Recently I was looking into ways I could streamline the beauty process for the Black Crystal. After combing through some of the documentation on the Foundry’s website., I found this notice about machine learning I looked into it some more, and apparently, it’s a very new thing in Nuke 13, that will let you train a network to do the brunt of your clean-up.
I watched a few tutorials, and I realized that the entire process is still very new, and it’s not a simple as it sounds. You still need to create 2 clean plate references for the network to train on. So that was the first step that counts as prep work. Another unexpected step were the environment variables.
This was something the Nuke tutorial didn’t talk about (but was in the documentation if you dug further). Apparently, for the CopyCat node to process, it needs about 2-4 GB’s or GPU. So, I had to alter the system’s environment variables to “make room” in a sense. This makes sense given that it is an intense process.
I am not sure I fit will finish or not. However, I did figure out another trouble-shooting error. I kept getting “Inconsistent channel inputs across sets” whenever I would try and train the network. I realized that if one input has an alpha channel, and the other doesn’t it will not work. So I went and removed the alpha, and that did the trick. After some searching I realized this hadn’t been documented yet.
Comments