H3 Face Transform Info

H3 Face Transform Info is the pack's debug node: it prints the per-frame tracking boxes so you can sanity-check what H3 Face Track + Crop did before you spend GPU time on a sampling pass. That's its whole job, and it's worth having on hand.

This is the node you reach for when things look wrong and you don't know where. The H3 face-refinement pipeline is a chain - track, crop, inject latent, denoise, stitch - and when the final video has a face that wobbles or lands misaligned, the fault could be anywhere downstream. This node tells you whether the tracking itself was good, cheaply and before any model work.

How it works

It's deliberately dumb, in the good way. It reads the transform from H3 Face Track + Crop and prints, per frame: the box's x/y/width/height, plus the magnification (canvas height ÷ box height). max_rows (default 12) controls how many lines you see - it samples the clip rather than dumping all ~200 frames, so you get a spread of the whole clip at a glance. The first line summarises frame count, canvas size and source size, so you also get an immediate check that the transform matches what you think you fed in.

The output is an info string, and it's an output node, meaning it displays directly in the graph - no preview hookup required. Just drag it in, wire the transform, and read.

When to use it

Two situations, both from the README's own guidance:

• The stitch looks misaligned. Before you start changing feather and colour_match on H3 Face Stitch Back, confirm the boxes themselves are sitting where you expect. If the transform is off, no amount of feather tuning fixes it.
• A clip where the subject turns away. Check that gap-filling behaved - that the interpolation across undetected frames produced a smooth path rather than a jump, which would show up as a visible box lurch in these numbers.

It's also a nice sanity check on magnification: if you spot mag < 1.0x across the clip, the tracker is downscaling crops into the canvas and discarding real detail - raise the canvas or skip the clip, because a refinement with nothing to work from is wasted GPU time.

Inputs and install

Just two inputs: transform (from H3 Face Track + Crop) and max_rows (12). No other wiring, no models, no extra dependencies.

cd ComfyUI/custom_nodes
git clone https://github.com/Carasibana/ComfyUI-H3-FaceRefine.git

Restart ComfyUI and it shows up under MiniMax H3/Face Refine (or install via ComfyUI Manager by searching "ComfyUI-H3-FaceRefine"). You'll want the tracker installed anyway since this node is useless without its transform - but if you're debugging someone else's workflow that already has the pack, this is the first node to drag in. Five seconds to wire, and it can save you a multi-minute sampling pass on a tracking bug.