SAM3DBody_FaceExpression - ComfyUI Built-in Node Documentation - ComfyUI

This node adds facial expressions to a SAM3D body by detecting faces in an image with the MediaPipe Face Landmarker, matching each detected face to a tracked person, and mapping the 52 ARKit blendshapes onto MHR’s 72-axis expression parameters. It then re-runs the body model so the output mesh vertices and keypoints match the new expression.

Inputs

Parameter	Description	Data Type	Required	Range
`sam3d_body_model`	The SAM3D body model containing the face landmark detector used to detect faces and regenerate the body mesh.	SAM3D_BODY_MODEL	Yes	-
`mhr_pose_data`	Pose data containing per-frame tracked persons with bounding boxes, keypoints, and expression parameters. The node matches each detected face to a person and writes updated expression parameters into this data.	MHR_POSE_DATA	Yes	-
`image`	Image frames used to detect faces. If the image batch has fewer frames than the pose data, the last frame is reused for the remaining frames.	IMAGE	Yes	-
`strength`	Global multiplier on all blendshapes. >1 exaggerates. Default: 1.0.	FLOAT	No	0.0 to 4.0 (step 0.05, default 1.0)
`mouth_strength`	Multiplier on mouth/jaw shapes. MediaPipe’s jawOpen saturates near 1.0. Default: 1.0.	FLOAT	No	0.0 to 4.0 (step 0.05, default 1.0)
`eye_strength`	Multiplier on eye shapes. MediaPipe rarely exceeds 0.5; 2-3x often needed. Default: 2.0.	FLOAT	No	0.0 to 4.0 (step 0.05, default 2.0)
`brow_strength`	Multiplier on brow/cheek/sneer shapes. MediaPipe outputs ~0.1-0.3; 2-3x. Default: 2.0.	FLOAT	No	0.0 to 4.0 (step 0.05, default 2.0)
`input_threshold`	Deadzone on MediaPipe’s raw output (below = zero, above = linear remap). Default: 0.02.	FLOAT	No	0.0 to 0.5 (step 0.01, default 0.02)
`blendshape_smooth_window`	Gaussian window on MediaPipe’s per-frame signal before MHR mapping. MediaPipe’s raw output swings 30-70% frame-to-frame on static faces. 1 = disabled. Use odd values. Default: 7.	INT	No	1 to 31 (step 2, default 7)

Note: A per-clip baseline subtraction is applied only when at least 30 frames in the clip contain detected persons. Detection gaps of up to 12 frames per person are filled by interpolation.

Outputs

Output Name	Description	Data Type
`mhr_pose_data`	The updated pose data. Each tracked person’s expression parameters are replaced with the mapped face expression, and the mesh vertices and keypoints are regenerated to match.	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): b2299e51be3556e639d5b04fcbee541ecf41e0d84c2c8a0fd4e211b2f6caba0b