DWpose Keypoint Diff

Pose keypoint data gets shuffled around a lot in this pack - converted between absolute and normalized coordinates, cleaned, reversed, re-batched. Most of the time nothing breaks. When something does break, you want to know before it silently ruins a whole video render. This node compares two POSE_KEYPOINT datasets and tells you exactly what differs, down to which frame, which person, and which body part.

It's a verification tool, not a transformer. You feed it two datasets of the same length and it walks through every (x, y, confidence) triplet in every frame, comparing keypoints_1 (the baseline) against keypoints_2. That's the whole job.

The inputs

• keypoints_1 - your baseline dataset.
• keypoints_2 - the dataset you're checking against it.
• coordinate_threshold (default 0.05) - how much coordinate drift counts as a mismatch. Because it auto-detects whether your data is absolute (pixel) or normalized (0–1), it scales the threshold appropriately; 0.05 means 5% of the canvas in normalized space.
• confidence_threshold (default 0.01) - how much confidence-value drift counts as a mismatch.

Tiny rounding differences under 0.0001 are ignored automatically, so you don't get false alarms from float noise after a round-trip conversion.

What it tells you

The single output, diff_info, is a STRING report with the frame count of both batches and a line-by-line list of mismatches: frame count differences, people-count mismatches per frame, point-count mismatches per subset, and per-keypoint coordinate/confidence deltas. Wire it into a text display node to read it, or just watch the console - it logs a clean info line for a perfect match and a warning if anything disagrees.

The canonical use case is after a round-trip: convert absolute → normalized → absolute, or run your keypoints through ReverseKeypoints twice, then diff the result against the original. If diff_info is clean, your pipeline didn't corrupt anything. If it's not, you now know exactly where.

There's one gotcha to respect: the comparison is structural and positional. Two datasets with the same frame count but people detected in a different order will flag as "different" even if the poses are identical - the node compares person index to person index, not by matching identities. So use it on data from the same detection run, not as a general "are these the same person" tool.

Installing it

Part of the DWposeDeluxe pack:

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

or search DWposeDeluxe in ComfyUI Manager. No models, no inference - pure number crunching, so it runs instantly even on huge batches. And since this is a young pack, if you're diffing data that came from another node pack, be aware that its JSON structure may not match what this pack expects; if the report looks like garbage, that's usually why.