Video Generation Evaluation
Overview
The field measures whether generated videos look good and, crucially, whether they follow prompts and depict coherent events over time. It combines perceptual metrics such as FVD and human preference ratings with benchmarks for motion, physics, geometry, long-form continuity, and audio-video alignment. Recent approaches use task-specific rubrics, vision-language model judges, and interpretable diagnostics to assess failures that aggregate scores can hide. A central challenge is that automated evaluators may disagree with people, while visual realism often fails to reflect physical reasoning, causal consistency, or successful task completion.
Latest papers
192 papers34 papers in the last four weeks, up 113% on the four weeks before. 0.2% of all new papers.