Thanks for the insight!
@wiltonmts Iām loving this! How are you incorporating CorridorKey? Is it a separate install or are you bundling it w/ Sammie? As Iām stuck to only a Mac, all the MLX stuff is really interesting so if there was a way to point this to that implementation that would be ace.
On that note, I just want to mention that I did the speed comparison out of curiosity @cnoellert and not to throw shade on your contribution. It does go to show how much of a bottleneck PyBox is which anyone who has ever used the BLG renderer can attest to.
Appreciate that @kyleobley but Iām not so thin-skinnedā¦
I do want to check out that repo though. Come hell or high water I couldnāt get mlx running faster than three seconds inference at the full tensor so I do have some curiosity around that though. 1024 was going in one second though which would match with your speed so itās on my list.
thanks so much for this, i just got it running ![]()
where is the matte you are feeding into the pybox coming from?
It was just a simple automatte from a gmask tracer node but any rough key will do.
my test results are quiet mixed so far and i thought it might had to do with how precise my input matte is, how good does it have to be in general, especialy when there are several objects? Also will it work on blue screen as well as on green screen?
I only ever tested with green screen to be honest and the results are fine most of thr time. For most of the greenscreen work I do I honestly have a setup I prefer using additive keying so I havenāt ended up using it that much. But different folks different strokes.
birefnet gives me great results on hair and fine details, but a choke-matte in sometimes needed.