Subject Compositing

For shots nobody can film and no generator produces convincingly: him in a spaceship, in a clown car, on the moon. A real take is segmented out of its real surroundings and laid over replacement footage, with a prop in front. It lives or dies on the source take — settle the footage before you promise the shot.

Requires

pip install 'video-studio-engine[vision] @ https://github.com/scrollmark/social-skills/archive/refs/heads/master.tar.gz', then video-studio composite_subject … and video-studio track_pointing …. Neither is bundled here — both import MediaPipe, NumPy and OpenCV and download model files on first run, so neither can ship as a plain file inside a skill; [vision] is the heaviest extra in the package. Without it nothing here executes: no matte, no feathered occlusion seam, no arm punch-through, no gesture timing. What still holds is the footage guidance below — it tells the user what to shoot, and is worth giving before they shoot. The pointing grammar lives in video-formats.

What the Segmenter Needs From the Take