Reproduce “DAVID-XR1: Detecting AI-Generated Videos with Explainable Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifeng Gao, Yifan Ding, Hongyu Su, Juncheng Li, Yunhan Zhao, Lin Luo, Zixing Chen, Li Wang and 4 more
DAVID-XR1 from Fudan University presents an explainable method for detecting AI-generated videos, utilizing a new dataset, DAVID-X, which includes spatio-temporal defect annotations and natural language explanations. The model achieved 76.7% detection accuracy on videos from unseen AI generators and provided interpretable, localized explanations for its verdicts, showing improved performance over its backbone model and other commercial vision-language models.
Key insights
- Creating a dataset with fine-grained defect-level, spatio-temporal annotations and natural language explanations is essential for developing explainable AI-generated video detection systems.
- Distilling visual Chain-of-Thought (CoT) from powerful pre-trained models effectively guides smaller vision-language models to produce interpretable reasoning paths.
- Jointly optimizing language modeling and classification objectives during fine-tuning ensures that the model's internal reasoning process is strongly linked to its detection decision, enhancing both accuracy and interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.