I really don’t understand what is happening here at all.
1 Like
The article seems to claim it’s breaking a live-action image down into render passes via ML that it then reassmbles via node tree thereby allowing you to adjust the live action as though it were a CG comp.
Looks like a complicated way to make a roto for an out of focus grassy area in the distance outside of a car window and pull a key for a well shot greenscreen plate.
The article was written such that it basically says nothing other than “you can do shit with copycat.”
The screen shots are equally useless. The only thing of any interest is →
…and based on the info on the GitHub for the project, you can build the source with little fanfare. From there one could run a handful of frames out as ground truth and match-up against the originals as source. Run 20k epochs and see what shit comes out the other side of the copycat sausage factory.