Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Rethinking the Readout".
Jane: Here's the gist of the paper. AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames A detector that captures such artifacts should therefore benefit from video pretrained backbones over image only ones In practice, however, video backbones with standard global readouts often fail to outperform strong image pretrained probes on AIGV benchmarks We attribute this gap to excessive spatiotemporal aggregation in the readout Video pretrained backbones tend to compress each frame into a single global descriptor This compression suppresses local patch level temporal dynamics and discards inter patch relations, which are precisely the cues that AIGV detection most reliably depends on Based on this, we propose Velocity Gated Patch Velocity Profiling (V-PVP), a lightweight readout that replaces only the aggregation layer with two parallel streams over the patch velocity field V-PVP serves as a general plug-and-play module that consistently improves performance across diverse video backbones under both end-to-end fine-tuning and linear probing settings Our method reaches 95.28 AUC on AIGVDBench while keeping the backbone fully frozen The results show that simply replacing the aggregation layer reactivates the temporal potential of frozen video backbones, restoring their advantage on AIGV detection Identifying the Bottleneck The authors identify the readout as an important bottleneck of video backbones in AIGV detection They provide quantitative evidence that standard aggregation discards two complementary forms of patch level evidence on which AIGV detection relies This loss is formalized as two complementary losses, namely the loss of inter patch relations and the dilution of per channel velocity magnitudes The standard readout is designed for joint spatiotemporal and semantic aggregation, which by its very design erases inter patch structure and dilutes channel level magnitudes The Proposed Solution V-PVP addresses the two losses identified above with two parallel streams over the patch velocity field One stream uses velocity gated patch attention to focus on patches with notable temporal variations The other stream aggregates patch velocities to preserve their per channel magnitudes through an element wise nonlinearity This new readout introduces only about 0.5M trainable parameters, roughly 0.57% of the backbone Mechanism of Improvement The velocity-gated aggregation z˜t conditions patch attention on v(p)t2 to carry inter patch heterogeneity into the pooled output The channel faithful stream rt preserves per channel magnitudes through an element wise absolute value before the spatial aggregation The two streams are projected independently by linear maps Wz and Wr and their outputs are concatenated at each time step Empirical Validation V-PVP consistently surpasses same backbone readouts to a large extent With VideoMAE as the backbone, our method reaches 95.28 AUC on AIGVDBench while keeping the backbone fully frozen V-PVP exceeds end to end VideoMAE fine tuning (about 87M trainable parameters) with the standard readout by +6.94 AUC This suggests that updating backbone parameters can amplify overfitting to the training generator’s spatial fingerprint rather than help Broader Implications The development of AI-generated video detection methods such as V-PVP has important societal and ethical implications Our work supports video authenticity verification as synthetic videos become increasingly realistic and easier to distribute The lightweight readout, with only about 0.5M trainable parameters, also makes recalibration more accessible to smaller research groups Future Directions A natural next step is to evaluate the method under realistic video propagation pipelines, including compression, resizing, frame-rate conversion, cropping Another promising direction is adaptive temporal evidence aggregation V-PVP may also benefit from lightweight domain adaptation, such as generator-specific recalibration, LoRA-style adapters Limitations First, generators that deliberately mimic real video statistics can shrink the inter patch separability of v(p)t2 and drive velocity gating toward uniform attention Second, very short clips with T < 4 leave too few temporal samples of patch velocity and increase the variance of z˜t and rt Finally, the readout bottleneck still exists on CNN based backbones but in a softened form Derivation Details The first stream is therefore the minimal repair operator that arises from combining Step 1, the residual decomposition of Step 2 and the two constraints on the conditioning variable The minimal repair operator for Loss 2 is therefore rt = 1/P X p v(p)t The ordering · > ReLU > x2 » id matches the prediction The identity map collapses by 59 percent, a direct empirical instantiation of the Jensen lower bound on S since opposite-signed values cancel along each channel before the absolute value is taken Experimental Setup The default V-PVP training configuration uses VideoMAE-Base pretrained on Kinetics-400 The total trainable parameters for V-PVP is 492,674 (∼0.57% of backbone) Cross-Family Testing V-PVP delivers consistent gains on CNN backbones, improving AUC by +2.91 on I3D-R50 and +2.71 on SlowFast-R50 The size of the improvement tracks how much patch level evidence survives until the readout, exactly as the readout bottleneck hypothesis predicts We identify the readout layer as the principal bottleneck that prevents video pretrained backbones from realizing their temporal advantage on AIGV detection The standard CLS or GAP aggregator performs a joint spatiotemporal and semantic collapse over patch tokens, which erases inter patch relations and dilutes per channel velocity magnitudes Both losses remove the cues that AIGV detection most reliably depends on Final Takeaway The right axis to invest capacity is the readout rather than the backbone The frozen backbone V-PVP exceeds end to end fine tuning with the standard readout at a fraction of the parameter cost References
1: OpenAI.
