VL-JEPA (Vision-Language Joint Embedding Predictive Architecture) in MLX

Reliable, minimal and scalable library for evaluate and conduct world model research

Experiments in Joint Embedding Predictive Architectures (JEPAs).

👆PyTorch Implementation of JEDi Metric described in "Beyond FVD: Enhanced Evaluation Metrics for Video Generation Quality"

This VL-JEPA implimentation takes direct insperation from the original VL-JEPA paper

A Video Joint Embedding Predictive Architecture (JEPA) that runs on a personal computer.

An open-source attempt at training a variant of LeCun's energy-based models (EBM) to reason in latent space and solve Sudoku.

A simple and efficient implementation of Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA)

Train a JEPA world model on a set of pre-collected trajectories from an environment involving an agent in two rooms.

Project for Yann Lecun's Deep Learning class. In this project, we train a JEPA world model on a set of pre-collected trajectories from a toy environment involving an agent in two rooms.

A PyTorch implementation of Latent Embedding JEPA for learning world models in continuous environments without latent collapse.

A practical explainer of JEPA, Meta AI’s Joint Embedding Predictive Architecture, with diagrams and insights comparing JEPA and Transformers.

Joint Embedding Predictive Architecture (JEPA) world model trained on agent trajectories to predict future latent states from pixel inputs and actions. Uses VICReg loss with RNN dynamics to evaluate how well learned embeddings reflect spatial behavior in toy environments.

PointJEPA-based, label-efficient 3D grasp joint-angle prediction (IROS 2025 FMRD Workshop).

This project focuses on the implementation of inverting I-JEAP, a new architecture designed to simulate human intelligence through self-supervised learning. Our goal is to invert the embeddings to demonstrate that such architectures can be vulnerable to inversion attacks

A research-oriented project focused on the practical application of the Joint Embedding Predictive Architecture (JEPA).

Demo implementations of JEPA World Models to support research

A minimal JEPA-based language model demonstrating latent-space reasoning on GSM8K using a single decoder-only Transformer.

Training backend for Cell Observatory models

Curiosity-driven Actor–Critic using JEPA latent dynamics.