V-JEPA
A video understanding model developed by Meta that adopts the Joint Embedding Predictive Architecture (JEPA). It aims to capture structured causal relationships within video content and represents a significant exploration in world model research.
Core Facts
Timeline (last 90 days)
Models such as OpenAI's Sora and Meta's V-JEPA have demonstrated early forms of world models in video prediction and physical reasoning.
Meta AI Research has released specific implementations such as I-JEPA and V-JEPA, achieving competitive results on visual self-supervised learning tasks.
Google DeepMind's Genie series, Meta's V-JEPA, and NVIDIA's Cosmos platform are all exploring technical approaches to embedding physical constraints into generative models.
Meta's research team's V-JEPA model, Google DeepMind's Genie, and other projects are exploring ways to enable AI to build an internal simulation of the physical world.
Yann LeCun proposed the JEPA framework, whose core idea is to perform prediction in an abstract representation space rather than pixel space, and Meta has released early implementations such as V-JEPA.
V-JEPA was released by Meta AI in 2024, training models to predict masked regions in representation space by masking up to 90% of the spatiotemporal extent of a video.
All Facts (7)
Both I-JEPA (the image version) and V-JEPA (the video version) were released and open-sourced by Meta AI
80%UnverifiedModels such as OpenAI's Sora and Meta's V-JEPA have demonstrated early forms of world models in video prediction and physical reasoning.
50%UnverifiedGoogle DeepMind's Genie series, Meta's V-JEPA, and NVIDIA's Cosmos platform are all exploring technical approaches to embedding physical constraints into generative models.
50%UnverifiedMeta's research team's V-JEPA model, Google DeepMind's Genie, and other projects are exploring ways to enable AI to build an internal simulation of the physical world.
50%UnverifiedYann LeCun proposed the JEPA framework, whose core idea is to perform prediction in an abstract representation space rather than pixel space, and Meta has released early implementations such as V-JEPA.
50%UnverifiedMeta AI Research has released specific implementations such as I-JEPA and V-JEPA, achieving competitive results on visual self-supervised learning tasks.
50%UnverifiedV-JEPA was released by Meta AI in 2024, training models to predict masked regions in representation space by masking up to 90% of the spatiotemporal extent of a video.
50%