VideoScore
Building automatic metrics for video generation quality via fine-grained human feedback. Covers visual quality, motion smoothness, text alignment, and factual consistency.
VideoScore is an automatic evaluation metric for AI-generated videos, trained on fine-grained human annotations. It covers five quality dimensions: visual quality, temporal consistency, dynamic degree, text-video alignment, and factual consistency.
Key contributions:
- EvalVid-QA: large-scale human-annotated video evaluation dataset
- VideoScore model trained to predict multi-dimensional quality scores
- High correlation with human judgments across diverse video generation models
- Enables scalable automated evaluation without human involvement