Skip to content
← Models
Open weightsZehong Shen, Huaijin Pi, Yan Xia et al. (Zhejiang University, State Key Lab of CAD&CG)/10 trustVerified

GVHMR

Monocular video to world-grounded SMPL motion: gravity-view coordinates recover where the person actually walked.

Visit official page →

About

Most 3D human-recovery models output people in camera space, which means the person slides around whenever the camera moves. GVHMR fixes this with gravity-view coordinates: a representation of human motion relative to gravity rather than to the camera, letting it recover SMPL/SMPL-X motion in both camera and world frames from a single video.

The SIGGRAPH Asia 2024 paper reports state-of-the-art results on the EMDB and RICH benchmarks for world-grounded motion, and the code ships with visual-odometry camera estimation so it works on moving-camera footage. A journal extension is slated for TPAMI 2026. The license is research/non-profit only, and the checkpoint downloads separately via Google Drive.

Specs

Context window

Video input (monocular RGB; SMPL/SMPL-X motion output)

Pricing

Free for research/non-profit use; commercial use needs separate permission

Free quota

N/A (self-hosted)

API access

No hosted API — self-hosted code; checkpoints via Google Drive; community Colab and Hugging Face demos exist

Weights

Open weights

Released

Not disclosed

Benchmarks

Provider-claimed (SIGGRAPH Asia 2024 paper): state-of-the-art on EMDB and RICH datasets for world-grounded human motion recovery. No independent published benchmarks were found.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

video-motion-captureworld-grounded-motion-recoverysmpl-motion-estimationcharacter-animationsports-motion-analysis

Trust score — 7/10

/10 trustVerified

Scored 7/10 (verified, metrics-v1). Evidence: peer-reviewed SIGGRAPH Asia 2024 paper (arXiv 2409.06662) by named Zhejiang University researchers; official zju3dv/GVHMR code is public with a Colab and Hugging Face demo. Strengths: strong venue, genuinely novel representation, public code and demos. Docked for: (1) license permits only educational/research/non-profit use — no commercial use; (2) accuracy figures are provider-reported with no independent benchmarks. License: custom research/non-profit-only license (commercial needs permission).

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
5.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · SIGGRAPH Asia 2024 (Shen, Pi et al.; Zhejiang University); gravity-view coordinates for world-grounded SMPL/SMPL-X recovery.
  • · Provider-claimed SOTA on EMDB and RICH; no independent benchmarks.
  • · Checkpoint via Google Drive; SMPL/SMPL-X body models fetched separately; TPAMI 2026 extension announced.

Sources