V-JEPA 2

World models · 2025

The idea

Can the JEPA bet survive video - motion, causality, object permanence - at the scale of internet footage?

V-JEPA brought joint-embedding prediction to video. V-JEPA 2 (Assran et al., 2025, arXiv:2506.09985) scales it: pretrain on more than a million hours of video (plus images), then evaluate the frozen encoder. Reported headlines include strong motion understanding (77.3 top-1 on Something-Something v2) and state-of-the-art action anticipation on Epic-Kitchens-100 (39.7 recall-at-5), with a further LLM-aligned variant for video question answering at the 8B scale.

The recipe is still the scaffolding JEPA inherited: an EMA target encoder, a predictor, stop-gradient. Collapse is prevented by asymmetry between branches, not by an explicit distributional constraint. That is exactly what LeJEPA and LeVJEPA later delete.

The misconception to name

Threads often call V-JEPA 2 “a world model” as if the pretrained encoder already plans. It does not. Phase one is an action-free video representation learner. The planner is phase two - a separate action-conditioned predictor trained on robot data - on the next page.

Verdict

What it fixedSelf-supervised video at internet scale: strong frozen motion understanding and a reusable encoder for a later action-conditioned planner.

What it still costsStill uses teacher-student scaffolding (EMA target, predictor, stop-gradient). The encoder alone does not plan - that is the next page.

The story, animated

Same scaffolding as JEPA, now fed by video - then the one misconception this page exists to correct.

Where it sits

Worth having first: JEPA

Unlocks: V-JEPA 2-AC · LeVJEPA

Source-checked against Assran et al., V-JEPA 2 (arXiv:2506.09985). Code: facebookresearch/vjepa2.