V-JEPA 2 Self-Supervised Training

What problem does it solve?

Enables end-to-end setup and execution of the V-JEPA 2 self-supervised training workflow, including context/predictor architectures, EMA target encoder management, loss computation, and checkpointing for reproducible experiments.

Core Features & Use Cases

• Self-supervised training orchestration: Seamlessly train V-JEPA 2 models using latent space predictions and EMA targets.
• EMA target management: Integrates an exponentially moving average target encoder to stabilize training and prevent collapse.
• Configurable DROID support: Supports DROID-style fine-tuning with autoregressive rollout and optional normalization of representations.
• Checkpointing & resumption: Handles save/load of all components (encoder, predictor, EMA target, optimizer, scaler) for reliable experiments.
• Use Case: Prototype a JEPA training loop on synthetic data, then scale to real video datasets with progressively larger setups.

Quick Start

Run a minimal training loop with a tiny encoder and predictor to validate the JEPA workflow.