Tom: Video generation models as world simulators.
Jane: OpenAI Sora technical report, 2024
5: Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou
7: Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu
14: Long Ma, Zihao Xue, Yan Wang, Zhiyuan Yan, Jin Xu
33: Will Kay, Joao Carreira, Karen Simonyan
4: Weijie Kong, Qi Tian, Zijian Zhang, Rox Min
15: Shuhai Zhang, ZiHao Lian, Jiahao Yang, Daiyuan Li
16: Taehoon Kim, Jongwook Choi, Yonghyun Jeong, Haeun Noh
34: Davide Alessandro Coccomini, Nicola Messina, Claudio Gennaro
38: Christian Internò, Robert Geirhos, Markus Olhofer Page 1 --- Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection Manni Cui1, Ziheng Qin2, ZiAn Wang3 Ruiqi Liu2 Dianyuan Zou1 Jianglan Wei1 Han Zhou1 Yu Liu1 Jingrui Xu1 Wenhao Wang4,† Zhenyu Zhang1,† Pruning redundant reasoning tokens AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames A detector that captures such artifacts should therefore benefit from video pretrained backbones over image only ones How the training signal is built The paper proposes Velocity Gated Patch Velocity Profiling (V-PVP), a lightweight readout that replaces only the aggregation layer with two parallel streams over the patch velocity field This method addresses the two losses identified above with two parallel streams over the patch velocity field One stream uses velocity gated patch attention to focus on patches with notable temporal variations The other stream aggregates patch velocities to preserve their per channel magnitudes through an element wise nonlinearity Where the gains come from The two losses share the same source The standard readout is designed for joint spatiotemporal and semantic aggregation, which by its very design erases inter patch structure and dilutes channel level magnitudes Recovering both requires intervening at the readout rather than within the backbone The minimal repair operator for Loss 2 is therefore rt = 1/P X p v(p)t We identify the readout layer as the principal bottleneck that prevents video pretrained backbones from realizing their temporal advantage on AIGV detection The standard CLS or GAP aggregator performs a joint spatiotemporal and semantic collapse over patch tokens, which erases inter patch relations and dilutes per channel velocity magnitudes Both losses remove the cues that AIGV detection most reliably depends on The right axis to invest capacity is the readout rather than the backbone References
1: OpenAI.
Tom: Video generation models as world simulators.
Paper summary: Lu: To wrap up, the core finding of "Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection" is that the standard CLS or GAP aggregator in video backbones destroys both inter-patch relations and per channel velocity magnitudes.
Meng: So, if we take what they said about Loss one, that spectral condition number of the patch velocity matrix being high for generated videos means that the signal we need is gone after aggregation.
Tom: And they show this by saying that V-PVP has two parallel streams: one focusing on velocity gated patch attention to keep those inter-patch relations, and another stream preserving per channel magnitudes through an element wise nonlinearity.
Jane: The authors conclude that the right axis for investment is the readout layer rather than trying to overhaul the entire backbone structure.
