On August 23, ComfyUI quietly gained the ability to turn a single photograph of a person into a rigged 3D body mesh — no custom node pack, no pip installs, no Blender plugin. Kijai’s integration of Meta’s SAM 3D Body was merged into ComfyUI core, shipping GLB and BVH export, video pose tracking and facial expression driving.
If you have ever fought to get consistent character poses out of an image model, or wanted to retarget a real person’s movement onto a 3D character without a mocap suit, this is the free path. The weights are 2 GB and not gated. There is also a browser demo if you would rather download nothing at all.
At a Glance
- What it is: Meta’s SAM 3D Body — a promptable model that reconstructs a full-body 3D human mesh (body, hands, feet) from one still image, now a native ComfyUI node set.
- What’s free: Everything. Open weights (1.99 GB int8 or 2.83 GB bf16), open inference code, and a hosted browser demo at Meta’s Segment Anything Playground.
- What it costs: $0. The license is Meta’s SAM License, which is royalty-free and does not ban commercial use.
- The payoff: Export a body mesh as GLB or an animation as BVH, tested against Blender import.
- The one catch: The merge landed after the last tagged ComfyUI release (v0.33.1, August 13). You need the nightly/master build today, or you wait for the next tag.
What SAM 3D Body Actually Does
SAM 3D Body — 3DB in the paper — came out of Meta Superintelligence Labs alongside SAM 3D Objects, with the research paper on arXiv as arXiv:2602.15989. It takes a single natural image and predicts a parametric human mesh: pose, shape, skeleton and camera.
The interesting choice is the mesh format. Instead of SMPL, which most human mesh recovery work has used for years, 3DB predicts parameters for the Momentum Human Rig (MHR) — Meta’s open-source body model, the one behind its Codec Avatars work. MHR decouples the skeleton from the soft-tissue surface shape, so you can change a character’s build without scrambling their pose. Meta released MHR under what it calls a permissive commercial license.
It is also promptable in the SAM sense: feed it 2D keypoints or a mask to steer the prediction when the automatic result misreads an occluded limb. Meta says it was trained on roughly 8 million curated images, mined from a far larger pool by a data engine that deliberately hunted for unusual poses and awkward capture conditions.
The Nodes You Get
The integration is dependency-free — pose rendering is pure PyTorch, so there is no rendering stack to install. Five core nodes:
SAM3DBody_Loader— loads the checkpointSAM3DBody_Predict— runs the mesh recoverySAM3DBody_Render— renders as a mesh, or in OpenPose / SCAIL-pose formatSAM3DBody_Smooth— temporal smoothing so video pose does not jitter frame to frameSAM3DBody_FaceExpression— drives facial expression through MediaPipe blendshapes
Three extras ship alongside: LoadMoGeModel and MoGeGeometryToFOV estimate camera field of view from the image, which matters if you want the mesh to sit correctly in a scene, and SAM3_VideoTrack pulls SAM 3.1 video tracks for multiple people across a clip.
Because SAM3DBody_Render can emit OpenPose-format skeletons, the output drops straight into any ControlNet pose workflow you already run — a better source of pose control than a 2D estimator, since the skeleton comes from an actual 3D reconstruction. It also feeds SCAIL-2 character animation workflows. If you have been building motion control chains, it pairs naturally with the setup we covered in Tencent’s SCoPE camera control release.
How It Compares
| Option | What you get | Free? | Runs locally | License |
|---|---|---|---|---|
| SAM 3D Body (ComfyUI) | 3D body mesh with hands and feet, GLB + BVH export | Yes, open weights | Yes — 1.99–2.83 GB | SAM License, commercial use permitted |
| Segment Anything Playground | Same model, in a browser, no install | Yes | No, hosted by Meta | Meta demo terms |
| DWPose / OpenPose ControlNet | 2D skeleton only — no mesh, no depth | Yes | Yes — under 1 GB | Apache-2.0 (DWPose) |
| Meshy, Rodin and similar | Textured 3D asset, not a rigged human body | Free tier, credit-limited | No, cloud SaaS | Commercial terms of service |
Try It Free
Zero install: open Meta’s Segment Anything Playground, click Playground, upload an image and reconstruct a person in 3D. Fastest way to find out whether the model handles your kind of subject before committing disk space.
Local, in ComfyUI: the weights live at Comfy-Org/sam-3d-body on Hugging Face. Grab one of the two files from the detection folder:
sam_3d_body_dinov3_int8_convrot.safetensors— 1.99 GB, the one to take if VRAM is tightsam_3d_body_dinov3_bf16.safetensors— 2.83 GB, full precision
Drop it in ComfyUI/models/detection/. No separate text encoder or VAE to chase — the checkpoint is self-contained. On a manual install, update first:
cd ComfyUI
git pull
python main.pyThat git pull is not optional. The PR merged on August 23; the most recent tagged release is v0.33.1 from August 13. On a stable Desktop build or a portable package downloaded before the merge, the SAM3DBody_ nodes will not appear in your node search at all. Then load the official test workflow JSON from the PR, which wires up detection, prediction, smoothing, face expression driving and 3D preview end to end.
