What it does
PoseCapture: extract, track and retarget character motion from video, entirely inside ComfyUI.
Give it a side-view video of a person doing a move. You get back a clean skeleton track of that move, fitted to the proportions of your own character, as a pose video of 3D cylinders: the input SCAIL-Preview needs to animate your character with the near arm and the near leg in front on every frame.
Extract: DWPose with the heels and toes (which the usual DWPose output drops), frame by frame, as a ComfyUI node.
Track: left/right label flips repaired, lost points filled, the figure turned to face right, both feet on one ground line, planted feet locked so they don't slide. Walk cycles become seamless loops; kicks, punches and step-ins stay as they are.
Retarget: the track takes the bone lengths of your character (measured with DWPose on one picture), then renders as z-buffered cylinders.
Three workflows. You extract a move once and fit it to as many characters as you like, so the chain comes in two parts; the third runs both in one go:
Extract poses from video (once per move): Load Video -> Scan Video Poses -> Repair Pose Scan -> Track from Pose Scan -> Save Skeleton Track, with a live Preview Skeleton Track.
Fit track to character (once per character): Load Skeleton Track + your character picture -> Fit Track to Character -> Render SCAIL Cylinders -> mp4 + frames.
Video to character in one graph: both at once, video in, fitted cylinder pose video out.
Everything runs as ComfyUI nodes: no scripts, no command line. Besides DWPose there is no model, and it all runs on a CPU.
The example
"Stationary high stepping alternating sprint kicks (side view)" by Wellman Wellness Training (CC BY 3.0, the excerpt is in the zip), retargeted to two characters: my cop and my neutral grey mannequin. Same track, each with its own proportions. The cop is AI-generated.
I checked that this download reproduces my results exactly: a fresh ComfyUI with PoseCapture 1.2.3 installed from the ComfyUI Registry gave the same track files byte for byte and every pose frame pixel-identical, for the two-part chain and for the one-graph workflow.
Install
ComfyUI Manager: Custom Nodes Manager -> search PoseCapture -> Install. Or load a workflow and click Install Missing Custom Nodes.
CLI:
comfy node install PoseCaptureManual:
git clone https://github.com/Eloque/PoseCaptureintoComfyUI/custom_nodes/
The workflows also use comfyui_controlnet_aux (DWPose) and ComfyUI-VideoHelperSuite. DWPose downloads its two models by itself on first use.
Source and documentation: https://github.com/Eloque/PoseCapture (GPL-3.0)
How to run it
Copy the files of
inputs/from the zip intoComfyUI/input/.Load
1_extract_poses_from_video.json, press Run (about 1.5 minutes on a CPU). The track is saved toComfyUI/output/pss_tracks/sprint_kicks_v1/.Load
2_fit_track_to_character.json, press Run (3 to 4 minutes on a CPU). Out come the cylinder pose video and frames.Pick
mannequin_on_white.pngin Load Image and run workflow 2 again: same move, other proportions.Or load
3_video_to_character_one_graph.jsonand press Run (about 4 minutes on a CPU): the same result in one go.
Your own move: one person seen from the side, whole body in frame, static camera. Choose one-shot for a move done in place or walk cycle for a walk.
Your own character: a 768 x 1024 picture of the whole character seen from the side, facing right, on plain white.
What it is for
The cylinder pose video drives SCAIL-Preview, which animates a character from one picture and keeps the limb order right in side views. My Posed Sprite Sheet Generator V2 shows that render step and the sprite sheets with the cop. PoseCapture is the motion half: bring any move from a video.
Honest limits
Side views only, one person, static camera.
Very close cameras show the far foot higher than the near one (floor perspective); the track moves it down onto the shared ground line.
DWPose can lose fast hands or feet for a frame; the repair interpolates them, but look at the Preview Skeleton Track before you use a move.
Credits
DWPose (via comfyui_controlnet_aux), VideoHelperSuite, ComfyUI.
Example video: Wellman Wellness Training, CC BY 3.0, https://www.youtube.com/watch?v=fpSV9sRewKI
Description
First release, for PoseCapture 1.2.3 (https://github.com/Eloque/PoseCapture). In the zip: three workflows (extract poses from video; fit track to character; both in one graph), the CC BY 3.0 sprint-kick clip with its attribution, the cop and the mannequin pictures, and a README with install and run steps. Attached: the extract chain on the example clip, one track retargeted to two characters, the skeleton preview, the cover.