Hand Mask Node 🕒🅡🅣🅝

Hand tracking gives you a skeleton, but skeletons don't composite over video. What you actually want for real-time work is a cut-out - a mask of the hand you can blur, tint, replace, or gate an effect on. Hand Mask Node is the bridge between MediaPipe's 21 landmarks and a usable MASK.

It consumes the HAND_DATA output from this pack's Hand Tracking Node and builds a filled mask from the landmarks. The clever part is mask_type. Pick full_hand and it computes the convex hull of all 21 points - a chunky silhouette of the whole hand. Pick palm_only and it uses just the wrist and the five knuckle bases (landmarks 0, 1, 5, 9, 13, 17) to isolate the palm. Or go granular: fingers_only masks all five fingers, and thumb, index, middle, ring, pinky give you a single finger each - those use the classic MediaPipe finger index ranges (thumb 1–4, index 5–8, middle 9–12, ring 13–16, pinky 17–20).

Mechanically it's OpenCV doing what OpenCV is good at: scale the normalized landmarks to the image dimensions, take a convexHull of the selected points, and fillConvexPoly it into a zeroed mask. It loops over both detected hands, so a two-hand frame gets both silhouettes. If no hands are present, you get an empty (all-black) mask rather than an error - which is the right behavior when the output drives a gate.

Inputs and outputs

• image - the frame you tracked. It's passed through unchanged.
• hand_data - the HAND_DATA from Hand Tracking Node. This is not an image; don't try to feed the debug image.
• mask_type - the eight-choice enum above. This is the whole show.

Outputs are IMAGE (your input, untouched) and MASK. That passthrough is convenient: one branch of your graph gets the original frame, the other gets the hand cut-out, and you don't need a separate reroute.

What to use it for

Real-time compositing mostly. Hand masks are great for painting "your hand is the subject" effects onto a live stream - the kind of thing you see in the pack's motion-detection demos. They also make decent inpainting masks if you're regenerating just the hand region. One honest caveat: these are hard-edged filled polygons, not soft alpha mattes. There's no feathering, no anti-aliasing. If you need a soft edge, blur the mask downstream - a small Gaussian is enough.

Install & gotchas

Same pack, same route: ryanontheinside/ComfyUI_RealtimeNodes via ComfyUI Manager, or

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

Two things to know. The hand pipeline is built on the legacy MediaPipe solutions API, and it downloads the hand-landmark model (~5 MB) automatically on first run - no manual model fetch, but the first frame will hang for a second. And the whole hand-tracking feature is explicitly marked experimental in the source, so treat it as "works, but young." If your mask looks like it's in the wrong place, double-check that the input image and the frame the hand data was computed from are the same size and aspect ratio - the landmarks are normalized, so a mismatch scales them wrong.