Full Body: Multi-HMR

Full Body: Multi-HMR is, bluntly, the node most people should try first in this pack. One forward pass over your image and you get a complete SMPL-X body - pose, hands, face expression, and real fitted shape parameters - with no assembly. No separate hand node, no face node, no grafting. That's the whole appeal of Multi-HMR (NAVER, ECCV 2024): it's the rare whole-body estimator that predicts everything jointly, so the hands actually integrate with the body instead of being pasted on afterward.

It's also the tolerant one. Unlike NLF, this estimator runs on CPU if you're stuck without a GPU, and its single quality knob is easy to understand. For someone who wants "photo in, posed 3D person out, minimal fuss," this is the path - which is why the pack's quick-start workflow (example_workflows/smplx_multihmr_example.json) uses it.

Inputs and outputs

Required:

• model - the MULTIHMR_MODEL from Load Multi-HMR.
• image - your photo.
• det_thresh - person detection confidence, default 0.3 (range 0.05–0.9). The knob you'll actually touch: if the person isn't detected, drop it; if you're picking up background clutter as a person, raise it. Multi-HMR detects people in the scene and estimates the dominant one.

Outputs are smplx (the full SMPL-X parameter bundle - feed it straight into the SMPL-X Editor) and a preview image (a 512px pose render for a quick check).

Installing

Standard pack install - ComfyUI Manager, search SMPLx Estimator, install, restart - or:

cd ComfyUI/custom_nodes
git clone https://github.com/ameliacode/ComfyUI-SMPLx-Estimator.git
cd ComfyUI-SMPLx-Estimator
pip install -r requirements.txt
python install.py

Weights you need by hand: multiHMR_896_L.pt in models/multihmr/ (from naver/multi-hmr - you must accept their non-commercial license first) and the registration-walled SMPLX_NEUTRAL.npz in models/smplx/ (see the Load SMPLx article). Those two missing files account for most "it won't run" reports on this path.

Common issues

• Person not detected. Lower det_thresh a notch. The estimator's console prints which side of things it found, so glance at the terminal when results are empty.
• Wrong person picked. With multiple people in frame, Multi-HMR estimates the dominant one. If it grabbed the wrong subject, crop to your target or raise the threshold to drop the competition.
• It feels slow on CPU. It runs, which is the miracle, but a full SMPL-X forward pass on CPU is not quick. GPU is the comfortable home for this estimator.
• Research-only weights. NAVER non-commercial. The pack's wrapper is MIT, but the checkpoint isn't.

If you want the least-friction route to a full expressive body - and you should want that, because it's what makes the editor and the ControlNet maps sing - start here. It's the recommended path for a reason.