1 articles with this tag
LeVJEPA trains a single video encoder with SIGReg, matching V-JEPA 2 at 5.6 to 20.8x less compute and beating image-pretrained DINOv2 on motion by ~2x.
