-dataset generalization, we also evaluated on ViFBench with fixed 8 FPS × 8 frames. The test set covers 19 model/task settings, with 165 real and 2,995 generated videos (3,160 in total). AUROC (%):

MethodAUROC
WaveRep76.68
DeCoF72.44
NPR67.16
DeMamba64.61
TALL64.57
STIL63.03
ReStraV58.39
D354.88
NSG-VD49.97
DELTA67.34

This benchmark is challenging for all methods. DELTA still outperforms most of the compared detectors.

Overall, DELTA is competitive on all three benchmarks while training only a 0.21M-parameter head on a frozen encoder.

In the revised PDF, we have added the fixed-FPS results on GenVideo, FakeParts, and ViFBench to Appendix J.

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

DELTA proposes multi-scale generative-latent transition representation for temporal localization and video-level detection. It extracts frame-level latent representations and computes multi-scale temporal differences to capture distinct evolution patterns between real and generated videos. DELTA introduces a training-free, multi-scale generative latent difference representation that moves beyond conventional frame-level discriminative features. The experimental results demonstrate the validity of generative latent differences.

Critical strengths

1.The manuscript proposes a training-free, multi-scale temporal difference representation to distinguish real and forged videos by capturing temporal inconsistencies. 2.It introduces feature processing mechanisms based on statistical descriptors and Finite Scalar Quantization (FSQ), which is to improve the performance of detection accuracy and robustness. 3.The proposed method achieves strong performance in temporal forgery localization and cross-generator detection.

Critical weaknesses

1.The method shows performance degradation when evaluated on different manipulation methods. 2.Global statistical aggregation may suppress forgery cues when manipulated regions occupy only a small portion of the video frames. 3.The manuscript only considers first-order temporal differences, while differences between real and generated videos may also involve higher-order temporal variations.

Questions for authors

As shown in critical weaknesses part.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Libv (1/2)

We thank the reviewer for the careful reading and constructive feedback, and we are glad that the reviewer finds value in our work. We have uploaded a revised PDF with all changes marked in blue, and we address each concern below.

W1: About degradation across manipulations

We thank the reviewer for raising this issue. We agree that localization mAP drops markedly under cross-manipulation transfer, and our analysis shows that the drop comes mainly from boundary precision.

Frame-level discrimination and interval detection are largely preserved. On RIFE and SVD-XT, DELTA's frame-level AUROC remains 79.73 and 87.37 (Table 2), and at tIoU ≥ 0.5 DELTA finds 92.14% of the forged intervals on RIFE and 94.29% on SVD-XT. The transition cues learned on Framer thus transfer to unseen manipulations: the model can distinguish generated frames from real ones and find the approximate location of the forged intervals.

DomainFrame AUROC[email protected][email protected] / 0.85 / 0.95
Framer97.21100%93.05 / 91.05 / 81.47
RIFE79.7392.14%24.69 / 10.25 / 1.20
SVD-XT87.3794.29%28.91 / 18.76 / 3.70

The drop occurs mainly in boundary precision. The three manipulations are conditioned differently (Framer on the first and last frames of the interval, RIFE on the two adjacent frames outside it, and SVD-XT only on the preceding frame), so their boundary transition patterns may differ, and a localization head trained only on Framer has difficulty transferring precise boundaries. This degradation appears for all localization methods in Table 2 and under the official ActivityForensics cross-generator protocol, where DELTA still outperforms ActionFormer at [email protected] in both directions (41.16 / 21.39 vs. 32.68 / 11.31, see our response to Reviewer JKHe W2).

We have added this analysis to Appendix L.3.

W2: About global spatial aggregation and small manipulated regions

We thank the reviewer for raising this point. We agree that spatial aggregation can weaken the signal of small-area forgeries, and we added controlled experiments to quantify this effect.

Video-level detection with reduced area. On 100 videos from the same sources, we constructed RIFE forgeries with the temporal interval fixed and the manipulated region restricted to 100% / 50% / 25% of the frame, and evaluated video-level detection AUROC.

AUROC (%)100%50%25%
DELTA72.1965.0357.43
NPR49.6450.2249.96
DeMamba49.7648.0348.68

DELTA's AUROC decreases as the manipulated area shrinks, but at 100% and 50% it is clearly above both baselines (which stay near 50 at all areas).

Localization with reduced area. For the localization branch, we ran the same area control on 100 Framer videos not seen during training.

mAP (%)100%50%25%
DELTA90.5845.4610.90
ActionFormer51.7132.7613.44
TADiff37.9726.5210.47

All three localization methods degrade markedly as the manipulated area shrinks. DELTA leads at 100% and 50%. At 25%, all three fall to around 10 and DELTA shows a large drop, indicating that global spatial statistics are indeed sensitive to localized edits.

