keon

keon / jepa

Public

implementing minimal versions of joint-embedding predictive architecture (JEPA)

jepa pytorch representation-learning self-supervised-learning world-models
43
2
100% credibility
Found May 14, 2026 at 43 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A set of standalone scripts that demonstrate self-supervised AI models for predicting missing parts in images and videos using simple toy examples, including visualizations and step-by-step tutorials.

How It Works

1
🔍 Discover jepa

You stumble upon this fun collection of simple lessons that teach AI how to predict hidden parts of pictures and videos.

2
💻 Set up on your computer

Download the files and prepare a quiet spot on your machine where everything runs smoothly on its own.

3
Choose your adventure
🖼️
Image prediction

Start with still pictures like colorful objects to learn basic guessing.

🎥
Video prediction

Dive into moving clips like dancing digits for dynamic forecasting fun.

4
🚀 Launch the lesson

Hit go and watch the AI train itself, filling in blanks step by step – it feels like magic unfolding.

5
📊 Explore the visuals

Open the matching picture-maker file to see colorful charts, animations, and how well it learned.

🎉 Master AI prediction

You've successfully taught an AI to imagine hidden worlds in images and videos, ready to read the tutorials for deeper insights.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 43 to 43 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is jepa?

This repo delivers minimal PyTorch scripts for training JEPA models—I-JEPA on images, V-JEPA and V-JEPA 2 on videos, plus C-JEPA and LeWorldModel—self-supervised predictors that learn embeddings by forecasting masked patches or trajectories. Run a single command like `python ijepa.py` to train on CIFAR-10 or Moving MNIST; data auto-downloads, works on CPU/GPU/MPS. Extras generate viz like mask grids, loss curves, and linear probes to inspect features.

Why is it gaining traction?

Zero boilerplate: standalone files with pinned torch deps, no shared utils or complex setup—fork and tweak instantly. Tutorials match code line-for-line, linking jepa paper figures, making it prime for C-JEPA GitHub hunts or JEPA 2 GitHub prototypes. Outputs like PCA evals show collapse avoidance in action.

Who should use this?

ML researchers replicating self-supervised papers on toy data, students grokking video world models, or engineers prototyping JEPA predictors before scaling. Ideal for quick experiments with jepa model GitHub baselines.

Verdict

Strong pick for learning despite 1.0% credibility score and 43 stars—mature docs via tutorials, full reproducibility, but lacks tests or big-data scale. Use for education, not prod.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.