I initially thought it was only spatial, but in your last example, the actor with the glasses had the same ID. This suggests that over time, temporal weights might be applied, making it more consistent with Luma changes. It’s really interesting!
Could you , track the crypto ID so each frame it refreshes the ID asignation its stuck with the same alpha. I presume it will brek any comp. I think its a cool tool to learn from but we are not there yet.
@@juancamilo908 yeah, i guess you could. But its not terribly efficient. I have a video coming soon that does mattes and are REALLY good for much less effort than this. Im really bummed by segment anything’s performance but what I share about mattes next is an actual game changer. I just need time to record it.
I hope this gets better. However if we get a clean matte for single frames, then we could then use the output to train copycat.
Cheers
Agreed! I have something coming in the next couple of weeks that should be able to do just that :)
I initially thought it was only spatial, but in your last example, the actor with the glasses had the same ID. This suggests that over time, temporal weights might be applied, making it more consistent with Luma changes. It’s really interesting!
Great, thanks a lot! Do you also have a video on Depth-Anything?
@@prony5145 coming soon :)
Thx for tutorial
Interesting and cool stuff but I think i'll give the boffins a bit longer to cook on this one before diving in
My feelings exactly.
Could you , track the crypto ID so each frame it refreshes the ID asignation its stuck with the same alpha. I presume it will brek any comp. I think its a cool tool to learn from but we are not there yet.
@@juancamilo908 yeah, i guess you could. But its not terribly efficient. I have a video coming soon that does mattes and are REALLY good for much less effort than this. Im really bummed by segment anything’s performance but what I share about mattes next is an actual game changer. I just need time to record it.