Scope and limitations. The localization task in the paper and the benchmarks used (ActivityForensics and ours) concern temporal segments in which whole frames are replaced by generated content, and DELTA's advantage holds in this setting. For spatially small localized edits, however, the current global spatial aggregation weakens the signal. In the revised PDF, we have added these experiments to Appendix M.4, listed small-area edits as a limitation, and left region-level transition descriptors to future work.

AuthorsOpenReview ↗

Response to Reviewer Libv (2/2)

W3: About higher-order temporal differences

We thank the reviewer for the suggestion. We agree that differences between real and generated videos may in principle also appear in higher-order temporal variations, and we ran a controlled experiment.

Second-order differences bring no gain. Keeping all other settings unchanged, we replaced the first-order difference with the second-order difference Δh2zt=(zt+2h−2zt+h+zt)/h2\Delta_h^2 z_t = (z_{t+2h} - 2z_{t+h} + z_t)/h^2 at each horizon h∈{1,2,4,8}h \in \{1,2,4,8\}.

Descriptor (mAP)Intra-domainVidu
First-order (DELTA)84.0486.31
Second-order81.6982.68
First + second84.1283.79

Second-order differences are below first-order differences on both test sets. We also concatenated the first- and second-order descriptors. Compared with the first-order descriptor, this is 0.08 higher intra-domain and 2.52 lower on Vidu, giving no consistent gain.

Our analysis identifies two reasons for this result:

  1. Overlapping information. DELTA's multi-horizon first-order differences already capture changes over different temporal ranges, so the information that second-order differences can add partly overlaps with them.
  2. Sensitivity to perturbations. Second-order differences are more sensitive to per-frame perturbations: if each frame's latent contains independent noise, the noise variance of the second-order difference is three times that of the first-order difference. Moreover, a jump at a boundary appears as a single peak in the first-order difference but as a positive–negative double peak in the second-order difference, which may hinder precise boundary localization.

We have added this ablation to Sec. 6 and Appendix N.

We thank the reviewer again for the comments, which have helped us improve the paper. We hope our responses address the concerns raised, and we would be grateful if the reviewer would consider updating the score accordingly. We are happy to provide any further clarification.

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

This paper addresses temporal localization of partial video forgeries and full-video detection of generated content, where existing approaches derive temporal cues from frame-level discriminative representations. It argues that isolated representation states overlook how video content evolves over time, and introduces DELTA, a training-free, multi-scale descriptor characterizing local transitions along generative latent trajectories. For temporal localization, the descriptors feed an off-the-shelf localizer; for full-video detection, continuous spatial transitions are quantized into reusable discrete patterns for robustness. The paper further contributes a benchmark with paired real sources, exact interval annotations, and source-disjoint splits for evaluating localization and cross-manipulation transfer. DELTA achieves strong performance on both tasks — under the same ActionFormer localizer, replacing CLIP features with DELTA improves mean AP from 77.68 to 86.31 on unseen generators.

Critical strengths

1.This paper proposes to model local transitions along generative latent trajectories for partial-video forgery detection.

2.DELTA realizes this as a training-free, multi-scale descriptor of structured local transitions in generative latent space, with a quantization step that maps continuous spatial transitions into reusable discrete patterns for robust full-video detection.

3.It contributes a partial-forgery temporal localization benchmark with paired real sources, exact interval annotations, source-disjoint splits, and matched interval placement between RIFE and SVD-XT, and reports strong results on both tasks — under the same ActionFormer localizer, replacing CLIP features with DELTA improves mean AP from 77.68 to 86.31 on unseen generators.

Critical weaknesses
  1. The first contribution advertises DELTA as "training-free," but this property covers only descriptor extraction: the localization head (ActionFormer) and the video-level classifier are supervised-trained on each benchmark's training split. The contribution needs further explanation.

  2. The claim of capturing temporal transition structure is not clearly verified. h=1 alone already reaches 82.56 of 84.04 mAP. How to show its effectiveness needs to be clarified.

  3. Table 5 reports that differencing Wan latents raises mAP from 56.11 to 84.04, but the cause is only described. And Wan obtains a larger gain than CLIP. The authors should clarify the reason.

  4. The video-level branch uses only seven adjacent differences from eight uniformly sampled frames — a short edited interval can fall entirely between samples. The related analysis is needed.

  5. The largest gains occur at tIoU 0.95 (45.03→73.06 intra-domain), attributed only to "more precise boundary cues." Please provide a boundary-error analysis and a control that masks interval-boundary transitions. It needs clear explanations.

Questions for authors

Please see the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer qYFo (1/3)

We thank the reviewer for the careful reading and constructive feedback, and we are glad that the reviewer finds value in our work. We have uploaded a revised PDF with all changes marked in blue, and we address each concern below.

