Full Body: Multi-HMR2

Full Body: Multi-HMR2 is the newest and technically most interesting estimator in this pack, and it's a different beast from its predecessor. Multi-HMR2 (NAVER, 2026) is DETR-based and camera-centric: it detects people in the image and recovers each one in scene-camera space, then the pack bridges the best one into SMPL-X. "Best one" here means the largest person in view - the node prints how many people it found, and the biggest gets the SMPL-X treatment.

The bridge is the clever part, and it's why this node exists rather than just being "Multi-HMR again." Multi-HMR2 doesn't predict SMPL-X at all - it predicts the Anny body model. So the pack re-poses an Anny model built with SMPL-X topology (the two share 1:1 vertices, 10475 of them), producing a registered target mesh, then fits SMPL-X parameters to that mesh with dense vertex and joint losses - roughly 9mm mean vertex error on the demo image. The payoff is that the betas are real: fitted shape, not a neutral default, and the hands get fitted too if you ask for it.

Inputs and outputs

Required:

• model - the MULTIHMR2_MODEL from Load Multi-HMR2.
• image - your photo.
• det_thresh - detection confidence, default 0.4 (range 0.05–0.95). Higher than Multi-HMR's default because multi-person detection benefits from a bit more selectivity. Tune down if your subject is missed.
• fit_hands - boolean, default on. Fits finger pose from the predicted mesh. It's genuinely slower, so if you're iterating on pose, this is the knob to flip off until you're done.

Optional:

• smplx_hands - from Hand: WiLoR, overrides the fitted hand pose (wrist-relative).
• smplx_face - from Face: SMIRK. This one matters more here: Multi-HMR2 leaves the face flat - jaw and expression aren't fitted. Graft SMIRK in or pose it in the editor, or you get a blank face on top of an otherwise great body.

Outputs: smplx (the full parameter bundle, feeding the SMPL-X Editor) and a preview image.

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

The multihmr2.pt checkpoint auto-downloads from Naver on first load (needs wget + internet), and the first run builds Anny's blend-shape caches for ~2 minutes. The registration-walled SMPLX_NEUTRAL.npz is still on you. Full rundown in the Load Multi-HMR2 and Load SMPLx articles.

Common issues

• First run looks hung. That ~2-minute Anny cache build on first load is normal; subsequent runs are fast.
• Wrong person. Largest-person bias means a big foreground subject wins. Crop to your target or raise det_thresh.
• Extreme body shapes. Anny covers all ages including infants, and extreme phenotypes sit outside SMPL-X's shape space - the betas fit is best-effort there. The source says so plainly; don't fight it.
• Blank face. Flat by design - graft SMIRK or edit expression in the editor.

This is the path when you want the newest recovery plus real fitted shape and fingers, and you're okay with a slightly slower, heavier pipeline. The camera-awareness also makes it the one whose output geometry feels most grounded when you orbit it in the editor.