InstantHMR Runs Real-Time 3D Human Mesh Recovery on Phone NPUs: 5.5ms per Person, 12x Faster Than Peers
naturalpad · hf · 2026-11-10
NaturalPad and LAAS-CNRS release open-source InstantHMR, a family of single-image 3D human mesh recovery regressors built for mobile inference, based on the Momentum Human Rig (MHR) parametric model and trained solely on public SAM 3D Body annotations.
- Recovers a 6,890-vertex body mesh in 5.5ms per person on a Samsung Galaxy S23 NPU (0.98ms on RTX 4070)—12x faster than the only mobile-compatible competitor, with competitive accuracy (40.0mm PA-MPJPE on 3DPW, 48.8mm on EMDB, 41.7mm on RICH).
- Re-evaluates 11 SOTA methods under a unified protocol, including Meta's official MHR-to-SMPL conversion, and benchmarks latency across 8 targets from desktop TensorRT to two phone NPUs.
- Code and models are open-sourced on GitHub.
More from Embodied
- Nvidia's GR00T 2 robots complete 97/100 toy assemblies, recovering from slips and stalls — mark_k · 2026-11-10
- Norway runs 2,000+ underwater AI laser robots delousing 50 million salmon — jjvincent · 2026-11-10
- Stanford tennis courts are so 'bottmaxxed' that bots are lapping each other — jamdac · 2026-11-10
- Fable 5.2 plays checkers with a robot arm, self-corrects a 7mm sag mid-game — socialwithaayan · 2026-11-10
- Waymo ditches Uber, consolidating robotaxi rides into its own Waymo One app — 4310sy · 2026-11-10
- 21-hour timelapse: building the Asimov humanoid robot from parts to walking — victormustar · 2026-11-10