W1: About the scope of training-free

We thank the reviewer for pointing out this ambiguity. In the paper, "training-free" refers to the DELTA descriptor rather than the full detection pipeline. Specifically, the descriptor is built on a pretrained and frozen Wan2.1 VAE encoder, its construction (Eqs. 1–3) involves only multi-horizon differencing and four spatial statistics with no learnable parameters (Sec. 4.1), and FSQ uses no learned codebook. Therefore, extracting the descriptor requires no labels or target data.

In contrast, the task heads are trained with supervision, namely ActionFormer for localization and a 0.206M classifier for video-level detection (Secs. 4.2–4.3, Appendix D).

To avoid ambiguity, we have revised the abstract, introduction, and contribution statement to describe this as training-free descriptor extraction using a frozen generative encoder, and to state that the task heads are supervised.

W2: About single vs. multiple horizons

We thank the reviewer for the comment. We agree that h=1 is the strongest single horizon, and adjacent-frame latent motion is the primary source of evidence. The value of multiple horizons is better generalization to unseen generators at a very small cost.

The gain is largest on the unseen generator. Combining h∈{1,2,4,8} raises intra-domain mAP from 82.56 to 84.04 (+1.48), and raises mAP on the unseen generator Vidu from 81.47 to 86.31 (+4.84), more than three times the intra-domain gain. Each of h=2, 4, 8 alone is below h=1 (Table 6), so the gain comes from combining horizons rather than from any longer horizon by itself. Our reading is that the pattern of adjacent-frame motion may be more closely tied to the specific generator and already suffices on generators seen in training, which is why the intra-domain gain is limited. Differences over longer horizons describe slower motion that depends less on the specific generator, and therefore add more complementary information on an unseen one.

To test this further, we compared single- and multi-horizon descriptors under the official ActivityForensics cross-domain protocol (A = Wan / Sci-Fi / VACE, B = FCVG / LTX, with disjoint training and test generators).

DescriptorA→B mAPB→A mAPA→B [email protected]B→A [email protected]
h=156.4933.3937.8421.13
h∈{1,2,4,8}58.5733.8141.1621.39

Multi-scale is no worse than single-scale in either direction: on A→B it raises mAP by 2.08 and [email protected] by 3.32, while on B→A the two are essentially on par (+0.42 / +0.26).

The cost is small. Differences for h>1 are prefix sums of the cached adjacent differences (Eq. 6), so no additional encoder passes are needed and only descriptor computation grows. We timed each stage on an A100 with one 193-frame video per generator (six videos):

Stage (s/video)h=1h∈{1,2,4,8}
Decoding + preprocessing3.3333.333
Wan encoding1.4151.415
Adjacent differences0.0210.021
Descriptor0.0210.058
Localizer0.0280.030
Total4.8184.856

Multi-scale adds only 0.037 s of descriptor computation per video, which increases the total by less than 1% (+0.79%). The time is dominated by frame decoding and Wan encoding, neither of which depends on the number of horizons. We have added these results to Sec. 6 and Appendix K.

AuthorsOpenReview ↗

Response to Reviewer qYFo (2/3)

W3: About the gains from differencing in Table 5

We thank the reviewer for the question. By Table 5, differencing raises Wan from 56.11 to 84.04 (+27.93) and CLIP from 18.99 to 44.78 (+25.79), so the absolute gains are similar and differencing is not specific to Wan. What separates the two is the representation itself, since the Wan state (56.11) already exceeds the CLIP transition (44.78).

Why differencing helps. For localization, an interval boundary is itself a change point in how the representation evolves over time, so a transition-based input corresponds more directly to the target. Fig. 4 shows that the standardized fake–real transition distance is about 4× larger inside manipulated segments than outside, and the boundary-masking control (see W5) shows that removing boundary transitions causes a large drop in [email protected].

Why Wan transitions are more useful. Wan latents are 16×28×28 maps trained for reconstruction, so their adjacent differences give the local appearance change at each position of the latent grid. CLIP embeddings are global and semantic and change little between adjacent frames, so we conjecture that less forensic information survives differencing. Consistently, transitions from all three generative encoders (SD1.5, CogVideoX1.5, Wan2.1) clearly outperform CLIP transitions.

We have added this discussion to Sec. 6.

W4: About uniform 8-frame sampling and short forged intervals

We thank the reviewer for raising this point. We agree that this is a limitation of the 8-frame video-level branch that should be addressed, and we added three analyses.

Scope of the two branches. The video-level branch targets fully generated videos (GenVideo), where every frame is generated and sampling cannot miss the forged content. Partial forgeries are handled by the localization branch, which computes transitions on every frame (all 193 frames for ActivityForensics, 35 consecutive frames at 7 FPS for our benchmark), so no interval can fall between sampled frames. FakeParts is also evaluated with the video-level branch, but it provides no exact interval annotations, so the miss rate cannot be measured directly.

