SAM 3.1 Checkpoint (CRT AutoDL)
The segmentation workhorse behind SAM3D Body
- MODEL
- CLIP
- VAE
Somewhere between the RT-DETR person boxes and the actual 3D body prediction, someone has to tell the model where the person actually ends - and that's a segmentation job. That's what the SAM 3.1 checkpoint is doing in the SAM3D Body pipeline. RT-DETR finds the people (rough boxes), SAM 3.1 refines that into precise per-person masks, and only then does the SAM3D Body model reconstruct the mesh from properly isolated crops. Skip the masks and the body model has to guess where your person ends and the background begins - which is exactly how you get a blob with a face.
CRTAutoDLSAM3BodyCheckpoint downloads sam3.1_multiplex_fp16.safetensors into ComfyUI/models/checkpoints/ and loads it through ComfyUI's own CheckpointLoaderSimple path. That last bit is worth noting: it's a full checkpoint, so it behaves exactly like an SD checkpoint - one node, three sockets (MODEL, CLIP, VAE). The "multiplex" variant is SAM 3.1's multi-concept mode, which matters when a photo has more than one person: each person gets their own mask rather than the model gluing everyone into one blob.
The inputs and outputs
No inputs - no dropdown, no filename field. Outputs are the standard checkpoint triple:
- MODEL - the segmentation model, wired into the SAM3D Body masking step (and usable with any SAM3.1 masking workflow, since SAM3.1 is a general-purpose segmenter).
- CLIP and VAE - part of the full-checkpoint interface, present because it loads via the checkpoint path.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt
Restart ComfyUI, or use Manager and search CRT-Nodes. Because it calls the native CheckpointLoaderSimple, the ComfyUI version requirement here is more relaxed than the AutoDL 3D nodes - but the SAM3D Body pipeline you're feeding still needs a current ComfyUI.
Gotchas
- First run downloads a multi-GB checkpoint into
models/checkpoints/.CRT AutoDLprogress in the console, one time. And since it's in the standard checkpoints folder, you can load it with the plainCheckpointLoaderSimpleafterward - this node is a download helper, not a lock-in. - This is the masker, not the mesh model. People new to this family grab the "SAM 3.1" checkpoint expecting it to do the 3D. It doesn't - that's
CRTAutoDLSAM3BodyModel(SAM3D_BODY_MODEL). Think of this node as the precision tools and the other one as the sculptor. - SAM 3's license is a custom SAM Licence, not Apache like SAM/SAM 2. Fine for personal and most commercial use, but if you're shipping a product around it, read the terms rather than assuming.
- Updates. Red node or
NaNafter a pack update → right-click → Fix node (recreate).
The three-node SAM3DBody assembly (RT-DETR → this → SAM3D Body model) is genuinely impressive the first time you watch a photo become a posed 3D skeleton. This node is the middle gear - unglamorous, invisible when it works, and the reason the output has clean edges instead of a person-shaped smear.
Inputs (0)
No inputs
Outputs (3)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| CLIP | CLIP | — |
| VAE | VAE | — |