Research map / Video / Video segmentation and tracking
Preserving Forgery Artifacts: AI-Generated Video Detection at Native Scale
Zhengcen Li, Chenyang Jiang, Hang Zhao, Shiyang Zhou, Yunyang Mo, Feng Gao, Fan Yang, Qiben Shan, Shaocong Wu, Jingyong Su
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 |
|---|---|---|---|
| process frames at dynamic native resolution up to 720p instead of random crop to 224p (ablation) | Qwen2.5-ViT · Magic Videos + GenVideo avg ACC | avg ACC: 78.06 -> 89.92 | Improved |
| use dynamic resolution instead of random resize to 224p (ablation) | Qwen2.5-ViT · Magic Videos + GenVideo avg ACC | avg ACC: 84.61 -> 88.60 | Improved |
| raise the maximum resolution from 448p to 720p (ablation) | Qwen2.5-ViT · Magic Videos (avg ACC) | ACC: 81.19 -> 83.20 | Improved |
| increase sampled frames T from 2 to 8 (ablation) | Qwen2.5-ViT · Magic Videos + GenVideo avg ACC | avg ACC: 82.93 -> 88.60 | Improved |
| full fine-tuning instead of LoRA (r=16) (ablation) | Qwen2.5-ViT · Magic Videos + GenVideo avg ACC | avg ACC: 86.84 -> 88.60 | Improved |
| LoRA instead of linear probing (ablation) | Qwen2.5-ViT · Magic Videos + GenVideo avg ACC | avg ACC: 81.26 -> 86.84 | Improved |
| higher spatial/temporal fidelity on GenVideo (low-res short videos) versus Magic (ablation) | Qwen2.5-ViT · GenVideo | avg ACC: 93.50 -> 96.64 | Mixed |
| use image-based AI-generated image detectors trained on image data for video detection | RINE, FatFormer, B-Free, Effort, WaveRep · Magic Videos | mACC: 49.47 for RINE vs 83.20 for Qwen2.5-ViT | Hurt |
| train on higher-quality generators (higher VBench score) | cross-generator detectors · Cross-validation (14 generators) | cross-validation accuracy: Pearson rho = 0.86 | Improved |
| restrict generator quality correlation analysis to six DiT-based models | cross-generator detectors · Cross-validation (DiT models) | Pearson correlation: 0.86 -> 0.92 | Improved |
| evaluate at a resolution different from that used in training | video detectors · cross-resolution evaluation | detection performance | Hurt |
Closest papers by meaning
- Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.90)
- BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation (NeurIPS 2026 · similarity 0.89)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.89)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.88)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.88)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.87)
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.
