Zura · Person Motion Control

Zura · Person Motion Control takes a video and hands you back the performer pasted onto a flat grey background, plus the mask it used. That's it. It generates nothing, and it is not a compositing node - the description in the source is blunt about it: the result is conditioning, not a finished masked composite.

So what's it for? Changing the room. When you re-angle or relight a performer with an ID-V2V / Wan VACE style route, the model is being asked to invent a new environment. If the old environment is still visible in the conditioning, it comes back. Painting the room grey and feeding the model a clean foreground plate tells it "this is the person; the room is disposable."

Inputs

• images - the frames.
• segmentation_model - a dropdown built at node-load time from the files in ComfyUI's ultralytics model folders, filtered to names containing seg. Which is why the folder has to exist and contain something before the node is any use; it also registers the models/ultralytics root and any sibling of a checkpoints folder, so extra_model_paths installs work.
• background_colour - a six-digit hex string, default #929398 (that warm grey you'll see in the pack's example prompts). A bad value raises Background colour must be a six-digit hex value.
• mask_erode_px - default 2, shrinks the mask inward. Worth a couple of pixels so you don't hand the model a fringed halo of the old room around hair and shoulders.
• edge_softness_px - default 1.0, a Gaussian blur on the mask, which is what stops the blend looking like a paper cutout.

Outputs: control_frames (the grey-background plate you feed a VACE or video-to-video conditioning input) and person_masks (a real MASK, useful for previewing or for your own downstream work).

The mechanism, honestly

It runs Ultralytics YOLO per frame - person class only, imgsz=960, conf=0.25 - keeps the largest mask, binarises it, then throws away any stray components with a connected-components pass so a second person in the background can't leak into your foreground. Then erode, blur, composite: frame * mask + grey * (1 - mask).

That per-frame detection is where your wall-clock goes, and where it can fail: if no person is found in a frame it raises Person not detected in frame N; cannot safely preserve performance rather than silently producing a grey frame. Heavy occlusion, a performer half out of frame or an extreme angle will do it.

Two things to know before you build on it. It is deliberately pre-generation: it strips old-room information from the conditioning only, and the frames you finally deliver are always the full VAE decode. And the detector runs through ultralytics - the same AGPL-3.0 territory that covers its weights, the same dependency that reached ComfyUI users in the 2024 poisoned-release incident. Fine for personal work; think about it before you ship a product.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ZURAVFX/ComfyUI_zura_nodes
pip install -r ComfyUI_zura_nodes/requirements.txt   # brings ultralytics, face-alignment, requests, yt-dlp

Then put a person segmentation weight where ComfyUI looks for ultralytics models - models/ultralytics/segm/person_yolov8m-seg.pt is the name the pack's other nodes expect, and Impact Subpack can fetch it for you. Restart ComfyUI after dropping the file in: the dropdown is populated when the node's inputs are built, so a model added later won't appear until you reload.

In the V4 clean relight node this runs internally, with mask_erode_px pinned to 2 and edge_softness_px to 1.0 - the same values you'd probably pick by hand.