Beyond FVD: Enhanced Evaluation Metrics for Video Generation Quality
The Fréchet Video Distance (FVD) is limited by assumptions about feature space and sample size, making it unreliable for video generation evaluation. The authors propose JEDi, the JEPA Embedding Distance, which uses Maximum Mean Discrepancy with a polynomial kernel on V-JEPA features. Experiments show JEDi requires significantly fewer samples to converge and aligns better with human evaluations than FVD.