Roto is the oldest technique in this glossary and it has not been displaced. When a real product or a real performer has to sit inside a generated environment, roto is what keeps the real thing real.
Automatic matting has improved enormously and still fails in the same places: hair, motion blur, glass, and anything semi-transparent. Those are also the edges an audience looks at, so the last five per cent of the work is most of the perceived quality.
In a hybrid pipeline roto is often what lets you avoid the hardest generative problem entirely. Compositing a real label into a generated scene beats asking a model to render legible type, every time.
Is rotoscoping still needed with modern AI matting?
Yes, for the edges automatic matting still fails on: hair, motion blur, glass and semi-transparency. Those are the edges viewers actually look at.
How does roto help with generative work?
It lets you sidestep the hardest generative problems. Compositing real packaging type into a generated scene is more reliable than asking a model to render a readable label.