jepa

This repository contains minimal, educational reimplementations of a family of AI research methods called JEPA — Joint-Embedding Predictive Architectures. JEPA is an approach to self-supervised learning, which means training AI models to understand the world from raw data without needing human-labeled examples. The general idea is that the model learns by predicting parts of its input it hasn't seen, rather than reconstructing pixels directly — working in a compressed "embedding" space (a mathematical representation of meaning) instead.

The repo includes five implementations, each in a single short Python file: I-JEPA for learning from still images, V-JEPA for learning from video, V-JEPA 2 which adds action-conditioned prediction (predicting what happens after taking an action), C-JEPA which works with distinct objects tracked across video frames, and LeWorldModel which is an end-to-end world model trained directly from pixels.

Each implementation is deliberately small and self-contained — written to be read and understood, not to achieve peak performance. Each pairs with a written tutorial explaining how the algorithm works and maps to the code. The implementations use simplified datasets (like CIFAR-10 images and Moving MNIST videos) and small model sizes, not the massive compute resources the original research papers used.

This is aimed at researchers, students, and practitioners who want to understand these cutting-edge AI techniques by reading minimal working code rather than the full research implementations. The README explicitly notes where each simplified version differs from the original paper. MIT licensed.