OpenCV calcOpticalFlowFarneback_1

OpenCV calcOpticalFlowFarneback_1 is the same dense-optical-flow node as calcOpticalFlowFarneback_0, emitted as the second overload of cv2.calcOpticalFlowFarneback. The opencv-comfyui generator numbered every overload it found; this function has two; there is zero practical difference between them. Pick one and move on with your life.

What the function does, in one sentence: it takes two frames and produces a full per-pixel motion field - a two-channel array where every pixel carries how far it moved horizontally and vertically between the frames. Dense flow is the "where did everything go" primitive, and Farneback is the classic algorithm for it: local polynomial approximation, matched across an image pyramid, refined iteratively. It's the kind of deterministic CV computation the pack's author says these nodes are for - small, quick, non-AI transformations - except this one is not small: it needs grayscale input, a specific out-parameter wiring, and the output isn't a viewable image.

Inputs

• prev / next - NPARRAY frames, grayscale uint8. Feed RGB and you hit the README's own img.type() == CV_8UC1 assertion error; run cvtColor (code 6) first.
• flow - NPARRAY: the out-parameter, surfaced as a required input. It gets overwritten, but you must wire a (H, W, 2) float32 array into the slot anyway - one of the pack's roughest edges.
• pyr_scale (FLOAT) - 0.5 standard.
• levels (INT) - pyramid depth; more = bigger motions.
• winsize (INT) - averaging window; 15–21 typical.
• iterations (INT) - 5 typical.
• poly_n (INT) - polynomial window, 5 or 7.
• poly_sigma (FLOAT) - 1.1–1.5.
• flags (INT) - 0 for fresh, 4 to seed from an existing flow.

Output: one nparray, the (H, W, 2) float flow field. Not an image - two-channel arrays don't survive Nparrays2Image cleanly, so visualization is on you.

Install and verdict

ComfyUI Manager → opencv-comfyui, or clone https://github.com/geroldmeisinger/opencv-comfyui into ComfyUI/custom_nodes, restart; needs opencv-contrib-python. Same pack-wide rules (batch size 1, NPARRAY conventions).

This is the honest "for CV heads only" node. If you're doing motion analysis on frames, it's the genuine primitive - and the friction is the pack's, not yours. For motion-aware generation, the video ecosystems in ComfyUI have friendlier purpose-built tools. And the _1 tells you nothing - it's just the redundant twin.