CV Find Transform (ECC)

Direct (intensity-based) image registration: estimates the geometric transform that aligns 'image' onto 'template' by maximizing the ECC correlation (cv2.findTransformECC) - no keypoints needed, ideal for small shifts/rotations between frames of a burst or a golden template. Color inputs are converted to grayscale internally. Apply the result with cv2.warpAffine (or warpPerspective for MOTION_HOMOGRAPHY), dsize = the template's size, flags = 'INTER_LINEAR | WARP_INVERSE_MAP' to pull the image into the template's frame. Robust by design: raw cv2 RAISES when the iteration does not converge (e.g. featureless or unrelated images) - this node instead returns found=false with an identity matrix that warps to a no-op, so the workflow keeps running; branch on 'found' with a control-flow node (e.g. Basic data handling's 'if/else').

By bmad4ever·Created 3 months ago·Updated 2 days ago· 0

template
image
mask

warp_matrix
correlation
found

◄motion_typeMOTION_EUCLIDEAN►

◄iterations100►

◄epsilon0.0000►

◄gauss_filt_size5►

◄min_correlation0.00►

Categoryimage/CV/features

Inputs (8)

Name	Type	Default	Description
template	NPARRAY,IMAGE		Reference image the other one is aligned TO. Accepts a ComfyUI IMAGE/MASK directly (frame 0 of a batch) or an NPARRAY. Arithmetic ops (add, multiply, etc.) process the full IMAGE batch when both inputs have the same batch size.
image	NPARRAY,IMAGE		Image to align onto the template. Accepts a ComfyUI IMAGE/MASK directly (frame 0 of a batch) or an NPARRAY. Arithmetic ops (add, multiply, etc.) process the full IMAGE batch when both inputs have the same batch size.
motion_type	COMBO	MOTION_EUCLIDEAN	Transform model to estimate: TRANSLATION (2 dof), EUCLIDEAN (shift + rotation, 3 dof), AFFINE (6 dof) or HOMOGRAPHY (8 dof, 3x3). Fewer degrees of freedom converge more reliably.
iterations	INT	1001–5000	Maximum ECC refinement iterations.
epsilon	FLOAT	0.00000–1	Stop when the correlation improves by less than this between iterations (smaller = more precise, slower).
gauss_filt_size	INT	51–101	Odd size of the Gaussian blur applied internally before matching - smooths noise so the optimization does not chase it. 1 = no smoothing.
min_correlation	FLOAT	0.00-1–1	Minimum final ECC correlation for 'found' to be true (1 = perfect). 0 accepts any converged estimate; raise it to reject dubious alignments.
maskopt	NPARRAY,MASK		Optional mask on the IMAGE being aligned: non-zero pixels participate in the alignment. Accepts a ComfyUI IMAGE/MASK directly (frame 0 of a batch) or an NPARRAY. Arithmetic ops (add, multiply, etc.) process the full IMAGE batch when both inputs have the same batch size.

Outputs (3)

Name	Type	Description
warp_matrix	NPARRAY	2x3 float32 transform (3x3 for MOTION_HOMOGRAPHY) mapping template coordinates into the image - use WARP_INVERSE_MAP when warping the image back. Identity if not found.
correlation	FLOAT	Final ECC correlation coefficient (1 = perfect alignment; 0.0 when not found).
found	BOOLEAN	—