V-JEPA (Video Joint Embedding Predictive Architecture)

Title: V-JEPA (Video Joint Embedding Predictive Architecture) Slug: v-jepa

V-JEPA (Video Joint Embedding Predictive Architecture) refers to a family of self-supervised video foundation models that extend the Joint Embedding Predictive Architecture (jepa) to spatiotemporal video data. Like image-JEPA, V-JEPA models learn by predicting abstract latent-space representations of one view or segment of a video from another, rather than reconstructing raw pixels. The "V" prefix marks the video domain adaptation of the JEPA framework introduced by Yann LeCun. The architecture typically combines a vision-transformer encoder that processes video patches into a latent representation, a predictor network that maps one latent to another, and (in early variants) a target encoder updated via stop-gradient or exponential moving average to prevent representational collapse.

The core mechanics involve masking a high proportion of video patches (typically 60-90%) and training the predictor to reconstruct the latent representation of the masked patches from the visible context. Unlike pixel-reconstruction objectives (as in MAE-style video models), V-JEPA's latent prediction objective lets the model discard pixel-level details that don't matter for high-level video understanding (texture, lighting, exact motion blur) and focus on the underlying spatiotemporal dynamics. The result is a model that learns strong representations for downstream tasks like action recognition, video classification, robotic manipulation, and world-modeling.

V-JEPA 2 (Assran et al., 2025) scaled the approach to billion-parameter scale and demonstrated emergent capabilities in long-horizon video prediction and embodied planning. V-JEPA 2.1 (Mur-Labadia et al., 2026) added dense feature extraction. LeVJEPA (Kuhn et al., 2026) showed that the standard JEPA heuristics (target encoder, masked prediction, stop-gradient) can be replaced with a single regularization-loss-based objective while matching V-JEPA 2's performance at up to 20x less pretraining compute. Domain-specialized variants include Drive-JEPA (autonomous driving), WA-JEPA (world-action modeling), and MC-JEPA (motion capture). V-JEPA is the video-domain sibling of jepa-joint-embedding-predictive-architecture and shares its lineage with image-jepa, mc-jepa, and other JEPA-family architectures.