BLOG ILLUSTRATION / EXECUTION STORY

V-JEPA predicts held-out features

The context and target encoders operate in parallel. The predictor gives features for comparison with stop-gradient target features. The training update uses context-encoder parameters to update target-encoder weights with an exponential moving average (EMA). Target-position selection aligns the two feature sets.

V-JEPA predicts held-out featuresThe context and target encoders operate in parallel. The predictor gives features for comparison with stop-gradient target features. The training update uses context-encoder parameters to update target-encoder weights with an exponential moving average (EMA). Target-position selection aligns the two feature sets. Context branch Target branch Video context Masked clip Context encoder Trainable weights Predictor Mask positions Full clip Same video Target encoder EMA / stop-grad Target selection Held-out positions Feature loss EMA
V-JEPA learns to predict held-out target embeddings from visible video context. The target encoder uses an exponential moving average (EMA), and its features are stop-gradient targets. Author-redrawn mechanism from Bardes et al., without token grids and implementation shapes. Source: Bardes et al. Playback time is illustrative and is not measured runtime. Open full-size SVG