Load Pose KeyPoints

Load Pose KeyPoints is the entry point to the pack's pose-driven workflow - the one that answers "can I make DragNUWA follow a human pose over time?" instead of a hand-drawn path. It loads a JSON file of pose keypoints from ComfyUI's output directory and hands it to the rest of the pipeline as a POSE_KEYPOINT value, which Split Tracking Points then converts into the tracking-points format the run node wants.

The format is OpenPose-style: a list of frames, each containing a person's 2D pose keypoints with confidence scores. The full intended pipeline is genuinely clever: you run pose extraction on a reference video (using ComfyUI's standard pose-estimation nodes, which save their results to the output folder as JSON), load the resulting keypoint sequence here, split it into per-joint trajectories with Split Tracking Points, and feed those into DragNUWA Run as drag paths. The pose in the video literally becomes the drag trajectory - the character in your still follows the pose from the video.

Inputs

• file_name - a dropdown listing the JSON files in ComfyUI/output/. The default (PoseKeypoint_00001.json) is just a guess; pick the file your pose-estimation step actually wrote. Note the dropdown is populated from the output directory at node-refresh time, so if your new JSON doesn't show up, refresh the node or check you're looking in the right folder.

Output

A POSE_KEYPOINT value - the parsed JSON - feeding straight into Split Tracking Points.

What's worth knowing before you invest

This is the least-supported part of the pack, and it shows. There's no pose extractor included here - you're responsible for producing OpenPose-format keypoint JSON and getting it into the output folder, which means wiring up an entirely separate set of nodes (or running pose estimation in another tool) before this node becomes useful. The chain also assumes a single person (people[0] is what Split Tracking Points reads), so multi-person videos will confuse it. And as with everything DragNUWA, the quality ceiling is 2024-era SVD - pose-following works best on simple, clear movement; intricate choreography turns into smeared approximations.

The README doesn't document this workflow at all, which is a hint about its maturity: the pose path is the pack's most experimental corner, held together by the author's conventions. If you just want motion, the trajectory tool and the instant brushes will get you there faster and with less setup. If you specifically want a person's movement to drive the animation, this is where you start - just budget for assembling the pose-estimation half of the pipeline yourself.