Breaking

Fast SAM 3D Body Accelerates Meta's Full-Body Mesh Model by 10x, Hitting 65ms Per Frame

July 13, 2026 at 07:57 EDT

Open Source
Robotics
Research & Papers

A new open-source framework called Fast SAM 3D Body claims a 10.25x end-to-end speedup over Meta's SAM 3D Body model, bringing single-image full-body human mesh recovery down to roughly 65 milliseconds per frame and putting real-time humanoid teleoperation within reach of a single RGB camera.

ECCV 2026 · Open-Source Release

Fast SAM 3D Body: real-time human mesh recovery from one RGB camera

A training-free framework rebuilds Meta's SAM 3D Body pipeline for a 10.25× end-to-end speedup — cutting single-image full-body reconstruction to about 65 ms per frame and putting vision-only humanoid teleoperation within reach.

10.25×

end-to-end pipeline speedup (up to 10.9×)

65ms

per frame on an RTX 5090

80%

task success on Unitree G1 demo

Speed per frame — original vs. reconstructed

Same unit (relative frame time). Shorter is faster.

~10.25×

SAM 3D Body
several sec / frame

1×

Fast SAM 3D Body
~65 ms / frame

The single biggest gain

Replacing iterative optimization with a feedforward neural mapping for the Momentum Human Rig → SMPL step yields a reported speedup of more than 10,000× on that stage alone — with fidelity kept roughly on par, and exceeding the original on at least one benchmark.

Detect → Encode → Decode → Rig → SMPL mapping

Why it matters

Vision-only — no IMU sensors needed
Drives a robot in real time from commodity cameras
Training-free: torch.compile + optional TensorRT, MIT license

Still open

Limited long-term evaluation outside demos
Accuracy in crowded multi-person scenes untested
TensorRT setup overhead not yet reviewed

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.