Lu: They demonstrate this by showing that V-PVP reaches ninety-five point twenty-eight AUC on AIGVDBench while keeping a VideoMAE-Base backbone fully frozen, which is a strong result.
Meng: For deployment, they show that updating the backbone parameters yields less gain than just using the standard readout layer when you use V-PVP compared to end-to-end fine tuning.
Tom: So what this means for us is that we can keep our existing video backbones and just swap in this lightweight V-PVP module to get a significant boost in performance without a massive increase in training time or parameter count.
Jane: It supports the idea that smaller research groups can make real progress on these problems because they don't need to retrain the entire foundation model from scratch.
Conclusion: Tom: So, to wrap up this paper, it’s all about fixing what happens at the very end of these video models—the readout layer.
Jane: Right, so they’re saying that if you just use standard ways to summarize a whole video clip, like taking the average of everything, you lose two really important things: how different patches are from each other in time, and how strong the signal is for each color channel.
Lu: Exactly. The standard way collapses all that detail into one big number or descriptor, and that’s what kills the temporal information we need to spot those subtle artifacts in AI-generated videos.
Meng: From a deployment standpoint, losing those patch-level details means the model isn't learning the right things for real video stuff. It’s not robust enough for production where everything is messy.
Tom: So, they propose this Velocity Gated Patch Velocity Profiling, or V-PVP. It’s a small change—just swapping out that aggregation layer with two parallel streams instead of one big blob.
Jane: What does that actually mean for the listener? It means we can keep the really powerful video backbones we already have and just give them this smarter way to look at their data without having to retrain the whole thing from scratch.
Lu: They show that this works even when you freeze the backbone completely, which is huge. We got a ninety-five point twenty-eight accuracy on AIGVDBench while keeping the main model parameters totally frozen.
Meng: That’s interesting because usually, if you want that kind of performance, you have to fine-tune everything end-to-end and train millions of parameters. This cuts that training cost way down.
Tom: And they show it beats their own end-to-end fine tuning by a significant margin, about six point ninety-four AUC better on the same setup. That’s a big win for efficiency.
Jane: It really supports the idea that instead of constantly trying to build bigger and bigger backbones, sometimes making the summary layer smarter is actually more effective than just adding more layers to the backbone itself.
Lu: The authors point out that this change restores the temporal potential of frozen video backbones, which means they’re not losing what those pre-trained models were good at.
Meng: I see why they’re interested in this because it makes things much cheaper to run and adapt for smaller teams who don't have massive GPU budgets.
Tom: So the main point is that we should probably stop thinking about just the backbone structure and start focusing on how we aggregate that patch information.
Jane: It shifts our focus from building bigger models to making the summary layer smarter, which feels like a much more practical engineering move right now.
Lu: And they did flag a couple of limits, though—like if the generated video is too similar to real video, or if the clip is just too short.
Meng: Those are important caveats. We’ll need to see how this holds up when we introduce realistic compression or resizing in our pipelines before we can fully trust it for real-world use.
Tom: So, that's the high-level idea: stop looking at the whole video as one thing and start looking at the velocity of individual patches.
Jane: And that leads us perfectly into how this specific method handles those velocities, which is what we’re going to look at next on our show.
Manni Cui, Ziheng Qin, ZiAn Wang, Ruiqi Liu, Dianyuan Zou, Jianglan Wei, Han Zhou, Yu Liu
Huazhong University of Science and Technology Institute of Automation, Chinese Academy of Sciences Institute of Automation, Chinese Academy of Sciences Jilin University Vast Intelligence Lab
cs.CV
Submitted: 2026-07-16
Updated: 2026-11-02
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames A detector that captures such artifacts should…
Key concepts
- Readout Bottleneck
- The final aggregation layer in video backbones is the main problem. Standard aggregators collapse patch information into a single descriptor, which loses essential temporal cues needed to spot subtle artifacts in AI-generated videos.
