eb_jepa
An open source library for learning representations using energy-based joint-embedding predictive architectures.
Learn representations for image, video, and action-conditioned planning using JEPAs.
eb_jepa is an open source library providing community examples of Energy-Based Joint-Embedding Predictive Architectures (JEPAs) from Meta AI Research. It includes code and tutorials for learning representations from images, video, and action-conditioned video, along with planning capabilities. It suits AI researchers and civil engineering machine learning specialists building predictive world models.
What it is
The library offers self-contained examples for Image JEPA, Video JEPA, and Action-Conditioned Video JEPA. It supports training on images, predicting next-frame video representations, and world modeling with planning in multi-room environments. It includes unified folder structuring and SLURM launcher utilities for managing training and hyperparameter sweeps.
Who it's for
It suits machine learning engineers and researchers developing predictive models and learning representations from visual data. It does not suit civil engineering professionals seeking ready-to-use BIM applications or end-user software without Python programming expertise.
Key features
Image JEPA
Learn self-supervised representations from unlabeled images on datasets like CIFAR-10 with classification evaluation.
Video JEPA
Predict next image representations sequentially across video frames.
Action-Conditioned Video
Build world models and handle environment planning tasks using action-conditioned video architectures.
SLURM Integration
Launch single jobs or full hyperparameter sweeps using integrated SLURM scripts and Weights & Biases tracking.
Up and running in 3 steps
Our honest verdict
What's great
- Includes self-contained examples for image, video, and action-conditioned video JEPAs
- Unified folder structure automatically encodes key hyperparameters to avoid collisions
- Built-in SLURM launcher supports automated hyperparameter sweeps and seed averaging via Weights & Biases
- Uses modern uv package management alongside standard virtual environments
Where it falls short
- Default configurations are tuned for high-end H100 GPUs and require memory adjustments on older hardware
- Requires writing and executing Python scripts with manual environment configuration
- Dataset paths and checkpoint directories must be manually exported via shell variables
Pricing & license
Free under the Apache-2.0 license with no paid hosted options provided. The only cost is running it yourself on your own GPU compute infrastructure.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
What hardware is required to run the training examples?
Default configs are tuned for H100 GPUs, but older GPUs like A100 or V100 can be used by reducing batch sizes.
How are hyperparameter sweeps handled?
You can launch full sweeps defined in YAML configs using the SLURM launcher with the --full-sweep flag.
