Latent Video Prediction for World Modeling: An Evaluation Uncovering Intriguing Favorable Evidence
The study compared four video foundation models (V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2) across five robustness metrics for video world modeling. Results show that latent prediction models consistently outperform others in feature discriminability, corruption robustness, fine-grained action discrimination, occlusion handling, and temporal direction sensitivity. This suggests latent prediction is superior for robust world modeling.