a reel popped up of a person in a room with a detached virtual shadow, the shadow moving with its own physics while the person stands still, all running through an apple neural network stack that feeds a touchdesigner scene. naturally the comments were full of windows users asking for the equivalent, and the real answer is that this effect is not one model at all. it is a segmenter that cuts the person out and a compositor that decides how the shadow moves. cut the shadow grammar loose from the demos and it travels anywhere.
that collapsed the webcam thread i have been circling for two weeks into a single shape. the ideas of cutting myself out of a frame and placing characters in a 3d space were never two projects, they are the two halves of one pipeline: cut me out, put the cutout into a 3d scene, and composite so i look like i am standing inside it. virtual production is the boring industry word for it, self avatar is the other, and the whole thing is genuinely buildable on the machine i already own.
the cheapest real version is close to a day of work. run the segmentation in process, ship clean transparent frames into a game engine, plant every frame on a billboard that faces the camera, and let the pose keypoints drive a simple object in the space, like the shadow plane itself. the viral effect translated into plain boring parts. the fancier upgrade, rigging a full 3d body from tracked joints, can wait until the flat version already feels alive.