@TheTuringPost: A new paper from @ylecun, @randall_balestr and colleagues extending LeJEPA – this time to video. LeVJEPA's goal is to m…

X AI KOLs Timeline Papers
video-processing ai-research efficiency machine-learning computer-vision

Summary

LeVJEPA extends LeJEPA to video for efficient training, using one encoder and projector to process only 5% of video patches, matching or beating V-JEPA 2 with significantly less compute.

A new paper from @ylecun, @randall_balestr and colleagues extending LeJEPA – this time to video. LeVJEPA's goal is to make video training much cheaper and simpler. For this, it uses just one encoder + a small projector. The projector is thrown away afterward ↓ LeVJEPA works like this: 1. It takes the same 16-frame video clip and creates several versions of it: one full view and several cropped/augmented views. The same encoder processes all of them. 2. Then the model learns to do 2 things: - Recognize different views of the same video and keep their representations similar. - Avoid learning the same representation for everything. SIGReg (the regularization method used in LeJEPA) prevents representations from collapsing, keeping them varied. And these two design choices also make LeVJEPA more efficient: → It processes only 5% of the video patches during training. The other 95% are randomly dropped to cut compute. → Attention is causal across time: each frame uses only the current and previous frames, and the model updates its representation as new frames arrive. This matters for streaming video and world models that operate continuously and need to update their "understanding" as new observations arrive. Thanks to all this, LeVJEPA matches or beats V-JEPA 2 with 5.6–20.8× less training compute at matched epochs.
Original Article
View Cached Full Text

Cached at: 08/30/26, 12:18 PM

A new paper from @ylecun, @randall_balestr and colleagues extending LeJEPA – this time to video.

LeVJEPA’s goal is to make video training much cheaper and simpler.

For this, it uses just one encoder + a small projector. The projector is thrown away afterward ↓

LeVJEPA works like this:

  1. It takes the same 16-frame video clip and creates several versions of it: one full view and several cropped/augmented views. The same encoder processes all of them.

  2. Then the model learns to do 2 things:

  • Recognize different views of the same video and keep their representations similar.
  • Avoid learning the same representation for everything. SIGReg (the regularization method used in LeJEPA) prevents representations from collapsing, keeping them varied.

And these two design choices also make LeVJEPA more efficient:

→ It processes only 5% of the video patches during training. The other 95% are randomly dropped to cut compute.

→ Attention is causal across time: each frame uses only the current and previous frames, and the model updates its representation as new frames arrive.

This matters for streaming video and world models that operate continuously and need to update their “understanding” as new observations arrive.

Thanks to all this, LeVJEPA matches or beats V-JEPA 2 with 5.6–20.8× less training compute at matched epochs.

Similar Articles

@lukaskuhn77: We introduce LeVLJEPA: the first fully non-contrastive end-to-end vision-language pretraining method competitive with C…

X AI KOLs Following

LeVLJEPA is the first fully non-contrastive end-to-end vision-language pretraining method, achieving competitive performance with CLIP and SigLIP without negatives, temperature, or momentum encoder. It learns via cross-modal prediction with stop-gradient targets and per-modality distributional regularization, providing stronger dense semantic features for downstream tasks like VLM backbones and semantic segmentation.

The Annotated JEPA

Hacker News Top

A step-by-step annotated implementation and explanation of Joint Embedding Predictive Architectures (JEPA) for self-supervised learning, covering I-JEPA, V-JEPA, and LeJEPA.