Semantics-aware contrastive and explicitly disentangled representations for AI-generated video detection
Semantic-aware Contrasting with Explicitly Disentangled Representations for AI-Generated Video Detection
AI summary
The paper proposes an AI-generated video detection dataset SAVGE (2,369,822 videos) and a model-agnostic training framework EDSC. SAVGE uses real video captions as prompts to generate fake videos, forming real-fake semantic groups; EDSC suppresses semantic information in the artifact branch, making the detector focus on generation artifacts. On SAVGE and GenVideo, EDSC improves most backbone-metric combinations, enhancing robustness and cross-dataset generalization under challenging settings.