- 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
- Substantial gains in boundary precision with the localizer itself held fixed, which cleanly isolates the contribution of the proposed representation.
- 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
- 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.
- 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.
- 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.
- 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
- 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)?
- 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.
- 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
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.
| Method | Wan | Sci-Fi | VACE | FCVG | LTX | Vidu |
|---|---|---|---|---|---|---|
| ActionFormer | 57.37 | 50.77 | 74.40 | 74.40 | 88.29 | 77.68 |
| DELTA (SD1.5 VAE) | 39.31 | 76.64 | 84.35 | 81.37 | 92.33 | 74.69 |
| DELTA (CogVideoX1.5 VAE) | 49.41 | 78.50 | 87.64 | 85.53 | 94.42 | 79.77 |
| DELTA (Wan2.1 VAE) | 62.80 | 81.65 | 93.37 | 89.26 | 96.80 | 86.31 |
- On Sci-Fi, which is decoded by the CogVideoX VAE, the CogVideoX encoder (78.50) still trails the Wan encoder (81.65).
- Among the six generators, Wan2.1 DELTA obtains its lowest mAP on the Wan generator itself (62.80).
- 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).
| Method | A→B mAP | B→A mAP | A→B [email protected] | B→A [email protected] |
|---|---|---|---|---|
| ActionFormer | 61.10 | 31.78 | 32.68 | 11.31 |
| DELTA | 58.57 | 33.81 | 41.16 | 21.39 |
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.
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.
| Method | Params | MS | Morph | Moon | HotShot | Show-1 | Gen-2 | Crafter | LaVie | Sora | Wild | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DeMamba | 125.71M | 95.71 | 97.22 | 98.82 | 80.25 | 85.43 | 97.66 | 94.31 | 89.27 | 93.87 | 83.03 | 91.56 |
| NPR | 1.44M | 93.42 | 96.36 | 98.74 | 71.55 | 70.41 | 96.95 | 94.86 | 83.06 | 87.07 | 82.48 | 87.49 |
| TALL | 86.92M | 88.94 | 89.13 | 97.37 | 76.58 | 87.19 | 97.29 | 93.43 | 84.64 | 89.65 | 85.52 | 88.97 |
| STIL | 22.69M | 94.20 | 97.59 | 97.75 | 82.59 | 84.65 | 98.12 | 95.16 | 89.21 | 90.63 | 85.78 | 91.57 |
| NSG-VD | 0.26M | 72.29 | 72.62 | 89.02 | 61.60 | 60.59 | 79.10 | 84.37 | 58.11 | 63.90 | 71.01 | 71.26 |
| D3 | 0 | 77.69 | 79.55 | 89.54 | 79.99 | 85.42 | 90.57 | 87.15 | 84.35 | 82.74 | 75.28 | 83.23 |
| ReStraV | 0.0035M | 79.92 | 70.84 | 86.85 | 76.84 | 84.48 | 93.10 | 79.28 | 78.07 | 75.44 | 70.67 | 79.55 |
| WaveRep | 86.53M | 98.29 | 99.64 | 99.96 | 89.70 | 93.82 | 99.89 | 99.96 | 95.19 | 88.75 | 87.00 | 95.22 |
| DeCoF | 7.10M | 97.89 | 98.31 | 99.56 | 95.00 | 94.16 | 99.82 | 99.10 | 92.97 | 98.84 | 93.04 | 96.87 |
| DELTA | 0.21M | 76.50 | 98.01 | 99.70 | 97.23 | 99.40 | 98.07 | 98.84 | 98.67 | 72.96 | 95.67 | 93.50 |
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:
| Method | TI2V | T2V | Interp. | Extrap. | Inpaint. | Outpaint. | Face | Style | Avg |
|---|---|---|---|---|---|---|---|---|---|
| NPR | 93.18 | 88.19 | 85.58 | 44.13 | 57.55 | 67.35 | 66.71 | 84.06 | 73.34 |
| CNNDet | 94.51 | 89.71 | 69.56 | 56.88 | 55.50 | 59.38 | 82.66 | 87.71 | 74.49 |
| DeMamba | 90.92 | 87.24 | 66.89 | 48.21 | 55.72 | 57.73 | 82.22 | 89.46 | 72.30 |
| C2P-CLIP | 97.40 | 89.94 | 95.79 | 53.27 | 66.02 | 65.76 | 94.96 | 91.73 | 81.86 |
| D3 | 63.91 | 69.88 | 51.87 | 43.69 | 50.88 | 55.86 | 54.19 | 57.42 | 55.96 |
| WaveRep | 98.76 | 88.64 | 99.43 | 82.78 | 70.90 | 86.98 | 53.52 | 81.05 | 82.76 |
| DeCoF | 96.56 | 91.75 | 61.98 | 60.99 | 55.02 | 56.71 | 78.61 | 92.66 | 74.28 |
| DELTA | 90.46 | 88.55 | 90.70 | 65.78 | 57.27 | 87.47 | 54.54 | 91.80 | 78.32 |
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 (%):
| Method | AUROC |
|---|---|
| WaveRep | 76.68 |
| DeCoF | 72.44 |
| NPR | 67.16 |
| DeMamba | 64.61 |
| TALL | 64.57 |
| STIL | 63.03 |
| ReStraV | 58.39 |
| D3 | 54.88 |
| NSG-VD | 49.97 |
| DELTA | 67.34 |
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.
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.
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.
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.
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.