SAM 3D Body Model (CRT AutoDL)

Meta's SAM 3D Body is the model that does something that sounds impossible until you see it: take one photo of a person and recover a full 3D human mesh - body, hands, face - without multi-view input. It's SAM3's research sibling, and the community wrapped it for ComfyUI in late 2025 ("recover full 3D human meshes from a single image" was the release pitch) before Comfy Org ported the whole pipeline into core. This node is the CRT AutoDL way to load its centerpiece: the sam_3d_body_dinov3 model.

CRTAutoDLSAM3BodyModel downloads the SAM3D Body weights into ComfyUI/models/detection/ on first run and hands them to ComfyUI's native SAM3DBody_Loader. One small bit of history that makes this node nicer than it looks: the original Meta-hosted download was gated behind accepting terms. The Comfy-Org mirror this pack fetches from isn't, so you skip the "sign in to download a model file" dance entirely.

The input that matters

• model_name - BF16 (default) or INT8 ConvRot. The int8 variant is the RAM/VRAM-friendly option; the model's DINOv3 backbone is the heavy part, and on a 12GB card the quantized weights are the difference between fitting and not. Start with BF16 if you've got the headroom - it's the quality pick.

The one output

sam3d_body_model - feeds into the native SAM3DBody_Predict node, which is where the photo (plus person bounding boxes and an optional FOV) becomes mhr_pose_data - the actual 3D pose/mesh data. From there you can render it, export it, or (with BuildPoseFile) write it out.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt

Restart ComfyUI, or install via Manager (search CRT-Nodes). The SAM3DBody_Loader it delegates to lives in comfy_extras, so this needs a current ComfyUI - update ComfyUI if the node doesn't exist in your search.

Gotchas

• First run downloads several GB. The CRT AutoDL progress bar is in the console. It's one-time, and the file lands in models/detection/ where the native loader can reuse it.
• This is a body model, not a generic segmenter. The SAM3D_BODY_MODEL socket only plugs into the SAM3D Body pipeline. The general segmentation job in this pack is handled by the SAM 3.1 checkpoint node (CRTAutoDLSAM3BodyCheckpoint), which outputs a regular MODEL/CLIP/VAE triple. Don't confuse the two.
• The full workflow needs its friends. The SAM3D Body pipeline is a small assembly line: RT-DETR detects the persons, the SAM 3.1 checkpoint produces masks, MoGe-2 estimates the camera FOV for correct scale, and only then does this model predict the mesh. Grab the whole CRT/AutoDL/SAM3Body family and you're set.
• Quality caveat from the 3D beat. The mesh this produces is untextured and the geometry is body-pose grade - great for a starting point or for body-shape work, not a finished game-ready asset. If you need a rendered character, you'll texture and clean it up downstream. That's true of every open 3D generator and it's worth knowing before you're disappointed.
• Updates. Red node or NaN after a pack update? Right-click → Fix node (recreate).

If you're doing anything that needs a real human body in 3D - reference, posing, body-shape analysis - this is the single most impressive node in the SAM3DBody family, and the AutoDL wrapper is the most frictionless way to get to it.