- Loss of Inter-Patch Relations
- This loss occurs when the standard readout erases how different patches relate to each other in time. V-PVP addresses this by using velocity gated attention to focus on patches with significant temporal changes, preserving these structural connections.
- Dilution of Per-Channel Velocity Magnitudes
- The standard readout washes out the distinct magnitudes of velocity signals across channels. The second stream in V-PVP uses an element-wise nonlinearity to aggregate patch velocities, ensuring that the important per-channel velocity information is maintained.
Terminology
Summary
AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames A detector that captures such artifacts should therefore benefit from video pretrained backbones over image only ones In practice, however, video backbones with standard global readouts often fail to outperform strong image pretrained probes on AIGV benchmarks We attribute this gap to excessive spatiotemporal aggregation in the readout Video pretrained backbones tend to compress each frame into a single global descriptor This compression suppresses local patch level temporal dynamics and discards inter patch relations, which are precisely the cues that AIGV detection most reliably depends on Based on this, we propose Velocity Gated Patch Velocity Profiling (V-PVP), a lightweight readout that replaces only the aggregation layer with two parallel streams over the patch velocity field V-PVP serves as a general plug-and-play module that consistently improves performance across diverse video backbones under both end-to-end fine-tuning and linear probing settings Our method reaches 95. p. 0 Abstract
AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise
from inter-frame inconsistencies rather than within individual frames. A detector
that captures such artifacts should therefore benefit from video pretrained backbones over image only ones.Where the paper says it
Identifying the Bottleneck
The authors identify the readout as an important bottleneck of video backbones in AIGV detection They provide quantitative evidence that standard aggregation discards two complementary forms of patch level evidence on which AIGV detection relies This loss is formalized as two complementary losses, namely the loss of inter patch relations and the dilution of per channel velocity magnitudes The standard readout is designed for joint spatiotemporal and semantic aggregation, which by its very design erases inter patch structure and dilutes channel level magnitudes
The Proposed Solution
V-PVP addresses the two losses identified above with two parallel streams over the patch velocity field One stream uses velocity gated patch attention to focus on patches with notable temporal variations The other stream aggregates patch velocities to preserve their per channel magnitudes through an element wise nonlinearity This new readout introduces only about 0.5M trainable parameters, roughly 0. p. 1 The
new readout introduces only about 0. 5M trainable parameters, roughly 0.Where the paper says it
Mechanism of Improvement
The velocity-gated aggregation z˜t conditions patch attention on v(p)t2 to carry inter patch heterogeneity into the pooled output The channel faithful stream rt preserves per channel magnitudes through an element wise absolute value before the spatial aggregation The two streams are projected independently by linear maps Wz and Wr and their outputs are concatenated at each time step
Empirical Validation
V-PVP consistently surpasses same backbone readouts to a large extent With VideoMAE as the backbone, our method reaches 95.28 AUC on AIGVDBench while keeping the backbone fully frozen V-PVP exceeds end to end VideoMAE fine tuning (about 87M trainable parameters) with the standard readout by +6. p. 6 Under the standard readout,
some fully fine tuned video backbone in Table 1 sits below the best image level baseline, indicating
that releasing backbone parameters alone does not unlock the temporal advantage promised by video
pretraining. Our frozen backbone V-PVP exceeds end to end VideoMAE fine tuning (about 87M
trainable parameters) with the standard readout by +6.Where the paper says it
Broader Implications
The development of AI-generated video detection methods such as V-PVP has important societal and ethical implications Our work supports video authenticity verification as synthetic videos become increasingly realistic and easier to distribute The lightweight readout, with only about 0.5M trainable parameters, also makes recalibration more accessible to smaller research groups
Future Directions
A natural next step is to evaluate the method under realistic video propagation pipelines, including compression, resizing, frame-rate conversion, cropping Another promising direction is adaptive temporal evidence aggregation V-PVP may also benefit from lightweight domain adaptation, such as generator-specific recalibration, LoRA-style adapters
