acceptodds
Withdrawn from ICLR 2027

Universal Representations for Robust AI-Generated Video Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Methods for detecting AI-generated video fall into three long-standing categories: low-level analysis of generation artifacts, high-level tests of physical plausibility, and learned classifiers over deep embeddings. Detectors in each category generalize poorly beyond their training distribution, and, we argue, published comparisons between them are confounded by backbone size discrepancies. We present, to our knowledge, the first architecture that explicitly unifies all three categories of cues—density-based anomaly scores over frozen encoder features, fused with spectral and motion-coherence streams—and name it SEAM. Trained on a single generator from AIGVDBench, SEAM reaches 0.92 AUC on 19 unseen open-source generators, beating the state-of-the-art (SOTA) performance, and matching SOTA performance (0.87 AUC) on 11 unseen commercial generators, including Sora, Kling, and Pika. We further introduce SLOP, a dataset of 6,000 real and AI-generated videos from Instagram, TikTok, and X, and a taxonomy of such content. SEAM achieves the best in-domain performance on SLOP (0.96 AUC), and a SEAM trained on SLOP transfers to unseen commercial generators better than any baseline trained on the same data (0.89 AUC).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deepfake Detection, Multimodal Representations, Media Integrity

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VySxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SEAM, which combines spectral, semantic, and temporal anomaly cues for AI-generated video detection. It also introduces SLOP, a social-media dataset, and evaluates cross-generator generalization and robustness to post-processing.

For me, this paper is an unfinished work, lacking convincing motivation, coherent reasoning, and experimental validation.

Critical strengths

I found no strengths in this paper.

Critical weaknesses

W1 — The writing of this paper is terrible. For example, the Introduction devotes considerable space to describing the generalization problem in forgery detection without a coherent and well-justified explanation of the method’s motivation and design, while the Method section provides unnecessary details. These issues make the paper read more like a technical report than an academic paper. There are also numerous formatting errors. For example, Tables 1, 2, and 4 extend beyond the page margins.

W2 — The methodological novelty is limited. Four-stream feature extraction, Gaussian fitting, and MLP classifiers are all common techniques in forgery detection, while the design of the statistical descriptors is confusing.

W3 — The paper’s technical content and experimental evaluation are thin. The authors claim that their method is inspired by STALL but provide no comparison with it, and the robustness experiments include only a few baselines. Moreover, SEAM suffers a severe performance drop in Table 5, which contradicts the authors’ claims of strong performance.

Questions for authors

See weaknesses.

Reviewer_HA6JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose SEAM and SLOP. Experiments across existing benchmarks and the proposed dataset demonstrate competitive cross-generator generalization, with the strongest improvements observed on unseen open-source generators. However, the paper's methodological contribution and the necessity of its architectural components remain insufficiently justified, while the experimental results raise questions about the claimed robustness and generalization capabilities.

Critical strengths
  • The authors evaluate their approach against a range of competing detectors across in-distribution, unseen generator, and cross-dataset setting
  • SLOP provide a potentially useful evaluation resource by incorporating social media content that has undergone platform specific processing.
  • SEAM achieves strong generalization performance on unseen open-source generators
  • The frozen encoder design allows domain adaptation by updating the density estimators and lightweight classification head rather than retraining the entire backbone
Critical weaknesses
  • Lack of novelty. The proposed method primarily combines existing works. While combining complementary signals is reasonable, the paper does not sufficiently explain why this particular combination is necessary.
  • Ablation results do not convincingly support the necessity of all proposed components. For examples, Table 3 shows that combining semantic and physical features achieves an AUROC of 0.925 on OOD-C slightly higher than the full configurations' 0.923. A more detailed analysis of the complementary information provided by each component would strengthen the architectural justification.
  • Table 7 shows that SEAM trained on AIGVDBench achieves only 0.569 AUROC when evaluated on SLOP, suggesting clear domain shift. After training on SLOP, it's performance increases to 0.957 but Effort also achieves 0.954 when trained on the same dataset suggesting that the performance improvement may be driven substantially by access to domain-specific training data rather than the proposed method.
Questions for authors

See weaknesses

Reviewer_YZ6cOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper address the problem of AI generated video detectors of different domain such as real world post processing and social media content.

the paper also explain the the detector of today devide into 3 categories: low -level artifact detectors, semantic plausibility detectors and embedding based detectors.

it intrduce the unified detector "SEAM" that aim to merge all three categories into one pipeline.

in addition the authors created a dataset of 6k video from social media called "SLOP"

The paper do validation on AIGVDBench, trained on single generator(open-sora) SEAM report 0.923 AUC of 19 unseen open source generators and 0.871 of 11 commercial generators

ablation: they show the semantic and pysical streams carry most of the performance.they also show that open vision outperform dinoV2 /V3.

Robustness: they paper show robustness to re-comression (H.264)

Critical strengths

1.Dataset. this topic is very problematic in terms of datasets, the ai generated video is very wide and contrain many different kind of videos that REAL/FAKE is not clear. in addition the generatros themself are getting better and better and bench marks from one or two years ago are seem unrealavant now. so more dataset specially from social media is a good start. 2.Match backbone size in the comparision is costly but important for good comparision. 3.Open source generalization. SEAM reached 0.923 AUC on 19 unseen open source generators, with big margin on the next best

Critical weaknesses

1.Uniform 8-frame sampling underdetermine the temporal stream : SEAM sample 8 frame uniformly across the whole clip. for example in SLOP duration range betwen few second and several minuts. this gap between sample "consecutive" frames varies by orders of magnitude, so the temporal descriptors measure cant measure "physical plausibility" 2.SLOP construction not "solve" the diverse of real/fake videos. when real videos come largely from celebrity and professional creators and ai video labels come from AI-Persona it is very biased distribution and not solving the dataset problem need to this research area. 3.the paper claim "Rubust" are overstated based on the small amount of validation that was done in this paper 4.Novelty. this paper novelty is limited when you use 3 kind of AI-detectors from three solution into one pipeline. 5.writing quality and presentation are low. (e.g. "we find this combination...exciting(?)") ,result paragraph restate table values instead of interpreting them. there are ref errors "Popov & tarnopolsky 2023 a trilayer-graphene physics paper, is cited for physical implausibility in generated media (llm....), fig 1. labels the fusion block as a "spectral extractor", etc etc

Questions for authors

The Weaknesses section go beyond what can be resolve through clarrification in a rebuttal.

but for feadback and future research i would like to hear: 1.Results with fixed-stride temporal sampling, together with an analysis controlling for clip duration (given the class-dependent duration differences in Fig. 4). 2.Evidence that SLOP’s labels are not confounded by source style (celebrity/creator reals vs. AI-persona fakes), e.g., matched-source controls or per-category results. 3.Mean ± std over multiple seeds, with a clear statement of which model variant produces each reported number.

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

Related papers

Discussion (0)

Sign in to comment.