GVHMR — World-Grounded Human Motion Recovery

Checkpoints for GVHMR (Shen et al., World-Grounded Human Motion Recovery via Gravity-View Coordinates, SIGGRAPH Asia 2024). Given a video, GVHMR recovers SMPL/SMPL-X human motion in both the camera and world frames.

Usage

pip install gvhmr
gvhmr auth smpl          # one-time: your MPI login, to fetch the gated SMPL/SMPL-X body models

import gvhmr

pipe = gvhmr.pipeline("human-motion-recovery", model="ryanrudes/gvhmr", device="cuda")
result = pipe("dance.mp4")

result.smpl_params_world      # world-frame SMPL params (global_orient/body_pose/betas/transl)
result.joints_world           # (L, 24, 3) world-frame joints
result.render("overlay.mp4")  # in-cam ∥ world overlay video
result.save_npz("dance.npz")

Contents

file	what
gvhmr/gvhmr_siga24_release.ckpt	the trained GVHMR denoiser (the released SIGGRAPH-Asia'24 model)
hmr2/…, vitpose/…, yolo/…	preprocessing backbones (feature extractor, 2D pose, detector)

Body models are not included. SMPL/SMPL-X are registration-gated by the Max Planck Institute and cannot be redistributed — gvhmr auth smpl fetches them from the official source with your account.

License

Model weights follow the original GVHMR license (non-commercial research). SMPL/SMPL-X body models are governed by their own MPI licenses. See the linked repositories.