DIGIT Drift QC Preview

The DIGIT Drift Gate tells you a render failed. The DIGIT Drift QC Preview is what you look at while you figure out why - an interactive review node that lets you cycle between the reference, the generated image, the pixel-diff heatmap, the edge-diff map, and an annotated overlay, or blink A/B between reference and render to spot the drift with your own eyes.

It sits wired after a Drift Gate, consuming the artifacts that gate already produced. Where the gate is the automated judge, this node is the human's second opinion - and it exists because "confidence 62, reject" is not a useful message on its own. You need to see the badge that got mangled.

How it works

You feed it the outputs straight from a Drift Gate node: reference_aligned, generated_aligned, diff_heatmap, edge_diff, passed, and qc_json (plus an optional confidence float for display). It draws red hotspot circles on the failed regions, parsed from the grid data in qc_json, so your eye goes straight to where the drift is instead of hunting the whole image.

The interactive layer is the point - cycle through layers (reference, generated, pixel diff, edge diff, annotated) or A/B blink between reference and generated to see the difference move. Outputs: annotated_image (the overlay with hotspots, useful for saving or wiring onward), passed (passed through from the gate), and layer_manifest (a string describing what's in the view - handy as a log line).

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/thedepartmentofexternalservices/comfyui-digit.git
cd comfyui-digit
pip install -r requirements.txt

Or ComfyUI Manager → search comfyui-digit → install → restart. Fully local, no API keys - it's purely a visualization layer over data the gate already computed.

How to actually use it

Build the chain Drift Gate → Drift QC Preview and wire all the outputs across. The pattern that works: run the gate, and when it rejects, don't just read the score - hit the preview, blink A/B a few times, and look at where the red circles land. Nine times out of ten you'll immediately see the drifted logo or the relit badge, and now you know whether it's a prompt fix, a seed change, or a case for a stronger detail_weight on the next pass. The annotated output is also the thing to save alongside the gate's QC sheet when a rejected render needs to be explained to a client or a supervisor - a circle-drawn render is a lot more convincing than a confidence number.