SAM3DBody_Predict - ComfyUI Built-in Node Documentation - ComfyUI

SAM3D Body Prediction runs 3D body and hand pose estimation on input images, detecting one or more people per frame. Tracking data or bounding boxes can be supplied to improve detection; when neither is provided, the node falls back to full-frame single-person detection.

Inputs

Parameter	Description	Data Type	Required	Range
`sam3d_body_model`	The SAM3D body model to use for prediction.	SAM3D_BODY_MODEL	Yes	—
`image`	Image or batch of images to run body prediction on.	IMAGE	Yes	—
`track_data`	Tracking data from SAM3 Video Track, required for multi-person detection.	SAM3_TRACK_DATA	No	—
`bboxes`	Per-frame bounding boxes used for better detection. Can be used as an alternative to tracking data.	BBOX	No	—
`run_hand_refinement`	Improves hand pose at the cost of extra inference time and memory use. Default: true.	BOOLEAN	No	true false
`fov`	Vertical FoV in degrees. Affects predicted depth and absolute scale. 0 = fall back to ~53° (16:9). Default: 0.0.	FLOAT	No	0.0 or greater
`batch_size`	Max person crops to process as a batch. Larger values use more VRAM for faster inference. Default: 64.	INT	No	1 to 512

Note: When track_data is provided, it takes precedence over bboxes. If neither track_data nor bboxes is provided, the node falls back to single-person full-frame detection. Bounding boxes can be provided for a single frame (applied to every frame) or per frame.

Outputs

Output Name	Description	Data Type
`mhr_pose_data`	Body pose data bundle containing per-frame pose detection results, face geometry, the input image size, canonical vertex colors, and a hand vertex mask.	MHR_POSE_DATA

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): f1039349cd2809423053bffde1c7d119c7c42f217327d23c608b1224d183770e