Note V-JEPA 2 ViT-L encoder (64 frames, 8192 tokens) + masked-latent predictor
Note V-JEPA 2 ViT-L SSv2 fine-tune (174-way action classification)