Point Preview

Every tracking setup has one awkward moment: telling the tracker what to track. With Point Tracker you'd be typing x and y coordinates blind. This node makes it a click. Feed it the first frame of your clip, click on the spot you want tracked (the character's mouth, usually), and it hands the coordinates to the tracker as clean x and y integers.

How it works

It's a preview node with a purpose. Required inputs are image (your first frame, or any frame you want to pick on) and the x / y coordinates themselves, which update as you click on the image. There's a crosshair_size slider (default 15) that controls how big the crosshair arms are - bump it up on small frames so you can actually see what you've picked.

What comes out is exactly what Point Tracker needs: the annotated preview image and the x / y integers, ready to wire straight into Point Tracker's start_x / start_y. That's the whole loop: click, wire, track. If you're hand-typing coordinates instead, the crosshair lets you verify visually that you aimed where you think you did.

A practical detail: pick a point with high local contrast. The tracker follows a small template patch of the image, and a spot on a smooth, textureless surface (the middle of a cheek) gives the optical flow nothing to hold onto. Click the corner of the mouth, a nostril edge, or a spot where a feature crosses the jawline, and the tracking will be dramatically more stable. You can still aim the compositor's placement with its own offset_x / offset_y knobs, so you don't need the click to land exactly on the mouth - just near a feature that tracks well.

Installing it

Part of TrentNodes - one install covers every node in this pack. ComfyUI Manager (search "Trent Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes
pip install -r requirements.txt

If Manager flags the pack as "unsafe", that's a known artifact of the author renaming the repo on day one and leaving two registry entries; the manual clone above sidesteps it. This node needs no model downloads - it's a click-and-output utility.

The one thing that trips people up

The x/y outputs only change when you interact with the preview, so if you've queued the whole graph and want to re-pick a point, do it before you queue. And remember the output point is in image pixels of that input frame - if your downstream video is a different resolution, Point Preview's coordinates won't line up, so pick on the actual frame you're tracking from.