qft-333/V-JEPA-Hand

V-JEPA Hand

This repository contains the model of the paper V-JEPA Hand: Simple, Accurate and Fast 3D Hand and Camera Trajectory Recovery from Egocentric Videos.

V-JEPA Hand jointly reconstructs bilateral 3D hand motion and camera trajectories from egocentric video in a single forward pass.

Project page · Code

Video demo

After installing the inference code and preparing your licensed MANO files:

python demo.py \
  --mano-dir /path/to/mano \
  --input input.mp4 \
  --output output.mp4

The demo automatically downloads these weights and outputs a silent video with hand-mesh overlays: blue for the left hand and orange for the right hand.

See the code repository for installation instructions and the Gradio interface.

Model files

• model.safetensors: inference weights, including the fine-tuned V-JEPA 2.1 encoder.
• config.json: model configuration and weight checksum.

MANO

MANO assets are not included. Obtain them separately from the official MANO website under its license.