Skip to content

JEPA

A joint-embedding predictive architecture encodes a context and a target, and trains a predictor to map the context's embedding to the target's. Because the loss lives in embedding space, the model spends its capacity on what is predictable instead of on texture and noise. When the context is the present, the target is the future and the predictor reads the action taken, a JEPA becomes a latent world model.

All models in this family live in torch_world.models.jepa and are exported from torch_world.models.

Images

  • I-JEPA

    Predict the features of masked image blocks from one visible context block.

    EMA target · smooth L1 · multi-block masks

  • LeJEPA

    Every view predicts the mean of the global views; SIGReg keeps embeddings Gaussian. No teacher, no stop-gradient, no schedules.

    SIGReg · multi-view

Video

  • V-JEPA

    Predict the features of masked spatio-temporal tubes. Covers V-JEPA 2 with 3D RoPE.

    EMA target · L1 · tube masks

Vision-language

  • VL-JEPA

    Predict the embedding of a text answer from visual tokens and a question; decode text only on demand.

    frozen X-encoder · InfoNCE

World models

Action-conditioned JEPAs. They subclass WorldModel, so rollout, cost and CEM planning work without extra code.

  • LeWM

    End-to-end from pixels: next-embedding MSE plus SIGReg, with an adaLN-conditioned causal predictor.

    trained encoder · SIGReg

  • DINO-WM

    Dynamics over the patch features of a frozen DINOv2 encoder, with proprioception.

    frozen encoder · MSE

Robotics

World models for robot control: several cameras, proprioceptive state, and action spaces that differ between embodiments. They are WorldModels too, and plan with the same CEM, conditioned on the robot's state and embodiment.

  • RoboJEPA

    One predictor for many robots and cameras, in the latent space of a frozen V-JEPA 2.1 encoder.

    frozen encoder · L1 + rollout

How they compare

model predicts from anti-collapse loss
I-JEPA features of masked image blocks a visible context block EMA target smooth L1
LeJEPA mean embedding of global views every view SIGReg MSE + SIGReg
V-JEPA features of masked video tubes the unmasked tokens EMA target L1 (+ variance hinge)
VL-JEPA embedding of the target text visual tokens + query contrastive InfoNCE
LeWM next frame embedding past embeddings + actions SIGReg MSE + SIGReg
DINO-WM next frame patch features past features + actions frozen encoder MSE
RoboJEPA next frame features of every view past features + actions + states frozen encoder L1, teacher forcing + rollout

Papers and code

model paper reference code
IJEPA Assran et al., 2023 facebookresearch/ijepa
LeJEPA Balestriero & LeCun, 2025 galilai-group/lejepa
VJEPA Bardes et al., 2024, Assran et al., 2025 facebookresearch/jepa, facebookresearch/vjepa2
VLJEPA Chen et al., 2026 not released
LeWM Maes et al., 2026 lucas-maes/le-wm
DINOWM Zhou et al., 2025 gaoyuezhou/dino_wm
RoboJEPA Zholus et al., 2026 facebookresearch/robo_jepa