An often used technique to improve ML mattes is to run them through an edge refiner model. A popular being VitMatte.
It needs a trimap of the edge and will then refine the matte further.
Here’s an .inf model for VitMatte, and Flame setup to use it.
In my testing I’ve found though that AutoMatte actually has a better result w/o VitMatte, meaning it doesn’t seem to have any benefit. To the contrary actually.
That’s a big kudos to AutoMatte then.
But feel free to test it, YMMV…
And you might find use in other mattes that didn’t originate with AutoMatte.
Yeah I’ve worked at Nuke places running SAM then VitMatte, but it’s usually not as good as MatteAnyone(2) let alone VideoMaMa. They left it behind when I showed them SammieRoto.
I often start with the whole frame via Sammie then patch in bits where Automatte is better afterwards.
BorisFX are really doing outstanding things these days.
I’m hoping for FBX or alembic output from their head tracker.
If they can do that then it’s trivially obvious that they can do projection receiving geometry too (which is truly something that was always in the realm of the possible for flame).
I used SmartRoto earlier this week again. Doesn’t get any cleaner…
Of course, it depends on the complexity of the shapes and movement. Sometimes AI roto can be faster. Always mix and match. Never limit yourself to one way if speed and quality matter.
Silhouette wins on any face segmentation roto you may need for color.
