AI & Creative Tools

A 32-Million-Parameter Model Now Captures Full 3D Body Motion From a Webcam Inside ComfyUI

InstantHMR fits a complete 3D body — pose, shape, 70 keypoints, a full mesh — from a photo, video, or live camera feed in about a millisecond per person on a desktop GPU, and it slots straight into ComfyUI's existing SAM 3D Body pipeline.

Motion capture has historically meant either an expensive suit covered in markers or a slow, GPU-hungry model that can’t run live. InstantHMR, a small model released in early November by Naturalpad Studio, aims at the gap between those two: a complete 3D human body — pose, shape, 70 2D and 3D keypoints, and a full mesh — recovered from a single ordinary camera, running in roughly 1 millisecond per person on a desktop GPU and about 5 milliseconds on a phone’s NPU. A new ComfyUI node pack, ComfyUI-InstantHMR, now puts that model directly in front of ComfyUI’s existing native pipeline for rendering and exporting that kind of body data.

What it actually recovers

InstantHMR comes in four open sizes — nano at 20 million parameters up to medium at 32.2 million — plus a medium-commercial variant trained only on footage cleared for commercial use. All of them output the same thing: a body in the Momentum Human Rig (MHR) format, which is also what ComfyUI’s native SAM 3D Body nodes already consume. That compatibility is the whole point of the integration — smoothing a shaky capture, rendering a mesh or OpenPose skeleton, and exporting an animated GLB or BVH file all keep working unchanged, because InstantHMR is just a faster, cheaper way to produce input those nodes already know how to handle.

The ComfyUI pack adds three new nodes in front of the native ones: a model loader, a prediction node that turns a person’s bounding box into a full MHR body, and a Live Capture node that runs the whole pipeline against a webcam or a phone camera streamed over the network in real time.

Live capture, not just a photo pipeline

The live-capture mode is where this gets interesting for anyone doing performance or interactive work rather than offline rendering. Point a webcam at yourself (or let a phone stream over a local QR-code pairing), and InstantHMR tracks the body continuously, driving a rigged character or recording clips as it goes — no suit, no markers, no dedicated capture volume. A smoothing parameter steadies the output without noticeably lagging fast movement, and a detect_every setting lets the slower person-detection step run less often than the pose model itself, keeping the preview fluid.

Each checkpoint trades size for accuracy in a predictable way: nano is the fastest and least precise, medium (the default) is the most accurate of the open checkpoints, and the whole lineup also supports an optional kinematic fit — a zero-parameter refinement step, run inside the same graph, that nudges the predicted pose to match the model’s own more-accurate 2D keypoints before anything downstream sees it.

Why the ComfyUI integration matters more than the model alone

A fast, open body-capture model is useful on its own, but bolting it onto ComfyUI’s existing SAM 3D Body pipeline is what makes it immediately usable rather than another research checkpoint someone has to write glue code for. Anyone already rendering MHR bodies, exporting rigs, or chaining pose data into other ComfyUI graphs gets live webcam-driven motion capture as a drop-in upstream source, for free. For solo creators and small studios specifically, that collapses a workflow that used to require either paid mocap hardware or a much heavier model into something that runs on a single GPU in real time.