Load Multi-HMR2

Load Multi-HMR2 is the loader for the newest estimator in this pack - and the one with the least manual setup, because unlike every other loader here it fetches its own weights. Multi-HMR2 (the 2026 follow-up from NAVER) is a DETR-based, camera-centric multi-person model: it detects people, recovers a mesh for each in scene-camera space, and picks the largest person to bridge into SMPL-X. The "2" buys you real fitted shape parameters and a proper scene camera rather than a fixed frontal assumption.

The loader is also the simplest to use, which is a nice change of pace after fighting with registration walls. You give it two things:

• model_path - a folder for the checkpoint, ComfyUI/models/multihmr2/, expecting multihmr2.pt. Here's the trick: if the file isn't there, the loader downloads it automatically from Naver's servers on first load. It needs wget and internet access, and it's a big file, so the first run just sits there for a while. That's normal.
• smplx_model - required, from Load SMPLx. Same as every estimator: the SMPL-X body model is the substrate everything deforms.
• device - auto / cuda / cpu.

Output is a model connection (type MULTIHMR2_MODEL) that feeds Full Body: Multi-HMR2.

Installing

Pack install is the standard song: 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 install.py step clones the Multi-HMR2 source into vendor/multi-hmr2 (override with the MULTIHMR2_DIR env var) and the requirements install the anny package and warp-lang - that's the new dependency this loader drags in versus the rest of the pack.

Gotchas - and there are real ones

• First-run caches take ~2 minutes. The very first time you load, the anny package builds blend-shape caches and fetches the Anny↔SMPL-X topology data (non-commercial). The README and source both call this out: expect a couple of minutes of apparent nothing on first load, then it's cached in ~/.cache/anny and fast after.
• The auto-download is easy to miss. If wget is missing or the network is flaky, the loader fails in a way that reads like a missing-file error. Check models/multihmr2/multihmr2.pt - if it's not there, the download didn't happen, and that's your first port of call.
• Biggest-person bias. Multi-HMR2 detects multiple people but the SMPL-X bridge keeps the largest person in frame. The estimator prints how many it found, so glance at the console when the output isn't the person you wanted.
• Non-commercial weights, again. The checkpoint follows Naver's release terms and the Anny SMPL-X topology data is non-commercial. Research use only, same as the rest of the pack.

If you want the newest, camera-aware estimation with zero manual weight-downloading, this is the loader. Just budget for the silent two-minute first load and the wget dependency.