LeVJEPA Matches V-JEPA 2 With 20x Less Video Pretraining Compute

The new pretraining method replaces target encoders and reconstruction machinery with a single encoder trained using invariance and SIGReg. This end-to-end approach achieves performance equivalent to V-JEPA 2 while reducing compute costs by 5.6 to 20.8 times.

LeVJEPA utilizes a stable recipe to maintain appearance and motion understanding, even when randomly dropping 95% of video tokens. The method is open source and removes the need for masked predictions, stop-gradients, or teacher-student schedules.