Chapter 3 — The JEPA family, without the name soup

One surname, different tools

A workshop has a jigsaw board, a video reader, a card sorter, and a route planner. Their labels all include “JEPA.” The label alone is no help. Four questions are: What does it eat? What does it compare? Which pieces learn? What can the finished tool do?

Those four questions clear up most of the family confusion.

Meet the relatives by their computation graphs

JEPA is the broad idea: use one part of the available information to constrain or predict another part in a learned representation space. The idea does not prescribe one encoder, mask, or anti-collapse mechanism.

I-JEPA works with still images. A context encoder plus predictor estimates target-encoder representations of selected image regions. V-JEPA moves this pattern to video: a context encoder and predictor estimate masked space-time target features, while an EMA target encoder and stop-gradient make the training graph asymmetric.

V-JEPA 2 scales video pretraining. Keep its action-free encoder/predictor separate from V-JEPA 2-AC, the later model trained with robot trajectories. Evidence about planning belongs to the action-conditioned predictor and MPC loop.

LeJEPA does not mean “large JEPA.” The paper expands the name as Latent-Euclidean JEPA and describes its recipe as lean. It combines same-source view matching with SIGReg, using an explicit distribution constraint to rule out collapse instead of teacher–student asymmetry.

LeVJEPA applies that lean objective to video. Global and local views all travel through one encoder and one projector; their [CLS] embeddings enter the loss. There is no target encoder, predictor, stopped gradient, or masked-token reconstruction in that objective. The EMA encoder maintained by the training code is an evaluation copy outside the loss graph.

LeWorldModel (LeWM) adds action-conditioned latent dynamics and a planning loop. It asks, “What might follow if I take this action?” LeVJEPA asks, “How can I encode the video I am observing?” The two ideas can meet, but one is not simply the next version number of the other. A causal encoder only prevents a frame from reading future frames; it does not invent actions.

A practical name test follows. Masked target tokens + predictor + EMA teacher looks like V-JEPA. Shared encoder/projector + same-clip [CLS] matching + SIGReg is the LeVJEPA studied here.

Repair a mislabeled sketch

Suppose a diagram titled “LeVJEPA” shows online encoder → predictor → stopped-gradient EMA teacher. Cross out the predictor and target branch. Feed every view through one encoder and projector; compute invariance and SIGReg from [CLS]. Draw EMA outside the objective and label it evaluation checkpoint only.

Family records

• I-JEPA paper and Meta research page
• V-JEPA paper and Meta research page
• V-JEPA 2 paper and official project page
• LeJEPA v3
• LeVJEPA v1 and pinned repository
• LeWorldModel v3

Keep these labels straight

1. JEPA is a representation-space prediction blueprint, not one fixed graph.
2. LeVJEPA’s objective contains shared encoder/projector paths, [CLS] matching, and SIGReg.
3. Planning results from V-JEPA 2-AC or LeWM cannot be credited to LeVJEPA.

Name test

1. Does “Le” in LeJEPA mean “large”?
2. Is there a predictor in the LeVJEPA training objective?
3. Is block-causal attention enough to turn the encoder into a planner?