@jon_barron
Jon Barron
1 year
Clever new paper from Netflix: sidestep the hard problem of collecting alpha matte training data by recording in front of a green screen and lighting the subject with only blue+red light. The green channel records alpha, and ML hallucinates real-green.
9
68
427

Replies

@nnnnicholas
nnnnicholas β›±
1 year
0
0
1
@KevinVFX
Kevin Comerford
1 year
@jon_barron This is like a modern version of the sodium vapor process. Instead of adding a 4th color to use as a matte, they are just using the green channel.
1
0
7
@efebic
efebic
1 year
@jon_barron @artixels Are we just calling everything ML models do hallucination now?
0
0
1
@ChrisVranos
Chris Vranos
1 year
@jon_barron @Yenaphe I played with this about a decade ago (just the shooting part, not the ai), and it felt like having no green light on the talent was unnecessary. We ended up using ~25%. The key was just as easy when comparing 0% to 25%, but repairing the green was much easier.
3
1
18
@Tom_C_Hall
Tom C. Hall πŸŽ₯
1 year
@jon_barron β€œNew” this process is ancient, goes back to photochemical processes
0
0
3
@wha_hoppen
MikeLaFontaine
1 year
@jon_barron This is a terrible idea. Filmmakers need to be able to pursue their art. Not possible if they can't see a representative image through the camera. Use AI tools to improve upon the existing extraction workflow.
0
0
1
@Reticulatas
Fuller
1 year
@jon_barron If you're filming in front of a flat green screen, why not just have a depth camera and mask the final by depth instead of color?
1
0
0
@nicksushkevich
Monsieur S.
1 year
@jon_barron good and fast way to key for low budget ads or infomercials and 95 netflix produced content. not for serious cinematography.
0
0
0