One detail worth untangling, because it will save you an afternoon: Meta’s original repository is gated behind manual approval and ships a 2.11 GB model.ckpt plus a separate 0.70 GB MHR asset. The Comfy-Org repackage is ungated and already in safetensors. Unless you need Meta’s own inference code, go straight to Comfy-Org.
The Honest Trade-Offs
Meta’s benchmark claims are Meta’s. The blog says SAM 3D Body outperforms previous models on multiple 3D benchmarks; the paper claims state of the art. There is no independent leaderboard for single-image human mesh recovery, so nobody outside Meta has verified it. And the frequently quoted 5:1 human-preference win rate belongs to SAM 3D Objects, a different model in the same release — not a SAM 3D Body number.
One person at a time. Meta states plainly that the model processes each individual separately and does not reason about multi-person or human-object interaction. SAM3_VideoTrack gets you multiple subjects; it does not get you a scene that is physically coherent between them.
Hands are the weak point. Meta acknowledges that while whole-body hand estimation improved substantially, accuracy still does not surpass specialised hand-only methods. If your work lives or dies on finger articulation, budget for cleanup.
Permissive, but not Apache-2.0. The SAM License grants a royalty-free worldwide right to use, reproduce, distribute and modify, with no commercial-use ban and no user-count cap. But redistribute the weights or anything derived from them and you must pass the license along, plus there are trade-control and privacy-law obligations and a no-reverse-engineering clause. Friendlier than the geographic restrictions buried in some recent video model licenses, but read it.
Hardware: neither Meta nor Comfy-Org has published a VRAM figure. From file size, a 1.99 GB int8 checkpoint is small by 2026 standards — the same weight class as the 3 GB Anima build we covered earlier this month — and should sit on an 8 GB card next to a normal workflow. That is an inference from size, not a measured benchmark.
Five Things To Do Differently
- Run MoGe first on any shot with perspective. If the camera is wide-angle or low, skipping FoV estimation puts the mesh at the wrong depth and scale.
LoadMoGeModelintoMoGeGeometryToFOVcosts you one node and fixes it. - Always insert
SAM3DBody_Smoothfor video. Per-frame mesh recovery jitters. The smoothing node is the difference between usable motion and a mesh that vibrates. - Use it as a pose source, not just a mesh source. Render to OpenPose format and drive your existing ControlNet workflow — a 3D-derived skeleton beats a 2D estimator on foreshortened and occluded limbs.
- Prompt it when it fails. When an arm behind a torso comes out wrong, correcting it with a few 2D keypoints is faster than re-shooting or retouching.
- Export BVH, not GLB, when you want motion. GLB gives you the mesh; BVH gives you animation curves Blender can retarget onto a different rig. The PR notes BVH export was tested against Blender import specifically.
FAQ
Is SAM 3D Body free?
Yes. The model weights, the inference code and the MHR body model are all free downloads, and Meta’s Segment Anything Playground lets you run it in a browser at no cost. There is no paid tier.
Can I use SAM 3D Body commercially?
The SAM License grants a royalty-free worldwide license to use, reproduce, distribute and modify the materials, with no clause prohibiting commercial use and no user-count threshold. If you redistribute the weights or derivatives you must include a copy of the license, and you must comply with trade-control and privacy laws. Read the license yourself before shipping a product on it.
What are SAM 3D Body’s system requirements in ComfyUI?
You need a ComfyUI build from master dated August 23, 2026 or later — the nodes are not in release v0.33.1. Disk space is 1.99 GB for the int8 checkpoint or 2.83 GB for bf16, placed in ComfyUI/models/detection/. No VRAM figure has been published by Meta or Comfy-Org.
Do I need a custom node pack for SAM 3D Body?
No. The implementation was merged into ComfyUI core and is dependency-free. Update ComfyUI and the SAM3DBody_ nodes appear in the node search.
Can SAM 3D Body handle video, or only single images?
The model reconstructs from a single image, but the ComfyUI integration adds video handling on top: temporal smoothing, SAM 3.1 video tracks for multi-person detection, and BVH export. It is per-frame recovery stabilised into a sequence, not a native video model.
Worth Your Afternoon?
Yes, with one condition: check that you can run master. If you are locked to a stable build for a client pipeline, wait for the next tagged release rather than breaking a working install. Everyone else gets a photo-to-rigged-body pipeline for 2 GB of disk and a git pull — a genuinely new capability for creators who do not own a mocap rig. Weighing it against paid image-to-3D services? We tested Meshy for image-to-3D on the Creative AI Show channel. Meshy invents an asset; SAM 3D Body reconstructs a real person.
Accurate as of August 25, 2026 — model availability, licensing terms and ComfyUI release tags move fast.
No responses yet