fenderwq
Is it U2-Net model? Very nice results (although I would prefer to mostly cut off the guy).
What scene and minute it's? I would like to compare it with RoboustVideoMatting model.
"This seems far more practical."
I wouldn't be so sure I can transcode 5k video at 100fps so pretty fast but it's heavily optimized code running on rtx 3090. Would be probably hard to run it on other machine (people would have to compile two c++ projects [or trust my binaries] and download bunch of tools from nvidia [cuda + cudnn + tensorRT])