(Down)Load SAM 3D Body Model

Every SAM3DBody workflow starts here. LoadSAM3DBodyModel downloads and loads Meta's SAM 3D Body checkpoint into memory and hands you back a model handle that every Process node in the pack needs. It's the loader, full stop - no image goes in, no mesh comes out, it just gets the brain into GPU memory so the actual work nodes can use it.

SAM 3D Body is Meta's late-2025 release for full-body 3D human mesh recovery from a single photo - no multi-camera rig, no turntable capture, just one image in and a posed 3D mesh out. ComfyUI-SAM3DBody, wrapped by Andrea Pozzetti, went up on GitHub and the ComfyUI Manager the same week Meta shipped it and got picked up fast on r/StableDiffusion. This node is the reason: everything downstream depends on it having successfully pulled the weights.

How it works

Drop it on the canvas, pick a precision, run it once, and wire its model output into whichever Process node you're using. It's not something you re-run per image - load it once, reuse the handle across as many SAM3DBodyProcess / SAM3DBodyProcessAdvanced / SAM3DBodyProcessMultiple calls as you want in the same session. There's no per-run cost to keeping it loaded beyond the VRAM it occupies.

The input and output that matter

There's exactly one input worth knowing: precision, a dropdown defaulting to fp32. If you're tight on VRAM, check what else is on that dropdown before assuming you need a bigger card - reported real-world usage from people running this pack is modest; one user on the release thread got it working fine on an 8GB card without it being noticeably slow.

The single output is model (SAM3D_MODEL) - a typed handle that only this pack's own nodes understand. It doesn't slot into anything from Impact Pack or other segmentation tools; it's specific to SAM3DBody's Process, ProcessAdvanced, ProcessMultiple, and AddMeshToSkeleton nodes.

How to install it

Easiest path is ComfyUI Manager - search SAM3DBody, install the top result (try nightly if the stable version doesn't show up), restart. If you'd rather go manual, or the Manager route stalls out:

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-SAM3DBody.git
cd ComfyUI-SAM3DBody
pip install -r requirements.txt --upgrade
python install.py

On macOS, if the install errors out on building xtcocotools, the install script is supposed to handle it - if you're doing this by hand instead, pip install --no-build-isolation xtcocotools fixes it.

Common issues & troubleshooting

This is where the Hugging Face gate bites. Meta ships the SAM 3D Body checkpoint behind an access-request wall on Hugging Face, and this is the node that tries to pull it. The community reaction on release day was pretty consistent - people got hyped by the demo, then hit "approval required" the moment they ran this node. The good news from the same thread: approval reportedly comes through in about half an hour for most people, and what you write in the request form doesn't seem to matter much. If you're denied outright, mirrors of the weights have shown up on ModelScope and Hugging Face reuploads - search around before assuming you're stuck.

First run is slow, that's the download, not a hang. The checkpoint is a real chunk of weights; give it time on first execution before assuming something's broken.

"Missing nodes" even after a clean install. A handful of people on the release thread reported the nodes not showing up in ComfyUI despite the Manager saying everything installed and pip install -r requirements.txt reporting nothing to do. If that happens, try the fully manual git-clone path above rather than trusting a partial Manager install, and make sure you actually restarted ComfyUI rather than just refreshing the browser tab.

Dual license, worth knowing before you ship anything. The wrapper code (this node included) is MIT, but the vendored SAM 3D Body library underneath it carries Meta's own SAM License - permissive, commercial use is fine, but if you publish results you're expected to credit SAM 3D Body per that license's terms.