Hi there , after upgrading to 2025.2.1 from 2024, i am playing with inference node and the onnx libraries. Particularly with depth anything v2.
The idea is to “convert” depth map to a usable mask. So after the creation of the zdepth, i did a Camera displace. I added a solid color and adjusted it on the z axis and some roation to cut bg information.
The resolution of the surface must be at least 2 and a little softness on the displace map helps with the edges. Then in the output tab created a matte output. Is as stable is the zDepth.
It has some flickering and holes here and there, but i think its a quick and dirty way to mask foreground stuff.
I hope you find it usefull.
4 Likes
This is neat. Also, there are a couple of matte extraction ONNX models on the Portal. They are specifically trained for portraits and can create much better mattes. Especially for details like hair curls.
1 Like
Thanks Sinan. Is there a discription to what each model do? Besides the obvious of course.
Currently there are less than 10 models on the Logik Portal. When you click their names a simple explanation is shown, if there is any of course.
2 Likes
Jeez. I slept on this. So cool that we have more of these to check out now!