Limitations
First, generators that deliberately mimic real video statistics can shrink the inter patch separability of v(p)t2 and drive velocity gating toward uniform attention Second, very short clips with T < 4 leave too few temporal samples of patch velocity and increase the variance of z˜t and rt Finally, the readout bottleneck still exists on CNN based backbones but in a softened form
Derivation Details
The first stream is therefore the minimal repair operator that arises from combining Step 1, the residual decomposition of Step 2 and the two constraints on the conditioning variable The minimal repair operator for Loss 2 is therefore rt = 1/P X p v(p)t The ordering · > ReLU > x2 » id matches the prediction The identity map collapses by 59 percent, a direct empirical instantiation of the Jensen lower bound on S since opposite-signed values cancel along each channel before the absolute value is taken
Experimental Setup
The default V-PVP training configuration uses VideoMAE-Base pretrained on Kinetics-400 The total trainable parameters for V-PVP is 492,674 (∼0.57% of backbone)
Cross-Family Testing
V-PVP delivers consistent gains on CNN backbones, improving AUC by +2.91 on I3D-R50 and +2. p. 8 Finally, the readout bottleneck still exists on
CNN based backbones but in a softened form, since internal convolutions already perform part of the
spatial aggregation before the readout, partially absorbing the patch level evidence that V-PVP relies
on. Appendix C confirms this on I3D-R50 and SlowFast-R50, but a thorough study of CNN specific
readout design is left to future work.Where the paper says it
Conclusion Summary
We identify the readout layer as the principal bottleneck that prevents video pretrained backbones from realizing their temporal advantage on AIGV detection The standard CLS or GAP aggregator performs a joint spatiotemporal and semantic collapse over patch tokens, which erases inter patch relations and dilutes per channel velocity magnitudes Both losses remove the cues that AIGV detection most reliably depends on
Final Takeaway
The right axis to invest capacity is the readout rather than the backbone The frozen backbone V-PVP exceeds end to end fine tuning with the standard readout at a fraction of the parameter cost
References
[1] OpenAI. Video generation models as world simulators. p. 9 References
[1] OpenAI. Video generation models as world simulators. References
[1] OpenAI. Video generation models as world simulators.Where the paper says it
Where the paper says it
--- Page 1 ---
Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection
Manni Cui1, Ziheng Qin2, ZiAn Wang3 Ruiqi Liu2 Dianyuan Zou1 Jianglan Wei1 Han Zhou1 Yu Liu1 Jingrui Xu1 Wenhao Wang4,† Zhenyu Zhang1,†
Pruning redundant reasoning tokens
AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames A detector that captures such artifacts should therefore benefit from video pretrained backbones over image only ones
How the training signal is built
The paper proposes Velocity Gated Patch Velocity Profiling (V-PVP), a lightweight readout that replaces only the aggregation layer with two parallel streams over the patch velocity field This method addresses the two losses identified above with two parallel streams over the patch velocity field One stream uses velocity gated patch attention to focus on patches with notable temporal variations The other stream aggregates patch velocities to preserve their per channel magnitudes through an element wise nonlinearity
Where the gains come from
The two losses share the same source The standard readout is designed for joint spatiotemporal and semantic aggregation, which by its very design erases inter patch structure and dilutes channel level magnitudes Recovering both requires intervening at the readout rather than within the backbone The minimal repair operator for Loss 2 is therefore rt = 1/P X p v(p)t
We identify the readout layer as the principal bottleneck that prevents video pretrained backbones from realizing their temporal advantage on AIGV detection The standard CLS or GAP aggregator performs a joint spatiotemporal and semantic collapse over patch tokens, which erases inter patch relations and dilutes per channel velocity magnitudes Both losses remove the cues that AIGV detection most reliably depends on The right axis to invest capacity is the readout rather than the backbone
[1] OpenAI. Video generation models as world simulators. OpenAI Sora technical report, 2024 [5] Zangwei Zheng et al. p. 9 OpenAI Sora technical report, 2024. [2] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. [5] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun
Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. Hunyuanvideo: A systematic framework for large video
