Background removal without a green screen
Left is the footage as shot, on a car park floor. Right is the same frames with the subject separated out, then lit into a different scene. No green screen, no hand-drawn mask and no prompt — the separation is inferred from the footage itself, which is what makes it usable on material that was never prepared for it. Her braids keep their strand detail and her fingers resolve individually instead of fusing into a blob. Two things the separation does not do, both handled after it: it drops the ground shadow, so a contact shadow is rebuilt from the subject's own footprint, and it does not relight — she still carries the light of the place she was actually standing in, so the foreground is graded into the new scene. That grade is measured rather than eyeballed: the street's median luminance is 42.7 against a mean of 73.3, because neon signs and wet asphalt pull the average up, and the ambient cast read from that scene is cold blue.
Measured on the delivered file: 367 frames at 1080×1920 · subject 29–36% of frame · the edge moves 7.1/255 per frame against the picture's 8.3/255, so it is quieter than the footage and does not shimmer · foreground matched to a scene whose median luminance (42.7) is under two‑thirds of its mean (73.3)