Optical Flow Mask Modulation ⚡🅡🅞🅣🅘

This one's a different shape from the Flex family, despite living conceptually next to it: it doesn't take a FEATURE input at all. Instead, it computes its own motion signal directly from your frame sequence using optical flow - the classic computer-vision technique for tracking how pixels move between consecutive frames - and uses that to modulate a mask you feed in. Give it a video and a base mask, and it can grow, reveal, or leave a decaying trail behind whatever's actually moving in the shot.

How it works

Pick a flow_method: Farneback (dense, a solid general-purpose default), LucasKanade or PyramidalLK (sparse point-tracking, better for large clean motions), or DIS (dense inverse search). The node computes flow between frames, filters it through flow_threshold and magnitude_threshold so small or noisy motion gets ignored, then turns the surviving flow magnitude into a modulation signal for your input masks. modulation_strength and blur_radius shape how strong and how smooth that signal is; subtract_original and invert control how it combines with your base mask.

The trail system is the more distinctive part: trail_length (1-20 frames) keeps a memory of recent motion, decay_factor controls how fast that memory fades, and decay_style decides whether the trail fades in opacity (fade) or shrinks in width (thickness, capped by max_thickness). That's what gets you a comet-tail effect behind a moving subject rather than just a mask that flickers on and off with the motion.

Inputs and outputs that matter

• masks and images - the base mask and the frame sequence flow is computed from.
• flow_method, flow_threshold, magnitude_threshold - how motion gets detected and filtered.
• modulation_strength, blur_radius, strength - how hard and how smooth the effect lands.
• trail_length, decay_factor, decay_style, max_thickness - the motion-trail behavior.
• invert, subtract_original, grow_with_blur - how it composites with your input mask.

Output is a single MASK.

Where you'd use it

Auto-generating a mask that follows a moving subject - a dancer, a hand, a moving vehicle - without rotoscoping every frame by hand, then feeding that mask into a masked inpaint or composite step. The trail options are worth exploring separately for a stylized motion-streak look even outside a strict "follow the subject" use case.

Installing it

ComfyUI Manager: search RyanOnTheInside, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside
cd ComfyUI_RyanOnTheInside
pip install -r requirements.txt

Flow computation is OpenCV work (opencv-python is in this pack's requirements.txt) and it's CPU-bound - it can get slow on long or high-resolution clips, especially with the denser algorithms across many frames.

Where people get burned

With this many knobs, start simple: Farneback at modulation_strength = 1 before touching the decay/thickness controls. If the output mask comes back blank, flow_threshold or magnitude_threshold is almost always the culprit - they're probably set too high for your footage's actual motion scale, so turn them down first rather than assuming the node is broken. And this assumes a continuous shot: a hard cut or scene change between frames registers as a huge, spurious burst of "flow" since the algorithm has no idea it's not real motion - don't feed it a montage and expect a sane result at the cut points.