DWposeDeluxe Keypoint Converter

Pose data is only useful if everyone downstream agrees on what it means, and this pack's own nodes will happily eat either normalized (0–1) or absolute (pixel) coordinates. That's a footgun: mix the two and your skeleton silently lands in the wrong place. The KeypointConverter is the node you run your data through when you want it clean, consistent, and saved - it converts coordinate systems, filters low-confidence junk, and writes the result to disk.

It's less glamorous than the estimator, but it's the one node that rescues you from "my pose is 4 inches to the left and everything's broken" debugging.

The settings that matter

• padding_mode - auto-padding fills in a canvas size if your data doesn't carry one; no-padding leaves it alone. The pack's geometry nodes need canvas dimensions, so if your data came from outside the pack and lacks them, this is where you fix that.
• force_normalized - converts everything to normalized (0–1) coordinates. Flip this when you're feeding data into a pipeline that demands relative coords and you don't want to trust auto-detection.
• confidence_threshold (default 0.25) - drops keypoints below this confidence. Good for scrubbing noise; raises it and you get a cleaner but sparser skeleton.
• reset_confidence - sets valid points to confidence 1.0 and invalid ones to 0.0. Binary on/off instead of fuzzy values, which some downstream tools prefer.
• pretty_json - human-readable, indented JSON output. The author is blunt: this makes very large files, so only use it on small batches when you genuinely need to eyeball the data.
• save_output - writes the converted data to a file in your ComfyUI output directory.

Outputs

• pose_keypoints - the processed, cleaned, converted dataset.
• input_keypoint_info - a STRING summary of what went in (frame count, people, detected format). The pack's UnifiedKeypointHandler also auto-detects normalized vs. absolute, which is what most of the other nodes use to stay consistent.

The workflow

Run the estimator → run the output through this converter once (normalize, filter, save) → load the saved file later with LoadPoseKeypoints → edit with the pack's other nodes. One clean pass up front beats chasing format bugs in five places downstream.

Installing it

Ships with DWposeDeluxe:

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 needed. Two honest caveats: the pretty JSON warning is real - a long video batch can balloon into a multi-megabyte wall of whitespace, so leave it off unless you're debugging. And if your source data is a weird or proprietary structure, the converter can only normalize what it recognizes; data from other packs may need manual cleanup first.