generative models. arXiv preprint arXiv:2412. Hunyuanvideo: A systematic framework for large video
generative models. arXiv preprint arXiv:2412. 20404, 2024. [6] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu,
Haiming Zhao, Jianxiao Yang, et al. [7] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan,
Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on
million-scale genvideo benchmark. Demamba: Ai-generated video detection on
million-scale genvideo benchmark. arXiv preprint arXiv:2405. arXiv preprint arXiv:2405. 19707, 2024. [14] Long Ma, Zihao Xue, Yan Wang, Zhiyuan Yan, Jin Xu, Xiaorui Jiang, Haiyang Yu, Yong
Liao, and Zhen Bi. Your one-stop solution for ai-generated video detection. Your one-stop solution for ai-generated video detection. arXiv preprint
arXiv:2601. arXiv preprint
arXiv:2601. 11035, 2026. Real vs fake statistics on all 63k AIGVDBench test videos. 2 Video Backbones and Temporal Readouts
Common video backbones such as TimeSformer [27], VideoMAE [28], UniFormerV2 [29], SlowFast
[30], I3D [31], and XCLIP [32] are pretrained on action recognition datasets such as Kinetics-400
[33]. Cogvideo: Large-scale
pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205. 06072, 2024. [4] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin
Li, Bo Wu, Jianwei Zhang, et al. [4] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin
Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video
generative models. 11035, 2026. [15] Shuhai Zhang, ZiHao Lian, Jiahao Yang, Daiyuan Li, Guoxuan Pang, Feng Liu, Bo Han, Shutao
Li, and Mingkui Tan. [15] Shuhai Zhang, ZiHao Lian, Jiahao Yang, Daiyuan Li, Guoxuan Pang, Feng Liu, Bo Han, Shutao
Li, and Mingkui Tan. Physics-driven spatiotemporal modeling for ai-generated video detection. Physics-driven spatiotemporal modeling for ai-generated video detection. arXiv preprint arXiv:2510. 08073, 2025. [16] Taehoon Kim, Jongwook Choi, Yonghyun Jeong, Haeun Noh, Jaejun Yoo, Seungryul Baek,
and Jongwon Choi. [16] Taehoon Kim, Jongwook Choi, Yonghyun Jeong, Haeun Noh, Jaejun Yoo, Seungryul Baek,
and Jongwon Choi. Beyond spatial frequency: Pixel-wise temporal frequency-based deepfake
video detection. Beyond spatial frequency: Pixel-wise temporal frequency-based deepfake
video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision,
pages 11198–11207, 2025. Beyond spatial frequency: Pixel-wise temporal frequency-based deepfake
video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision,
pages 11198–11207, 2025. Beyond spatial frequency: Pixel-wise temporal frequency-based deepfake
video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision,
pages 11198–11207, 2025. [8] Jianfa Bai, Man Lin, Gang Cao, and Zijie Lou. Ai-generated video detection via spatial-temporal
anomaly learning. Cogvideo: Large-scale
pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 4 ---
low-rank coherence of Vt (log σ1 − log σ10)
Cohen's d = 0.640 real (n=3000)
fake (n=60,69)
(a) Loss 1. Spectral condition number of the patch velocity matrix Vt ∈ R P ×D, defined as κspec = log σ1 − log σ10 Generated videos are systematically higher, with Cohen’s d ≈ 0. p. 2 Spectral condition number of the patch
velocity matrix Vt ∈ R
P ×D, defined as κspec =
log σ1 − log σ10. Generated videos are systematically
higher, with Cohen’s d ≈ 0.Where the paper says it
Loss Analysis
Loss 1 tests whether some choice of ω(p) within S can preserve inter-patch relations This conclusion is a structural property of the family S and holds equally for the uniform weights of GAP, the input-dependent attention weights of CLS, and the position-heterogeneous weights of fixed-query pooling Loss 2 tests whether some choice of ϕ within gϕ(v)[d] can preserve per-channel magnitudes The minimal repair operator for Loss 2 is therefore rt = 1/P X p v(p)t
Boundary Comparison of the Three Criteria
The choice · satisfies all three criteria The ordering · > ReLU > x2 » id matches the prediction The identity map collapses by 59 percent, a direct empirical instantiation of the Jensen lower bound on S since opposite-signed values cancel along each channel before the absolute value is taken
Sample-Level Correlation
The ϕ coefficient is well below the strong-correlation reference of 0.5, indicating that the errors of the two streams are weakly correlated The marginal gain is largest on Q1 and decreases toward Q4, matching the predicted division of labor that z˜ is sufficient when the Loss 1 evidence is strong and r becomes necessary when it is weak
Implementation Details
The default V-PVP training configuration uses VideoMAE-Base pretrained on Kinetics-400 The total trainable parameters for V-PVP is 492,674 (∼0.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Velocity Gated Patch Velocity Profiling (V-PVP) readout replacement allows video backbones to retain their temporal advantage on AIGV detection by addressing
excessive spatiotemporal aggregation in the readout.
