What is jepa?

This repository contains minimal, educational reimplementations of a family of AI research methods called JEPA, short for 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 has not seen, rather than reconstructing pixels directly, and it does this prediction in a compressed embedding space, a mathematical representation of meaning, instead of in raw pixel space. 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, meaning predicting what happens after taking an action, C-JEPA which works with distinct objects tracked across video frames, and LeWorldModel, an end-to-end world model trained directly from pixels. Each implementation is deliberately small and self-contained, written to be read and understood rather than to achieve top performance. Every algorithm pairs with a written tutorial explaining how it works and how it maps to the code. The implementations use simplified datasets, such as CIFAR-10 images and Moving MNIST videos, and small model sizes, not the large compute resources the original research papers used. This is aimed at researchers, students, and practitioners who want to understand these 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 it is based on. The project is released under the MIT license.