Shortening the forged interval. We therefore constructed RIFE forgeries with interval lengths of 16 / 8 / 4 / 2 frames on 100 videos from the same sources, kept uniform 8-frame sampling, and report video-level AUROC (%). "Videos hit" counts videos with at least one sampled forged frame.

Length (frames)DELTANPRDeMambaDeCoFVideos hit
1672.1949.6449.7652.90100 / 100
862.6549.1849.7852.16100 / 100
457.5149.6649.9451.6085 / 100
252.6650.0750.0450.5934 / 100

DELTA's detection performance does decrease as the interval shortens and approaches chance at 2 frames, while all three baselines stay near chance at all lengths.

Increasing the number of sampled frames. On FakeParts, we keep the model unchanged, increase the input from 8 to 16 and 32 uniformly sampled frames, and compare with the 8-frame results on the same videos. The table reports AUROC on the FakeParts Interpolation task, which specifically targets partial-frame forgery.

FramesAUROCGain
877.63—
1684.33+6.70
3284.80+7.17

Sparse sampling does limit the video-level branch on short partial forgeries. Performance decreases as the interval shortens, and on the FakeParts Interpolation task, increasing the number of sampled frames to 16 or 32 improves AUROC by about 7 points. We have stated this in the Limitations and added the results to Appendix M.

In addition, our benchmark was built with short intervals in mind: each Framer interval spans only 4–14 frames, and the median interval for RIFE and SVD-XT is 8 frames (about 1.14 s). On such short forgeries, DELTA reaches 88.52 localization mAP on Framer (Table 2), and 78.19 mAP on the ActivityForensics group whose forged segments cover only 1.00%–18.30% of the video (see our response to Reviewer JKHe W4). Therefore, the frame-level localization branch handles short partial forgeries effectively.

AuthorsOpenReview ↗

Response to Reviewer qYFo (3/3)

W5: About the gains at high tIoU and boundary evidence

We thank the reviewer for this important suggestion. We added a boundary-masking control and a boundary-error analysis.

Boundary-masking control. For each temporal position and each horizon h, if the differencing window crosses the start or end of a ground-truth interval, we zero the 64 descriptor dimensions corresponding to that horizon, with the model and inference settings unchanged. As a control, for each video and each horizon we randomly zero the same number of non-boundary components and average over five random draws.

Setting[email protected]
Original73.73
Boundary zeroed2.37
Non-boundary zeroed68.48

Zeroing boundary transitions lowers [email protected] from 73.73 to 2.37, whereas zeroing the same number of non-boundary transitions lowers it only to 68.48. This indicates that localization precision at high tIoU relies mainly on transition evidence near interval boundaries, not merely on the amount of evidence removed.

Boundary-error analysis. We rank proposals by confidence, match them one-to-one to ground-truth intervals (IoU ≥ 0.5), and compare the absolute start and end errors on intervals matched by both methods. The last column reports [email protected] for DELTA and ActionFormer, respectively.

Test setDELTA error (start / end)ActionFormer error (start / end)[email protected]
Intra-domain0.090 / 0.107 s0.168 / 0.157 s99.63% / 99.07%
Vidu0.105 / 0.109 s0.132 / 0.212 s100% / 99.62%

Both methods find almost all forged intervals ([email protected] above 99% for both), and the difference lies in boundary accuracy. DELTA has smaller median boundary errors, consistent with its advantage at [email protected] in the main results.

Taken together, the two analyses indicate that DELTA's advantage at tIoU 0.95 comes mainly from more accurate start and end boundaries, which rely on transition evidence near the interval boundaries. We have added both analyses to Sec. 6 and Appendix L.

We thank the reviewer again for the comments, which have helped us improve the paper. We hope our responses address the concerns raised, and we would be grateful if the reviewer would consider updating the score accordingly. We are happy to provide any further clarification.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback. We have posted point-by-point responses and uploaded a revised PDF with all changes marked in blue. The main updates are:

  1. Scope of "training-free" clarified in the abstract, introduction, and contributions (JKHe, qYFo).
  2. Comparisons with D3, ReStraV, DeCoF, and WaveRep under fixed-FPS sampling on GenVideo, FakeParts, and ViFBench (JKHe).
  3. Per-generator and official cross-domain results on ActivityForensics (JKHe, qYFo).
  4. Boundary-masking and boundary-error analyses, with controls on interval length, manipulated area, and higher-order differences (qYFo, Libv).

As the discussion period is approaching its end, we would greatly appreciate it if the reviewers could take a look at our responses and let us know whether any concerns remain. We are happy to provide further clarification.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.