This enables detectors to capturesubtle temporal artifacts that arise from inter-frame inconsistencies
such as flicker, jitter, and drift. -
The system can detect AI-generated videos with high accuracy while keeping the backbone fully frozen, achieving a mean per-generator AUC of 95.28 on AIGVDBench using only about 0.5 million trainable parameters for the readout layer. This makes recalibration more accessible to smaller research groups and public organizations.
-
V-PVP can distinguish real from fake videos by addressing two complementary losses:
the loss of inter patch relations
through velocity gated patch attention andthe dilution of per channel velocity magnitudes
via a channel faithful stream that applies an element wise absolute value before spatial summation.
Abstract
AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames. A detector that captures such artifacts should therefore benefit from video pretrained backbones over image only ones. In practice, however, video backbones with standard global readouts often fail to outperform strong image pretrained probes on AIGV benchmarks. We attribute this gap to excessive spatiotemporal aggregation in the readout. Video pretrained backbones tend to compress each frame into a single global descriptor. This compression suppresses local patch level temporal dynamics and discards inter patch relations, which are precisely the cues that AIGV detection most reliably depends on. Based on this, we propose Velocity Gated Patch Velocity Profiling (V-PVP), a lightweight readout that replaces only the aggregation layer with two parallel streams over the patch velocity field, adding only about 0.5 M trainable parameters. V-PVP serves as a general plug-and-play module that consistently improves performance across diverse video backbones under both end-to-end fine-tuning and linear probing settings. Our method reaches 95.28 AUC on AIGVDBench while keeping the backbone fully frozen. The results show that simply replacing the aggregation layer reactivates the temporal potential of frozen video backbones, restoring their advantage on AIGV detection. Code is available at https://github.com/manic-cui/Rethinking-the-readout.
Sources
- CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers
- CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer
- HunyuanVideo: A Systematic Framework For Large Video Generative Models
- Open-Sora: Democratizing Efficient Video Production for All
- Wan: Open and Advanced Large-Scale Video Generative Models
- DeMamba: AI-Generated Video Detection on Million-Scale GenVideo Benchmark
- Turns Out I'm Not Real: Towards Robust Detection of AI-Generated Videos
- Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods
- Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable
- Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection
- Your One-Stop Solution for AI-Generated Video Detection
- Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection
- Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning
- Ivy-Fake: A Unified Explainable Framework and Benchmark for Image and Video AIGC Detection
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection
- Detecting AI-Generated Images via Contextual Anomaly Estimation in Masked AutoEncoders
- MIRROR: Manifold Ideal Reference ReconstructOR for Generalizable AI-Generated Image Detection
- UniFormerV2: Spatiotemporal Learning by Arming Image ViTs with Video UniFormer
- The Kinetics Human Action Video Dataset
- Deepfake Detection Scheme Based on Vision Transformer and Distillation
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs