Research map / Video / Video-language understanding
AvatarShield: Visual Reinforcement Learning for Human-Centric Synthetic Video Detection
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add residual extractor (ablation) | AvatarShield · FakeHumanVid mean (in-domain) | AUC: 0.7989 -> 0.9381 | Improved |
| use residuals instead of raw VAE reconstructions (ablation) | AvatarShield · FakeHumanVid mean (in-domain) | AUC: 0.8609 -> 0.9381 | Improved |
| add temporal compensation reward (ablation) | AvatarShield · FakeHumanVid mean (in-domain) | AUC: 0.8451 -> 0.9381 | Improved |
| train with GRPO instead of SFT (ablation) | AvatarShield · FakeHumanVid mean (in-domain) | AUC: 0.8301 -> 0.9381 | Improved |
| add temporal compensation reward (ablation) | AvatarShield · FakeHumanVid StableAnimator (in-domain) | AUC: 0.7522 -> 0.9573 | Improved |
| add temporal compensation reward (ablation) | AvatarShield · FakeHumanVid CogVideo (in-domain) | AUC: 0.8210 -> 0.9571 | Improved |
| add residual extractor (ablation) | AvatarShield · FakeHumanVid StableAnimator (in-domain) | AUC: 0.7176 -> 0.9573 | Improved |
| AvatarShield instead of MM-Det | AvatarShield · FakeHumanVid mean (in-domain) | AUC: 0.8979 -> 0.9381 | Improved |
| AvatarShield instead of MM-Det, trained on pose data | AvatarShield · FakeHumanVid Hallo3 (cross-domain, trained on pose) | AUC: 0.7505 -> 0.8592 | Improved |
| AvatarShield instead of MM-Det, trained on text data | AvatarShield · FakeHumanVid Wanx (cross-domain, trained on text) | AUC: 0.7999 -> 0.9134 | Improved |
| AvatarShield instead of MM-Det, trained on mixed data | AvatarShield · FakeHumanVid cross-domain mean (trained on mixed) | AUC: 0.8114 -> 0.8783 | Improved |
| AvatarShield instead of RECCE | AvatarShield · FakeHumanVid audio-driven mean (Hallo3, HelloMeme) | AUC: 0.8601 -> 0.9353 | Improved |
Closest papers by meaning
- VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning (ICML 2026 · similarity 0.84)
- Reinforcement Learning for Visual Object Detection (CVPR 2016 · similarity 0.83)
- Visual Reinforcement Learning with Residual Action (AAAI 2025 · similarity 0.82)
- RLGF: Reinforcement Learning with Geometric Feedback for Autonomous Driving Video Generation (NeurIPS 2025 · similarity 0.81)
- Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content (CVPR 2025 · similarity 0.81)
- Video In-context Learning: Autoregressive Transformers are Zero-Shot Video Imitators (ICLR 2025 · similarity 0.81)
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.
