Research map / Video / Video segmentation and tracking
Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes V-PVP, a lightweight frozen-backbone readout that replaces global GAP/CLS pooling over video patch tokens with two streams over the patch velocity field (velocity-gated attention and channel-faithful absolute-velocity pooling), reaching 95.28 mean AUC on AIGVDBench with VideoMAE.
Problem
Video pretrained backbones with standard global readouts (GAP/CLS) fail to outperform image-pretrained probes for AI-generated video detection, because aggregation suppresses patch-level temporal artifacts and inter-patch relations.
Approach
Keep the video backbone frozen and replace only the aggregation layer with two parallel streams over patch velocities: a velocity-gated learnable attention pooling conditioned on patch velocity magnitude, and a mean-of-absolute-velocity stream that avoids sign cancellation before pooling.
Key results
- V-PVP with frozen VideoMAE reaches 95.28 mean per-generator AUC on AIGVDBench open-source split, first among compared methods.
- V-PVP reaches 93.75 AUC on GenVidBench-143k, ranking first.
- Frozen V-PVP (~0.5M trainable params) exceeds full fine-tuning of VideoMAE with standard readout (88.34) by about 6.9 AUC.
Surprising findings
- Image-pretrained-style detectors and fully fine-tuned video backbones with standard readouts can underperform frozen image probes, indicating readout rather than representation is the bottleneck.
- Heads with more spatial capacity (Conv2D, Conv3D, PatchMLP) generalize worse across generators than the structured velocity-gated readout.
Limitations
- Generators that mimic real video statistics (e.g., injected jitter, preserved high-frequency detail) may reduce velocity-based separability and push gating toward uniform attention.
- Very short clips (T<4) provide too few temporal samples, increasing variance.
- Gains are smaller on CNN backbones (I3D-R50, SlowFast-R50), and frozen tokens may be too weak when the backbone's pretraining domain is distant from the target.
Experimental setup
Datasets: AIGVDBench GenVidBench-143k Kinetics-400
Benchmarks: AIGVDBench GenVidBench-143k
Metrics: mean per-generator AUC
Baselines: NPR UnivFD Effort DDA MIRROR DeMamba DeCoF ReSTrav NSG-VD I3D SlowFast TimeSformer
Builds on: VideoMAE AIGVDBench protocol
Closest papers by meaning
- Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection (NeurIPS 2025 · similarity 0.87)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.84)
- AI-Generated Video Detection via Perceptual Straightening (NeurIPS 2025 · similarity 0.84)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.83)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.83)
- Smooth Regularization for Efficient Video Recognition (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.